{"cached_at":"2026-09-29T09:54:56.623718+00:00","cl_docket_id":"74659430","docket":{"resource_uri":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","id":74659430,"court":"https://www.courtlistener.com/api/rest/v4/courts/txwd/","court_id":"txwd","original_court_info":null,"idb_data":null,"clusters":[],"audio_files":[],"assigned_to":"https://www.courtlistener.com/api/rest/v4/people/9604/","referred_to":null,"bankruptcy_information":null,"absolute_url":"/docket/74659430/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T13:27:54.778569-07:00","date_modified":"2026-09-25T15:45:29.416696-07:00","source":1,"appeal_from_str":"","assigned_to_str":"Leon Schydlower","referred_to_str":"","panel_str":"","date_last_index":null,"date_cert_granted":null,"date_cert_denied":null,"date_argued":null,"date_reargued":null,"date_reargument_denied":null,"date_filed":"2026-08-17","date_terminated":null,"date_last_filing":"2026-08-31","case_name_short":"","case_name":"Neural AI, LLC v. Tesla Inc.","case_name_full":"","slug":"neural-ai-llc-v-tesla-inc","docket_number":"7:26-cv-00318","docket_number_core":"2600318","docket_number_raw":"7:26-mc-00318","docket_number_source":0,"federal_dn_office_code":"7","federal_dn_case_type":"mc","federal_dn_judge_initials_assigned":"LS","federal_dn_judge_initials_referred":"","federal_defendant_number":null,"pacer_case_id":"1172927763","cause":"Civil Miscellaneous Case","nature_of_suit":"890 Other Statutory Actions","jury_demand":"None","jurisdiction_type":"Federal Question","appellate_fee_status":"","appellate_case_type_information":"","mdl_status":"","filepath_ia":"","filepath_ia_json":"","ia_upload_failure_count":null,"ia_needs_upload":true,"ia_date_first_change":"2026-08-17T13:27:54.767203-07:00","date_blocked":null,"blocked":false,"appeal_from":null,"parent_docket":null,"tags":[],"panel":[]},"parties":[],"entries":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/476458976/","id":476458976,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/492021611/","id":492021611,"tags":[],"absolute_url":"/docket/74659430/12/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-31T16:05:31.417625-07:00","date_modified":"2026-09-08T22:35:40.222873-07:00","sha1":"326116cd13f3b47303ad0dfa47d3ca92778c2d0e","page_count":8,"file_size":360593,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.12.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.12.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 1 of 8\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n Misc. Case No. 7 :26-mc-00318\n Petitioner,\n Principal case pending in W estem District of\n V. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\nTESLA, INC.,\n\n Respondent.\n\n\n NEURAL Al'S REPLY IN SUPPORT OF ITS MOTION TO COMPEL\n COMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY TESLA, INC.\n\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 2 of 8\n\n\n\n\n Tesla's opposition asks this Court to believe that NAI agreed to accept a five-paragraph\n\ndeclaration listing GPU model numbers and call it a day. No document supports that claim: not\n\nthe subpoena, not the draft declaration NAI provided, not any document in the record. What the\n\nrecord does show is that NAI offered Tesla a streamlined path to resolve both subpoenas by\n\naddressing fifteen substantive questions in declaration form, and NAI would forgo broader\n\ndocument discovery \"subject to resolving any material gaps.\" Ex. 10 at 1; Tesla took the shortcut\n\nbut skipped the substance, then unilaterally declared the matter \"concluded.\" Ex. 21 at 2. It is not.\n\nI. TESLA REWRITES THE PARTIES' MEET AND CONFER EFFORTS\n\n Tesla's self-serving account of the parties' meet-and-confer is contradicted by the written\n\n record. Tesla's central premise-that NAI agreed to limit discovery to identifying hardware and\n\n software products (Opp. at 4-6}-has no support. NAI offered to avoid more intrusive discovery\n\n into Tesla's systems if Tesla confirmed that it used NVIDIA software on NVIDIA hardware\n\n without source-code modification and that relevant functionalities were present. Ex. 10 at 1; Ex.\n\n 12. Otherwise, NAI needed information about Tesla's modifications and their effect on\n\n infringement. Ex. 10 at 1. The \"out-of-the-box\" discussion was a path to narrowing discovery\n\n through those substantive confirmations, and not an agreement to accept a product list.\n\n The draft declaration's scope is explicit: fifteen paragraphs covering GPU identification\n\n (Ex. 12 ,-i 3); three categories of NVIDIA software (,-r,-r 4--6); whether Tesla uses that software\n\n without modification to source code (,-r 7); whether hardware and software function, to the best\n\n of Tesla's knowledge, as designed by NVIDIA (,-r 8); use of NVIDIA-distributed pretrained\n\n models (,-r 9); CPU/GPU memory architecture (,-r,-r 10---11); standard data paths, including\n\n GPUDirect (,Ml 12-13); U.S. operations (,-r 14); and frequency of use (,-r 15). That is the scope\n\n NAI proposed and the information needed to resolve the subpoenas.\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 3 of 8\n\n\n\n\n Tesla states N AI' s draft declaration sought \"pre-written admissions.\" Opp. at 1. Not true.\n\n NAI's draft declaration was a template addressing the information needed to resolve the\n\n subpoenas and did not dictate any language. Indeed, NAI expressly said \"Tesla may revise to\n\n ensure its accuracy.\" Ex. 10 at 1. But any revision still needed to address any modification Tesla\n\n made to the NVIDIA hardware and software, if any, relevant to NAI's infringement theories.\n\n Tesla's own declarant admits that NAI only agreed to accept the declaration \"subject to\n\n any material gaps.\" Fawzy Deel. 1 13. NAI expressly preserved its right to seek additional\n\n information, and the declaration's omissions produced this dispute.\n\nII. TESLA'S DECLARATION IS MATERIALLY DEFICIENT\n\n Tesla recasts the dispute as one over \"length.\" Opp. at 7. That is a straw man. The problem\n\nis what Tesla omitted, not how much it said. Its five-paragraph declaration sidesteps the technical\n\nfacts needed to resolve the subpoenas.\n\n Rather than revising NAI's draft declaration to accurately address the core subjects, Tesla\n\nprovided non-answers. Ex. 20. The declaration identifies six GPU models (id. 1 3), a partial\n\nsoftware list (14), and states that some software is used \"as provided\" (15). It omits:\n\n 1. Whether Tesla uses all NVIDIA software without modification to code (Ex. 12, 17);\n\n 2. Whether hardware/software functions as designed by NVIDIA (id. 18);\n\n 3. Use of NVIDIA-distributed pretrained models (id.19);\n\n 4. CPU/GPU memory architecture (id. ffl0--11);\n\n 5. Standard data paths including GPUDirect (id. ffl2-13);\n\n 6. U.S. operations (id. 114); and\n\n 7. Frequency of use (id. 115).\n\nThose omissions are material.\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 4 of 8\n\n\n\n\n Indeed, Tesla's qualification that only \"some software\" is used \"as provided\" creates\n\nprecisely the uncertainty that further discovery must resolve. Ex. 20 ,r 5. That wording necessarily\n\nleaves open whether Tesla modifies other NVIDIA software, what those modifications are, and\n\nwhether they affect the accused functionality. Tesla offers no explanation. Those unanswered\n\nquestions make a Tesla deposition even more necessary-not less-to determine how Tesla\n\nactually uses NVIDIA's software and to develop the third-party evidence bearing on NAI's\n\ninfringement claims against NVIDIA.\n\n Moreover, NAI asked only for Tesla's best knowledge and not a guarantee that NVIDIA's\n\nproducts work \"as designed and intended.\" Opp. at 5. If Tesla knows otherwise, that fact warrants\n\nfurther discovery on induced infringement. Tesla cannot evade a material factual question by\n\nrelabeling it a request for a legal conclusion. It is not.\n\nIII. THE MOTION IS TIMELY\n\n Tesla's timeliness argument fails on every level.\n\n First, the motion was timely under the stipulated deadline. NVIDIA agreed that third-\n\nparty document motions could be filed through August 18, 2026. Laiche Deel. ,r 28. NAI filed on\n\nAugust 17. Tesla protests it was \"not a party to that stipulation\" (Opp. at 10, n.4 ), but the stipulation\n\ngoverns the schedule in the underlying litigation. Tesla has no standing to veto a scheduling\n\nagreement between the actual parties.\n\n Second, the motion is timely even without the stipulation. Local Rule CV -16(e) permits\n\nmotions filed ''within 14 days after the discovery deadline [if they] pertain to conduct occurring\n\nduring the final 7 days of discovery.\" Tesla served its materially inadequate declaration on August\n\n10, the day before discovery closed. NAI filed on August 17, within fourteen days of the original\n\ndeadline and Tesla's noncompliance. The motion is therefore independently timely.\n\n\n\n\n 3\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 5 of 8\n\n\n\n\n Third, service of the subpoenas was timely. Tesla argues NAI served them too late because\n\nthey were issued in June 2026 rather than October 2025, when NAI served other customers. But\n\nTesla cites no case holding that a subpoena served nearly two months before the close of document\n\ndiscovery-and three months before the close of deposition discovery-is untimely. It is not. The\n\nsubpoenas were timely, and Tesla had more than adequate time to respond but chose not to.\n\n IV. TESLA'S DIVISIONAL OBJECTION IS MERITLESS\n\n NAI complied with Rule 45 and the Clerk's express instructions. The subpoenas designate\n\nAustin as the place of compliance, and NAI filed its motion in the Western District of Texas, which\n\nis the \"district where compliance is required.\" Fed. R. Civ. P. 45(d)(2)(B)(i). Rule 45 requires\n\nfiling in the proper district, not a particular division. Rule 45(c)'s 100-mile limitation governs\n\nwhere Tesla must comply, not the courthouse hearing the motion.\n\n Moreover, NAI was instructed to direct its motion to Midland/Odessa. Because the original\n\nmotion had no divisional header, the Clerk issued a deficiency notice directing NAI to revise the\n\nheader to read \"Midland/Odessa Division\" and refile. See Ex. A. NAI followed that instruction\n\nexactly. Tesla's objection thus identifies, at most, an administrative assignment issue the Court\n\nmay correct. It is no basis to deny an otherwise timely motion filed in the correct federal district,\n\nmuch less without leave to refile.\n\n V. TESLA'S CROSS-MOTION TO QUASH SHOULD BE DENIED\n\n Tesla's burden objections are boilerplate. It identifies no specific cost or hardship and does\n\nnot explain why answering fifteen factual questions about its own GPU infrastructure is unduly\n\nburdensome. That is insufficient. See Waller v. Jet Specialty, Inc., 2024 WL 7050192, at *3 (W.D.\n\nTex. 2024) (\"The resisting party 'must show how the requested discovery is overly broad, unduly\n\nburdensome, or oppressive by submitting affidavits or offering evidence revealing the nature of\n\n\n\n\n 4\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 6 of 8\n\n\n\n\nthe burden.\"') (citation omitted).\n\n Tesla's overbreadth objection fares no better. NAI's twelve requests target discrete steps\n\nin the accused method and specific technical elements of its infringement claims-software\n\nselection, data paths, memory architecture, and GPU computation scheduling-not a fishing\n\nexpedition. N AI further narrowed the requests by offering a declaration alternative that would have\n\neliminated document production entirely. Tesla refused that accommodation and now complains\n\nabout the scope of the discovery it forced NAI to pursue. NAI has made every effort to minimize\n\nTesla's burden. It offered to accept a declaration in lieu of document production. It provided\n\nfocused technical questions. It narrowed scope at every tum. Tesla rejected each accommodation.\n\n Moreover, NVIDIA itself told this Court that it lacks knowledge of how customers deploy\n\nits products and that NAI must seek this information directly from customers. Laiche Deel. ,r 24.\n\nTesla is the only entity that can answer these questions.\n\n Finally, to be clear, NAI does not need access to Tesla's unique trade-secret code. NAI\n\nneeds to know how Tesla is incorporating NVIDIA 's code, what functionalities it is invoking, and\n\nwhat modifications, if any, it is making. Tesla repeatedly frames NAI's requests as seeking all of\n\nTesla's proprietary information, but NAI has repeatedly explained its focus is on Tesla's use and\n\nintegration of NVIDIA products-not Tesla's independent innovations-and the draft declaration\n\nconfirms the same scope. The Court should ignore Tesla's baseless confidentiality objections.\n\nVI. CONCLUSION\n\n For the foregoing reasons, the Court should grant NAI's Motion to Compel and deny\n\n Tesla's Cross-Motion to Quash. At minimum, it should compel Tesla to produce documents\n\n sufficient to establish the facts addressed in NAI's draft declaration and designate a witness to\n\n testify regarding the declaration Tesla served and any remaining gaps. Ex. 14.\n\n\n\n\n 5\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 7 of 8\n\n\n\n\nDated: August 30, 2026\n\n\n Respectfully submitted,\n\n Isl Rocco Magni\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Attorneys for Petitioner Neural AI, LLC\n\n\n\n\n 6\n\f Case 7:26-mc-00318-LS Document 12 Filed 08/31/26 Page 8 of 8\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that on August 30, 2026, a true and correct copy of\n\nthe foregoing document was served on counsel for Tesla, Inc. and all counsel of record in the\n\nunderlying action.\n\n\n Isl Rocco Magni\n Rocco Magni\n\n\n\n\n 7\n\f","ocr_status":2,"date_upload":"2026-09-01T09:05:00.463458-07:00","document_number":"12","attachment_number":null,"pacer_doc_id":"181037306929","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Reply to Response to Motion","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/492089796/","id":492089796,"tags":[],"absolute_url":"/docket/74659430/12/1/neural-ai-llc-v-tesla-inc/","date_created":"2026-09-01T08:55:43.644020-07:00","date_modified":"2026-09-08T22:35:20.959139-07:00","sha1":"b4e1fa872838663aaac7b604b4d151cb48b7a846","page_count":3,"file_size":2020875,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.12.1.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.12.1.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 12-1 Filed 08/31/26 Page 1 of 3\n\n\n\n\n EXHIBIT\n\n A\n\f Case 7:26-mc-00318-LS Document 12-1 Filed 08/31/26 Page 2 of 3\n\n\n Sunday, August 30, 2026 at 6:25:35 PM Pacific Daylight Time\n\n\nSubject: Activity in Case 7:26-mc-00318-DC Neural Al, LLC v. Tesla Inc. Deficiency Notice\nDate: Tuesday, August 18, 2026 at 12:15:01 PM Pacific Daylight Time\nFrom: TXW_USDC_Notice@txwd.uscourts.gov\nTo: cmecf_notices@txwd.uscourts.gov\n\n\n\nEXTERNAL Email\n\nThis is an automatic e-mail message generated by the CM/ECF system. Please DO NOT\nRESPOND to this e-mail because the mail box is unattended.\n***NOTE TO PUBLIC ACCESS USERS*** Judicial Conference of the United States\npolicy permits attorneys of record and parties in a case (including pro se litigants) to\nreceive one free electronic copy of all documents filed electronically, if receipt is required by\nlaw or directed by the filer. PACER access fees apply to all other users. To avoid later\ncharges, download a copy of each document during this first viewing. However, if the\nreferenced document is a transcript, the free copy and 30 page limit do not apply.\n\n U.S. District Court [LIVE]\n\n Western District of Texas\n\nNotice of Electronic Filing\n\nThe following transaction was entered on 8/18/2026 at 2: 13 PM CDT and filed on 8/18/2026\nCase Name: Neural Al, LLC v. Tesla Inc.\nCase Number: 7:26-mc-00318-DC\nFiler:\nDocument Number: No document attached\n\n\nDocket Text:\nDEFICIENCY NOTICE: re [1] MOTION to Compel Compliance with Subpoena\nServed on Third-Party Tesla, Inc. (Reason for deficiency, i.e. Header does not read\nMidland/Odessa Division). Please correct and refile. In the docket text, type\nCorrected Motion to Compel. This deficient document will not be forwarded to the\nassigned judge forconsideration. (ktm)\n\n\n7:26-mc-00318-DC Notice has been electronically mailed to:\n\nBrian D. Melton bmelton@susmangodfrey.com , ecf-\n4 748d80593a6@ecf.r:2acerr:2ro.com , haley..:9ratzer-0160@ecf.r:2aceq~ro.com\n\nMax L. Tribble , Jr mtribble@susmangodfrey'..com, ksP-eedy_@susmangodfrey'..com ,\nsschulze@susmangodfrey'..com\n\nRocco Magni rmagni@susmangodfrey.com, rachel-solis-3173@ecf.pacerpro.com ,\n\n\n 1 of 2\n\f Case 7:26-mc-00318-LS Document 12-1 Filed 08/31/26 Page 3 of 3\n\n\nrocco-magni-susman-godfrey-984 7@ecf.pacerpro.com, rsol is@susmangodfrey.com\n\nTanner H. Laiche tlaiche@susmangodfrey.com, tdenio@susmangodfrey.com\n\n7:26-mc-00318-DC Notice has been delivered by other means to:\n\n\n\n\n 2 of 2\n\f","ocr_status":1,"date_upload":"2026-09-01T09:03:43.440698-07:00","document_number":"12","attachment_number":1,"pacer_doc_id":"181037306930","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit A","acms_document_guid":""}],"date_created":"2026-08-31T16:05:31.397909-07:00","date_modified":"2026-08-31T16:22:42.285803-07:00","date_filed":"2026-08-31","time_filed":"17:20:00","entry_number":12,"recap_sequence_number":"2026-08-31.001","pacer_sequence_number":42,"description":"REPLY to Response to Motion, filed by Neural AI, LLC, re 6 CORRECTED MOTION to Compel Compliance With Subpoena Served on Third Party Tesla, Inc. filed by Petitioner Neural AI, LLC (Attachments: # 1 Exhibit A)(Magni, Rocco) (Entered: 08/31/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/475630485/","id":475630485,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491169122/","id":491169122,"tags":[],"absolute_url":"/docket/74659430/11/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-24T17:06:26.345413-07:00","date_modified":"2026-09-08T09:08:56.823827-07:00","sha1":"03756b0e88941f6ff021d24a0fa8ca56fb84db1d","page_count":15,"file_size":289218,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 1 of 15\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC\n\n Petitioner, Misc. Case No. 7:26-mc-00318-LS\nv.\n\nTESLA, INC.,\n\n Respondent.\n\n\n\n\n NON-PARTY TESLA, INC.\u2019S OPPOSITION TO NEURAL AI, LLC\u2019S\n MOTION TO COMPEL COMPLIANCE WITH SUBPOENA AND\n CROSS MOTION TO QUASH NEURAL AI, LLC\u2019S SUBPOENAS\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 2 of 15\n\n\n\n\n TABLE OF CONTENTS\n\n\nI. INTRODUCTION .............................................................................................................. 1\nII. FACTUAL BACKGROUND ............................................................................................. 2\n A. NAI\u2019s Dilatory Third-Party Campaign and the Sweeping Subpoenas to Tesla ............... 2\n B. Tesla\u2019s Objections, Meet-and-Confers, Investigation, and Declaration .......................... 2\nIII. ARGUMENT ...................................................................................................................... 4\n A. Under NAI\u2019s Own Theory, Customer Discovery Ends at Identification and As-Provided\n Use\u2014Information Tesla Already Provided...................................................................... 4\n B. Tesla Fully Performed the Only Compromise NAI Actually Negotiated ........................ 6\n C. The Remaining Requests Are Overbroad, Unduly Burdensome, and Seek Trade Secrets\n Disproportionate to Any Need ......................................................................................... 6\n D. NAI Served Its Subpoenas Too Late and Filed Its Motion Out of Time ......................... 9\n E. NAI\u2019s Motion Should Have Been Filed in Austin\u2014the Place of Compliance .............. 10\nIV. CONCLUSION ................................................................................................................. 10\n\n\n\n\n i\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 3 of 15\n\n\n\n\n TABLE OF AUTHORITIES\n\nCases Page(s)\nBurns v. Bank of America,\n No. 03-cv-1685 (RMB) (JCF), 2007 WL 1589437 (S.D.N.Y. June 4, 2007) ............................. 9\nDays Inn Worldwide, Inc. v. Sonia Invs.,\n 237 F.R.D. 395 (N.D. Tex. 2006) ......................................................................................... 9, 10\nHoeflein v. Crescent Drilling & Prod., Inc.,\n No. SA 19-CV-01194-FB, 2020 WL 7643122 (W.D. Tex. Dec. 23, 2020) ............................... 4\nThomas v. IEM, Inc.,\n No. 06-886-B-M2, 2008 WL 695230 (M.D. La. Mar. 12, 2008)................................................ 9\n\n\nRules\nFed. R. Civ. P., Rule 26 .............................................................................................................. 6, 9\nFed. R. Civ. P., Rule 29 ..................................................................................................................10\nFed. R. Civ. P., Rule 37 ................................................................................................................ 10\nFed. R. Civ. P., Rule 45 ......................................................................................................... passim\n\n\n\n\n ii\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 4 of 15\n\n\n\n\nI. INTRODUCTION\n\n Neural AI, LLC (\u201cNAI\u201d) asks this Court to conscript Tesla, Inc. (\u201cTesla\u201d), a non-party, into\n\nproving NAI\u2019s infringement case against NVIDIA. Far from typical of customer subpoenas, NAI\n\ndemands granular Tesla confidential trade secrets, including Tesla-specific: AI infrastructure,\n\nsource code, system architecture, memory management, data paths, buffer reuse, scheduling, and\n\nrelated 30(b)(6) testimony\u2014across essentially any Tesla system that touches an NVIDIA GPU.\n\nSome document requests and deposition topics even implicate non-NVIDIA GPUs or software.\n\nThis is facially improper third-party discovery.\n\n Tesla responded exactly as Rule 45 contemplates. It timely served written objections; met\n\nand conferred twice within two weeks; explained the basis for its objections; and agreed to\n\ninvestigate NAI\u2019s narrowed discovery despite the late hour: (1) identification of the NVIDIA GPUs\n\nand software Tesla uses and (2) whether Tesla uses the software as provided. In reliance on NAI\u2019s\n\nnarrowed scope, Tesla conducted a thorough investigation\u2014through its supply chain, IT\n\ninfrastructure, and engineering departments\u2014and produced a sworn declaration providing\n\nprecisely the facts NAI requested.\n\n That should have ended the dispute. Instead, one day before the close of fact discovery,\n\nNAI rejected Tesla\u2019s declaration because Tesla refused to adopt NAI\u2019s pre-written admissions\n\nreciting claim language and filed this Motion six days after the written discovery deadline. NAI\u2019s\n\ndissatisfaction with a completed, good-faith compromise is not \u201cnon-compliance.\u201d\n\n NAI\u2019s Motion should be denied. First, under NAI\u2019s own indirect infringement theory, only\n\nproduct identification and as-provided use are relevant to its claims\u2014and Tesla already provided\n\nboth. Second, Tesla complied in good faith and fully performed the compromise reached during\n\nthe meet-and-confers. Third, the remaining requests and topics are overbroad, unduly burdensome,\n\nand seek trade secrets out of proportion to any legitimate need. Fourth, the Motion was belatedly\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 5 of 15\n\n\n\n\nfiled after the written discovery deadline. Finally, it should have been filed in Austin, the place of\n\ncompliance, not Midland. In the alternative, the Court should modify the subpoenas to cover only\n\nthe information Tesla already produced and quash the remainder, including any testimony.\n\nII. FACTUAL BACKGROUND\n\n A. NAI\u2019s Dilatory Third-Party Campaign and the Sweeping Subpoenas to Tesla\n\n NAI began third-party discovery against other NVIDIA customers as early as October 15,\n\n2025. 1 Exs. 22-23. 2 Yet NAI waited until June 25, 2026 to serve Tesla. Mot. 3.\n\n The scope of NAI\u2019s subpoenas is sweeping. Its twelve requests for documents (\u201cRFPs\u201d)\n\nseek confidential information of any Tesla system that uses an NVIDIA GPU, including: all\n\nsoftware, frameworks, source code, configurations, and custom code used on NVIDIA GPUs\n\n(RFPs 1, 5); whether and how Tesla uses or modifies NVDIA software, including details of Tesla-\n\nwritten software (RFPs 4-5); Tesla\u2019s software design, architecture, and data/control/execution flow\n\n(RFP 5); neural-network structure (RFP 6); pointer/buffer reuse (RFP 7); internal input/output data\n\npaths and memory management (RFPs 8-12); and scheduling/queueing/control of GPU\n\ncomputations, including through \u201ccustom software\u201d (RFP 11). Ex. 5 at 16-18. NAI\u2019s five\n\ndeposition topics (\u201cTopics\u201d) mirror that scope. Ex. 5 at 27. Several RFPs and Topics even implicate\n\nnon-NVIDA GPUs or software. See, e.g., RFPs 1, 4-12; Topics 4-5.\n\n B. Tesla\u2019s Objections, Meet-and-Confers, Investigation, and Declaration\n\n After an agreed extension, Tesla served its objections and responses on July 21, 2026,\n\nobjecting on relevance, overbreadth, undue burden, proportionality, trade secrets, and availability\n\n\n\n\n1\n In addition to Amazon and Microsoft, NAI has moved to compel at least the following third-\nparties in this district: xAI, Meta, CoreWeave, Google, and Oracle. See, e.g., Case Nos. 7:26-mc-\n00319, 7:26-mc-00322, 7:26-mc-00323, 7:26-mc-00324, 7:26-mc-00325, 7:26-mc-00327.\n2\n Tesla cites exhibits from NAI\u2019s Motion, with additional exhibits numbered consecutively.\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 6 of 15\n\n\n\n\nfrom NVIDIA, and offering to confer. Ex. 9 (O&Rs). The parties conferred twice, on July 28 and\n\nAugust 7, 2026. Fawzy Decl. \u00b6\u00b6 7, 15. At the first conference, Tesla explained that the subpoenas\n\nswept far past NVIDIA hardware and software to reach Tesla\u2019s own software and non-NVIDIA\n\nGPUs, and that their unbounded categories made an investigation burdensome and time\n\nconsuming\u2014particularly when NAI had served no infringement contentions or any other basis for\n\nsuch broad technical discovery. Id. \u00b6\u00b6 7-9. NAI offered to reduce that burden and expedite\n\ndiscovery by accepting a declaration in lieu of a document production, identifying what NVIDIA\n\nhardware and software Tesla uses and whether Tesla uses that software \u201cout of the box\u201d (i.e., as\n\nprovided). Tesla agreed to consider the narrowed scope, and because Tesla\u2019s counsel was on leave\n\nand traveling abroad, the parties agreed Tesla would respond the following week. Id.\n\n In reliance, Tesla conducted a reasonable investigation, including through supply chain, IT\n\ninfrastructure, and engineering teams. Id. \u00b6 10. On August 4\u2014while that investigation was\n\nongoing\u2014NAI sent a technical questionnaire (\u201cQuestionnaire\u201d) and a pre-written draft declaration\n\n(\u201cDraft Decl.\u201d), stating that it was intended to \u201cguide [Tesla\u2019s] investigation\u201d and that Tesla \u201cmay\n\nrevise to ensure its accuracy.\u201d Ex. 10 at 1. NAI also threatened Tesla that an August 11 discovery\n\ndeadline \u201cleaves [NAI] no practical alternative but to move to compel by the end of this week or,\n\nat the latest, August 10, to preserve its rights.\u201d Id. NAI\u2019s Draft Decl. and Questionnaire went far\n\nbeyond any scope discussed in the meet-and-confer or information Tesla agreed to investigate.\n\nThey sought sworn admissions reciting language of asserted patent claims and encompassed broad\n\ntechnical information sought in NAI\u2019s sweeping subpoenas. Ex. 12 (Draft Decl.) \u00b6\u00b6 7\u201315; Ex. 11.\n\n At the August 7 conference, Tesla declined to commit to NAI\u2019s overbroad Draft Decl. and\n\nreiterated that it would investigate what it had agreed to at the first conference: the NVIDIA\n\nhardware and software Tesla uses and whether that software is used off the shelf. Ex. 21 at 2. NAI\n\n\n\n\n 3\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 7 of 15\n\n\n\n\nagreed that would satisfy its subpoenas, \u201csubject to any material gaps.\u201d Fawzy Decl. \u00b6 13. Working\n\nthrough the weekend, Tesla served the Declaration of Alon Daks on August 10, identifying (1) the\n\nNVIDIA GPUs and software Tesla uses and (2) which software it uses as provided. Id. \u00b6 14; Ex.\n\n20; Ex. 21 at 2. Nothing was missing from the scope discussed on August 7 and confirmed in\n\nTesla\u2019s post-conference email. Ex. 21 at 2.\n\n On August 11 (i.e., one day after service of the Daks Decl. and the last day of document\n\ndiscovery), NAI responded by unilaterally declaring the Daks Decl. \u201cmaterially insufficient\u201d and\n\ndemanding further information never discussed during either meet-and-confer, including\n\nparagraphs 7\u201315 of the Draft Decl. which were written by NAI, document production as to all\n\nRFPs, and a 30(b)(6) deposition. Ex. 21 at 1. Without any further meet-and-confer, NAI filed this\n\nMotion on August 17\u2014six days after the written-discovery deadline.\n\nIII. ARGUMENT\n\n Tesla opposes NAI\u2019s Motion and also moves to quash NAI\u2019s subpoenas for undue burden,\n\nas shown below. Fed. R. Civ. P. 45(d)(3)(A) (requiring a court to quash or modify a subpoena that\n\nimposes an undue burden); Hoeflein v. Crescent Drilling & Prod., Inc., No. SA 19-cv-01194-FB,\n\n2020 WL 7643122, at *3 (W.D. Tex. Dec. 23, 2020) (Undue burden may be determined based on\n\nfactors such as relevance, overbreadth, need, time period, specificity, and burden.).\n\n A. Under NAI\u2019s Own Theory, Customer Discovery Ends at Identification and As-\n Provided Use\u2014Information Tesla Already Provided\n\n NAI\u2019s Motion confirms the limited role of customer discovery\u2014only real-world\n\ndeployment is relevant to NAI\u2019s allegation that NVIDIA induces or contributes to infringement by\n\nencouraging others to use the accused products. Mot. 2. The only Tesla facts with a potential nexus\n\nare (1) NVIDIA GPUs and software Tesla uses, and (2) whether Tesla uses them as provided. Tesla\n\nhas provided both under oath. Ex. 20 (Daks Decl.) \u00b6\u00b6 3\u20135. Everything else NAI seeks is irrelevant\n\n\n\n 4\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 8 of 15\n\n\n\n\nand NAI has never provided Tesla any information to substantiate further discovery:\n\n Tesla custom software, software identity, and modification mechanics (RFPs 1, 4;\n\nTopics 3\u20134). Tesla\u2019s proprietary software and wrappers are not NVIDIA products. Once Tesla has\n\nconfirmed as-provided use of the identified NVIDIA libraries, \u201chow\u201d Tesla builds around them is\n\na Tesla design choice, unrelated to NVIDIA or to any claim between NAI and NVIDIA. Deep\n\ncustomization discovery is, if anything, the opposite of NAI\u2019s own \u201cas provided\u201d theory: the more\n\nTesla has customized, the less Tesla\u2019s systems prove about what NVIDIA may have induced.\n\n Architecture, neural-net design, memory, data paths, and scheduling (RFPs 5\u201312;\n\nTopic 5). NAI admits its requests \u201ctrack the accused computation \u2026 to establish infringement.\u201d\n\nMot. 7. That is a method-claim chart aimed at Tesla\u2019s systems. If, under its purported infringement\n\ntheory, NAI is entitled to discovery into the functionality of NVIDIA\u2019s provided stack, then it\n\nshould obtain that information from NVIDIA\u2019s documents and party discovery. It should not\n\nburden a customer, nor can it force a non-party to sign a declaration that tracks the language of the\n\nasserted patent claims. Ex. 12 \u00b6\u00b6 7\u201315.\n\n NAI\u2019s Draft Decl. \u00b6\u00b6 7\u201315 and Questionnaire. In its August 11 email, NAI claims there\n\nare \u201cmaterial gaps\u201d in the Daks Declaration because it doesn\u2019t swear that Tesla\u2019s systems:\n\n\u201cfunction \u2026 as designed and intended by NVIDIA\u201d (\u00b6 8); freeze pretrained model and network\n\nstructure (\u00b6 9); avoid GDS/UVM/unified memory and use certain memory management (\u00b6\u00b6 10-\n\n12); use a single default data path without custom code (\u00b6 13), and are in use daily (\u00b6 15). Ex. 21\n\nat 1; Ex. 12 (Draft Decl.) \u00b6\u00b6 7\u201315. This is not product identification or as-provided use evidence.\n\nKnowledge of NVIDIA\u2019s designs and intent should be sought from NVIDIA.\n\n NAI\u2019s argument that the information is uniquely in Tesla\u2019s possession is misplaced. Mot.\n\n7. That Tesla possesses knowledge of its own confidential trade secrets (such as its own memory\n\n\n\n\n 5\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 9 of 15\n\n\n\n\ntopology and compiler choices) does not make it relevant or discoverable. Rule 26 asks whether\n\nthe matter bears on claims and defenses, which is not the case here.\n\n B. Tesla Fully Performed the Only Compromise NAI Actually Negotiated\n\n NAI portrays Tesla as a non-party that hid behind objections (Mot. 9), but the opposite is\n\ntrue. See supra \u00a7 II.B. Tesla explained at the July 28 conference why the subpoena\u2019s unbounded\n\nscope reached far beyond NVIDIA hardware and software; NAI never answered that point and\n\nnever identified why such broad Tesla technical information mattered to its claims or defenses\n\nagainst NVIDIA. NAI instead agreed to narrow its requests, acknowledging it was short on time\n\nbefore the close of fact discovery. Fawzy Decl. \u00b6 8. Despite its counsel being on leave and out of\n\nthe country, Tesla committed to investigate which NVIDIA GPUs and software it uses and whether\n\nit uses them as provided. Id. \u00b6 9. It consulted its supply chain, IT infrastructure, and software\n\nengineering teams, conferred again on August 7, and served a sworn declaration on August 10\n\nsupplying precisely those facts. Id. \u00b6\u00b6 10, 13; Ex. 20 (Daks Decl.). That is good-faith compliance\n\nand full performance of Tesla\u2019s end of the bargain. NAI\u2019s dissatisfaction with Tesla\u2019s refusal to\n\nadopt its Draft Decl. does not justify compelling Tesla\u2019s trade secrets.\n\n NAI\u2019s oversimplification that Tesla provided \u201ca five-paragraph declaration\u201d is misleading.\n\nMot. 4. Length is not the test; substance is. The Daks Decl. provided the narrowed scope of\n\ninformation that NAI requested and Tesla agreed to investigate, i.e., which NVIDIA GPUs and\n\nsoftware Tesla uses and which ones were used as provided. Ex. 20. Indeed, NAI invited a\n\ndeclaration \u201cin lieu of\u201d broader discovery. It cannot later manufacture unilateral \u201cmaterial gaps.\u201d\n\nEx. 10 at 1; Ex. 21 at 1. Tesla should not be penalized for pursuing a reasonable compromise.\n\n C. The Remaining Requests Are Overbroad, Unduly Burdensome, and Seek Trade\n Secrets Disproportionate to Any Need\n\n Contrary to Neural AI\u2019s assertion, its subpoenas are not limited to customer-deployment\n\n\n\n 6\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 10 of 15\n\n\n\n\ndiscovery. Rather, as explained above, RFPs 1\u201312 and deposition Topics 1\u20135 demand a full\n\ntechnical autopsy of any Tesla system that touches an NVIDIA GPU and even non-NVIDIA GPU\n\nin certain requests, including source code, custom code, architecture, memory topology, data paths,\n\nand scheduling. Ex. 5. That is overbroad, unduly burdensome on a non-party under Rule 45(d),\n\nand seeks trade secrets far beyond what is proportional to NAI\u2019s infringement claims.\n\n For example, RFP 7 asks Tesla to disclose any use of \u201ca pointer\u201d and its use in buffer\n\nmanagement\u2014a fundamental concept used in almost all computer programming, without any\n\nmeaningful limitations. 3 Ex. 5 at 17. It potentially implicates any Tesla system that uses GPUs,\n\nincluding non-NVIDIA GPUs. Similarly, RFPs 5, 8-12 and Topics 3-5 make sweeping requests for\n\nTesla\u2019s confidential information about its software architecture, input/output data paths design,\n\nmemory management, and computations scheduling. Ex. 5 at 16-18. Simply put, these RFPs and\n\nTopics either seek the functionality of NVIDIA software, in which case, NAI must obtain that\n\ninformation from NVIDIA, or it seeks the functionality of Tesla software, which is irrelevant,\n\noverbroad, and overly burdensome to allegations of infringement by NVIDIA software.\n\n The burden of searching for and collecting this information would be substantial.\n\nCompliance with RFPs 5\u201312 and Topics 3\u20135 could encompass an investigation, across nearly\n\nseven years and company-wide operations, including across systems, workflows, and codebases\n\nthat use NVIDIA GPUs \u201cto perform computations\u201d; locating the engineers and custodians\n\nknowledgeable about each such system; collecting architecture, design, data-flow, control-flow,\n\nand execution-flow materials; and reviewing highly sensitive source code and internal technical\n\ndocuments for responsiveness and privilege. Fawzy Decl. \u00b6 4. That effort would cut across multiple\n\n\n\n3\n See, e.g., https://www.geeksforgeeks.org/dsa/pointer-in-programming/ (\u201cPointer is a variable\nwhich stores the memory address of another variable as its value. . . Pointers allows low-level\nmemory access, dynamic memory allocation, and many other functionality.\u201d).\n\n\n 7\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 11 of 15\n\n\n\n\norganizations and product areas and would divert engineers from ordinary business to potentially\n\nreconstruct implementations, memory layouts, buffer strategies, data-transfer paths, and\n\nscheduling behavior. Id. It is a multi-team technical investigation into core AI infrastructure, with\n\nthe attendant costs of collection, review, redaction, confidentiality designations, and potential\n\nsource-code logistics. Id. Rules 45(d)(1) and (d)(3) require the Court to protect a non-party from\n\nprecisely this kind of significant expense and disruption.\n\n Further, NAI cannot show \u201csubstantial need\u201d for the information it seeks in RFPs 5\u201312 and\n\nTopics 3\u20135 that \u201ccannot be otherwise met\u201d under Rule 45(d)(3)(C), given discovery from NVIDIA,\n\npublic information, and Tesla\u2019s Daks Decl. This is especially so because NAI\u2019s subpoenas ask for\n\nconfidential Tesla trade secrets about the company\u2019s AI operations, including regarding source\n\ncode and related internal architecture, design, and data flow, input/output data handling, memory\n\nlayout, buffer strategies, and scheduling strategies. But NAI has not shown, during the meet-and-\n\nconfers or in its Motion, that it is entitled to so such highly confidential trade secret information\n\nwithout violating Rule 45(d)(3)(B)(i).\n\n Finally, NAI\u2019s pre-written Draft Decl. and Questionnaire are not \u201cfocused\u201d as it alleges.\n\nRather, they require Tesla to map\u2014and swear to\u2014its GPU memory hierarchy, input/output data\n\npaths, buffer reuse, compilation toolchain, model structure, and supposed conformity with\n\n\u201cNVIDIA\u2019s design.\u201d Those topics are the same trade-secret, overbroad demands embodied in RFPs\n\n4\u201312 and Topics 3\u20135. See Ex. 12 (Draft Decl.), \u00b6\u00b6 9-13; Ex. 11 (sections re \u201cInput Data Path\u201d and\n\n\u201cOutput Data Path and Memory Transfers\u201d). NAI\u2019s Draft Decl. even asks Tesla to make legal\n\nconclusions and speculations, such as declaring that \u201cthe hardware and NVIDIA software function\n\ntogether as designed and intended by NVIDIA.\u201d Ex. 12, \u00b6 8. Rule 45 allows proportionate third-\n\nparty discovery. It does not allow a plaintiff to conscript a non-party into proving its infringement\n\n\n\n\n 8\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 12 of 15\n\n\n\n\ncase. Tesla\u2019s declaration already provided the former; the Court should refuse the latter.\n\n D. NAI Served Its Subpoenas Too Late and Filed Its Motion Out of Time\n\n NAI\u2019s subpoenas are dilatory. A subpoenaing party must pursue third-party discovery with\n\ndiligence; merely serving a subpoena before the deadline is not enough where the requests could\n\nnot reasonably be answered, narrowed, collected, and produced before discovery closed. Fed. R.\n\nCiv. P. 45(d)(1). Courts refuse to enforce last-minute third-party subpoenas that leave no realistic\n\npath to complete discovery by the cutoff. Indeed, Rule 45 is a discovery device, not an \u201cend-run\u201d\n\naround the discovery process. See Thomas v. IEM, Inc., No. 06-886-B-M2, 2008 WL 695230 (M.D.\n\nLa. Mar. 12, 2008); Burns v. Bank of America, No. 03-cv-1685, 2007 WL 1589437 (S.D.N.Y. June\n\n4, 2007); Fed. R. Civ. P. 26(b)(2)(C)(ii), 45(d)(1). NAI began subpoenaing NVIDIA customers\n\nsuch as Amazon and Microsoft in October 2025, but waited until June 25, 2026, to serve Tesla\u2014\n\na publicly known NVIDIA customer. See Mot. 3, 7 (noting \u201cTesla is a significant [NVIDIA]\n\ncustomer\u201d); Exs. 22-23. NAI\u2019s August 4 email admitted it was racing that deadline and threatened\n\na motion \u201cto preserve its rights.\u201d Ex. 10 at 1. By the July 28 meet-and-confer, NAI knew it was\n\ntoo late to seek twelve deep technical RFPs and corporate testimony. Fawzy Decl. \u00b6\u00b6 7-8. NAI\u2019s\n\nown delay left no realistic time for meaningful non-party investigation, trade-secret review, meet-\n\nand-confers, or production of the vast scope of materials sought in RFPs 5\u201312.\n\n Further, NAI\u2019s Motion was untimely. Courts measure the timeliness of a motion to compel\n\nagainst the deadline for completion of discovery, not against a later motion deadline. In Days Inn\n\nWorldwide, Inc. v. Sonia Invs., the N.D. Texas court denied a motion to compel filed two weeks\n\nafter the discovery deadline even though it was filed before the \u201cdeadline for other motions.\u201d 237\n\nF.R.D. 395, 396 (N.D. Tex. 2006). In ruling so, the court noted that there was a consensus among\n\nthe 10th, 7th, 6th, and 1st Circuits and various district courts that \u201ccourts generally looked to the\n\ndeadline for completion of discovery in considering whether a motion to compel has been timely\n\n\n 9\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 13 of 15\n\n\n\n\nfiled.\u201d Id. at 397 (collecting cases). NAI filed the Motion six days after written discovery closed\u2014\n\nseeking post-deadline production from a non-party. NAI\u2019s lack of diligence in launching third-\n\nparty discovery cannot justify a lack of diligence in enforcing it\u2014and neither supports reopening\n\nwritten discovery against Tesla after the deadline. The Motion should be denied as untimely. 4\n\n E. NAI\u2019s Motion Should Have Been Filed in Austin\u2014the Place of Compliance\n\n NAI\u2019s Motion was improperly filed in Midland. Rule 45 requires compliance within 100\n\nmiles of where the subpoenaed person resides, works, or regularly conducts business, Fed. R. Civ.\n\nP. 45(c)(2)(A), and any motion to compel must be filed where compliance is required. Fed. R. Civ.\n\nP. 45(d)(2)(B)(i); see also Fed. R. Civ. P. 37(a)(2). NAI\u2019s subpoenas designated Austin as the place\n\nof compliance. See Ex. 5; Mot. 5. NAI, however, filed in Midland, more than 300 miles from\n\nTesla\u2019s Austin headquarters. The underlying case\u2019s location does not override Rule 45, and district-\n\nwide jurisdiction does not satisfy Rule 45\u2019s location requirement. Further, because the written\n\ndiscovery deadline has passed, the Court should deny the Motion without leave to re-file.\n\nIV. CONCLUSION\n\n For the foregoing reasons, Tesla respectfully requests that the Court (1) deny NAI\u2019s\n\nMotion in its entirety, and (2) quash NAI\u2019s document and deposition subpoenas. Alternatively, if\n\nthe Court does not deny the Motion in full, it should: (1) hold that the Daks Decl. satisfies Tesla\u2019s\n\nobligations as to identification and as-provided use; and (2) modify the subpoenas to those\n\nsubjects only and quash RFPs 1 and 3\u201312 and Topics 1\u20135.\n\n\n\n\n4\n NAI relies on a stipulation with NVIDIA permitting motions to compel third parties through\nAugust 18, 2026. Mot. 3. But Tesla was not a party to that stipulation, and it was never adopted by\ncourt order. Fed. R. Civ. P. 29 (stipulations regulate procedure between parties, not non-party\nburdens).\n\n\n 10\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 14 of 15\n\n\n\n\nDated: August 24, 2026 Respectfully submitted,\n\n /s/ Jun Zheng\n\n Jun Zheng\n TX Bar No. 24102681\n zhengjun@tesla.com\n Tesla, Inc.\n 1 Tesla Rd\n Austin, TX 78725\n (512) 417-3528\n\n Ashraf Fawzy\n DC Bar No. 989132\n afawzy@tesla.com\n Tesla, Inc.\n 800 Connecticut Ave. NW\n Washington, DC 20006\n (202) 905-9221\n\n Gina H. Cremona\n CA Bar No. 305392\n gcremona@tesla.com\n Tesla, Inc.\n 1501 Page Mill Rd.\n Palo Alto, CA 94304\n (650) 647-0015\n\n Counsel for Tesla, Inc.\n\f Case 7:26-mc-00318-LS Document 11 Filed 08/24/26 Page 15 of 15\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned hereby certifies that all counsel of record who are deemed to have\n\nconsented to electronic service are being served with a copy of this document via the Court\u2019s\n\nCM/ECF system when it is filed with the system.\n\n\n\n\n /s/ Jun Zheng\n Jun Zheng\n\f","ocr_status":2,"date_upload":"2026-08-25T09:53:40.632376-07:00","document_number":"11","attachment_number":null,"pacer_doc_id":"181037260029","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Response in Opposition to Motion","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491243847/","id":491243847,"tags":[],"absolute_url":"/docket/74659430/11/1/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-25T09:54:05.752738-07:00","date_modified":"2026-09-08T09:02:53.781041-07:00","sha1":"30ee48d0b0676b7c2789226fd0cd1b46ef48d91c","page_count":5,"file_size":167741,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.1.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.1.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 11-1 Filed 08/24/26 Page 1 of 5\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\nNEURAL AI, LLC,\n\n Petitioner,\n Misc. Case No. 7:26-mc-00318-LS\n v.\n\nTESLA, INC.,\n\n Respondent.\n\n\n\nDECLARATION OF ASHRAF FAWZY IN SUPPORT OF TESLA, INC.\u2019S OPPOSITION\n TO NEURAL AI, LLC\u2019S MOTION TO COMPEL COMPLIANCE WITH SUBPOENA\n AND CROSS MOTION TO QUASH NEURAL AI, LLC\u2019S SUBPOENAS\n\n\n I, Ashraf Fawzy, hereby declare as follows:\n\n 1. I am presently employed by Tesla, Inc. (\u201cTesla\u201d) and my official title is Managing\n\nCounsel, IP Litigation. I am an attorney duly licensed to practice in Washington D.C. and am\n\nadmitted to the Western District of Texas. I\u2019m counsel for Tesla in the above-captioned civil\n\naction.\n\n 2. I make this declaration in support of Tesla\u2019s Opposition to Neural AI\u2019s (\u201cNAI\u201d)\n\nMotion to Compel Compliance with Subpoena and Cross Motion to Quash Neural AI, LLC\u2019s\n\nSubpoenas. I have personal knowledge about the matters in this declaration and, if called to testify,\n\ncould and would testify competently to them.\n\n 3. NAI served Tesla subpoenas for document production and deposition testimony on\n\nJune 25, 2026. Based on publicly available information, NAI\u2019s action against NVIDIA was filed\n\non September 13, 2024, and NAI sought third-party discovery from NVIDIA customers such as\n\nAmazon and Microsoft in the underlying action as early as October 15, 2025.\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 11-1 Filed 08/24/26 Page 2 of 5\n\n\n\n\n 4. The burden to Tesla of searching for and collecting the information requested by\n\nNAI\u2019s subpoenas would be substantial. Compliance with the subpoenas entail identification and\n\ninvestigation, across nearly seven years (from September 2018 to present) and company-wide\n\noperations, including into different systems, workflows, and codebases that use NVIDIA GPUs\n\n\u201cto perform computations\u201d; locating the engineers and custodians knowledgeable about each such\n\nsystem; searching for and collecting architecture, design, data-flow, control-flow, and execution-\n\nflow materials; and reviewing highly sensitive source code and internal technical documents for\n\nresponsiveness and privilege. That effort would cut across multiple organizations and product\n\nareas and would divert engineers from ordinary business to potentially reconstruct historical\n\nimplementations, memory layouts, buffer strategies, data-transfer paths, and scheduling behavior.\n\nIt is a multi-team technical investigation into Tesla\u2019s core AI infrastructure, with the attendant\n\ncosts of collection, review, redaction, confidentiality designations, and potential source-code\n\nlogistics.\n\n 5. NAI\u2019s subpoena for document production set a compliance date of July 14, 2026,\n\njust over two weeks after the service date. Tesla\u2019s counsel promptly reached out to NAI\u2019s counsel\n\nfor a three-week extension. To which NAI\u2019s counsel responded that it could only agree to a one-\n\nweek extension \u201c[g]iving the upcoming close of fact discovery.\u201d\n\n 6. On July 21, 2026, Tesla timely served its written objections and responses to NAI\u2019s\n\nsubpoenas, objecting to, among others, relevance, overbreadth, undue burden, and seeking\n\nconfidential trade secrets unproportional to NAI\u2019s needs, and offering to meet and confer to narrow\n\nthe scope of the requests.\n\n 7. The following week, on July 28, 2026, Tesla held its first meet-and-confer with\n\nNAI\u2019s counsel. During the conference, Tesla began the call by explaining the basis for its\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 11-1 Filed 08/24/26 Page 3 of 5\n\n\n\n\nobjections, including that the subpoenas are overbroad, unduly burdensome, and seek Tesla\n\nconfidential information irrelevant to NAI\u2019s infringement claim against NVIDIA. Tesla further\n\nexplained how NAI\u2019s subpoenas were not just limited to NVIDA GPUs and software, but that they\n\nalso potentially implicated Tesla\u2019 own software and non-NIDIA GPUs. Tesla explained that the\n\nunbounded categories of information sought by NAI\u2019s subpoenas made the investigation\n\nburdensome and time consuming, particularly given the short amount of time NAI provided for\n\ncompliance. Tesla also explained that it if NAI wanted information as to what NVIDIA provided\n\nto its customers, it should seek such information from parties to the litigation and that Tesla did\n\nnot believe NAI had a basis for the breadth of information sought in its subpoena, particularly\n\nwithout any further substantiation.\n\n 8. In response, NAI acknowledged that it was short on time and stated that it was\n\nwilling to reduce Tesla\u2019s burden and expedite the discovery by accepting a declaration in lieu of a\n\ndocument production, and that the declaration would be based on the results of Tesla\u2019s\n\ninvestigation as to what NVIDIA GPUs and software Tesla uses and whether Tesla uses the\n\nsoftware \u201cout of the box.\u201d\n\n 9. Tesla agreed to consider NAI\u2019s narrowed scope. Tesla\u2019s counsel was on leave and\n\ntraveling out of the country at the time. Thus, the parties agreed that the parties would follow up\n\nby email the following week.\n\n 10. After the first meet-and-confer on July 28, 2026, and in reliance on NAI\u2019s\n\nstatements, Tesla began a thorough investigation, including through its supply chain, IT\n\ninfrastructure, and engineering teams, on the subjects the parties\u2019 agreed on during the conference,\n\ni.e., which NVIDIA GPUs and software Tesla uses and whether Tesla uses them as provided.\n\n 11. On August 4, 2026, while Tesla\u2019s investigation was still ongoing, NAI sent Tesla a\n\n\n\n\n 3\n\f Case 7:26-mc-00318-LS Document 11-1 Filed 08/24/26 Page 4 of 5\n\n\n\n\ntechnical questionnaire (\u201cQuestionnaire\u201d) and a pre-written draft declaration (\u201cDraft\n\nDeclaration\u201d), stating that it was intended to \u201cguide [Tesla\u2019s] investigation\u201d and that Tesla \u201cmay\n\nrevise to ensure its accuracy.\u201d\n\n 12. In the same August 4, 2026 email, NAI threatened Tesla that an August 11, 2026\n\ndiscovery deadline \u201cleaves Neural AI no practical alternative but to move to compel by the end of\n\nthis week or, at the latest, August 10, to preserve its rights.\u201d\n\n 13. Three days later, on Friday, August 7, 2026, Tesla had a second meet-and-confer\n\nwith NAI. During that conference, Tesla stated that it would not commit to NAI\u2019s overbroad Draft\n\nDeclaration and reiterated that it would investigate what it has agreed to during the first conference,\n\ni.e., which NVIDIA GPUs and software it uses and whether the software was used as provided by\n\nNVDIA. NAI agreed that such a search would satisfy its subpoenas, \u201csubject to any material gaps.\u201d\n\n 14. After the Friday meet-and-confer, Tesla promptly worked through the weekend,\n\ncollecting the relevant information and working with its declarant, a senior staff software engineer.\n\nAnd on Monday, August 10, 2026, a day before NAI\u2019s written discovery deadline, Tesla served\n\nthe declaration of Alon Daks.\n\n 15. The next day, on August 11, 2026, NAI responded to Tesla, unilaterally declaring\n\nthat the Daks declaration was \u201cmaterially insufficient\u201d and demanding Tesla to produce further\n\ninformation that was never discussed during either meet-and-confer, including paragraphs 7\u201315 of\n\nthe Draft Declaration pre-written by NAI, document production as to all document requests in the\n\nsubpoenas, and a 30(b)(6) deposition. Without any further meet-and-confer, NAI filed this Motion\n\non August 17, 2026.\n\n\n\n\n 4\n\f Case 7:26-mc-00318-LS Document 11-1 Filed 08/24/26 Page 5 of 5\n\n\n\n\n I declare under penalty of perjury under the laws of the United States that the foregoing is\n\ntrue and correct to the best of my knowledge.\n\n\n\n\nExecuted on August 24, 2026\n Ashraf Fawzy (Aug 24, 2026 19:27:26 EDT)\n\n Ashraf Fawzy\n\n\n\n\n 5\n\f","ocr_status":1,"date_upload":"2026-08-25T09:56:10.997238-07:00","document_number":"11","attachment_number":1,"pacer_doc_id":"181037260030","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Declaration of Ashraf Fawzy","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491243848/","id":491243848,"tags":[],"absolute_url":"/docket/74659430/11/2/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-25T09:54:05.787425-07:00","date_modified":"2026-09-08T09:06:28.006279-07:00","sha1":"fdf701491ae24f0db3e0cf31f45d742e0e832fa5","page_count":26,"file_size":1190028,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.2.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.2.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 11-2 Filed 08/24/26 Page 1 of 26\n\n\n\n\n EXHIBIT 22\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page12of\n of15\n 26\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n\nNEURAL AI, LLC,\n\n Petitioner, Case No. 7:26-mc-00241\n\n v.\n [Underlying Case: USDC\nAMAZON.COM, INC., Western District of Texas No.\n 7:24-cv-00221-ADA-DTG]\n Respondent.\n\n\n\n\n NEURAL AI\u2019S MEMORANDUM IN SUPPORT OF ITS MOTION TO COMPEL\nCOMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY AMAZON.COM, INC.\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page23of\n of15\n 26\n\n\n\n\n TABLE OF CONTENTS\n\nA. FACTUAL BACKGROUND ....................................................................................................1\n\n 1. The Underlying Litigation ...................................................................................................1\n 2. The Rule 45 Subpoena to Amazon and Amazon\u2019s Initial Objections .................................2\n 3. NAI\u2019s Meet-and-Confer Efforts and Narrowing and Amazon\u2019s Continued\n Non-Compliance ..................................................................................................................2\n 4. Procedural History ...............................................................................................................4\nB. THE COURT HAS JURISDICTION OVER THIS DISPUTE BECAUSE\n THE PLACE OF COMPLIANCE IN AUSTIN IS PROPER. ..................................................4\n\nC. AMAZON MUST PRODUCE DOCUMENTS RESPONSIVE TO THE SUBPOENA. .........6\n\n 1. The subpoenaed materials are relevant and proportional to the needs of the case. .............7\n 2. Amazon\u2019s burden objections are unsupported. ....................................................................9\n 3. Amazon cannot continue to defer production with vague promises. .................................10\n\n\n\n\n i\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page34of\n of15\n 26\n\n\n\n\n TABLE OF AUTHORITIES\n\n Page(s)\n\nCases\n\n611 Carpenter LLC v. Atlantic Casualty Ins. Co.,\n 2024 WL 1977160 (W.D. Tex. April 30, 2024) ....................................................................7, 9\n\nConservation L. Found., Inc. v. Equilon Enters. LLC,\n No. CV 17-396-WES, 2025 WL 2821238 (D.R.I. Oct. 3, 2025) ..............................................5\n\nLinet Americas, Inc. v. Hill-Rom Holdings, Inc.,\n No. 21-cv-6890, 2025 WL 889579 (N.D. Ill. Jan. 27, 2025).....................................................6\n\nMeritage Homes, LLC v. AIG Specialty Ins. Co.,\n No. 1:23-MC-00944-DII, 2024 WL 221448 (W.D. Tex. Jan. 18, 2024) ...................................4\n\nPhila. Indem. Ins. Co. v. Odessa Family YMCA,\n No. 7:20-CV-00134-DC, 2020 WL 6484069 (W.D. Tex. June 26, 2020) ................................4\n\nTrs. of Bos. Univ. v. Everlight Elecs. Co.,\n No. 12-CV-11935-PBS, 2014 WL 12792496 (D. Mass. Sept. 8, 2014)................................5, 6\n\nVelocity Pat. LLC v. FCA US LLC,\n No. 13 CV 8419, 2017 WL 11893112 (N.D. Ill. Nov. 2, 2017) ................................................5\n\nWaller v. Jet Specialty, Inc.,\n No. 23-CV-00121-DC-RCG, 2024 WL 7050192 (W.D. Tex. Nov. 19, 2024) .........................7\n\nRules\n\nFederal Rule of Civil Procedure 26 ...........................................................................................7, 10\n\nFederal Rule of Civil Procedure 45 ....................................................................................... passim\n\n\n\n\n ii\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page45of\n of15\n 26\n\n\n\n\n Despite eight months of good-faith efforts from petitioner Neural AI, LLC (\u201cNAI\u201d) to\n\nnegotiate with third-party subpoena recipient and respondent Amazon.com, Inc. (\u201cAmazon\u201d),\n\nAmazon still has not produced a single document in response to the subpoena NAI served on\n\nOctober 15, 2025. During that eight-month period, NAI sought discovery directly from the\n\ndefendant in the underlying case NVIDIA Corporation (\u201cNVIDIA\u201d), used information learned\n\nfrom NVIDIA to try to guide Amazon\u2019s search for responsive documents, provided additional\n\nexplanation of the infringing technology, and ultimately narrowed its subpoena to only 9 priority\n\nrequests for production. Still, Amazon has not committed to producing a single document and\n\ninstead only agreed generally to investigate the existence of possibly responsive documents and\n\ninformation. That sort of investigation is something that should have occurred months ago when\n\nAmazon first received the subpoena. Its vague promises to search now\u2014eight months after the\n\nsubpoena was served and less than two months before fact discovery closes in the underlying\n\ncase\u2014is too little too late. NAI respectfully requests that the Court issue an order compelling\n\nAmazon to comply with the Rule 45 subpoena NAI served on October 15 and requiring production\n\nof documents responsive to NAI\u2019s nine requests for production by a date certain prior to the close\n\nof fact discovery in the underlying case.\n\nA. FACTUAL BACKGROUND\n\n 1. The Underlying Litigation\n\n The underlying action\u2014Neural AI, LLC v. NVIDIA Corporation 7:24-cv-00221-ADA-\n\nDTG (W.D. Tex.)\u2014involves claims of direct, indirect, and induced patent infringement by\n\nNVIDIA relating to U.S. Patent Nos. 8,648,867; RE49,461; and RE48,438 (the \u201cPatents-in-Suit\u201d).\n\nThe Patents-in-Suit teach systems and methods for GPU-accelerated computing technology. NAI\n\nalleges that NVIDIA\u2019s hardware (i.e., its GPUs and servers) and software (i.e., NeMo, TensorRT,\n\nand cuDNN) infringe the Patents-in-Suit and that NAI encourages its customers to combine those\n\n\n 1\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page56of\n of15\n 26\n\n\n\n\nproducts in an infringing manner. Amazon is one of NVIDIA\u2019s largest customers and a real-world\n\nintegrator of the accused GPU-acceleration hardware and software. NAI has subpoenaed multiple\n\nof NVIDIA\u2019s customers seeking documents in their unique possession to support its allegations of\n\nindirect and induced infringement. NAI seeks production of those documents prior to August 11,\n\n2026, the close of fact discovery in the underlying case.\n\n 2. The Rule 45 Subpoena to Amazon and Amazon\u2019s Initial Objections\n\n NAI served its Rule 45 subpoena on Amazon on October 15, 2026. NAI noticed the place\n\nof compliance at 100 Congress Avenue, Suite 2000, Austin, Texas 78701. See Portuguese Decl.,\n\nExhibit A at 6. NAI chose this place of compliance because Amazon has a significant presence\n\nand conducts business in Austin, Texas through its corporate office buildings at 11501 Alterra\n\nParkway, Austin, Texas 78758. See Portuguese Decl., Exhibit B. This Amazon corporate office is\n\nonly 12 miles away from the place of compliance, and Amazon currently has 1058 job listings for\n\nin-person roles at its Austin location. See Portuguese Decl., Exhibit C. The initial subpoena\n\ncontained 20 requests relating to Amazon\u2019s purchase, use, incorporation, sale, or development of\n\nproducts containing or depending on the accused NVIDIA hardware and software. Exhibit A. The\n\nrequests were limited in time to the relevant damages period in the underlying case, from\n\nSeptember 13, 2018 to the present, and limited in scope to U.S.-based or U.S.-directed activity. Id.\n\n Amazon served objections on November 13, 2025. See Portuguese Decl., Exhibit D.\n\nAmazon objected to the place of compliance because it was more than 100 miles from Amazon\u2019s\n\nheadquarters in Seattle, Washington. Id. at 3. As to the substance of the requests, Amazon refused\n\nto search for or produce documents responsive to any request for production. See generally id.\n\n 3. NAI\u2019s Meet-and-Confer Efforts and Narrowing and Amazon\u2019s Continued\n Non-Compliance\n\n The parties first met and conferred on November 19, 2025. At that time and Amazon\u2019s\n\n\n\n 2\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page67of\n of15\n 26\n\n\n\n\nrequest, NAI agreed to seek additional information from NVIDIA first. NVIDIA subsequently\n\nproduced documents and information confirming that Amazon is a significant NVIDIA customer\n\nbut containing gaps about how Amazon actually used itself or bundled, configured, and sold the\n\naccused products to its customers.\n\n Shortly after NAI had received documents confirming that Amazon was a significant\n\nNVIDIA customer and partner, Neural AI re-engaged Amazon on April 3, 2026, and the parties\n\nconferred again on April 13, 2026. On April 27, 2026, NAI substantially narrowed the subpoena\n\nto nine priority requests (Nos. 5, 7-10, 12-14, and 19). See Portuguese Decl., Exhibit E at 4-6. At\n\nthe same time, NAI provided more detailed descriptions of the accused functionality and the type\n\nof bundling of NVIDIA hardware and software NAI is interested in and, to help Amazon in its\n\nsearch for responsive information, identified the specific NVIDIA hardware products Amazon had\n\nacquired during the relevant period. The parties met and conferred again on May 21 but, as of that\n\nmeet and confer, Amazon still had done little to no investigation into how it uses the NVIDIA\n\nproducts it purchased or what responsive documents it may have.\n\n On June 4, 2026, Amazon\u2019s counsel wrote by email that Amazon was still \u201cin the process\n\nof making our way through the orgs\u201d and \u201cfiguring out whether and where they have the\n\ninformation.\u201d See Exhibit E at 1. But for the first time\u2014over 7 months after Amazon received the\n\nsubpoena\u2014Amazon finally stated it \u201chad the lay of the land\u201d and would \u201ccomplete a reasonable\n\nsearch and provide you with what we\u2019re able to find three weeks from tomorrow.\u201d Id. The parties\n\nconferred on June 5, at which time NAI learned that Amazon\u2019s commitment to search for\n\ndocuments was illusory. Amazon still did not know if the information it was compiling was\n\n\u201cgarbage or not\u201d and could not commit that it would actually produce any responsive documents\n\nat the conclusion of its three-week search.\n\n\n\n\n 3\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page78of\n of15\n 26\n\n\n\n\n 4. Procedural History\n\n On June 1, 2026, while the parties were continuing to meet and confer, NAI sent Amazon\n\na dispute chart pursuant to Section IV of the Court\u2019s March 5, 2025, Standing Order Governing\n\nProceedings (OGP)\u2014Patent Cases (\u201cOGP\u201d). Because Amazon was a third party to the underlying\n\ndispute, NAI asked Amazon to respond to the dispute chart in 7 days, rather than the 3 days\n\ncontemplated in the OGP. On June 9, 2026, NAI sent an updated dispute chart taking into account\n\nthe information Amazon provided on the parties\u2019 most recent meet and confer. Amazon completed\n\nits portion of the dispute chart on June 12, 2026. It objected to the dispute chart process and\n\njurisdiction as threshold issues and on the merits.\n\n NAI submitted the dispute chart to the Court on June 15, 2026. The Court held a hearing\n\non the dispute chart on June 17, 2026. At the hearing, the Court instructed NAI to file a motion to\n\ncompel, rather than use the dispute chart process. See Portuguese Decl., Exhibit F at 37:14-38:6.\n\nB. THE COURT HAS JURISDICTION OVER THIS DISPUTE BECAUSE THE\n PLACE OF COMPLIANCE IN AUSTIN IS PROPER.\n\n The District Court for the Western District of Texas is the proper court to resolve Neural\n\nAI\u2019s motion to compel because Rule 45 directs the serving party to seek an order compelling\n\nproduction in \u201cthe court for the district where compliance is required.\u201d Fed. R. Civ. P.\n\n45(d)(2)(B)(i); see also Meritage Homes, LLC v. AIG Specialty Ins. Co., No. 1:23-MC-00944-DII,\n\n2024 WL 221448, at *4 (W.D. Tex. Jan. 18, 2024); Phila. Indem. Ins. Co. v. Odessa Family YMCA,\n\nNo. 7:20-CV-00134-DC, 2020 WL 6484069, at *1 (W.D. Tex. June 26, 2020). The place of\n\ncompliance for the subpoena at issue is located at 100 Congress Ave., Ste. 2000, Austin, Texas\n\n78701, which is located within this District. This Court\u2019s jurisdiction, then, turns on whether the\n\nplace of compliance listed in the subpoena is proper. See Exhibit F at 32:3-5 (Amazon agrees with\n\nNAI that the analysis \u201cboils down to whether the place of compliance is correct.\u201d).\n\n\n\n 4\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page89of\n of15\n 26\n\n\n\n\n For document subpoenas, Rule 45(c)(2)(A) permits production \u201cat a place within 100 miles\n\nof where the person resides, is employed, or regularly transacts business in person.\u201d The issuing\n\nparty is not limited to selecting a place of compliance only within 100 miles of the recipient\u2019s\n\nheadquarters. See, e.g., Conservation L. Found., Inc. v. Equilon Enters. LLC, No. CV 17-396-\n\nWES, 2025 WL 2821238, at *1 (D.R.I. Oct. 3, 2025) (rejecting argument that the place where an\n\nentity \u201cregularly transactions business in person\u201d is limited to the corporate headquarters because\n\nit \u201cignores the plain language of the Rule.\u201d). Rule 45 could have stated such a narrow requirement,\n\nbut it did not. Instead, the Rule allows for a place of compliance within 100 miles of any location\n\nwhere the recipient transacts business in person. For a company like Amazon that conducts\n\nsignificant business nationally, a party issuing a subpoena has many choices.\n\n The place of compliance is not limited to a location where potential document custodians\n\nare located. First, such a rule is logically non-sensical because the party serving the subpoena\n\ncannot know where the custodians possessing relevant documents are located before serving the\n\nsubpoena. Rule 45 cannot require that a party serving a subpoena on a large, nation-wide company\n\nplay a guessing game with the compliance location and cross its fingers that the custodian with\n\ndocuments responsive to its subpoena is located near the place of compliance, rather than at a\n\nregional office across the country. Second, courts routinely reject this very argument. See, e.g.,\n\nVelocity Pat. LLC v. FCA US LLC, No. 13 CV 8419, 2017 WL 11893112, at *4 (N.D. Ill. Nov. 2,\n\n2017) (holding that the place of compliance was proper within 100 miles of any of the subpoena\n\ntarget\u2019s regional offices or facilities and rejecting argument that, \u201cregardless of its other locations,\u201d\n\nits headquarters was the \u201conly location where it stores\u201d requested documents); Trs. of Bos. Univ.\n\nv. Everlight Elecs. Co., No. 12-CV-11935-PBS, 2014 WL 12792496, at *3 (D. Mass. Sept. 8,\n\n2014) (holding that place of compliance in Boston was proper because Apple had two offices and\n\n\n\n\n 5\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed 06/24/26\n 08/24/26 Page 9\n 10ofof15\n 26\n\n\n\n\nfour retail locations in Massachusetts and rejecting Apple\u2019s argument that its headquarters and\n\nrelevant documents are in Cupertino, California). \u201cRule 45(c) says nothing about the location of\n\ndocuments subpoenaed.\u201d Trs. of Bos. Univ., 2014 WL 12792496, at *3.\n\n Amazon\u2019s argument that it can produce documents only within 100 miles of its Seattle,\n\nWashington headquarters makes even less sense given its role as one of \u201cthe largest cloud\n\ninfrastructure providers in the country.\u201d Exhibit F at 36:11-15. Information stored on the cloud,\n\nrather than on local hard drives, can be accessed from anywhere, including by Amazon\u2019s\n\nemployees in the Austin location. Given recent technology advances, largely driven by Amazon\u2019s\n\nown cloud business, it is a fiction that Amazon would physically produce documents at the Austin\n\naddress listed as the place of compliance; of course, Amazon will send responsive documents to\n\nNAI electronically. But even if Amazon had to produce physical documents, its Austin-based\n\nemployees could access the documents remotely, print them out, and deliver them only 12 miles\n\nfrom Amazon\u2019s Austin office to the place of compliance.\n\n Here, the Austin place of compliance satisfies Rule 45(c)(2)(A). Amazon regularly\n\ntransacts business in person in Austin through its corporate office and technology hub on Alterra\n\nParkway, only 12 miles away from the place of compliance. This corporate office location is not\n\nempty or dormant. Amazon\u2019s own job postings reflected 1058 open Austin positions as of June\n\n23, 2026. See Exhibit C; see also Linet Americas, Inc. v. Hill-Rom Holdings, Inc., No. 21-cv-6890,\n\n2025 WL 889579, at *4 (N.D. Ill. Jan. 27, 2025) (relying on the subpoena recipient\u2019s job postings\n\nfor positions in Chicago to find that a place of compliance within 100 miles of Chicago was\n\nproper). Because the subpoena\u2019s listed place of compliance in Austin is proper, this Court has\n\njurisdiction to resolve this motion to compel.\n\nC. AMAZON MUST PRODUCE DOCUMENTS RESPONSIVE TO THE\n SUBPOENA.\n\n\n\n 6\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page10\n 11of\n of15\n 26\n\n\n\n\n Federal Rule of Civil Procedure 26 provides that a party may obtain discovery regarding\n\nany nonprivileged matter that is relevant to the parties\u2019 claims or defenses and proportional to the\n\nneeds of the case. Fed. R. Civ. P. 26(b)(1). Where, as here, a non-party refuses discovery in\n\nresponse to a validly issued subpoena, Federal Rule of Civil Procedure 45 provides the Court for\n\nthe district where compliance is required with broad discretion to compel the production of\n\ndocuments and information from third parties. Fed. R. Civ. P. 45(d)(2)(B)(i); Waller v. Jet\n\nSpecialty, Inc., No. 23-CV-00121-DC-RCG, 2024 WL 7050192, at *1 (W.D. Tex. Nov. 19, 2024).\n\nOnce a party moving to compel discovery establishes that the materials are relevant or will lead to\n\nthe discovery of admissible evidence, the burden rests upon the nonparty resisting discovery to\n\nsubstantiate its objections. 611 Carpenter LLC v. Atlantic Casualty Ins. Co., 2024 WL 1977160,\n\nat *1 (W.D. Tex. April 30, 2024) (granting party\u2019s motion to compel non-party subpoena). The\n\nnon-party \u201cmust state with specificity the objection and how it relates to the particular request\n\nbeing opposed, and not merely that it is overly broad and burdensome.\u201d Id.\n\n 1. The subpoenaed materials are relevant and proportional to the needs of the\n case.\n\n Neural AI\u2019s narrowed requests seek documents that are directly relevant to proving how\n\nNVIDIA\u2019s accused GPU-acceleration technology is deployed and used in real-world systems.\n\nAmazon is one of NVIDIA\u2019s most significant customers and a large-scale integrator of the accused\n\nhardware and software. Publicly available information on Amazon\u2019s website indicate that Amazon\n\nand NVIDIA have a deep partnership, collaborating on the deployment of more than one million\n\nNVIDIA GPUs across Amazon Web Services Regions.1 See Portuguese Decl., Exhibit G at 1.\n\n The narrowed requests target nine specific categories of documents:\n\n\n\n1 NVIDIA\u2019s confidential documents also demonstrate the close relationship between\nNVIDIA and Amazon, but NAI cites only public documents to avoid the need for sealing.\n\n\n 7\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page11\n 12of\n of15\n 26\n\n\n\n\n\uf0b7 RFP 5 seeks documents sufficient to identify Amazon\u2019s products or services that depend on\n the accused technology.\n\uf0b7 RFPs 7, 8, 9 and 10 seek technical documents, source code, configuration files, development\n notes and other documents sufficient to show how Amazon\u2019s products or services implement,\n incorporate, use, integrate, invoke, or interact with the accused NVIDIA products.\n\uf0b7 RFP 12 seeks communications with NVIDIA relating to the setup, integration, customization,\n support, or use of the accused NVIDIA products.\n\uf0b7 RFP 13 seeks Amazon\u2019s internal documents or reports reflecting the benefits or business value\n derived from its use of the accused NVIDIA products.\n\uf0b7 RFP 14 seeks the revenue, usage, or subscription data for Amazon\u2019s products or services that\n relied on accused NVIDIA products.\n\uf0b7 RFP 19 seeks internal engineering documentation sufficient to show the design, development,\n or operation of Amazon\u2019s products that use, incorporate, or were developed in connection with\n accused NVIDIA products.\nSee Exhibit A. Each category is directly relevant to proving how the accused products operate in\n\ncommercial deployments and how Amazon\u2019s systems interact with NVIDIA\u2019s GPU-acceleration\n\nsoftware\u2014including CUDA, TensorRT, and PyTorch with CUDA. In fact, the Court already\n\ndetermined that these same requests are relevant in the context of a discovery dispute with Dell,\n\nanother NVIDIA customer. See Exhibit F at 22:4-20 (\u201cAs I see these requests for production, I do\n\nbelieve that they are targeted to relevant information. . . . I believe the documents that identify and\n\ninclude that information, at least to an extent, are relevant to the underlying lawsuit. And that same\n\nthought permeates through all of these.\u201d).\n\n These materials are also uniquely in Amazon\u2019s possession. NVIDIA has already confirmed\n\nthat it does not possess information about how its customers use its products. See Portuguese Decl.,\n\nExhibit H at 22:2-11. Internal integration materials, architecture documents, implementation\n\nartifacts, internal communications, and revenue and usage data showing real-world deployment of\n\nthe accused technology exist only in Amazon\u2019s files. Neural AI cannot obtain equivalent\n\ninformation from any other source.\n\n\n\n\n 8\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page12\n 13of\n of15\n 26\n\n\n\n\n The requests are proportional to the needs of the case. Neural AI has narrowed from a\n\nbroader initial set to nine priority requests. Neural AI further narrowed the focus to three\n\ncombinations of NVIDIA products: use of an NVIDIA GPU in combination with (1) an original,\n\ncustom, or modified version of PyTorch using CUDA; (2) TensorRT; and (3) applications that\n\nutilize PyTorch with CUDA or TensorRT. See Exhibit E at 5. Meanwhile, the temporal scope\n\n(September 13, 2018 to present) tracks the relevant damages period, and the requests are limited\n\nto U.S.-based or -directed activity. Given the importance of the issues at stake and the amount in\n\ncontroversy in the underlying patent infringement action, the narrowed requests are proportional.\n\n 2. Amazon\u2019s burden objections are unsupported.\n\n To start, Amazon\u2019s written objections to burden are inadequate because they do not \u201cstate\n\nwith specificity\u201d the burden Amazon would face in producing responsive documents. 611\n\nCarpenter LLC, 2024 WL 1977160, at *1. The boilerplate objections, absent evidence of burden,\n\ndo not show that the burden of complying with the subpoena is undue and cannot outweigh the\n\nrelevance of the discovery sought.\n\n Further, NAI has taken reasonable steps to minimize Amazon\u2019s burden. NAI spent months\n\npursuing information directly from NVIDIA to avoid the need to obtain the same from Amazon.\n\nFor example, NAI withdrew the initial RFP 1 (\u201cDocuments sufficient to identify all types of\n\nNVIDIA GPU-Acceleration Hardware purchased, acquired, or deployed by You.\u201d) because it\n\nobtained data regarding Amazon\u2019s purchases from NVIDIA itself. Further, even for requests for\n\ndocuments only within Amazon\u2019s possession, NAI prioritized its requests and agreed to narrow\n\nthe subpoena to only 9 RFPs, most of which are requests only for documents \u201csufficient to show\u201d\n\nthe requested information. At the same time, NAI provided Amazon with information that NAI\n\nthought would facilitate the investigation, including (1) a detailed explanation of the specific\n\nsoftware and hardware combinations that NAI alleges infringes and (2) a list of the accused\n\n\n 9\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page13\n 14of\n of15\n 26\n\n\n\n\nproducts Amazon purchased from NVIDIA during the relevant period so that Amazon could search\n\nfor information about those specific products. NAI also regularly offered on meet and confers that\n\nit was willing to discuss and work through any burden-related issues Amazon encountered in its\n\ninvestigation, but to this day, Amazon has never articulated a specific hardship in responding to\n\nthe subpoena, as opposed to general allegations that the subpoena requests are too broad.\n\n 3. Amazon cannot continue to defer production with vague promises.\n\n As described above, NAI has been patient and cooperative with Amazon. But the fact\n\ndiscovery deadline in the underlying case is now less than two months away. Amazon\u2019s vague\n\npromise to look into the matter and search for undefined documents\u2014made for the first time on\n\nJune 4, 2026\u2014is insufficient. NAI had no choice but to seek the Court\u2019s intervention. Given the\n\nupcoming discovery deadline, NAI suggests that the Court require Amazon to begin producing\n\ndocuments within 7 days of the Court\u2019s order on this motion and to complete production by no\n\nlater than Friday, July 24, 2026.\n\n For the foregoing reasons, NAI respectfully requests that this Court (1) overrule Amazon\u2019s\n\nplace-of-compliance objections and hold that this Court has jurisdiction to rule on this motion to\n\ncompel; (2) compel Amazon to produce non-privileged documents responsive to Neural AI\u2019s nine\n\npriority subpoena requests\u2014RFPs 5, 7-10, 12-14, and 19; (3) require Amazon to begin rolling\n\nproduction within 7 days of the Court\u2019s order on this motion and to complete production by Friday,\n\nJuly 24, 2026; and (4) require Amazon to serve a privilege log compliant with Fed. R. Civ. P.\n\n45(e)(2) and 26(b)(5) for any responsive materials withheld on privilege or work-product grounds,\n\nconcurrent with its final production.\n\n\n\n\n 10\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page14\n 15of\n of15\n 26\n\n\n\n\nDated: June 24, 2026\n\n\n Respectfully submitted,\n\n /s/ Mark Siegmund\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHNSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n\n\n\n 11\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page15\n 16of\n of15\n 26\n\n\n\n\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Petitioner Neural AI, LLC\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that a true and correct copy of the foregoing\ndocument was served on all parties via electronic mail on this 24th day of June 2026.\n\n /s/ Mark D. Siegmund\n Mark D. Siegmund\n\n\n\n\n 12\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page117\n of of\n 1026\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n NEURAL AI, LLC,\n\n Petitioner, Case No. 7:26-mc-00241\n\n v. [Underlying Case: USDC\n Western District of Texas No. 7:24-cv-\n AMAZON.COM, INC., 00221-ADA-DTG]\n\n Respondent.\n\n\n MEMORANDUM IN RESPONSE TO PETITIONER\u2019S MOTION TO COMPEL\nCOMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY AMAZON.COM, INC.\n I. INTRODUCTION\n\n Neural AI, LLC\u2019s (\u201cNAI\u201d) Motion should be denied for two reasons. First, the subpoena\n\nis invalid on its face because it demands a place of compliance that is not within 100 miles of\n\nAmazon\u2019s headquarters or relevant document custodians; i.e., not where any responsive\n\ndocuments will be found. Second, no disputes are ripe for decision. In response to non-party\n\nAmazon\u2019s objections to NAI\u2019s overbroad subpoena, the parties conferred and discussed Amazon\n\nproducing one thing and only one thing: a spreadsheet containing a report of the volume of accused\n\nNVIDIA GPUs configured in one of three allegedly relevant ways. During the June 17 hearing,\n\nAmazon confirmed it had agreed to complete its production of this information by June 26.\n\nHowever, two days prior to June 26, NAI filed its Motion. Amazon produced the volume\n\ninformation on June 26, mooting NAI\u2019s motion as it relates to the volume information.\n\n Prior to filing the Motion, NAI did not confer with Amazon regarding any other issues.\n\nNAI\u2019s failure to meet and confer on these issues violates Local Civil Rule CV-7(g) (meet and\n\nconfer requirement) and Fed. R. Civ. P. 45(d)(1) (NAI must \u201ctake reasonable steps to avoid\n\nimposing undue burden or expense on\u201d Amazon). In the June 17 hearing, the Court ordered NAI\n\nto confer with Dell to more precisely define various disputes on similar requests). NAI has no\n\nexcuse for not conferring here. The Court should deny NAI\u2019s motion.\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page218\n of of\n 1026\n\n\n\n\n II. FACTUAL BACKGROUND\n\n On October 15, 2025, NAI served its subpoena setting the place of compliance as Austin,\n\nTexas. Dkt. 1-3, at 2. 1 NAI\u2019s subpoena contained 20 requests amounting to discovery akin to that\n\nwhich would be served on a party in patent litigation. Dkt. 1-3. On November 13, 2025, Amazon\n\nobjected that NAI\u2019s requests were facially overly broad, unduly burdensome, and sought irrelevant\n\ninformation. Importantly, Amazon stated that it needed more information before it could even\n\nconduct a search. See generally Dkt. 1-6. Amazon further objected \u201cto the subpoena as improper\n\nbecause the demanded place of production is not within 100 miles of Seattle, Washington, where\n\nAmazon resides and regularly transacts business in person.\u201d Id. at 4.\n\n On November 19, 2025, the parties conferred, and Amazon \u201cexplained that we needed\n\nmore information to conduct a search[,]\u201d including at a minimum the identification of relevant\n\nproducts and the information NAI sought about those products that was unavailable from NVIDIA.\n\nDkt. 1-7, at 9. NAI did not dispute that Amazon needed this information, and explained that NAI\n\nwas \u201cstill pursuing discovery from defendant and [was] working to better identify the products\n\n[NVIDIA] supplied . . . that were relevant, as well as what [NAI] needed from [Amazon] about\n\nthose products that [NAI] couldn\u2019t get from defendant.\u201d Id. After months of silence, on April 3,\n\n2026, NAI asked Amazon for an update, including whether \u201cAmazon . . . will search for and\n\nproduce documents responsive to each of the subpoena requests.\u201d Id. at 10. Amazon reminded\n\nNAI that Amazon was awaiting the information NAI agreed to provide last November. Id. at 9.\n\nNamely, Amazon stated that \u201c[w]hen we last spoke five months ago, we explained that we needed\n\nmore information to conduct a search\u201d and NAI \u201cagreed to circle back with [Amazon] once you\n\nhad that information.\u201d Id.\n\n On April 13, 2026, the parties again conferred, and Amazon again reminded NAI\u2019s counsel\n\nwhat NAI had agreed to provide back in November 2025. Dkt. 1-7, at 6. On April 27, 2026, for\n\nthe first time, NAI provided Amazon with the listing of relevant products, and the configuration\n\n\n1\n NAI is plaintiff in the underlying patent dispute with defendant NVIDIA.\n 2\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page319\n of of\n 1026\n\n\n\n\nthat NAI contended was relevant in its litigation with NVIDIA 2. Id. As a result, NAI spent five\n\nand half months\u2014from when NAI served its subpoena on October 13, 2025, to April 27, 2026\u2014\n\nto provide Amazon the information needed to conduct a search. Once Amazon had the information\n\nneeded to start a search, it did so.\n\n On May 21, 2026, Amazon shared its findings from Amazon\u2019s preliminary investigation:\n\nthe accused computer chips were potentially used in many places across the company, and tracking\n\ndown each one, how it was used, and how it was configured was not realistically possible. Byer\n\nDeclaration of Benjamin J. Byer (\u201cByer Decl.\u201d) \u00b6 3. Namely, asking a cloud provider to track\n\ndown this information for many thousands of GPUs is akin to asking an automative company to\n\ntrack down every bolt and produce documents showing how each was used. Id. Amazon asked\n\nNAI whether NAI could narrow or focus the requests in any way. NAI refused and stated NAI\n\nwould simply go to the court rather than meaningfully confer. Id. Amazon nonetheless agreed to\n\nlook for ways to provide information about the volume of the accused products that were used in\n\none of the three identified configurations. Id. NAI requested that Amazon provide a date certain\n\nfor this production, but NAI did not request any other information. Id. Amazon explained that\n\nsince NAI had only just provided the information needed to begin the search, Amazon could not\n\nyet commit to when it would be completed.\n\n On June 1, 2026, rather than engage in any discussion, NAI simply sent Amazon a\n\ndiscovery dispute chart and demanded Amazon either respond to the chart or \u201cconfirm in writing\n\nby June 8 what categories of documents [Amazon] will agree to search for and produce. . . along\n\nwith the timeline for completing that production.\u201d Dkt. 1-7, at 3. On June 4, 2026, Amazon\n\nresponded that Amazon would agree to compete the reasonable search the parties had discussed\n\nand provide the results \u201cthree weeks from tomorrow\u201d (i.e., June 26, 2026). In a final effort to look\n\nfor a cooperative resolution, Amazon again requested a meet and confer. Id. at 2.\n\n\n\n2\n NAI also dropped certain requests to which Amazon had objected. Id. Dropping improper\nrequests, however, did not meaningfully narrow the subpoena.\n 3\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page420\n of of\n 1026\n\n\n\n\n On June 5, 2026, the parties conferred, and Amazon explained that based on a reasonable\n\nsearch Amazon did not have documents that provided company-wide visibility into the usage NAI\n\nsought, but offered as an alternative to conduct a search and compile into a spreadsheet data\n\nshowing for each accused NVIDIA GPUs the volume configured in one of the three ways NAI\n\nidentified as relevant. Byer Decl. \u00b6 4. When Amazon asked NAI\u2019s counsel whether NAI felt\n\nAmazon should be doing anything more, \u201cNAI was unable to identify anything it believed\n\n[Amazon] should be doing that it hadn\u2019t already agreed to do.\u201d Byer Decl., Ex. A, at 1. Although\n\nAmazon was \u201cconducting the search [the parties] discussed\u2014tracking down where, how, and in\nwhat volume the accused GPUs are used[,]\u201d NAI refused to withdraw its discovery dispute chart.\n\nByer Decl, Ex. A, at 2.\n\n On June 17, 2026, the Court held a hearing for the discovery disputes NAI had with Dell,\n\nMicrosoft, and Amazon, all of whom NAI had served with similar subpoenas. Dkt. 1-8. When\n\naddressing NAI\u2019s dispute with Dell, the Court stated that although some requests to Dell appeared\n\nto cover some relevant information, \u201cthe scope of . . . many of them is probably more broad than\n\nnecessary\u201d and that the requested discovery \u201ccould easily become disproportionate to the benefits\n\nthat would be obtained from the information.\u201d Byer Decl., Ex. B (June 17, 2026 Tr. of Disc. H\u2019rg\n\n(\u201cTranscript\u201d) at 22:6-7; 23:3-5). It therefore ordered Dell and NAI to meet and confer on each\n\nrequest to more precisely define their dispute. Byer Decl., Ex. B (Transcript at 24:4-6).\n\n At the hearing, Amazon\u2019s counsel explained it had agreed to produce by June 26 as the\n\nparties had discussed and confirmed Amazon\u2019s commitment to work with NAI if NAI felt\n\nsomething was missing. Byer Decl., Ex. B (Transcript at 34:10-35:22). The Court also agreed with\n\nAmazon that the discovery dispute statement was the improper mechanism to hear the dispute\n\ngiven Amazon\u2019s objection to it.\n\n On Wednesday, June 24, 2026\u2014two days before Amazon\u2019s agreed production\u2014NAI filed\n\nthis Motion. On June 26, 2026, Amazon made its agreed production. Byer Decl. \u00b6 5. NAI has\n\nneither requested a meet and confer nor contacted Amazon stating that it believes something is\n\nmissing from Amazon\u2019s production. Byer Decl. \u00b6 6.\n\n 4\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page521\n of of\n 1026\n\n\n\n\n III. ARGUMENT\n\n NAI\u2019s Motion should be denied for two reasons. First, the underlying subpoena fails to\n\nidentify a place of compliance that satisfies Rule 45\u2019s restriction. Second, NAI fails to present a\n\nripe dispute for this Court.\n\n A. NAI Has Moved to Compel in The Wrong District.\n\n Rule 45 permits a requesting party to set the place of compliance \u201cwithin 100 miles of\n\nwhere the person resides, is employed, or regularly transacts business in person.\u201d Fed. R. Civ. P.\n\n45(c)(2)(A). NAI does not argue Amazon.com, Inc. (the entity it subpoenaed) is a resident of Texas\nor that it is somehow employed there. NAI instead claims Amazon.com, Inc. \u201cregularly transacts\n\nbusiness\u201d within 100 miles of Austin because Amazon generally has a corporate office and\n\nunrelated job postings in Austin. Rule 45 does not permit a requesting party to paint with such\n\nbroad brush, forcing a non-party to produce documents at a location having no connection to the\n\ndocuments requested.\n\n Under Rule 45, the relevant business activities are those tethered to the location \u201c\u2018[1] where\n\nthe corporation is headquartered or [2] the custodian of records resides, is employed, or regularly\n\ntransacts business in person.\u2019\u201d Cleary v. Kaleida Health, 2024 WL 1297708, at *3 (W.D.N.Y.\n\n2024) (quoting 9 Moore\u2019s Federal Practice \u00a7 45.25[2] (Matthew Bender 3d ed.)); Europlay Cap.\n\nAdvisors, LLC v. Does, 323 F.R.D. 628, 629 (C.D. Cal. 2018) (requiring that the subpoena have a\n\nplace of compliance at the location where \u201ccustodians of records reside, are employed, and\n\nregularly transact business in person.\u201d). In other words, the question is not where an entity is\n\nsomehow employed or regularly transacted business in person, but where the custodians of records\n\nare employed or regularly transacts business in person. Europlay, 323 F.R.D. at 629 (emphasis\n\nadded) (analyzing whether the non-party\u2019s \u201ccustodians of records reside, are employed, and\n\nregularly transact business in person\u201d). This follows from a plain reading of Rule 45, as it makes\n\nno sense to consider an entity\u2019s employment or where it conducts business \u201cin person.\u201d\n\n As a result, courts have squarely rejected the argument that the existence of an office within\n\na district makes that district a proper Rule 45 place of compliance. Dellaportas v. Shahin, 2025\n\n 5\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page622\n of of\n 1026\n\n\n\n\nWL 3019810, at *1\u20132 (S.D.N.Y. Oct. 29, 2025), adhered to on reconsideration, 2025 WL 3456400\n\n(S.D.N.Y. Dec. 2, 2025) (finding that a motion to compel should be heard in the district where the\n\n\u201cemployees who would be responsible for searching and producing information responsive to\n\nPlaintiff\u2019s subpoena are.\u201d). NAI also suggests that a place of compliance is where a party has job\n\npostings. Mot. at 6 (citing Dkt. 1-5). This argument misses the point. NAI cites no evidence that\n\nany of those job postings are for Amazon.com, Inc. (the entity subpoenaed), much less evidence\n\nsuggesting Amazon.com, Inc.\u2019s custodians of relevant information are located in Texas. NAI has\n\ntherefore failed to carry its burden to show it has satisfied Rule 45 or that its subpoena is\nenforceable in this District. See Cruz v. AerSale, Inc., 2025 WL 1426884, at *4\u20136 (D.N.M. 2025)\n\n(denying motion to compel where the movant failed to establish the proper court of compliance).\n\n B. NAI Fails to Present a Ripe Dispute.\n NAI\u2019s discovery requests fall generally into two relevant buckets: (1) requests for\n\ninformation on the volume of accused NVIDIA GPUs configured in the relevant way (the \u201cvolume\n\ninformation\u201d), and (2) requests for additional technical documents, internal emails, internal\n\ndocuments, and Amazon financial information. NAI\u2019s motion should be denied on both fronts.\n\nAmazon has already produced the volume information the parties discussed during the meet and\n\nconfers. On the remaining requests, NAI has not conferred with Amazon on these requests, fails\n\nto satisfy its burden to show relevance, and fails to refute Amazon\u2019s objections.\n\n 1. Amazon Has Already Produced the Volume Material the Parties\n Discussed.\n On June 5, 2026, Amazon explained that based on a reasonable search Amazon did not\n\nhave responsive documents kept in the ordinary course of business that provided company wide\n\ndata on the configurations NAI for which sought discovery. Byer Decl. \u00b6 4. Nevertheless, Amazon\n\nagreed to go beyond its obligation under Rule 45 and search for and compile data to create a\n\ndocument that would identify the accused NVIDIA GPUs configured in the accused manner, and\n\ntheir volumes. Byer Decl., \u00b6 4. During the parties\u2019 numerous conferrals, NAI never requested\n\nAmazon do anything else. See Byer Decl., Ex. A, at 2. On June 26, 2026, Amazon produced that\n\n 6\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page723\n of of\n 1026\n\n\n\n\ninformation. Byer Decl. \u00b6 5. To the extent NAI believed additional information was called for,\n\nAmazon also offered to meet and confer with NAI to understand what, if anything, NAI contended\n\nwas needed from Amazon to address NAI\u2019s technical requests (e.g., RFPs 5, 7, 8, 9, 10, and 19).\n\nTo date, NAI has not identified any deficiency in Amazon\u2019s production of volume information or\n\nrequested a meet and confer. Thus, Amazon has already gone beyond its duties under Rule 45 to\n\ncreate documents that it does not maintain in the ordinary course of business, and there is no\n\npending dispute about the sufficiency of Amazon\u2019s production on the volume of accused NVIDIA\n\nGPU configured in one of the three allegedly relevant ways\u2014the only issue on which the parties\nhave conferred.\n\n 2. NAI fails to Present Any Ripe Dispute on the Remaining Requests.\n NAI\u2019s Motion also appears to request Amazon be compelled to provide additional, non-\n\ntechnical information. Although its Motion does not discuss such materials, RFP 12 requests\n\n\u201ccommunications between [Amazon] and NVIDIA\u201d relating to the accused NVIDIA products,\n\nRFP 13 requests Amazon\u2019s \u201cinternal documents\u201d regarding the Accused Products, and RFP 14\n\nseeks \u201crevenue, usage, or subscription data\u201d for Amazon services or software platforms related to\n\nthe Accused Products. Dkt. 1-3, at 18\u201320. Amazon timely objected to these requests because,\n\namong other things, they sought information not relevant to either party\u2019s claims or defenses.\n\nDkt. 1-5, at 17\u201320. To the extent NAI contends its motion covers such materials, 3 that portion of\n\nits Motion would fail for three reasons.\n\n First, the parties have never conferred on Amazon\u2019s objections. The parties conferred over\n\ntwo periods of time. First, shortly after NAI served its subpoena, the parties discussed the\n\ninformation Amazon needed to begin searching, and NAI agreed to provide that. More than five\n\nmonths after serving the subpoena, NAI finally did so by identifying the relevant GPUs and the\n\nthree configurations of those GPUs it claimed was relevant. Second, the parties reinitiated their\n\n\n3\n NAI appears to have withdrawn its similar requests against Dell, so it is unclear whether NAI\nintends to pursue these request here, particularly since NAI\u2019s Motion makes no attempt to justify\nthem. Byer Decl., Ex. C.\n 7\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page824\n of of\n 1026\n\n\n\n\nconferrals shortly after NAI\u2019s five-month delay. These discussions focused solely on the timing of\n\nAmazon producing a report showing the volume of accused GPUs configured in one of the accused\n\nways. See supra \u00a7 II.B. The parties never discussed requests for emails, financial information, non-\n\ntechnical documents, or any other topic in the nine RFPs NAI\u2019s Motion cites without discussion.\n\n Local Rule CV-7(g) states that the Court \u201cmay refuse to hear or may deny a nondispositive\n\nmotion unless the movant advises the court within the body of the motion that counsel for the\n\nparties have conferred in a good-faith attempt to resolve the matter by agreement and certifies the\n\nspecific reason that no agreement could be made.\u201d Conferring in good faith \u201cmeans that the parties\nmust genuinely attempt to resolve the dispute without judicial intervention, and not to treat their\n\nnegotiations simply as a formal prerequisite for judicial review.\u201d Perkins v. United States Parcel\n\nServ. of Am., Inc., 2024 WL 1493808, at *1\u20132 (W.D. Tex. 2024). Courts deny motions to compel\n\nfor failing to meet and confer because they do not precisely present a dispute to the court. Id.; see\n\nalso Diaz v. Cuatro T Constr., Inc., 2021 WL 2709681, at *1 (W.D. Tex. 2021). Because the\n\nparties have never conferred on those other document requests, the portion of NAI\u2019s Motion\n\naddressing them should be denied.\n\n Remarkably, NAI ignores this Court\u2019s directive on the very same issue during the June 17,\n\n2026 hearing as it relates to NAI\u2019s subpoena to Dell. Byer Decl., Ex. B (Transcript at 24). Namely,\n\nthe Court \u201corder[ed] [Dell and NAI] to meet and confer to address what information\u2014figure out\n\nwhat information is out there and what can be produced[.]\u201d Id. at 23:7-9. NAI now skips the step\n\nthe Court expressly ordered\u2014a conferral \u201caddressed discreetly to each request for production.\u201d\n\nByer Decl., Ex. B. (Transcript at 24:8-9). It is unclear why NAI believes it can skip the same meet\n\nand confer requirement with Amazon. Indeed, during its meet and confer with Dell, NAI narrowed\n\nthe scope of its requests to eliminate information it is still seeking here including emails and\n\nrevenue data. Byer Decl, Ex. C, at 9\u201311. This highlights the policy behind the meet and confer\n\nrequirement and the unnecessary burden NAI places on the Court by ignoring it.\n\n\n\n\n 8\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document\n 11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page925\n of of\n 1026\n\n\n\n\n Second, Amazon has specific objections to these requests as covering irrelevant\n\ninformation. For example, RFP 12 requests a broad set of communications that NAI has not\n\ndemonstrated are relevant or proportional to the issues in this case. As Amazon stated in its\n\nobjection, requests for email communications are particularly burdensome. Dkt 1-6, at 17\u201318;\n\nHedgeye Risk Mgmt., LLC v. Dale, 2023 WL 4353076, at *2 (S.D.N.Y. 2023) (citation omitted)\n\n(\u201call . . . communications \u201d that relate to multiple categories of records \u201cis often a red flag for\n\noverbreadth and undue burden.\u201d); Chinitz v. Realogy Holdings Corp., 2020 WL 6265083, at *3\n\n(W.D. Tex. 2020) (similarly denying a motion to compel requests for \u201call communications\u201d as\n\u201cfacially overbroad\u201d). Indeed, even in party discovery, \u201c[e]mail discovery is not presumptively\n\nrelevant to [patent] litigation,\u201d and \u201c\u2018overbroad email production requests, carry staggering time\n\nand production costs that have a debilitating effect on litigation.\u2019\u201d Hoist Fitness Sys., Inc. v.\n\nTuffStuff Fitness Int\u2019l, Inc., 2019 WL 121195, at *3 (C.D. Cal. 2019) (quoting Introduction to\n\nModel Order Regarding E-Discovery in Patent Cases, at p. 2 (Fed. Cir. 2011)); see also Standing\n\nOrder Governing Proceedings (OGP) 4.4-Patent Cases, at p. 3 (noting that \u201cthe Court will not\n\nrequire general search and production of email or other electronically stored information (ESI)\n\nrelated to email (such as metadata), absent a showing of good cause\u201d). Amazon likewise objects\n\nthat its internal documents and financial records not available in any hypothetical negotiation\n\nbetween NAI and NVIDIA have no relevance to the underlying lawsuit. SPH Am., LLC v. AT&T\n\nMobility, L.L.C., 2016 WL 11783677, at *2 (S.D. Cal. 2016) (denying motion to compel because\n\nthe party failed to show any relevance as to the Georgia-Pacific factors). Indeed, although NAI\n\nseeks to compel the production of financial information here, NAI dropped its request for financial\n\ndata during its conferrals with Dell. Byer Decl., Ex. C, at 10.\n\n Third, NAI admits as the moving party it has the burden to establish \u201cthat the materials are\n\nrelevant or will lead to the discovery of admissible evidence,\u201d Mot. at 7, but NAI has failed to\n\ncarry it. Hobbs v. Petroplex Pipe & Constr., Inc., 2018 WL 3603074, at *2 (W.D. Tex. 2018)\n\n(noting that the moving party has the burden to establish relevance). NAI provides a single\n\nsentence alleging that its requests are \u201cdirectly relevant to proving how NVIDIA\u2019s accused GPU-\n\n 9\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00241-LS Document\n Document11-2\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page1026ofof1026\n\n\n\n\nacceleration technology is deployed and used in real-world systems[,]\u201d and \u201cto proving how the\n\naccused products operate in commercial deployments and how Amazon\u2019s systems interact with\n\nNVIDIA\u2019s GPU-acceleration software\u2014including CUDA, TensorRT, and PyTorch with CUDA.\u201d\n\nMot. at 7. At most, this justifies its technical requests that Amazon has produced on. NAI is entirely\n\nsilent about how internal email, internal documents, of Amazon\u2019s financial information would\n\nhave any relevance. 4 NAI also suggests the Court already found all its requests to be relevant.\n\nMot. at 8. But NAI cherry picks quotes while ignoring the Court\u2019s ultimate finding and order that\n\nthe parties confer on the issues. Byer Decl., Ex. B (Transcript at 24). Because the Court did not\n\nrule on relevance, NAI cannot skirt its burden to show that relevance \u201cdiscreetly to each request\n\nfor production.\u201d Id.; El Paso Disposal, LP v. Ecube Labs Co., 2025 WL 1879607, at *3\u20136 (W.D.\n\nTex. 2025) (denying a motion to compel because moving party made \u201cconclusory arguments about\n\nrelevance\u201d without sufficient explanation or consideration for the burden on the non-party).\n\n IV. CONCLUSION\n\n For the reasons stated above, the Court should deny NAI\u2019s Motion.\n\n\n DATED this 1st day of July, 2026.\n /s/ Darryl J. Adams\n Darryl J. Adams (Texas Bar No. 00796101)\n Slayden Grubert Beard PLLC\n 401 Congress Ave., Ste. 1650\n Austin, TX 78701\n Tel: 512.402.3562\n Email: dadams@sgbfirm.com\n\n Attorney for Respondent Amazon.com, Inc.\n\n\n\n4\n Rule 45 \u201cprovides additional protections where a subpoena seeks trade secret or confidential\ncommercial information from a nonparty,\u201d requiring a higher showing of \u201csubstantial need.\u201d\nVinton Steel, LLC. v. Com. Metals Co., 2023 WL 2518881, at *3 (W.D. Tex. 2023) (citation and\ninternal quotation marks omitted). Amazon objected because NAI\u2019s requests asked for Amazon\u2019s\nconfidential documents and trade secrets, Dkt. 1-6, at 5, 18\u201321, but NAI does not address its\n\u201csubstantial need\u201d for the information.\n\n 10\n\f","ocr_status":2,"date_upload":"2026-08-25T09:56:11.877293-07:00","document_number":"11","attachment_number":2,"pacer_doc_id":"181037260031","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 22","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491243849/","id":491243849,"tags":[],"absolute_url":"/docket/74659430/11/3/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-25T09:54:05.806613-07:00","date_modified":"2026-09-08T09:09:54.250967-07:00","sha1":"3f98c6111d8527798faa3ac261cf0675c84fc406","page_count":27,"file_size":1128778,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.3.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.3.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 11-3 Filed 08/24/26 Page 1 of 27\n\n\n\n\n EXHIBIT 23\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page12of\n of15\n 27\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n\nNEURAL AI, LLC,\n\n Petitioner, Case No. 7:26-mc-00242\n\n v.\n [Underlying Case: USDC\nMICROSOFT CORPORATION, Western District of Texas No.\n 7:24-cv-00221-ADA-DTG]\n Respondent.\n\n\n\n\n NEURAL AI\u2019S MEMORANDUM IN SUPPORT OF ITS MOTION TO COMPEL\n COMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY MICROSOFT\n CORPORATION\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page23of\n of15\n 27\n\n\n\n\n TABLE OF CONTENTS\n\nA. FACTUAL BACKGROUND ....................................................................................................1\n 1. The Underlying Litigation ...................................................................................................1\n 2. The Rule 45 Subpoena to Microsoft and Microsoft\u2019s Initial Objections .............................2\n 3. NAI\u2019s Meet-and-Confer Efforts and Narrowing and Microsoft\u2019s Continued\n Non-Compliance ..................................................................................................................2\n 4. Procedural History ...............................................................................................................3\nB. THE COURT HAS JURISDICTION OVER THIS DISPUTE BECAUSE\n THE PLACE OF COMPLIANCE IN AUSTIN IS PROPER. ..................................................4\nC. MICROSOFT MUST PRODUCE DOCUMENTS RESPONSIVE TO THE\n SUBPOENA. .............................................................................................................................6\n 1. The subpoenaed materials are relevant and proportional to the needs of the case. .............7\n 2. Microsoft\u2019s burden objections are unsupported. ..................................................................9\n 3. Microsoft cannot continue to defer production with vague promises................................10\n\n\n\n\n i\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page34of\n of15\n 27\n\n\n\n\n TABLE OF AUTHORITIES\n\n Page(s)\n\nCases\n\n611 Carpenter LLC v. Atlantic Casualty Ins. Co.,\n 2024 WL 1977160 (W.D. Tex. April 30, 2024) ....................................................................7, 9\n\nConservation L. Found., Inc. v. Equilon Enters. LLC,\n No. CV 17-396-WES, 2025 WL 2821238 (D.R.I. Oct. 3, 2025) ..............................................5\n\nLinet Americas, Inc. v. Hill-Rom Holdings, Inc.,\n No. 21-cv-6890, 2025 WL 889579 (N.D. Ill. Jan. 27, 2025).....................................................6\n\nMeritage Homes, LLC v. AIG Specialty Ins. Co.,\n No. 1:23-MC-00944-DII, 2024 WL 221448 (W.D. Tex. Jan. 18, 2024) ...................................4\n\nPhiladelphia Indem. Ins. Co. v. Odessa Family YMCA,\n No. 7:20-CV-00134-DC, 2020 WL 6484069 (W.D. Tex. June 26, 2020) ................................4\n\nTrs. of Bos. Univ. v. Everlight Elecs. Co.,\n No. 12-CV-11935-PBS, 2014 WL 12792496 (D. Mass. Sept. 8, 2014)................................5, 6\n\nVelocity Pat. LLC v. FCA US LLC,\n No. 13 CV 8419, 2017 WL 11893112 (N.D. Ill. Nov. 2, 2017) ................................................5\n\nWaller v. Jet Specialty, Inc.,\n No. 23-CV-00121-DC-RCG, 2024 WL 7050192 (W.D. Tex. Nov. 19, 2024) .........................7\n\nRules\n\nFederal Rule of Civil Procedure 26 .......................................................................................6, 7, 10\n\nFederal Rule of Civil Proecdure 45 ....................................................................................... passim\n\n\n\n\n ii\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page45of\n of15\n 27\n\n\n\n\n Despite eight months of good-faith efforts from petitioner Neural AI, LLC (\u201cNAI\u201d) to\n\nnegotiate with third-party subpoena recipient and respondent Microsoft Corporation\n\n(\u201cMicrosoft\u201d), Microsoft still has not produced a single document in response to the subpoena NAI\n\nserved on October 15, 2025. During that eight-month period, NAI sought discovery directly from\n\nthe defendant in the underlying case NVIDIA Corporation (\u201cNVIDIA\u201d), used information learned\n\nfrom NVIDIA to try to guide Microsoft\u2019s search for responsive documents, provided additional\n\nexplanation of the infringing technology, and ultimately narrowed its subpoena to only 9 priority\n\nrequests for production. Still, Microsoft has not committed to producing a single document and\n\ninstead only agreed generally to investigate the existence of possibly responsive documents and\n\ninformation. That sort of investigation is something that should have occurred months ago when\n\nMicrosoft first received the subpoena. Its vague promises to search now\u2014eight months after the\n\nsubpoena was served and less than two months before fact discovery closes in the underlying\n\ncase\u2014is too little too late. NAI respectfully requests that the Court issue an order compelling\n\nMicrosoft to comply with the Rule 45 subpoena NAI served on October 15 and requiring\n\nproduction of documents responsive to NAI\u2019s nine requests for production by a date certain prior\n\nto the close of fact discovery in the underlying case.\n\nA. FACTUAL BACKGROUND\n\n 1. The Underlying Litigation\n\n The underlying action\u2014Neural AI, LLC v. NVIDIA Corporation 7:24-cv-00221-ADA-\n\nDTG (W.D. Tex.)\u2014involves claims of direct, indirect, and induced patent infringement by\n\nNVIDIA relating to U.S. Patent Nos. 8,648,867; RE49,461; and RE48,438 (the \u201cPatents-in-Suit\u201d).\n\nThe Patents-in-Suit teach systems and methods for GPU-accelerated computing technology. NAI\n\nalleges that NVIDIA\u2019s hardware (i.e., its GPUs and servers) and software (i.e., NeMo, TensorRT,\n\nand cuDNN) infringe the Patents-in-Suit and that NAI encourages its customers to combine those\n\n\n 1\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page56of\n of15\n 27\n\n\n\n\nproducts in an infringing manner. Microsoft is one of NVIDIA\u2019s largest customers and a real-\n\nworld integrator of the accused GPU-acceleration hardware and software. NAI has subpoenaed\n\nmultiple of NVIDIA\u2019s customers seeking documents in their unique possession to support its\n\nallegations of indirect and induced infringement. NAI seeks production of those documents prior\n\nto August 11, 2026, the close of fact discovery in the underlying case.\n\n 2. The Rule 45 Subpoena to Microsoft and Microsoft\u2019s Initial Objections\n\n NAI served its Rule 45 subpoena on Microsoft on October 15, 2026. NAI noticed the place\n\nof compliance at 100 Congress Avenue, Suite 2000, Austin, Texas 78701. See Portuguese Decl.,\n\nExhibit A at 6. NAI chose this place of compliance because Microsoft has a significant presence\n\nand conducts business in Austin, Texas. For example, Microsoft currently has 46 job listings for\n\nin-person roles at its Austin, Texas location. See Portuguese Decl., Exhibit B. The initial subpoena\n\ncontained 20 requests relating to Microsoft\u2019s purchase, use, incorporation, sale, or development of\n\nproducts containing or depending on the accused NVIDIA hardware and software. Exhibit A. The\n\nrequests were limited in time to the relevant damages period in the underlying case, from\n\nSeptember 13, 2018 to the present, and limited in scope to U.S.-based or U.S.-directed activity. Id.\n\n Microsoft served objections on November 13, 2025. See Portuguese Decl., Exhibit C.\n\nMicrosoft objected to the place of compliance because it was more than 100 miles from\n\nMicrosoft\u2019s headquarters in Seattle, Washington. Id. at 4. As to the substance of the requests,\n\nMicrosoft refused to search for or produce documents responsive to any request for production.\n\nSee generally id.\n\n 3. NAI\u2019s Meet-and-Confer Efforts and Narrowing and Microsoft\u2019s Continued\n Non-Compliance\n\n The parties first met and conferred on November 19, 2025. At that time and Microsoft\u2019s\n\nrequest, NAI agreed to seek additional information from NVIDIA first. NVIDIA subsequently\n\n\n\n 2\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page67of\n of15\n 27\n\n\n\n\nproduced documents and information confirming that Microsoft is a significant NVIDIA customer\n\nbut containing gaps about how Microsoft actually used itself or bundled, configured, and sold the\n\naccused products to its customers.\n\n Shortly after NAI had received documents confirming that Microsoft was a significant\n\nNVIDIA customer and partner, Neural AI re-engaged Microsoft on April 3, 2026, and the parties\n\nconferred again on April 13, 2026. On April 27, 2026, NAI substantially narrowed the subpoena\n\nto nine priority requests (Nos. 5, 7-10, 12-14, and 19). See Portuguese Decl., Exhibit D at 4-6. At\n\nthe same time, NAI provided more detailed descriptions of the accused functionality and the type\n\nof bundling of NVIDIA hardware and software NAI is interested in and, to help Microsoft in its\n\nsearch for responsive information, identified the specific NVIDIA hardware products Microsoft\n\nhad acquired during the relevant period. The parties met and conferred again on May 21 but, as of\n\nthat meet and confer, Microsoft still had done little to no investigation into how it uses the NVIDIA\n\nproducts it purchased or what responsive documents it may have.\n\n On June 4, 2026, Microsoft\u2019s counsel wrote by email that Microsoft was still \u201cin the\n\nprocess of making our way through the orgs\u201d and \u201cfiguring out whether and where they have the\n\ninformation.\u201d See Exhibit D at 1. But for the first time\u2014over 7 months after Microsoft received\n\nthe subpoena\u2014Microsoft finally stated it \u201chad the lay of the land\u201d and would \u201ccomplete a\n\nreasonable search and provide you with what we\u2019re able to find three weeks from tomorrow.\u201d Id.\n\nThe parties conferred on June 5, at which time NAI learned that Microsoft\u2019s commitment to search\n\nfor documents was illusory. Microsoft still did not know if the information it was compiling was\n\n\u201cgarbage or not\u201d and could not commit that it would actually produce any responsive documents\n\nat the conclusion of its three-week search.\n\n 4. Procedural History\n\n On June 1, 2026, while the parties were continuing to meet and confer, NAI sent Microsoft\n\n\n 3\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page78of\n of15\n 27\n\n\n\n\na dispute chart pursuant to Section IV of the Court\u2019s March 5, 2025, Standing Order Governing\n\nProceedings (OGP)\u2014Patent Cases (\u201cOGP\u201d). Because Microsoft was a third party to the underlying\n\ndispute, NAI asked Microsoft to respond to the dispute chart in 7 days, rather than the 3 days\n\ncontemplated in the OGP. On June 9, 2026, NAI sent an updated dispute chart taking into account\n\nthe information Microsoft provided on the parties\u2019 most recent meet and confer. Microsoft\n\ncompleted its portion of the dispute chart on June 12, 2026. It objected to the dispute chart process\n\nand jurisdiction as threshold issues and on the merits.\n\n NAI submitted the dispute chart to the Court on June 15, 2026. The Court held a hearing\n\non the dispute chart on June 17, 2026. At the hearing, the Court instructed NAI to file a motion to\n\ncompel, rather than use the dispute chart process. See Portuguese Decl., Exhibit E at 37:14-38:6.\n\nB. THE COURT HAS JURISDICTION OVER THIS DISPUTE BECAUSE THE\n PLACE OF COMPLIANCE IN AUSTIN IS PROPER.\n\n The District Court for the Western District of Texas is the proper court to resolve Neural\n\nAI\u2019s motion to compel because Rule 45 directs the serving party to seek an order compelling\n\nproduction in \u201cthe court for the district where compliance is required.\u201d Fed. R. Civ. P.\n\n45(d)(2)(B)(i); see also Meritage Homes, LLC v. AIG Specialty Ins. Co., No. 1:23-MC-00944-DII,\n\n2024 WL 221448, at *4 (W.D. Tex. Jan. 18, 2024); Philadelphia Indem. Ins. Co. v. Odessa Family\n\nYMCA, No. 7:20-CV-00134-DC, 2020 WL 6484069, at *1 (W.D. Tex. June 26, 2020). The place\n\nof compliance for the subpoena at issue is Planet Depos \u2013 Downtown Austin c/o Lexitas Legal,\n\n100 Congress Ave., Ste. 2000, Austin, Texas 78701, which is located within this District. This\n\nCourt\u2019s jurisdiction, then, turns on whether the place of compliance listed in the subpoena is\n\nproper. See Exhibit E at 32:3-5 (Microsoft agrees with NAI that the analysis \u201cboils down to\n\nwhether the place of compliance is correct.\u201d).\n\n\n\n\n 4\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page89of\n of15\n 27\n\n\n\n\n For document subpoenas, Rule 45(c)(2)(A) permits production \u201cat a place within 100 miles\n\nof where the person resides, is employed, or regularly transacts business in person.\u201d The issuing\n\nparty is not limited to selecting a place of compliance only within 100 miles of the recipient\u2019s\n\nheadquarters. See, e.g., Conservation L. Found., Inc. v. Equilon Enters. LLC, No. CV 17-396-\n\nWES, 2025 WL 2821238, at *1 (D.R.I. Oct. 3, 2025) (rejecting argument that the place where an\n\nentity \u201cregularly transactions business in person\u201d is limited to the corporate headquarters because\n\nit \u201cignores the plain language of the Rule.\u201d). Rule 45 could have stated such a narrow requirement,\n\nbut it did not. Instead, the Rule allows for a place of compliance within 100 miles of any location\n\nwhere the recipient transacts business in person. For a company like Microsoft that conducts\n\nsignificant business nationally, a party issuing a subpoena has many choices.\n\n The place of compliance is not limited to a location where potential document custodians\n\nare located. First, such a rule is logically non-sensical because the party serving the subpoena\n\ncannot know where the custodians possessing relevant documents are located before serving the\n\nsubpoena. Rule 45 cannot require that a party serving a subpoena on a large, nation-wide company\n\nplay a guessing game with the compliance location and cross its fingers that the custodian with\n\ndocuments responsive to its subpoena is located near the place of compliance, rather than at a\n\nregional office across the country. Second, courts routinely reject this very argument. See, e.g.,\n\nVelocity Pat. LLC v. FCA US LLC, No. 13 CV 8419, 2017 WL 11893112, at *4 (N.D. Ill. Nov. 2,\n\n2017) (holding that the place of compliance was proper within 100 miles of any of the subpoena\n\ntarget\u2019s regional offices or facilities and rejecting argument that, \u201cregardless of its other locations,\u201d\n\nits headquarters was the \u201conly location where it stores\u201d requested documents); Trs. of Bos. Univ.\n\nv. Everlight Elecs. Co., No. 12-CV-11935-PBS, 2014 WL 12792496, at *3 (D. Mass. Sept. 8,\n\n2014) (holding that place of compliance in Boston was proper because Apple had two offices and\n\n\n\n\n 5\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed 06/24/26\n 08/24/26 Page 9\n 10ofof15\n 27\n\n\n\n\nfour retail locations in Massachusetts and rejecting Apple\u2019s argument that its headquarters and\n\nrelevant documents are in Cupertino, California). \u201cRule 45(c) says nothing about the location of\n\ndocuments subpoenaed.\u201d Trs. of Bos. Univ., 2014 WL 12792496, at *3.\n\n Microsoft\u2019s argument that it can produce documents only within 100 miles of its Seattle,\n\nWashington headquarters makes even less sense given its role as one of \u201cthe largest cloud\n\ninfrastructure providers in the country.\u201d Exhibit E at 36:11-15. Information stored on the cloud,\n\nrather than on local hard drives, can be accessed from anywhere, including by Microsoft\u2019s\n\nemployees in the Austin location. Given recent technology advances, largely driven by Microsoft\u2019s\n\nown cloud business, it is a fiction that Microsoft would physically produce documents at the Austin\n\naddress listed as the place of compliance; of course, Microsoft will send responsive documents to\n\nNAI electronically. But even if Microsoft had to produce physical documents, its Austin-based\n\nemployees could access the documents remotely, print them out, and deliver them within the same\n\ncity from Microsoft\u2019s Austin office to the place of compliance.\n\n Here, the Austin place of compliance satisfies Rule 45(c)(2)(A). Microsoft regularly\n\ntransacts business in person in Austin through its corporate office in the same city as the place of\n\ncompliance. This corporate office location is not empty or dormant. Microsoft\u2019s own job postings\n\nreflected 46 open Austin positions as of June 23, 2026. See Exhibit B; see also Linet Americas,\n\nInc. v. Hill-Rom Holdings, Inc., No. 21-cv-6890, 2025 WL 889579, at *4 (N.D. Ill. Jan. 27, 2025)\n\n(relying on the subpoena recipient\u2019s job postings for positions in Chicago to find that a place of\n\ncompliance within 100 miles of Chicago was proper). Because the subpoena\u2019s listed place of\n\ncompliance in Austin is proper, this Court has jurisdiction to resolve this motion to compel.\n\nC. MICROSOFT MUST PRODUCE DOCUMENTS RESPONSIVE TO THE\n SUBPOENA.\n\n Federal Rule of Civil Procedure 26 provides that a party may obtain discovery regarding\n\n\n\n 6\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page10\n 11of\n of15\n 27\n\n\n\n\nany nonprivileged matter that is relevant to the parties\u2019 claims or defenses and proportional to the\n\nneeds of the case. Fed. R. Civ. P. 26(b)(1). Where, as here, a non-party refuses discovery in\n\nresponse to a validly issued subpoena, Federal Rule of Civil Procedure 45 provides the Court for\n\nthe district where compliance is required with broad discretion to compel the production of\n\ndocuments and information from third parties. Fed. R. Civ. P. 45(d)(2)(B)(i); Waller v. Jet\n\nSpecialty, Inc., No. 23-CV-00121-DC-RCG, 2024 WL 7050192, at *1 (W.D. Tex. Nov. 19, 2024).\n\nOnce a party moving to compel discovery establishes that the materials are relevant or will lead to\n\nthe discovery of admissible evidence, the burden rests upon the nonparty resisting discovery to\n\nsubstantiate its objections. 611 Carpenter LLC v. Atlantic Casualty Ins. Co., 2024 WL 1977160,\n\nat *1 (W.D. Tex. April 30, 2024) (granting party\u2019s motion to compel non-party subpoena). The\n\nnon-party \u201cmust state with specificity the objection and how it relates to the particular request\n\nbeing opposed, and not merely that it is overly broad and burdensome.\u201d Id.\n\n 1. The subpoenaed materials are relevant and proportional to the needs of the\n case.\n\n Neural AI\u2019s narrowed requests seek documents that are directly relevant to proving how\n\nNVIDIA\u2019s accused GPU-acceleration technology is deployed and used in real-world systems.\n\nMicrosoft is one of NVIDIA\u2019s most significant customers and a large-scale integrator of the\n\naccused hardware and software. Publicly available information on Microsoft\u2019s website indicate\n\nthat Microsoft and NVIDIA have a deep partnership, collaborating to \u201cenable[] more\n\nconversational AI solutions, further integrating hardware and software solutions and making AI\n\nmore accessible and easier to use.\u201d1 See Portuguese Decl., Exhibit F at 1.\n\n The narrowed requests target nine specific categories of documents:\n\n\n\n1 NVIDIA\u2019s confidential documents also demonstrate the close relationship between\nNVIDIA and Microsoft, but NAI cites only public documents to avoid the need for sealing.\n\n\n 7\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page11\n 12of\n of15\n 27\n\n\n\n\n\uf0b7 RFP 5 seeks documents sufficient to identify Microsoft\u2019s products or services that depend on\n the accused technology.\n\uf0b7 RFPs 7, 8, 9 and 10 seek technical documents, source code, configuration files, development\n notes and other documents sufficient to show how Microsoft\u2019s products or services implement,\n incorporate, use, integrate, invoke, or interact with the accused NVIDIA products.\n\uf0b7 RFP 12 seeks communications with NVIDIA relating to the setup, integration, customization,\n support, or use of the accused NVIDIA products.\n\uf0b7 RFP 13 seeks Microsoft\u2019s internal documents or reports reflecting the benefits or business\n value derived from its use of the accused NVIDIA products.\n\uf0b7 RFP 14 seeks the revenue, usage, or subscription data for Microsoft\u2019s products or services that\n relied on accused NVIDIA products.\n\uf0b7 RFP 19 seeks internal engineering documentation sufficient to show the design, development,\n or operation of Microsoft\u2019s products that use, incorporate, or were developed in connection\n with accused NVIDIA products.\nSee Exhibit A. Each category is directly relevant to proving how the accused products operate in\n\ncommercial deployments and how Microsoft\u2019s systems interact with NVIDIA\u2019s GPU-acceleration\n\nsoftware\u2014including CUDA, TensorRT, and PyTorch with CUDA. In fact, the Court already\n\ndetermined that these same requests are relevant in the context of a discovery dispute with Dell,\n\nanother NVIDIA customer. See Exhibit E at 22:4-20 (\u201cAs I see these requests for production, I do\n\nbelieve that they are targeted to relevant information. . . . I believe the documents that identify and\n\ninclude that information, at least to an extent, are relevant to the underlying lawsuit. And that same\n\nthought permeates through all of these.\u201d).\n\n These materials are also uniquely in Microsoft\u2019s possession. NVIDIA has already\n\nconfirmed that it does not possess information about how its customers use its products. See\n\nPortuguese Decl., Exhibit G at 22:2-11. Internal integration materials, architecture documents,\n\nimplementation artifacts, internal communications, and revenue and usage data showing real-\n\nworld deployment of the accused technology exist only in Microsoft\u2019s files. Neural AI cannot\n\nobtain equivalent information from any other source.\n\n\n\n\n 8\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page12\n 13of\n of15\n 27\n\n\n\n\n The requests are proportional to the needs of the case. Neural AI has narrowed from a\n\nbroader initial set to nine priority requests. Neural AI further narrowed the focus to three\n\ncombinations of NVIDIA products: use of an NVIDIA GPU in combination with (1) an original,\n\ncustom, or modified version of PyTorch using CUDA; (2) TensorRT; and (3) applications that\n\nutilize PyTorch with CUDA or TensorRT. See Exhibit D at 5. Meanwhile, the temporal scope\n\n(September 13, 2018 to present) tracks the relevant damages period, and the requests are limited\n\nto U.S.-based or -directed activity. Given the importance of the issues at stake and the amount in\n\ncontroversy in the underlying patent infringement action, the narrowed requests are proportional.\n\n 2. Microsoft\u2019s burden objections are unsupported.\n\n To start, Microsoft\u2019s written objections to burden are inadequate because they do not \u201cstate\n\nwith specificity\u201d the burden Microsoft would face in producing responsive documents. 611\n\nCarpenter LLC, 2024 WL 1977160, at *1. The boilerplate objections, absent evidence of burden,\n\ndo not show that the burden of complying with the subpoena is undue and cannot outweigh the\n\nrelevance of the discovery sought.\n\n Further, NAI has taken reasonable steps to minimize Microsoft\u2019s burden. NAI spent\n\nmonths pursuing information directly from NVIDIA to avoid the need to obtain the same from\n\nMicrosoft. For example, NAI withdrew the initial RFP 1 (\u201cDocuments sufficient to identify all\n\ntypes of NVIDIA [hardware] purchased, acquired, or deployed by You.\u201d) because it obtained data\n\nregarding Microsoft\u2019s purchases from NVIDIA itself. Further, even for requests for documents\n\nonly within Microsoft\u2019s possession, NAI prioritized its requests and agreed to narrow the subpoena\n\nto only 9 RFPs, most of which are requests only for documents \u201csufficient to show\u201d the requested\n\ninformation. At the same time, NAI provided Microsoft with information that NAI thought would\n\nfacilitate the investigation, including (1) a detailed explanation of the specific software and\n\nhardware combinations that NAI alleges infringes and (2) a list of the accused products Microsoft\n\n\n 9\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page13\n 14of\n of15\n 27\n\n\n\n\npurchased from NVIDIA during the relevant period so that Microsoft could search for information\n\nabout those specific products. NAI also regularly offered that it was willing to discuss and work\n\nthrough any burden-related issues Microsoft encountered in its investigation, but to this day,\n\nMicrosoft has never articulated a specific hardship in responding to the subpoena, as opposed to\n\ngeneral allegations that the subpoena requests are too broad.\n\n 3. Microsoft cannot continue to defer production with vague promises.\n\n As described above, NAI has been patient and cooperative with Microsoft. But the fact\n\ndiscovery deadline in the underlying case is now less than two months away. Microsoft\u2019s vague\n\npromise to look into the matter and search for undefined documents\u2014made for the first time on\n\nJune 4, 2026\u2014is insufficient. NAI had no choice but to seek the Court\u2019s intervention. Given the\n\nupcoming discovery deadline, NAI suggests that the Court require Microsoft to begin producing\n\ndocuments within 7 days of the Court\u2019s order on this motion and to complete production by no\n\nlater than Friday, July 24, 2026.\n\n For the foregoing reasons, NAI respectfully requests that this Court (1) overrule\n\nMicrosoft\u2019s place-of-compliance objections and hold that this Court has jurisdiction over this\n\nmotion; (2) compel Microsoft to produce non-privileged documents responsive to Neural AI\u2019s nine\n\npriority subpoena requests\u2014RFPs 5, 7-10, 12-14, and 19; (3) require Microsoft to begin rolling\n\nproduction within 7 days of the Court\u2019s order on this motion and to complete production by Friday,\n\nJuly 24, 2026; and (4) require Microsoft to serve a privilege log compliant with Fed. R. Civ. P.\n\n45(e)(2) and 26(b)(5) for any responsive materials withheld on privilege or work-product grounds,\n\nconcurrent with its final production.\n\n\n\n\n 10\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page14\n 15of\n of15\n 27\n\n\n\n\nDated: June 24, 2026\n\n\n Respectfully submitted,\n\n /s/ Mark Siegmund\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHNSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n\n\n\n 11\n\fCase\nCase7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 1-1 Filed\n Filed06/24/26\n 08/24/26 Page\n Page15\n 16of\n of15\n 27\n\n\n\n\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Petitioner Neural AI, LLC\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that a true and correct copy of the foregoing\n\n 7:26\n document was served on all parties via electronic mail on this 24th day of June 2026.\n\n /s/ Mark D. Siegmund\n Mark D. Siegmund\n\n -\n mc-\n 242\n\n\n\n 12\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page117\n of of\n 1127\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n NEURAL AI, LLC,\n\n Petitioner, Case No. 7:26-mc-00242\n\n v. [Underlying Case: USDC\n Western District of Texas No. 7:24-cv-\n MICROSOFT CORPORATION, 00221-ADA-DTG]\n\n Respondent.\n\n\n MEMORANDUM IN RESPONSE TO PETITIONER\u2019S MOTION TO COMPEL\n COMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY MICROSOFT\n CORPORATION\n I. INTRODUCTION\n\n Neural AI, LLC\u2019s (\u201cNAI\u201d) Motion should be denied because (1) the subpoena is invalid\n\non its face, having a place of compliance in violation of Rule 45, and (2) it fails to present a ripe\n\ndispute before this Court on any of the requests. First, Rule 45 required NAI to set the place of\n\ncompliance in the Western District of Washington, where Microsoft Corporation (\u201cMicrosoft\u201d) is\n\nheadquartered and has the required information. Failing to do so makes the subpoena invalid on\n\nits face, and unenforceable. Second, NAI does not present a ripe dispute before this Court.\n\nMicrosoft\u2019s initial objections stated that it needed additional information to conduct a search. On\nthe parties\u2019 initial meet and confer, NAI did not dispute that and agreed to provide that information.\n\nAfter five months of silence, NAI provided that information on April 27, 2026. In the short period\n\nof time between then and when NAI first sought to compel Microsoft, the parties on multiple\n\ninstances met and conferred and discussed Microsoft\u2019s production. Namely, based on a reasonable\n\nsearch, because Microsoft did not keep documents containing the information NAI sought,\n\nMicrosoft offered to compile data and produce NAI a table identifying the approximate volumes\n\nof accused chips configured in the relevant ways. The parties\u2019 subsequent discussion focused\n\nsolely on the timing of that production, not its adequacy. Microsoft has since made that production,\nthus mooting the only issues on which the parties have met and conferred. To the extent NAI\n\n 1\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page218\n of of\n 1127\n\n\n\n\nintends its Motion to cover any other request, it fails to carry its burden to establish relevance and\n\nfails to fulfill its obligation to meet and confer. Accordingly, Microsoft requests the Court deny\n\nNAI\u2019s Motion.\n\n II. FACTUAL BACKGROUND\n\n A. The Parties\u2019 Meet and Confers and Microsoft\u2019s Search.\n On October 15, 2025, NAI served its subpoena containing 20 requests amounting to\n\ndiscovery akin to that which would be served on a party in patent litigation. Dkt. 1-3. For example,\n\nrequest 16 sought \u201cMarketing, customer-facing, or internal communications describing or\n\nreferencing Your reliance on NVIDIA GPU-Acceleration Software and NVIDIA GPU-\n\nAcceleration Hardware or toolkits for performance, scalability, or innovation\u201d and request 12\n\nsought \u201cCommunications between You and NVIDIA relating to the setup, integration,\n\ncustomization, support, or use of any NVIDIA GPU-Acceleration Software.\u201d Id. at 19\u201320. The\n\nsubpoena set the place of compliance as Austin, Texas. Id. at 2. On November 13, 2025, Microsoft\n\nobjected that NAI\u2019s requests were facially overly broad and unduly burdensome, sought irrelevant\n\ninformation, and importantly stated that Microsoft needed more information before it could even\n\nconduct a search. See generally Dkt. 1-5. Microsoft further objected \u201cto the subpoena as improper\n\nbecause the demanded place of production is not within 100 miles of Redmond, Washington,\n\nwhere Microsoft resides and regularly transacts business in person.\u201d Dkt. 1-5, at 5.\n On November 19, 2025, the parties met and conferred, and Microsoft \u201cexplained that we\n\nneeded more information to conduct a search[,]\u201d including at a minimum the identification of\n\nrelevant products and the information NAI sought about those products that was unavailable from\n\nNVIDIA. Dkt. 1-6, at 9. NAI responded that it was \u201cstill pursuing discovery from defendant and\n\nwere working to better identify the products it supplied to . . . that were relevant, as well as what\n\n[it] needed from [Microsoft] about those products that [it] couldn\u2019t get from defendant.\u201d Id. On\n\nApril 3, 2026, NAI asked Microsoft for an update, including whether \u201cMicrosoft will search for\n\nand produce documents responsive to each of the subpoena requests.\u201d Dkt. 1-6, at 10. On April 6,\n2026, Microsoft reminded NAI that it was awaiting the information NAI agreed to provide months.\n\n 2\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page319\n of of\n 1127\n\n\n\n\nId. at 9 (Microsoft stating that \u201c[w]hen we last spoke five months ago, we explained that we needed\n\nmore information to conduct a search\u201d and NAI \u201cagreed to circle back with us once you had that\n\ninformation.\u201d).\n\n On April 13, 2026, the parties again met and conferred, and Microsoft again reminded\n\nNAI\u2019s counsel what it had agreed to provide the previous November. Dkt. 1-6, at 6. On April 27,\n\n2026, for the first time, NAI provided Microsoft with the listing of relevant products, and the\n\nconfiguration that it contended was relevant in its litigation with NVIDIA. Id. at 5\u20136.\n\n On May 21, 2026, the parties met and conferred, and Microsoft shared its findings from its\n\npreliminary investigation: the accused computer chips were potentially used in many places across\n\nthe company, and tracking down each one, how it was used, and how it was configured was not\n\nmeaningfully possible. Declaration of Benjamin J. Byer (\u201cByer Decl.\u201d) \u00b6 3. Namely, asking a\n\ncloud provider to track down many thousands of GPUs is akin to asking an automative company\n\nto track down every wrench and produce documents showing how each was used. Microsoft asked\n\nNAI whether it could narrow or focus its requests in any way. Id. It refused and stated it would\n\nsimply go to the court rather than meaningfully confer. Id. Microsoft nonetheless agreed to look\n\nfor ways to provide information about the approximate volume of the accused products that were\n\nused in one of the three identified configurations. Id. NAI requested it provide a date certain it\n\nwould have this volume information, but did not request any other information. Microsoft\nexplained that since NAI had only just provided the information needed to begin the search,\n\nMicrosoft could not commit to when it would be completed.\n\n B. NAI Files a Discovery Dispute In the Underlying Case and Then Files This\n Motion.\n On June 1, 2026, true to its word, rather than engage in any discussion, NAI simply sent\n\nMicrosoft a discovery dispute chart and demanded Microsoft either respond to the chart or\n\n\u201cconfirm in writing by June 8 what categories of documents [it] will agree to search for and\n\nproduce . . . along with the timeline for completing that production.\u201d Dkt. 1-6, at 2\u20133. On June 4,\n2026, Microsoft responded that it would agree to compete the reasonable search the parties had\n\n\n 3\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page420\n of of\n 1127\n\n\n\n\ndiscussed and provide the results \u201cthree weeks from tomorrow [i.e., June 26, 2026].\u201d Id. at 2. In a\n\nfinal effort to look for a cooperative resolution, Microsoft again requested a meet and confer. Id.\n\n On June 5, 2026, the parties met and conferred, and Microsoft explained that based on a\n\nreasonable search it did not have documents that provided company-wide visibility into the usage\n\nNAI sought, but offered as an alternative to conduct a search and compile into a spreadsheet data\n\nshowing for accused NVIDIA GPUs the approximate volume configured in one of the three\n\nconfirmations NAI identified as relevant. Byer Decl. \u00b6 4. When Microsoft asked NAI\u2019s counsel\n\nwhether it felt Microsoft should be doing anything more, \u201cNAI was unable to identify anything it\n\nbelieved [Microsoft] should be doing that it hadn\u2019t already agreed to do.\u201d Byer Decl., Ex. A, at 1.\n\nAlthough Microsoft was \u201cconducting the search [the parties] discussed\u2014tracking down where,\n\nhow, and in what volume the accused GPUs are used[,]\u201d NAI refused to withdraw its discovery\n\ndispute chart. Byer Decl., Ex. A, at 2.\n\n On June 17, 2026, the Court held a hearing for the discovery disputes NAI had with Dell,\n\nAmazon, and Microsoft, all of whom NAI had served with similar subpoenas. Dkt. 1-7. When\n\naddressing NAI\u2019s dispute with Dell, the Court stated that \u201cthe scope of some of [the requests] \u2013 in\n\nmany of them is probably more broad than necessary\u201d and that the requested discovery \u201ccould\n\neasily become disproportionate to the benefits that would be obtained from the information.\u201d Byer\n\nDecl., Ex. B (June 17, 2026 Tr. of Disc. H\u2019rg (\u201cTranscript\u201d) at 22:6-7; 23:3-5). It therefore ordered\nDell and NAI to meet and confer to more precisely define their dispute. Byer Decl., Ex. B\n\n(Transcript at 24:4-6).\n\n At the hearing, Microsoft\u2019s counsel explained it had agreed to produce by June 26 as the\n\nparties had discussed and confirmed its commitment to work with NAI if it felt it had missed\n\nsomething. Byer Decl., Ex. B (Transcript at 34:10-35:22). The Court also agreed with Microsoft\n\nthat the discovery dispute statement was the improper mechanism to hear the dispute given\n\nMicrosoft\u2019s objection to it. On Wednesday, June 24, 2026\u2014two days before Microsoft\u2019s agreed\n\nproduction\u2014NAI filed this Motion. On June 26, 2026, Microsoft made its agreed production. Byer\n\n\n\n 4\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page521\n of of\n 1127\n\n\n\n\nDecl. \u00b6 5. NAI has neither requested a meet and confer nor identified anything it believes is missing\n\nfrom Microsoft\u2019s production.\n\n III. ARGUMENT\n\n NAI\u2019s Motion should be denied for two reasons. First, the underlying subpoena fails to\n\nidentify a place of compliance that satisfies Rule 45\u2019s restriction. Second, it fails to present a ripe\n\ndispute for this Court.\n\n A. NAI Has Moved to Compel in The Wrong District.\n\n Rule 45 permits a requesting party to set the place of compliance \u201cwithin 100 miles of\n\nwhere the person resides, is employed, or regularly transacts business in person.\u201d Fed. R. Civ. P.\n\n45(c)(2)(A). NAI does not argue Microsoft is a resident of Texas or that it is somehow employed\n\nthere. It instead claims Microsoft \u201cregularly transacts business\u201d within 100 miles of Austin because\n\nit has a corporate office and unrelated job postings in Austin. Rule 45 does not permit a requesting\n\nparty to paint with such broad brush, forcing a non-party to produce documents at a location having\n\nno connection to the documents requested.\n\n Under Rule 45, the only relevant business activities are those tethered to the location of the\n\nnonparty\u2019s headquarters and the location where \u201ccustodians of records reside, are employed, and\n\nregularly transact business in person.\u201d Europlay Cap. Advisors, LLC v. Does, 323 F.R.D. 628, 629\n\n(C.D. Cal. 2018) (emphasis added) (motion to compel against Google to heard in the District where\nnon-party was headquartered and custodians of records reside in that District); see also Procaps\n\nS.A. v. Patheon Inc., 2015 WL 1722481, at *3 (S.D. Fla. 2015) (\u201cBecause [the non-party\n\ncorporation] is headquartered in Parsippany, N.J., [the place of compliance] is the District of New\n\nJersey.\u201d); Burnett v. Wahlburgers Franchising LLC, 2018 WL 10466827, at *2 (E.D.N.Y. 2018)\n\n(concluding that \u201cthe proper forum for the motion to compel would be in the district in California\n\nwhere the nonparty\u2019s headquarters are located, not where the files are to be produced.\u201d). Courts in\n\nTexas follow this rule. In re Xiaomi Tech. Netherlands B.V., 2025 WL 3068736, at *12 (E.D. Tex.\n\n2025) (following Europlay\u2019s analysis and concluding that compliance was proper at the non-\nparty\u2019s \u201cprincipal place of business\u201d). In other words, the question is not where an entity is\n\n 5\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page622\n of of\n 1127\n\n\n\n\nsomehow employed or regularly transacted business in person, but where the custodians of records\n\nare employed or regularly transacted business in person. This follows from a plain reading of Rule\n\n45, as it makes no sense to consider an entity\u2019s employment or where it conducts business \u201cin\n\nperson.\u201d\n\n As a result, courts have squarely rejected the argument that the existence of a Microsoft\n\noffice within a district makes that district a proper Rule 45 place of compliance, concluding instead\n\nthat compliance for Microsoft must be set in \u201cthe Western District of Washington.\u201d Dellaportas v.\n\nShahin, 2025 WL 3019810, at *1\u20132 (S.D.N.Y. Oct. 29, 2025), adhered to on reconsideration, 2025\n\nWL 3456400 (S.D.N.Y. Dec. 2, 2025) (finding that a motion to compel against Microsoft should\n\nbe heard in the district where the \u201cemployees who would be responsible for searching and\n\nproducing information responsive to Plaintiff\u2019s subpoena are.\u201d). NAI argues that a place of\n\ncompliance is where a party has job postings. Mot. at 6 (citing Dkt. 1-4). This argument misses\n\nthe point. NAI cites no evidence that any of those job posting suggest custodians of relevant\n\ninformation are located in Texas. NAI has failed to carry its burden to show it has satisfied Rule\n\n45 and its subpoena is enforceable in this District. See Cruz v. AerSale, Inc., 2025 WL 1426884,\n\nat *4\u20136 (D.N.M. 2025) (denying motion to compel where the movant failed to establish the court\n\nof compliance).\n\n B. NAI Fails to Present a Ripe Dispute.\n NAI\u2019s discovery requests fall generally into two buckets: (1) requests for technical\n\ninformation regarding Microsoft\u2019s use of the accused GPUs, and (2) requests for internal emails,\n\ninternal documents, and Microsoft financial information. NAI\u2019s motion should be denied on both\n\nfronts. Microsoft has already produced the technical information the parties discussed during the\n\nmeet and confers. On the remaining requests, NAI fails to satisfy its burden to show relevance,\n\nfails to refute Microsoft\u2019s objections, and the parties have never met and conferred on these\n\nrequests.\n\n\n\n\n 6\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page723\n of of\n 1127\n\n\n\n\n 1. Microsoft Has Already Produced the Sole Technical Material the\n Parties Discussed.\n On June 5, 2026, Microsoft explained that based on a reasonable search it did not have\n\nresponsive documents kept in the ordinary course of business that provided company wide data on\n\nthe configurations NAI sought discovery on. Byer Decl. \u00b6 4. But rather than stand on its objections,\n\nMicrosoft agreed to satisfy NAI\u2019s technical requests by instead searching for and compiling data\n\nto create a document that would identify the accused NVIDIA GPUs used in the accused manner,\n\nand their relevant volumes. Id. NAI never requested Microsoft do anything else. Byer Decl., Ex.\n\nA, at 1. On June 26, 2026, Microsoft produced that information. Byer Decl. \u00b6 5. To the extent NAI\nbelieved additional information was called for, Microsoft also offered to meet and confer with NAI\n\nto understand what, if anything, it contended it needed from Microsoft to address its technical\n\nrequests (e.g., RFPs 5, 7, 8, 9, 10, and 19). NAI has neither identified any deficient in Microsoft\u2019s\n\nproduction, or requested a meet and confer. Byer Decl. \u00b6 6. Although Microsoft remains willing\n\nto do so, any dispute falling from such a hypothetical future meet and confer falls outside a motion\n\nNAI strategically filed before Microsoft\u2019s production.\n\n 2. NAI fails to Present Any Ripe Dispute on the Remaining Requests.\n NAI\u2019s Motion also appears to request Microsoft be compelled to provide additional, non-\n\ntechnical information. Although its motion does not discuss such materials, RFP 12 requests\n\n\u201ccommunications between [Microsoft] and NVIDIA\u201d relating to the accused NVIDIA products,\n\nRFP 13 requests Microsoft\u2019s \u201c[i]nternal documents\u201d regarding the Accused Products, and RFP 14\nseeks \u201c[r]evenue, usage, or subscription data\u201d for Microsoft services or software platforms related\n\nto the Accused Products. Dkt. 1-3, at 18\u201320. Microsoft timely objected to these requests because,\n\namong other things, they sought information not relevant to either party\u2019s claims or defenses.\n\nDkt. 1-5, at 19-20. To the extent NAI contends its motion covers such materials,1 that portion of\n\nits Motion would fail for three reasons.\n\n\n1\n NAI appears to have withdrawn its similar requests against Dell, so it is unclear whether it intends\nto pursue these requests here, particularly since its Motion does not acknowledge they exist, much\nless make any attempt to justify them. Byer Decl., Ex. C.\n 7\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page824\n of of\n 1127\n\n\n\n\n First, Microsoft has specific objections to these requests as covering irrelevant information.\n\nFor example, RFP 12 requests a broad set of communications that NAI has not demonstrated are\n\nrelevant or proportional to the issues in this case. As Microsoft stated in its objection, requests for\n\nemail communications are particularly burdensome. Dkt 1-5, at 18-19; Hedgeye Risk Mgmt., LLC\n\nv. Dale, 2023 WL 4353076, at *2 (S.D.N.Y. 2023) (\u201call . . . communications\u201d that relate to multiple\n\ncategories of records \u201coften is a red flag for overbreadth and undue burden.\u201d (citation omitted));\n\nChinitz v. Realogy Holdings Corp., 2020 WL 6265083, at *3 (W.D. Tex. 2020) (similarly denying\n\na motion to compel requests for \u201call communications\u201d as \u201cfacially overbroad\u201d). Indeed, even in\n\nparty discovery, \u201c[e]mail discovery is not presumptively relevant to [patent] litigation,\u201d and\n\n\u201c\u2018overbroad email production requests, carry staggering time and production costs that have a\n\ndebilitating effect on litigation.\u2019\u201d Hoist Fitness Sys., Inc. v. TuffStuff Fitness Int\u2019l, Inc., 2019 WL\n\n121195, at *3 (C.D. Cal. 2019) (quoting Introduction to Model Order Regarding E-Discovery in\n\nPatent Cases at p. 2 (Fed. Cir. 2011)); see also Standing Order Governing Proceedings (OGP) 4.4-\n\nPatent Cases, p. 3 (noting that \u201cthe Court will not require general search and production of email\n\nor other electronically stored information (ESI) related to email (such as metadata), absent a\n\nshowing of good cause.\u201d). Microsoft likewise objects that its internal documents and financial\n\nrecords that would post-date any hypothetical negotiation of a royalty between NAI and NVIDIA\n\nhave no relevance to the underlying lawsuit. SPH Am., LLC v. AT&T Mobility, L.L.C., 2016 WL\n11783677, at *2 (S.D. Cal. 2016) (denying motion to compel because the party failed to show any\n\nrelevance as to the Georgia-Pacific factors).\n\n Second, the parties have never met and conferred on Microsoft\u2019s objections. All of the\n\nparties meet and confers focused solely on the timing of Microsoft producing a report showing the\n\napproximate volume of accused GPUs configured in the accused way. See supra \u00a7 II.B. Local Rule\n\nCV-7(g) states that the Court \u201cmay refuse to hear or may deny a nondispositive motion unless the\n\nmovant advises the court within the body of the motion that counsel for the parties have conferred\n\nin a good-faith attempt to resolve the matter by agreement and certifies the specific reason that no\nagreement could be made.\u201d Conferring in good faith \u201cmeans that the parties must genuinely\n\n 8\n\f Case\n Case\n 7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document\n 11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page925\n of of\n 1127\n\n\n\n\nattempt to resolve the dispute without judicial intervention, and not to treat their negotiations\n\nsimply as a formal prerequisite for judicial review.\u201d Perkins v. United States Parcel Serv. of Am.,\n\nInc., 2024 WL 1493808, at *1\u20132 (W.D. Tex. 2024) (citation omitted). Courts deny motions to\n\ncompel for failing to meet and confer because they do not precisely present a dispute to the court.\n\nId.; see also Diaz v. Cuatro T Constr., Inc., 2021 WL 2709681, at *1 n.1 (W.D. Tex. 2021).\n\nBecause the parties have never conferred on those other document requests, the portion of NAI\u2019s\n\nMotion addressing them should be denied.\n\n Third, NAI admits as the moving party it has the burden to establish \u201cthat the materials are\n\nrelevant or will lead to the discovery of admissible evidence,\u201d Mot. at 7, but it has failed to carry\n\nits burden. Hobbs v. Petroplex Pipe & Constr., Inc., 2018 WL 3603074, at *2 (W.D. Tex. 2018)\n\n(noting that the moving party has the burden to establish relevance). NAI\u2019s sole relevance\n\nargument is a single sentence alleging that its requests are \u201cdirectly relevant to proving how\n\nNVIDIA\u2019s accused GPU-acceleration technology is deployed and used in real-world systems[,]\u201d\n\nand \u201cto proving how the accused products operate in commercial deployments and how\n\nMicrosoft\u2019s systems interact with NVIDIA\u2019s GPU-acceleration software\u2014including CUDA,\n\nTensorRT, and PyTorch with CUDA.\u201d Mot. at 7. At most, this justifies its technical requests that\n\nMicrosoft has produced on. NAI is entirely silent about how internal email, internal documents, of\n\nMicrosoft\u2019s financial information would have any relevance.2 Although NAI cites the RFPs that\nencompass those documents, it does not identify what they cover or make any effort to carry its\n\nburden to establish relevance. To the extent NAI intends to move on requests encompassing\n\ninternal documents, email, and financial information, this portion of its motion should also be\n\ndenied. El Paso Disposal, LP v. Ecube Labs Co., 2025 WL 1879607, at *3\u20136 (W.D. Tex. 2025)\n\n\n2\n Rule 45 \u201cprovides additional protections where a subpoena seeks trade secret or confidential\ncommercial information from a nonparty,\u201d requiring a higher showing of \u201csubstantial need.\u201d\nVinton Steel, LLC. v. Com. Metals Co., 2023 WL 2518881, at *3 (W.D. Tex. 2023) (citation and\ninternal quotation marks omitted). Microsoft objected because NAI\u2019s requests asked for\nMicrosoft\u2019s confidential documents and trade secrets, Dkt. 1-5, at 5, 18\u201321, but NAI does not\naddress its \u201csubstantial need\u201d for the information.\n\n 9\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page1026ofof1127\n\n\n\n\n(denying a motion to compel because moving party made \u201cconclusory arguments about relevance\u201d\n\nwithout sufficient explanation or consideration for the burden on the non-party).\n\n IV. CONCLUSION\n\n For the reasons stated above, the Court should deny NAI\u2019s Motion.\n\n\n DATED this 1st day of July, 2026.\n\n Davis Wright Tremaine LLP\n Attorneys for Microsoft Corporation\n\n\n By: /s/ Andrew T. Gorham\n Andrew Thompson (\u201cTom\u201d) Gorham\n State Bar No. 24012715\n GILLAM & SMITH, L.L.P.\n 7232 Crosswater Avenue\n Tyler, Texas 75703\n Telephone: (903) 934-8450\n Facsimile: (903) 934-9257\n Email: tom@gillamsmithlaw.com\n\n Ben Byer (pro hac vice forthcoming)\n WSBA # 38206\n Angelo Marchesini (pro hac vice forthcoming)\n WSBA # 57051\n\n\n\n\n 10\n\f Case\n Case7:26-mc-00318-LS\n 7:26-mc-00242-LS Document\n Document11-3\n 7 Filed\n Filed07/01/26\n 08/24/26 Page\n Page1127ofof1127\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned hereby certifies that a true and correct copy of the above and foregoing\n\ndocument has been served on this the 1st day of July, 2026 to all counsel of record who are\n\ndeemed to have consented to electronic service via the Court\u2019s CM/ECF system.\n\n\n\n /s/ Andrew T. Gorham\n\n\n\n\n 11\n\f","ocr_status":2,"date_upload":"2026-08-25T09:56:14.331494-07:00","document_number":"11","attachment_number":3,"pacer_doc_id":"181037260032","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 23","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491243850/","id":491243850,"tags":[],"absolute_url":"/docket/74659430/11/4/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-25T09:54:05.826714-07:00","date_modified":"2026-09-08T09:12:28.266928-07:00","sha1":"165dfe722d01ac3c577708603bae106e0d718e4b","page_count":1,"file_size":128514,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.4.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.11.4.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 11-4 Filed 08/24/26 Page 1 of 1\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\nNEURAL AI, LLC,\n\n Petitioner,\n Misc. Case No. 7:26-mc-00318-LS\n v.\n\nTESLA, INC.,\n\n Respondent.\n\n\n\n [PROPOSED] ORDER DENYING NEURAL AI, LLC\u2019S\n MOTION TO COMPEL COMPLIANCE WITH SUBPOENA SERVED ON\n THIRD-PARTY TESLA, INC. AND GRANTING TESLA, INC.\u2019S\n CROSS MOTION TO QUASH NEURAL AI, LLC\u2019S SUBPOENAS\n\n Before the Court is Neural AI, LLC\u2019s Motion to Compel Compliance with Subpoena\n\n Served on Third-Party Tesla, Inc. (\u201cMotion\u201d) and Tesla\u2019s Cross Motion to Quash Neural AI\n\n LLC\u2019s Subpoenas (\u201cCross Motion\u201d). Having considered the parties\u2019 briefing, relevant facts,\n\n and applicable law, the Court finds that Neural AI\u2019s Motion should be and hereby is DENIED\n\n and Tesla\u2019s Cross Motion should be and hereby is GRANTED.\n\n IT IS FURTHER ORDERED that Neural AI\u2019s subpoena to produce documents and\n\n subpoena to testify at a deposition served on Tesla are QUASHED.\n\n\n\n\nSO ORDERED: this ____ day of ___________, 2026\n\n\n\n\n United States District Judge\n\n\n\n\n 1\n\f","ocr_status":2,"date_upload":"2026-08-25T09:56:14.158862-07:00","document_number":"11","attachment_number":4,"pacer_doc_id":"181037260033","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Proposed Order","acms_document_guid":""}],"date_created":"2026-08-24T17:06:26.316478-07:00","date_modified":"2026-08-25T09:53:48.454202-07:00","date_filed":"2026-08-24","time_filed":"18:51:39","entry_number":11,"recap_sequence_number":"2026-08-24.001","pacer_sequence_number":39,"description":"Response in Opposition to Motion, filed by Tesla Inc., re 6 CORRECTED MOTION to Compel Compliance With Subpoena Served on Third Party Tesla, Inc. filed by Petitioner Neural AI, LLC (Attachments: # 1 Declaration of Ashraf Fawzy, # 2 Exhibit 22, # 3 Exhibit 23, # 4 Proposed Order)(Zheng, Jun) (Entered: 08/24/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/475591892/","id":475591892,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491128820/","id":491128820,"tags":[],"absolute_url":"/docket/74659430/10/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-24T13:19:39.967896-07:00","date_modified":"2026-08-24T13:19:39.980710-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"10","attachment_number":null,"pacer_doc_id":"181037257324","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Notice of Attorney Appearance","acms_document_guid":""}],"date_created":"2026-08-24T13:19:39.884089-07:00","date_modified":"2026-08-24T15:43:41.623141-07:00","date_filed":"2026-08-24","time_filed":"15:11:18","entry_number":10,"recap_sequence_number":"2026-08-24.003","pacer_sequence_number":36,"description":"NOTICE of Attorney Appearance by Ashraf Fawzy on behalf of Tesla Inc.. Attorney Ashraf Fawzy added to party Tesla Inc.(pty:res) (Fawzy, Ashraf) (Entered: 08/24/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/475587556/","id":475587556,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491124457/","id":491124457,"tags":[],"absolute_url":"/docket/74659430/8/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-24T13:04:24.656414-07:00","date_modified":"2026-08-24T13:04:24.667144-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"8","attachment_number":null,"pacer_doc_id":"181037257070","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Notice of Attorney Appearance","acms_document_guid":""}],"date_created":"2026-08-24T13:04:24.612130-07:00","date_modified":"2026-08-24T15:43:41.577037-07:00","date_filed":"2026-08-24","time_filed":"14:55:32","entry_number":8,"recap_sequence_number":"2026-08-24.001","pacer_sequence_number":30,"description":"NOTICE of Attorney Appearance by Jun Zheng on behalf of Tesla Inc.. Attorney Jun Zheng added to party Tesla Inc.(pty:res) (Zheng, Jun) (Entered: 08/24/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/475587481/","id":475587481,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/491124382/","id":491124382,"tags":[],"absolute_url":"/docket/74659430/9/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-24T13:04:18.210576-07:00","date_modified":"2026-08-24T13:04:18.231249-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"9","attachment_number":null,"pacer_doc_id":"181037257138","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Notice of Attorney Appearance","acms_document_guid":""}],"date_created":"2026-08-24T13:04:18.152042-07:00","date_modified":"2026-08-24T15:43:41.607864-07:00","date_filed":"2026-08-24","time_filed":"14:58:41","entry_number":9,"recap_sequence_number":"2026-08-24.002","pacer_sequence_number":33,"description":"NOTICE of Attorney Appearance by Gina Cremona on behalf of Tesla Inc.. Attorney Gina Cremona added to party Tesla Inc.(pty:res) (Cremona, Gina) (Entered: 08/24/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/475263590/","id":475263590,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490790630/","id":490790630,"tags":[],"absolute_url":"/docket/74659430/7/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-20T15:04:59.174066-07:00","date_modified":"2026-08-23T04:40:01.110342-07:00","sha1":"ffa9595ca9a7c420571e88d8743ed70a6b470c28","page_count":1,"file_size":128474,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.7.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.7.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 7 Filed 08/20/26 Page 1 of 1\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n NEURAL AI, LLC, \u00a7\n Plaintiff, \u00a7\n \u00a7\n v. \u00a7\n \u00a7 NO: MO:26-MC-00318\n TESLA INC., \u00a7 (Principal Case: MO:24-CV-221-LS-DTG)\n Defendant. \u00a7\n \u00a7\n\n\n\n ORDER TO TRANSFER\n\n The above-named and numbered case is transferred from the docket of U.S. District\nJudge David Counts to the docket of U.S. District Judge Leon Schydlower both judges having\nconsented to the transfer.\n It is therefore ORDERED the above-named and numbered case is hereby\nTRANSFERRED to the docket of U.S. District Judge Leon Schydlower. Pursuant to the Order\nAssigning the Business of the Court, the Clerk shall credit this case to the percentage of business\nof the receiving Judge. All Orders shall remain in effect unless otherwise ordered by Judge Leon\nSchydlower.\n It is so ORDERED.\n SIGNED this 20th day of August, 2026.\n\n\n\n\n DAVID COUNTS\n UNITED STATES DISTRICT JUDGE\n\f","ocr_status":2,"date_upload":"2026-08-20T15:37:10.431990-07:00","document_number":"7","attachment_number":null,"pacer_doc_id":"181037242095","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Order Reassigning Case","acms_document_guid":""}],"date_created":"2026-08-20T15:04:59.154866-07:00","date_modified":"2026-08-20T15:30:05.050427-07:00","date_filed":"2026-08-20","time_filed":"16:34:19","entry_number":7,"recap_sequence_number":"2026-08-20.001","pacer_sequence_number":27,"description":"ORDER TO TRANSFER. Case reassigned to District Judge Leon Schydlower for all proceedings. Judge David Counts no longer assigned to case. Signed by Judge David Counts. (kg) (Entered: 08/20/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474974578/","id":474974578,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490491595/","id":490491595,"tags":[],"absolute_url":"/docket/74659430/6/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-18T18:05:13.979259-07:00","date_modified":"2026-08-22T23:58:57.219362-07:00","sha1":"b8203f3469f95ca7d1fbf14bfc35dc38450dd2a6","page_count":15,"file_size":293260,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 1 of 15\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n Misc. Case No. 7:26-mc-00318-DC\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\nTESLA, INC.,\n\n Respondent.\n\n\n NEURAL AI\u2019S CORRECTED MOTION TO COMPEL COMPLIANCE WITH\n SUBPOENA SERVED ON THIRD-PARTY TESLA, INC.\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 2 of 15\n\n\n\n\n TABLE OF CONTENTS\n\nI. FACTUAL BACKGROUND ............................................................................................. 2\n\n A. The Underlying Litigation ...................................................................................... 2\n\n B. The Rule 45 Subpoena to Tesla and Tesla\u2019s Initial Objections .............................. 3\n\n C. NAI\u2019s Meet-and-Confer Efforts and Tesla\u2019s Continued Non-Compliance ............ 4\n\nII. THE COURT HAS JURISDICTION BECAUSE THE PLACE OF COMPLIANCE IN\n AUSTIN, TEXAS IS PROPER. ......................................................................................... 5\n\nIII. TESLA MUST PRODUCE DOCUMENTS RESPONSIVE TO THE SUBPOENA. ....... 6\n\n A. The Requested Discovery Goes to the Heart of NAI\u2019s Infringement Claims......... 7\n\n B. Tesla Has Not Substantiated Burden and NAI Offered Narrower Alternatives. .... 9\n\nIV. CONCLUSION ................................................................................................................. 10\n\n\n\n\n i\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 3 of 15\n\n\n\n\n TABLE OF AUTHORITIES\n\n Page(s)\n\nCases\n\n611 Carpenter LLC v. Atlantic Casualty Ins. Co.,\n 2024 WL 1977160 (W.D. Tex. April 30, 2024) ....................................................................6, 9\n\nConservation L. Found., Inc. v. Equilon Enters. LLC,\n 2025 WL 2821238 (D.R.I. Oct. 3, 2025) ...................................................................................5\n\nFG SRC LLC v. Xilinx, Inc.,\n 2022 WL 22997130 (D. Del. Apr. 11, 2022) .............................................................................8\n\nFlores v. Lowes Home Centers, L.L.C.,\n 2023 WL 3959399 (W.D. Tex. June 12, 2023) .........................................................................6\n\nKim v. NuVasive, Inc.,\n 2011 WL 3844106 (S.D. Cal. Aug. 29, 2011) ...........................................................................8\n\nLucent Technologies, Inc. v. Gateway, Inc.,\n 580 F.3d 1301 (Fed. Cir. 2009)..................................................................................................8\n\nMeritage Homes, LLC v. AIG Specialty Ins. Co.,\n 2024 WL 221448 (W.D. Tex. Jan. 18, 2024) ............................................................................5\n\nNeural AI, LLC v. NVIDIA Corp.,\n Case No. 7:24-cv-00221-LS-DTG (W.D. Tex.) ................................................................1, 2, 3\n\nVelocity Pat. LLC v. FCA US LLC,\n 2017 WL 11893112 (N.D. Ill. Nov. 2, 2017) ............................................................................5\n\nWaller v. Jet Specialty, Inc.,\n 2024 WL 7050192 (W.D. Tex. Nov. 19, 2024) .........................................................................6\n\nRules\n\nFed. R. Civ. P. 26 .............................................................................................................................6\n\nFed. R. Civ. P. 26(b)(1)....................................................................................................................6\n\nFed. R. Civ. P. 45 .................................................................................................................1, 3, 5, 6\n\nFed. R. Civ. P. 45(c)(2)(A) ..........................................................................................................5, 6\n\nFed. R. Civ. P. 45(d)(2)(B)(i) ......................................................................................................5, 6\n\n\n\n\n ii\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 4 of 15\n\n\n\n\n This motion arises from Neural AI, LLC v. NVIDIA Corp., Case No. 7:24-cv-00221-LS-\n\nDTG (W.D. Tex.) (the \u201cUnderlying Action\u201d), a patent-infringement action pending in this District.\n\nNeural AI, LLC (\u201cNAI\u201d) alleges that NVIDIA Corporation (\u201cNVIDIA\u201d) infringes through its\n\nGPU-accelerated hardware and software: U.S. Patent No. 8,648,867 (the \u201c\u2019867 Patent\u201d), Reissue\n\nPatent No. RE48,438 (the \u201c\u2019438 Patent\u201d), and Reissue Patent No. RE49,461 (the \u201c\u2019461 Patent\u201d)\n\n(collectively, the \u201cAsserted Patents\u201d). See Exhibits 1\u20134.\n\n NVIDIA has made customer deployment central to the Underlying Action. It disputes\n\nwhether customers configure and use the accused products as NAI alleges and contends that NAI\n\nmust obtain customer evidence to prove indirect infringement and its extent. Tesla, Inc. (\u201cTesla\u201d),\n\na major NVIDIA customer and end user, therefore possesses evidence NVIDIA says NAI must\n\nobtain. See Exhibits 8, 13\u201316; Laiche Decl. \u00b6\u00b6 24\u201325.\n\n NAI served Tesla on June 25, 2026, with Rule 45 subpoenas for documents and corporate\n\ntestimony. See Exhibits 5\u20136. The subpoenas seek targeted information about how Tesla configures,\n\nintegrates, and uses NVIDIA GPUs and software. Id. Tesla objected to every request and topic,\n\nproduced no documents, and designated no witness. See Exhibit 9.\n\n NAI met and conferred repeatedly, identified implicated Tesla systems, supplied focused\n\ntechnical questions, and offered a revised declaration in lieu of broader discovery. See Exhibits\n\n10\u201312. Tesla instead unilaterally sent a five-paragraph declaration on August 10 identifying six\n\nGPU models and partial software, while omitting the core deployment facts. See Exhibit 20. Tesla\n\nthen declared the matter concluded. NAI identified the material gaps on August 11, but Tesla did\n\nnot respond. See Exhibit 21.\n\n NAI therefore respectfully requests an order compelling Tesla to produce documents\n\nresponsive to Requests 1\u201312 on a rolling basis by dates certain and to designate and produce a\n\n\n\n 1\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 5 of 15\n\n\n\n\nknowledgeable witness on Deposition Topics 1\u20135. Tesla\u2019s unilateral declaration does not moot\n\nthat relief, and only an order can secure the discovery Tesla has refused to provide.\n\nI. FACTUAL BACKGROUND\n\n A. The Underlying Litigation\n\n The Underlying Action involves direct, indirect, and induced infringement claims\n\nconcerning claims 16\u201319 of the \u2019867 Patent; claims 1, 3\u20136, 8\u20139, 12, 14, 17\u201318, 21\u201323, 29\u201330, 32,\n\n40, 43\u201344, 46, 48, 51\u201352, and 55\u201357 of the \u2019438 Patent; and claims 21\u201325 and 27\u201330 of the \u2019461\n\nPatent (collectively, the \u201cAsserted Claims\u201d). See Exhibit 1. The Asserted Patents teach GPU-\n\naccelerated computing systems and methods. See Exhibits 2\u20134.\n\n NAI\u2019s January 20, 2026 Final Infringement Contentions and the Subpoenas identify\n\nintegrated combinations of NVIDIA hardware\u2014including GPUs, super computers and servers\u2014and\n\nGPU-accelerated software, including CUDA, cuDNN, TensorRT, and higher-level frameworks, as\n\nthe Accused Products. See Exhibit 7 at 2\u20139. NAI alleges NVIDIA encourages customers to combine,\n\nconfigure, and use those products in an infringing manner. See id. at 12\u201313.\n\n NVIDIA disputes that selling the Accused Products proves customers configure and deploy\n\nthem as NAI alleges. Although NVIDIA admits it sells the products, partners with resellers,\n\nmanaged-service providers, and data-center providers, and supports customers and partners, it\n\ndenies infringement and disclaims knowledge of customer use. See Exhibit 8 \u00b6\u00b6 18, 20\u201323, 87,\n\n131, 171. NVIDIA has stated throughout discovery that NAI must seek evidence of actual\n\ndeployment from customers and end users. Laiche Decl. \u00b6\u00b6 24\u201325. Customer evidence is therefore\n\ndirectly relevant to NAI\u2019s infringement claims.\n\n Tesla is a major NVIDIA customer and end user. It unveiled a supercomputer with 5,760\n\nNVIDIA A100 GPUs in 2021, expanded it to 7,360 NVIDIA A100 GPUs in 2022, and later\n\n\n\n\n 2\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 6 of 15\n\n\n\n\ndeployed Cortex, a training cluster of approximately 50,000 H100 GPUs at Gigafactory Texas.\n\nSee Exhibits 13\u201315. Tesla uses these systems to train neural networks for Full Self-Driving,\n\nAutopilot, and other AI applications. See Exhibits 13\u201316. Tesla\u2019s records concerning those\n\nsystems\u2019 configuration, deployment, and use bear directly on whether the accused products are\n\nused in an infringing manner and whether NVIDIA induced or contributed to that use.\n\n NAI has subpoenaed multiple NVIDIA customers for evidence uniquely within their\n\npossession. Document discovery in the Underlying Action closed August 11, 2026, and deposition\n\ndiscovery closes September 16, 2026. See Neural AI, LLC v. NVIDIA Corp., Case No. 7:24-cv-\n\n00221-LS-DTG, Dkt. 181 (W.D. Tex.). NVIDIA stipulated that motions to compel third parties\n\nrelated to document discovery could be filed by August 18, 2026. Laiche Decl. \u00b6 28.\n\n B. The Rule 45 Subpoena to Tesla and Tesla\u2019s Initial Objections\n\n NAI served Tesla on June 25, 2026, with Rule 45 subpoenas for documents and corporate\n\ntestimony, designating Austin as the place of compliance. See Exhibits 5\u20136. Tesla\u2019s global\n\nheadquarters and principal place of business are in Austin; its Gigafactory Texas occupies over 10\n\nmillion square feet and employs approximately 20,000 people. See Exhibits 17\u201319.\n\n The document subpoena contains twelve targeted requests, and the accompanying\n\ndeposition subpoena contains five topics, concerning Tesla\u2019s configuration, integration, and use of\n\nNVIDIA GPUs and related software. See Exhibit 5. Both are limited to the relevant damages\n\nperiod\u2014September 13, 2018 to the present\u2014and to U.S.-based activity or activity supporting or\n\ndirected toward U.S. operations. Id. at 15.\n\n Tesla\u2019s responses were due July 14, 2026. At Tesla\u2019s request, NAI extended the deadline\n\nto July 21. See Exhibit 10 at 10. Tesla then objected to every document request and deposition\n\ntopic and declined to produce documents or designate a witness. See Exhibit 9. For each request\n\n\n\n\n 3\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 7 of 15\n\n\n\n\nand topic, Tesla said it was merely \u201cwilling to meet and confer regarding the scope of this Request\n\nand the burden it imposes on Tesla\u201d. Id.\n\n C. NAI\u2019s Meet-and-Confer Efforts and Tesla\u2019s Continued Non-Compliance\n\n The parties first met and conferred on July 28. NAI explained the targeted discovery,\n\nidentified implicated Tesla systems, and discussed categories likely to satisfy the requests. Tesla\n\nsaid it was still investigating, needed technical personnel, and objected to the Subpoenas\u2019 breadth\n\nand requests for confidential technical information. Exhibit 10 at 3\u20135.\n\n On August 4, NAI supplied focused technical questions designed to identify Tesla\u2019s actual\n\nNVIDIA products, configurations, data paths, memory use, runtime compilation and TensorRT\n\nuse, frequency, scale, and U.S. nexus, and to narrow collection. See Exhibits 10, 11.\n\n NAI also supplied a draft declaration as an alternative to broader document production and\n\ndeposition testimony. Exhibit 12. It invited a knowledgeable Tesla declarant to revise the draft\n\nafter a reasonable investigation and address the NVIDIA products and software Tesla uses,\n\noperation as designed, CPU/GPU memory, input and output paths, pretrained models, custom code\n\nor configurations, and frequency. Id. NAI said it would consider an executed declaration in lieu of\n\nfurther discovery, subject to resolving material gaps. See Exhibits 10 at 1.\n\n The parties met and conferred again on August 7, 2026. Tesla responded that its investigation\n\nremained ongoing, that it was not prepared to provide substantive answers, and that it could not\n\ncommit to providing a declaration with sufficient specificity to address the relevant issues.\n\n On August 10, Tesla sent NAI a five-paragraph declaration signed by Alon Daks, Tesla\u2019s\n\nSenior Staff Software Engineer, without giving NAI an opportunity to review it before submission.\n\nSee Exhibit 20. Tesla\u2019s counsel stated that \u201cTesla considers this matter concluded.\u201d Exhibits 21 at\n\n2. The declaration, however, was materially deficient. It merely identifies six NVIDIA GPU\n\n\n\n\n 4\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 8 of 15\n\n\n\n\nmodels and a partial software list, but omits unmodified use of NVIDIA software, whether the\n\nhardware and software function as designed, NVIDIA-distributed pretrained models, CPU/GPU\n\nmemory architecture, and the standard data path including GPUDirect. Exhibit 20. It also collapses\n\nthe software categories and misidentifies Triton, without confirming whether PyTorch or Triton\n\nare used on NVIDIA GPUs. Id.\n\n On August 11, NAI advised Tesla that the declaration was materially insufficient,\n\nidentified these gaps and other deficiencies, and requested another meet and confer. See Exhibit\n\n21 at 1. Tesla did not respond, supplement the declaration, produce documents, designate a\n\nwitness, or propose a conference.\n\n After explaining the requests, narrowing the issues, and offering alternatives, NAI now\n\nseeks court intervention.\n\nII. THE COURT HAS JURISDICTION BECAUSE THE PLACE OF COMPLIANCE\n IN AUSTIN, TEXAS IS PROPER.\n\n Rule 45 directs a party seeking to compel compliance to apply to \u201cthe court for the district\n\nwhere compliance is required.\u201d Fed. R. Civ. P. 45(d)(2)(B)(i); see also Meritage Homes, LLC v.\n\nAIG Specialty Ins. Co., 2024 WL 221448, at *4 (W.D. Tex. Jan. 18, 2024). The Subpoenas\n\ndesignate Planet Depos, Downtown Austin, 100 Congress Ave., Ste. 2000, Austin, TX 78701, as\n\nthe place of compliance.\n\n Rule 45(c)(2)(A) permits document production at a place within 100 miles of where the\n\nperson resides, is employed, or regularly transacts business in person. The rule does not limit the\n\nplace of compliance to an entity\u2019s headquarters or the location of particular documents or custodians.\n\nSee, e.g., Conservation L. Found., Inc. v. Equilon Enters. LLC, 2025 WL 2821238, at *1 (D.R.I. Oct.\n\n3, 2025); Velocity Pat. LLC v. FCA US LLC, 2017 WL 11893112, at *4 (N.D. Ill. Nov. 2, 2017).\n\n Austin satisfies Rule 45(c)(2)(A). Tesla\u2019s global headquarters is at 1 Tesla Road, Austin,\n\n\n\n 5\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 9 of 15\n\n\n\n\nTexas 78725; its Gigafactory Texas covers 2,500 acres and more than 10 million square feet; and\n\nTesla employs approximately 20,000 people there. See Exhibits 17\u201319. The designated place of\n\ncompliance is proper, so this Court has jurisdiction to resolve the motion.\n\nIII. TESLA MUST PRODUCE DOCUMENTS RESPONSIVE TO THE SUBPOENA.\n\n Federal Rule of Civil Procedure 26 provides that a party may obtain discovery regarding\n\nany nonprivileged matter that is relevant to the parties\u2019 claims or defenses and proportional to the\n\nneeds of the case. Fed. R. Civ. P. 26(b)(1). \u201cAt the discovery stage, relevancy is broadly\n\nconstrued.\u201d Flores v. Lowes Home Centers, L.L.C., 2023 WL 3959399, at *2 (W.D. Tex. June 12,\n\n2023). \u201c\u2018[A] request for discovery should be considered relevant if there is any possibility that the\n\ninformation sought may be relevant to the claim or defense of any party.\u2019\u201d Id.\n\n Where, as here, a non-party refuses discovery in response to a validly issued subpoena,\n\nFederal Rule of Civil Procedure 45 provides the Court for the district where compliance is required\n\nwith broad discretion to compel the production of documents and information from third parties.\n\nFed. R. Civ. P. 45(d)(2)(B)(i). Once a party moving to compel discovery establishes that the\n\nmaterials are relevant, \u201c[t]he party opposing discovery bears the burden of stating \u2018with specificity\n\nthe grounds for objecting to the request.\u2019\u201d Waller v. Jet Specialty, Inc., 2024 WL 7050192, at *1\n\n(W.D. Tex. Nov. 19, 2024); see also 611 Carpenter LLC v. Atlantic Casualty Ins. Co., 2024 WL\n\n1977160, at *1 (W.D. Tex. April 30, 2024) (granting party\u2019s motion to compel non-party\n\nsubpoena). The non-party \u201cmust state with specificity the objection and how it relates to the\n\nparticular request being opposed, and not merely that it is overly broad and burdensome.\u201d 611\n\nCarpenter, 2024 WL 1977160, at *1.\n\n\n\n\n 6\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 10 of 15\n\n\n\n\n A. The Requested Discovery Goes to the Heart of NAI\u2019s Infringement Claims.\n\n NVIDIA has made customer deployment a disputed issue. Tesla is a significant customer\n\nand end user. It built an NVIDIA-based supercomputer with 5,760 A100 GPUs, expanded it to 7,360\n\nA100s, and later deployed Cortex with approximately 50,000 H100s and 16,000 H200s. See Exhibits\n\n13\u201315. Tesla uses these systems to train neural networks for Full Self-Driving, Autopilot, and other\n\nAI applications. See Exhibits 13\u201316. Tesla\u2019s records and testimony can show whether and how often\n\nthe accused methods are actually performed.\n\n The requested information is also uniquely within Tesla\u2019s possession. NVIDIA may\n\nknow what it designed and distributed, but according to NVIDIA, only Tesla knows what it\n\nselected, configured, deployed, and actually ran. Laiche Decl. \u00b6\u00b6 24\u201325. NVIDIA claims it does\n\nnot possess Tesla\u2019s internal architecture diagrams, environment configurations, profiler traces,\n\nor memory-allocation records, and NVIDIA has repeatedly disclaimed meaningful insight into\n\ncustomer deployments. Id. At a discovery hearing, NVIDIA told the Court that NAI would need\n\nto seek real-world deployment information from its customers and end users. See Laiche Decl.\n\n\u00b6 24. NAI has done so.\n\n The twelve document requests track the accused computation from software selection\n\nthrough input, execution, memory, and output to establish infringement. See Exhibit 5.\n\n \u2022 Requests 1\u20134 seek documents sufficient to identify the software, frameworks, libraries,\n APIs, scripts, configurations, and custom code Tesla uses on NVIDIA GPUs; the NVIDIA\n software and sample code it uses; and how Tesla calls, interfaces with, wraps, depends on,\n modifies, or extends that functionality. They show which accused products Tesla deployed\n and whether Tesla used them as provided and intended or materially modified them.\n\n \u2022 Requests 5\u20136 seek architecture, design, data-flow, control-flow, and execution-flow\n materials and documents showing whether the relevant computations involve neural\n networks, layers, and outputs used as inputs to later neurons or layers. They show whether\n Tesla\u2019s systems perform the claimed GPU-accelerated operations and how the accused\n hardware and software work together as relevant to the Asserted Claims.\n\n\n\n\n 7\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 11 of 15\n\n\n\n\n \u2022 Requests 7\u201310 address the claimed memory and data paths: pointer, buffer, and\n intermediate-result reuse; GPU-memory allocation or partitioning; movement of inputs\n from CPU, host, storage, sensor, camera, or network sources into GPU memory; and\n storage, transfer, accumulation, reuse, or return of outputs and intermediate results. They\n target the memory-management and data-transfer limitations at the heart of the Asserted\n Claims.\n\n \u2022 Requests 11\u201312 seek documents sufficient to show how GPU computations are scheduled,\n ordered, queued, synchronized, parallelized, launched, interrupted, resumed, and executed,\n and how inputs, commands, model or parameter changes, interruptions, or output changes\n affect GPU-queue placement. They target the claimed coordination and control of GPU\n computations.\n\nThese requests thus target the facts relevant to whether Tesla\u2019s systems perform the asserted\n\nmethod steps and whether NVIDIA\u2019s software, instructions, and support cause or encourage that\n\nuse. See Lucent Technologies, Inc. v. Gateway, Inc., 580 F.3d 1301, 1321\u201322, 1333-34 (Fed. Cir.\n\n2009) (stating patentee must show all steps of claimed method were performed to prove indirect\n\ninfringement and that damages \u201cought to be correlated, in some respect, to the extent the infringing\n\nmethod is used by\u201d directly infringing third parties); FG SRC LLC v. Xilinx, Inc., 2022 WL\n\n22997130, at *5 (D. Del. Apr. 11, 2022) (\u201c[I]t is clear that SRC has a need for information from\n\nthird parties demonstrating that the parties practice the asserted claims.\u201d); Kim v. NuVasive, Inc.,\n\n2011 WL 3844106, at *3 (S.D. Cal. Aug. 29, 2011) (\u201cThe Court finds the information sought in\n\nthe subpoenas is relevant to NuVasive's claim of induced infringement and damages therefrom\u201d).\n\n The five deposition topics seek testimony on the same subjects as the document requests\n\nand directs Tesla to designate one or more knowledgeable persons on five topics. See Exhibit 5.\n\n \u2022 Topic 1: the NVIDIA software and libraries Tesla uses to perform computations on\n NVIDIA GPUs;\n\n \u2022 Topic 2: NVIDIA sample source code Tesla uses, in whole or in part;\n\n \u2022 Topic 3: Tesla\u2019s customizations and data inputs that alter how NVIDIA software\n performs computations;\n\n\n\n\n 8\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 12 of 15\n\n\n\n\n \u2022 Topic 4: identification of Tesla software that uses NVIDIA GPUs to perform\n computations; and\n\n \u2022 Topic 5: how output and intermediate GPU results are stored, referenced by pointers,\n transferred, copied, streamed, written back, returned, accumulated, reused, or otherwise\n made available between GPU memory and CPU, host, system, storage, display,\n network, or other memory locations.\n\n NAI also offered the draft declaration as a less burdensome alternative and said it would\n\nconsider it in lieu of document production and testimony, subject to resolving material gaps. See\n\nExhibits 10, 12. Tesla\u2019s five-paragraph response was materially deficient and omitted core\n\ndeployment facts and therefore did not eliminate the need for discovery. See Exhibits 20\u201321.\n\n The information is central to the issues in dispute, uniquely within Tesla\u2019s possession, and\n\nsought through targeted, time-limited requests and topics. Tesla should be compelled to produce it.\n\n B. Tesla Has Not Substantiated Burden and NAI Offered Narrower Alternatives.\n\n Tesla\u2019s burden objections do not state with specificity the burden associated with any\n\nrequest or topic. See 611 Carpenter, 2024 WL 1977160, at *1. Tesla repeated the same objections\n\nacross all twelve requests and five topics and said only that it was \u201cwilling to meet and confer\n\nregarding the scope of this Request and the burden it imposes on Tesla.\u201d Exhibit 9. Boilerplate\n\nobjections do not substantiate undue burden.\n\n Tesla also objected that the definition of \u201cNVIDIA GPUs\u201d encompasses multiple\n\narchitectures and \u201chundreds of individual product SKUs.\u201d See Exhibit 9 at 6. That reflects the\n\nscope of NVIDIA\u2019s accused product line, which NAI\u2019s infringement contentions identify as\n\nAccused Products. See Exhibit 7 at 2\u20139. The subpoenas seek documents \u201csufficient to show\u201d\n\nidentified facts about products Tesla actually deploys, and they do not require Tesla to identify\n\nevery GPU it has ever used. See Exhibit 5 at 15.\n\n NAI minimized burden by agreeing to a one-week extension, identifying implicated\n\nproducts and systems, supplying focused technical questions, and offering a draft declaration as\n\n\n 9\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 13 of 15\n\n\n\n\nan alternative to broader discovery. See Exhibits 10\u201312. NAI also invited Tesla to identify specific\n\nburdens. Tesla has offered no concrete search, production date, revised declaration, or witness.\n\nLaiche Decl. \u00b6 27.\n\n Tesla\u2019s August 10 declaration does not satisfy the subpoenas or the compromise NAI\n\noffered. It is only five paragraphs and covers the bare minimum: six NVIDIA GPU models and a\n\npartial list of software and libraries. Exhibit 20. It does not address the core technical issues within\n\nthe subpoenas, which are unmodified use of NVIDIA software; whether the hardware and software\n\nfunction as NVIDIA designed; NVIDIA-distributed pretrained models; CPU/GPU memory\n\narchitecture; and the standard CPU-memory-to-GPU-memory data path, including GPUDirect.\n\nCompare Exhibit 12 with Exhibit 20. Tesla compounded its noncompliance by submitting the\n\ndeclaration unilaterally, without giving NAI an opportunity to review it for completeness. NAI\u2019s\n\noffer was expressly conditional, and stated NAI would consider a declaration in lieu of broader\n\ndiscovery only \u201csubject to resolving any material gaps.\u201d Exhibit 10 at 1. Tesla had no basis to\n\nconvert that conditional compromise into a unilateral declaration that its obligations were satisfied.\n\nIV. CONCLUSION\n\n For the foregoing reasons, NAI respectfully requests that this Court (1) compel Tesla to\n\nproduce nonprivileged documents responsive to Requests 1\u201320 on a rolling basis to be completed\n\nwithin twenty-one (21) days of the Court\u2019s order, and (2) compel Tesla to designate and produce a\n\nknowledgeable witness on Deposition Topics 1\u20135 within thirty (30) days of the Court\u2019s order.\n\n\n\n\n 10\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 14 of 15\n\n\n\n\nDated: August 18, 2026\n\n\n Respectfully submitted,\n\n /s/ Rocco Magni\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Attorneys for Petitioner Neural AI, LLC\n\n\n\n\n 11\n\f Case 7:26-mc-00318-DC Document 6 Filed 08/18/26 Page 15 of 15\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that on August 18, 2026, a true and correct copy of\n\nthe foregoing document was served on counsel for Tesla, Inc. and all counsel of record in the\n\nunderlying action.\n\n\n /s/ Rocco Magni\n Rocco Magni\n\n\n\n CERTIFICATE OF CONFERENCE\n\n The undersigned certifies that counsel for Neural AI, LLC conferred in good faith with\n\ncounsel for non-party Tesla, Inc. regarding the issues raised in this Motion, including through\n\nZoom conferences on or about July 28 and August 7, 2026, and related email correspondence.\n\nDespite those efforts, the parties were unable to resolve the dispute.\n\n /s/ Rocco Magni\n Rocco Magni\n\n\n\n\n 12\n\f","ocr_status":2,"date_upload":"2026-08-19T09:19:29.063904-07:00","document_number":"6","attachment_number":null,"pacer_doc_id":"181037220267","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Compel","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554418/","id":490554418,"tags":[],"absolute_url":"/docket/74659430/6/1/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.109770-07:00","date_modified":"2026-08-23T09:39:26.374213-07:00","sha1":"65f7e608b06fbd3a1c40b6cc6afe00edf896e5df","page_count":5,"file_size":176339,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.1.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.1.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 6-1 Filed 08/18/26 Page 1 of 5\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n NEURAL AI, LLC,\n Misc. Case No. 7:26-mc-00318-DC\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\n TESLA, INC.,\n\n Respondent.\n\n\n DECLARATION OF TANNER LAICHE IN SUPPORT OF NEURAL AI\u2019S\n MOTION TO COMPEL COMPLIANCE WITH SUBPOENA\n SERVED ON THIRD-PARTY TESLA, INC.\n\n I, Tanner Laiche, declare as follows:\n\n 1. I am an attorney duly licensed to practice in the States of California and Washington\n\nand am admitted to the Western District of Texas. I am an associate at the law firm of Susman\n\nGodfrey LLP, and am a counsel of record for Petitioner Neural AI, LLC (\u201cNAI\u201d) in the above-\n\ncaptioned matter. I make this declaration in support of NAI\u2019s Motion to Compel Compliance with\n\nSubpoena Served on Third-Party Tesla, Inc. (\u201cTesla\u201d). Unless otherwise stated, I have personal\n\nknowledge of the facts set forth herein and could competently testify thereto.\n\n 2. Attached hereto as Exhibit 1 is a true and correct copy of NAI\u2019s First Amended\n\nComplaint for Patent Infringement filed in the Underlying Action on December 12, 2024.\n\n 3. Attached hereto as Exhibit 2 is a true and correct copy of U.S. Patent No.\n\n8,648,867, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n 4. Attached hereto as Exhibit 3 is a true and correct copy of Reissue Patent No.\n\nRE48,438, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 6-1 Filed 08/18/26 Page 2 of 5\n\n\n\n\n 5. Attached hereto as Exhibit 4 is a true and correct copy of Reissue Patent No.\n\nRE49,461, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n 6. Attached hereto as Exhibit 5 is a true and correct copy of NAI\u2019s Subpoenas to\n\nProduce Documents and to Testify at a Deposition served on Tesla, Inc., served on June 25, 2026,\n\nincluding the accompanying definitions, instructions, requests for production, and deposition\n\ntopics.\n\n 7. Attached hereto as Exhibit 6 is a true and correct copy of the Affidavit of Service\n\nconfirming service of the subpoenas on Tesla\u2019s registered agent on June 25, 2026.\n\n 8. Attached hereto as Exhibit 7 is a true and correct copy of NAI\u2019s January 20, 2026\n\nFinal Infringement Contentions Cover Page filed in the Underlying Action.\n\n 9. Attached hereto as Exhibit 8 is a true and correct copy of NVIDIA\u2019s Amended\n\nAnswer to NAI\u2019s Amended Complaint, filed on November 25, 2025, as docketed at Dkt. 130 in\n\nthe Underlying Action (public, redacted version).\n\n 10. Attached hereto as Exhibit 9 is a true and correct copy of Non-Party Tesla, Inc.\u2019s\n\nObjections and Responses to Plaintiff Neural AI, LLC\u2019s Subpoena, dated July 21, 2026.\n\n 11. Attached hereto as Exhibit 10 is a true and correct copy of the meet-and-confer and\n\nextension correspondence between counsel for NAI and counsel for Tesla regarding NAI\u2019s\n\nsubpoenas.\n\n 12. Attached hereto as Exhibit 11 is a true and correct copy of NAI\u2019s Third-Party\n\nQuestions provided to Tesla on August 4, 2026.\n\n 13. Attached hereto as Exhibit 12 is a true and correct copy of NAI\u2019s Draft Third-Party\n\nDeclaration provided to Tesla on August 4, 2026.\n\n 14. Attached hereto as Exhibit 13 is a true and correct copy of an NVIDIA Blog post\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-1 Filed 08/18/26 Page 3 of 5\n\n\n\n\ntitled \u201cTesla Unveils Supercomputer Powered by NVIDIA GPUs,\u201d printed from\n\nhttps://blogs.nvidia.com. I obtained this document from NVIDIA\u2019s publicly accessible website on\n\nAugust 17, 2026.\n\n 15. Attached hereto as Exhibit 14 is a true and correct copy of a Tom\u2019s Hardware\n\narticle titled \u201cTesla Brags About In-House Supercomputer, Now With 7,360 A100 GPUs,\u201d printed\n\nfrom https://www.tomshardware.com. I obtained this document from Tom\u2019s Hardware\u2019s publicly\n\naccessible website on August 17, 2026.\n\n 16. Attached hereto as Exhibit 15 is a true and correct copy of a TechCrunch article\n\ntitled \u201cTesla Dojo: The rise and fall of Elon Musk\u2019s AI supercomputer,\u201d printed from\n\nhttps://techcrunch.com. I obtained this document from TechCrunch\u2019s publicly accessible website\n\non August 17, 2026.\n\n 17. Attached hereto as Exhibit 16 is a true and correct copy of a Popular Science article\n\ntitled \u201cWhat we know about Tesla\u2019s supercomputer,\u201d printed from https://www.popsci.com. I\n\nobtained this document from Popular Science\u2019s publicly accessible website on August 17, 2026.\n\n 18. Attached hereto as Exhibit 17 is a true and correct copy of the Tesla Careers\n\nwebpage, printed from https://www.tesla.com/careers. I obtained this document from Tesla\u2019s\n\npublicly accessible website on August 17, 2026.\n\n 19. Attached hereto as Exhibit 18 is a true and correct copy of an article titled \u201cTesla\n\nemploys 20,000 in Austin, could triple amid Cybertruck ramp-up.\u201d I obtained this document from\n\na publicly accessible website on August 17, 2026.\n\n 20. Attached hereto as Exhibit 19 is a true and correct copy of the Giga Texas webpage\n\nfrom Tesla\u2019s website, printed from https://www.tesla.com. I obtained this document from Tesla\u2019s\n\npublicly accessible website on August 17, 2026.\n\n\n\n\n 3\n\f Case 7:26-mc-00318-LS Document 6-1 Filed 08/18/26 Page 4 of 5\n\n\n\n\n 21. Attached hereto as Exhibit 20 is a true and correct copy of the Declaration of Alon\n\nDaks, a Senior Staff Software Engineer at Tesla, Inc., executed on August 10, 2026, regarding\n\nTesla\u2019s use of NVIDIA GPUs and software identified in NAI\u2019s subpoena.\n\n 22. Attached hereto as Exhibit 21 is a true and correct copy of additional meet-and-\n\nconfer correspondence between counsel for NAI and counsel for Tesla, dated August 10\u201311, 2026,\n\nregarding Tesla\u2019s declaration and the outstanding subpoena issues.\n\n 23. The Underlying Action\u2014Neural AI, LLC v. NVIDIA Corporation, Case No. 7:24-\n\ncv-00221-LS-DTG (W.D. Tex.)\u2014is a patent infringement action in which NAI alleges that\n\nNVIDIA, Corp.\u2019s (\u201cNVIDIA\u201d) GPU-accelerated computing hardware and software infringe U.S.\n\nPatent No. 8,648,867 (the \u201c\u2019867 Patent\u201d), Reissue Patent No. RE48,438 (the \u201c\u2019438 Patent\u201d), and\n\nReissue Patent No. RE49,461 (the \u201c\u2019461 Patent\u201d) (collectively, the \u201cAsserted Patents\u201d).\n\n 24. Throughout discovery in the above-captioned action, NVIDIA has taken the\n\nposition that NAI must obtain evidence from NVIDIA\u2019s customers and end users to prove how the\n\naccused products are actually deployed and configured. During a sealed discovery conference on\n\nSeptember 8, 2025, NVIDIA\u2019s counsel represented to the Court that it only provides tools for end-\n\nuser companies to build their own AI applications on NVIDIA hardware and that if NAI wanted\n\nto learn how the accused software was deployed in real-world systems, NAI would need to seek\n\nthat information directly from NVIDIA\u2019s customers and end users.\n\n 25. Throughout discovery NVIDIA has also disclaimed knowledge concerning how its\n\ncustomers, partners, and end users ultimately deploy, configure, and operate NVIDIA hardware\n\nand software. NAI accordingly turned to third-party discovery, including subpoenas to Tesla, to\n\nobtain the evidence NVIDIA contends NAI must have.\n\n 26. On August 10, 2026, Tesla provided a declaration from Alon Daks, a Senior Staff\n\n\n\n\n 4\n\f Case 7:26-mc-00318-LS Document 6-1 Filed 08/18/26 Page 5 of 5\n\n\n\n\nSoftware Engineer (Exhibit 20). However, as set forth in Exhibit 21, the declaration is materially\n\ninsufficient because it addresses only GPU identification and partial software identification while\n\nomitting entirely the substantive technical topics set forth in NAI\u2019s draft declaration (draft\n\nparagraphs 7\u201315), including unmodified use of NVIDIA software, hardware/software functioning\n\nas designed by NVIDIA, use of NVIDIA-distributed pretrained models, CPU/GPU memory\n\narchitecture, and the standard data flow.\n\n 27. As of the date of this declaration, Tesla has not: (a) identified any specific document\n\nrequests to which it will respond; (b) agreed to produce any responsive documents; (c) fully\n\nanswered NAI\u2019s technical questions; (d) designated a witness for deposition; or (e) committed to\n\nany date by which it will do any of the foregoing.\n\n 28. Document discovery in the Underlying Action closed on August 11, 2026.\n\nDefendant NVIDIA, Corp. in the underlying action stipulated to extend the deadline for third-party\n\nmotions to compel to August 18, 2026. Deposition discovery closes on September 16, 2026.\n\n I declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct.\n\n Executed on August 18, 2026, in Seattle, Washington.\n\n\n\n\n Tanner Laiche\n\n\n\n\n 5\n\f","ocr_status":2,"date_upload":"2026-08-20T15:37:39.915735-07:00","document_number":"6","attachment_number":1,"pacer_doc_id":"181037220268","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Affidavit Declaration of Tanner Laiche","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554420/","id":490554420,"tags":[],"absolute_url":"/docket/74659430/6/2/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.161711-07:00","date_modified":"2026-08-23T04:39:58.617338-07:00","sha1":"49c8ce52e85ffc154f6be005dfaf86494abb946e","page_count":95,"file_size":2297307,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.2.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.2.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-2 Filed 08/18/26 Page 1 of 95\n\n\n\n\n EXHIBIT\n\n 1\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 2 of1 95\n of 94\n\n\n\n\n UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND-ODESSA DIVISION\n\nNEURAL AI, LLC )\n )\n )\n Plaintiff, )\nv. ) Civil Action No. 7:24-cv-00221\n )\nNVIDIA CORPORATION )\n ) JURY TRIAL DEMANDED\n )\n Defendant. )\n )\n\n FIRST AMENDED COMPLAINT FOR PATENT INFRINGEMENT\n\n Neural AI, LLC (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) alleges against Defendant Nvidia Corporation\n\n(\u201cNvidia\u201d or \u201cDefendant\u201d) the following:\n\n 1. This case involves patented technologies that revolutionized, and have become\n\nwidely adopted in, the field of graphical processor unit (\u201cGPU\u201d)-accelerated computing for\n\nartificial intelligence, machine learning, and complex numerical simulations. GPU-accelerated\n\ncomputing powers many of the most advanced and powerful forms of artificial intelligence that\n\nhave exploded over the past decade.\n\n 2. Highly complex numerical simulations, such as the prediction of protein chains,\n\ngenetic sequences and cryptographic sequences, and advanced machine learning techniques such\n\nas deep learning neural networks, require hardware capable of a high degree of parallel processing\n\nfor efficient computation. GPUs, which generally have hundreds to thousands more computational\n\nprocessors or \u201ccores\u201d than central processing units (\u201cCPUs\u201d), are the preferred hardware for\n\nexecuting such simulations and machine learning techniques. Indeed, the parallel operation of\n\nthousands of high-performance GPUs have become a basic necessity for the execution and training\n\n\n\n\n 1\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 3 of2 95\n of 94\n\n\n\n\nof complex natural language and image-generation models, such as ChatGPT\u2019s GPT-4 and\n\nSora.AI. (See https://www.fierceelectronics.com/sensors/chatgpt-runs-10k-nvidia-training-gpus-\n\npotential-thousands-more.)\n\n 3. Before Plaintiff\u2019s innovations, the conventional wisdom in the field of GPU-\n\naccelerated computing was that the exchange of intermediate outputs between a GPU and a CPU\n\nwas too computationally expensive. This was so because the GPU, adapted for highly parallel\n\nprocessing tasks (e.g., graphically modeling a physics engine or rendering complex moving\n\nimages), was ill-suited for handling operations better left to the CPU, like interacting with a user\u2019s\n\nmouse and keyboard or sending and receiving simple datasets. Plaintiff\u2019s foundational technology\n\nchanged this by inventing techniques that leveraged the unique advantages of both the CPU and\n\nthe GPU to enable their efficient interplay in hardware-accelerated computing.\n\n 4. Plaintiff\u2019s patented technologies are enshrined in U.S. Patent Nos. 8,648,867 (\u201cthe\n\n\u2019867 Patent\u201d), RE49,461 (\u201cthe \u2019461 Patent\u201d), and RE48,438 (\u201cthe \u2019438 Patent\u201d) (collectively, \u201cthe\n\nAsserted Patents\u201d or \u201cThe GPU-Based Acceleration Patents\u201d).\n\n NATURE OF THE CASE\n\n 5. Plaintiff brings claims under the patent laws of the United States, 35 U.S.C. \u00a7 1, et\n\nseq., for infringement of the Asserted Patents. Defendant has infringed and continues to infringe\n\neach of the Asserted Patents under at least 35 U.S.C. \u00a7\u00a7271(a), 271(b) and 271(c).\n\n THE PARTIES\n\n 6. Plaintiff Neural AI, LLC, is the owner by assignment of each of the Asserted\n\nPatents.\n\n 7. The technology of the Asserted Patents underpins multiple artificial intelligence\n\nand accelerated computing products that incorporate the patented technology, such as Neurala,\n\n\n\n\n 2\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 4 of3 95\n of 94\n\n\n\n\nInc.\u2019s Vision Inspection Automation (VIA), Vision AI software, and Brain Builder platform.\n\n 8. Neural AI is a Texas limited liability company and is a registered business in Texas.\n\nNeural AI maintains its principal office in this District, at 510 Austin Avenue, Suite 2554, Waco,\n\nTX 76701.\n\n 9. Defendant Nvidia Corporation is a Delaware corporation with its headquarters and\n\nprincipal place of business in Santa Clara, California. (See https://investor.nvidia.com/financial-\n\ninfo/sec-filings/sec-filings-details/default.aspx?FilingId=17293267, U.S. Securities and Exchange\n\nCommission Form 10-K for Fiscal Year Ended January 28, 2024;\n\nhttps://nvidianews.nvidia.com/multimedia/santa-clara-headquarters.) Defendant Nvidia\n\nCorporation is registered with the Secretary of State to conduct business in Texas. Nvidia has an\n\noffice in this District located in Austin, Texas. (See https://www.nvidia.com/en-us/contac.)\n\n JURISDICTION & VENUE\n\n 10. This action arises under the Patent Laws of the United States, 35 U.S.C. \u00a7 1, et seq.\n\nThe Court has subject matter jurisdiction pursuant to 28 U.S.C. \u00a7\u00a7 1331 and 1338(a).\n\n 11. This Court has personal jurisdiction over Defendant because it regularly conducts\n\nbusiness in the State of Texas and in this District. This business includes operating systems, using\n\nand/or providing computer hardware, software, firmware, and platforms, and/or providing services\n\nand/or engaging in activities in Texas and in this District that infringe one or more claims of the\n\nAsserted Patents, as well as inducing and contributing to the direct infringement of others through\n\nacts in this District.\n\n 12. Nvidia has also, directly and through its extensive network of partnerships,\n\nincluding with local IT service providers, purposefully and voluntarily placed products and/or\n\nprovided services that practice and/or implement the methods, systems, and apparatuses claimed\n\n\n\n\n 3\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 5 of4 95\n of 94\n\n\n\n\nin the Asserted Patents into the stream of commerce with the intention and expectation that they\n\nwill be purchased and used by customers in this District, as detailed below. (See\n\nhttps://www.nvidia.com/en-us/about-nvidia/partners/.)\n\n 13. Defendant has also acknowledged that this Court has personal jurisdiction over it\n\nin cases filed against it in this District. (See, e.g., Vantage Micro LLC v. NVIDIA Corporation,\n\nCase No. 6:19-cv-00582-RP, ECF 22 (W.D. Tex., Jan. 4, 2020) (admitting to personal\n\njurisdiction); Ocean Semiconductor LLC v. NVIDIA Corporation, Case No. 6:20-cv-01211-ADA,\n\nECF 14 (W.D. Tex., Mar. 12, 2021) (same).) Defendant has admitted \u201cit is subject to this Court\u2019s\n\ngeneral personal jurisdiction.\u201d (Id.)\n\n 14. Venue is proper in this District pursuant to 28 U.S.C. \u00a7\u00a7 1391(b) and (c) and 28\n\nU.S.C. \u00a7 1400(b) because Defendant Nvidia Corporation has regular and systematic contacts\n\nwithin this District and has committed acts of infringement within this District.\n\n 15. Defendant Nvidia Corporation is a registered business in Texas and has regular and\n\nestablished places of business in this District. Nvidia has an office in this District located at 11001\n\nLakeline Blvd, Suite 100 Bldg. 2, Austin, Texas 78717. (See https://craft.co/nvidia.) Nvidia\u2019s\n\nAustin office has \u201c54,000 SF of new shell office and DVS labs\u201d and \u201c35,000 SF of offices, testing\n\nand software labs.\u201d (See https://kiddgrp.com/project/nvidia-corporation/.)\n\n 16. Defendant Nvidia Corporation has hundreds of employees in this District\u2014\n\nincluding positions in engineering, sales, marketing, and finance. LinkedIn lists approximately 792\n\npersons associated with Nvidia and identified as being located in the Austin or Austin metropolitan\n\narea. (See\n\nhttps://www.linkedin.com/company/nvidia/people/?facetGeoRegion=104472865%2C90000064.)\n\nLinkedIn also lists approximately 1,158 persons associated with Nvidia and identified as being\n\n\n\n\n 4\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 6 of5 95\n of 94\n\n\n\n\nlocated in the State of Texas. (See id.)\n\n 17. In addition, Defendant Nvidia Corporation has over 100 jobs posted for the State\n\nof Texas on its affiliated Workday page with approximately 93 of those jobs\u2014the vast majority of\n\nwhich are engineering jobs\u2014listed for Austin, Texas. (See\n\nhttps://nvidia.wd5.myworkdayjobs.com/NVIDIAExternalCareerSite?locations=91336993fab910\n\naf6d702939a7fcc2d9&locations=91336993fab910af6d702b631b94c2de (approximately 111\n\nNvidia job postings for Texas).) These jobs are particularly relevant to the Asserted Patents and\n\nAccused Products, as defined below, because they pertain to artificial intelligence, machine\n\nlearning, deep learning, data centers, accelerated computing, high performance computing\n\n(\u201cHPC\u201d), and related hardware, software, and/or firmware\u2014including Nvidia\u2019s GPUs, CPUs,\n\nsystems-on-a-chip (\u201cSoCs\u201d), platforms, and application programming interfaces.\n\n 18. Nvidia\u2019s operations in this District include client outreach and sales for each of the\n\nAccused Products and related or supporting services. As detailed above, Nvidia has customer-\n\nfacing personnel and operations in this District. Nvidia also provides technical support to partners\n\nand customers for its products in the District.\n\n 19. Nvidia has committed acts of infringement within this District. Nvidia uses the\n\nAccused Products in this District in manners that practice the Asserted Patents, including by testing\n\nthe Accused Products and by using the Accused Products at its offices and premises in this District.\n\n 20. Defendant makes, uses, advertises, offers for sale, and/or sells hardware for\n\naccelerated computing, including GPUs, CPUs, and SoCs; computers for accelerated computing\n\n(e.g., supercomputers, servers, and data centers for high performance computing); and computer\n\nplatform software-as-a-service (\u201cSaaS\u201d) that implements accelerated computing (including the\n\nAccused Products) in the State of Texas and in this District directly and/or through its partnerships\n\n\n\n\n 5\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 7 of6 95\n of 94\n\n\n\n\nwith businesses in the State of Texas and in this District. Defendant also provides data center and\n\nHPC services that practice the Asserted Patents in the State of Texas and in this District directly\n\nand/or through its partnerships with businesses in the State of Texas and in this District.\n\n 21. Nvidia sells, offers for sale, advertises, makes, installs, and/or otherwise provides\n\nhardware, software, firmware, and/or computer platforms for accelerated computing and data\n\ncenter and HPC services, including the Accused Products, the use of which infringes the Asserted\n\nPatents in this District and the State of Texas. (See https://www.nvidia.com/en-us/data-\n\ncenter/solutions/accelerated-computing/.) Nvidia performs these acts directly and/or through its\n\npartnerships with other entities. (See id. (\u201cNVIDIA has defined a range of accelerated platforms\n\nthat each consist of hardware systems designed according to the needs of the use case as well as\n\nthe software stack that enables the operation and management of the business applications. These\n\nhardware systems and software are available from NVIDIA and our partners.\u201d).)\n\n 22. Nvidia also uses a network of partners, which comprise re-sellers, managed service\n\nproviders, and product and solution experts, to provide the Accused Products and implementation\n\nservices for the Accused Products to customers in this District. Each of these partners sells, offers\n\nfor sale, installs, and/or implements Nvidia\u2019s accelerated computing hardware, software, and/or\n\ncomputer platform services. (See https://www.nvidia.com/en-us/about-nvidia/partners/.)\n\n 23. Nvidia\u2019s partners include \u201cData Center Provider[s].\u201d (See\n\nhttps://www.nvidia.com/en-us/about-nvidia/partners/.) Nvidia\u2019s Data Center Provider partners\n\n\u201coffer colocation services such as high-density data center facilities, interconnected infrastructure,\n\nand state-of-art cooling technologies for hosting NVIDIA DGX\u2122 servers globally.\u201d (See id.)\n\nNvidia\u2019s Data Center Provider partners in the \u201cNVIDIA DGX-Ready Data Center program, built\n\non the NVIDIA DGX\u2122 platform and delivered by NVIDIA partners,\u201d help \u201caccelerate the scaling\n\n\n\n\n 6\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 8 of7 95\n of 94\n\n\n\n\nof AI across [a customer\u2019s] organization.\u201d (See https://www.nvidia.com/en-us/data-\n\ncenter/colocation-partners/#aligned-energy.)\n\n 24. As further detailed below, Nvidia engages in activities that directly infringe the\n\nAsserted Patents within this District. For example, Nvidia\u2019s operation and use of its accelerated\n\ncomputing hardware, software, and/or computer platform services, including its data center-scale\n\naccelerated computing platforms, within this District infringe the Asserted Patents.\n\n 25. Nvidia also infringes (directly or indirectly) the Asserted Patents by providing\n\nservices in connection with the Accused Products including installing, maintaining, supporting,\n\noperating, providing instructions, and/or advertising Nvidia\u2019s computer platform, data center, and\n\nHPC services within this District. For example, under Nvidia\u2019s cloud and data center line of\n\nproducts and services, the Nvidia DGX platform is a \u201ca fully integrated hardware and software AI\n\nplatform\u201d and \u201ccombines the best of NVIDIA software, infrastructure, and expertise in a modern,\n\nunified AI development solution.\u201d (See https://www.nvidia.com/en-us/data-center/dgx-platform/.)\n\nIndeed, \u201cDGX infrastructure is a complete AI solution, and includes NVIDIA AI Enterprise\n\nsoftware to accelerate data science pipelines and streamline development and deployment of\n\nproduction-grade AI applications.\u201d (See id.) Nvidia platform user and partner customers infringe\n\nthe Asserted Patents by installing and operating Nvidia\u2019s computer platform software, which\n\nperforms the claimed methods in the Asserted Patents within this District. (See also, e.g.,\n\nhttps://www.nvidia.com/en-us/data-center/products/ai-enterprise/ (Nvidia AI Enterprise);\n\nhttps://developer.nvidia.com/cuda-zone (Nvidia CUDA Toolkit); https://www.nvidia.com/en-\n\nus/data-center/gpu-cloud-computing/ (GPU Cloud Computing).)\n\n 26. Defendant encourages and induces its customers of the Accused Products to\n\nperform the methods claimed in the Asserted Patents. For example, Nvidia makes its accelerated\n\n\n\n\n 7\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 9 of8 95\n of 94\n\n\n\n\ncomputing platforms and services available on its website, widely advertises those platforms and\n\nservices, provides applications that allow partners and users to access those platforms and services,\n\nprovides instructions for installing, and maintaining those platforms and services and supporting\n\nsoftware and/or firmware, and provides technical support to users. (See\n\nhttps://www.nvidia.com/en-us/data-center/dgx-support/.)\n\n 27. Nvidia further encourages and induces its customers to operate Nvidia\u2019s hardware\n\nand software in an infringing manner, and to use Nvidia\u2019s infringing computer platforms, by\n\nproviding directions for and encouraging customers to install software, such as software for\n\nNVIDIA AI Enterprise and CUDA, (see https://docs.nvidia.com/ai-enterprise/deployment-guide-\n\nvmware/0.1.0/software.html; https://developer.nvidia.com/cuda-downloads), which offers\n\nevaluation, installation, configuration, customization, and development of Nvidia\u2019s infringing\n\nsoftware products and services.\n\n 28. Defendant also contributes to the infringement of its customers and end users of the\n\nAccused Products by offering within the United States or importing into the United States the\n\nAccused Products, which are for use in practicing, and under normal operation practice, one or\n\nmore of the methods claimed in the Asserted Patents, constituting a material part of the inventions\n\nclaimed, and not a staple article or commodity of commerce suitable for substantial non-infringing\n\nuses. Indeed, as shown herein, the Accused Products and the example functionality described\n\nbelow have no substantial non-infringing uses and are specifically designed to practice the methods\n\nclaimed in the Asserted Patents.\n\n 29. On information and belief, Defendant has not disputed that venue is proper in this\n\nDistrict in cases filed against it in this District. (See, e.g., Vantage Micro LLC v. NVIDIA Corp.,\n\nNo. 6:19-cv-00582, ECF 22; Polaris Innovations Ltd. v. Dell Inc. et al., No. 5:16-cv-00451, ECF\n\n\n\n\n 8\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30Filed\n Filed\n 08/18/26\n 12/12/24Page\n Page\n 10 of\n 9 of\n 9594\n\n\n\n\n19; Cirrus Logic, Inc. v. ATI Techs., et al., No. 1:03-cv-00302, ECF 6.)\n\n 30. Defendant\u2019s infringement adversely impacts Plaintiff in this District.\n\n PLAINTIFF\u2019S PATENTED INNOVATIONS\n\n 31. The Asserted Patents pioneered the adaptation of GPU-acceleration technology to\n\nthe supervised execution of complex artificial intelligence algorithms and numerical simulations,\n\nsuch that it became possible for the first time to dynamically supervise, review, and correct\n\nintermediate \u201csolutions\u201d that were produced by these accelerated algorithms and simulations\n\nwithout performance loss.\n\n The GPU-Based Acceleration Patents\n U.S. Patent Nos. 8,648,867, RE49,461, and RE48,438\n\n 32. The \u2019867, \u2019461, and \u2019438 Patents are part of the same patent family and generally\n\ndisclose and claim systems and methods related to the accelerated execution of numerical\n\nsimulations and neural networks such that the intermediate outputs of a given execution \u201cstep\u201d can\n\nbe dynamically transferred from the GPU to the CPU, reviewed, and corrected within the same\n\ncomputational cycle before being fed as inputs to the next execution step.\n\n 33. The \u2019867 Patent is entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on September 24, 2007, and was duly and legally issued by the United States\n\nPatent and Trademark Office (\u201cUSPTO\u201d) on February 11, 2014. The \u2019867 Patent claims priority\n\nto Provisional Application No. 60/826,892, filed on September 25, 2006. A true and correct copy\n\nof the \u2019867 Patent is attached as Exhibit 1.\n\n 34. The \u2019438 Patent is entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on November 9, 2017, and was duly and legally issued by the USPTO on\n\nFebruary 16, 2021. The \u2019438 Patent is a re-issue of the \u2019867 Patent and claims priority to\n\nProvisional Application No. 60/826,892, filed on September 25, 2006. A true and correct copy of\n\n\n\n 9\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 11 10\n of 95\n of 94\n\n\n\n\nthe \u2019438 Patent is attached as Exhibit 2.\n\n 35. The \u2019461 Patent is also entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on December 29, 2020, and was duly and legally issued by the USPTO on\n\nMarch 14, 2023. The \u2019461 Patent is a re-issue of the \u2019867 Patent and claims priority to Provisional\n\nApplication No. 60/826,892, filed on September 25, 2006. A true and correct copy of the \u2019461\n\nPatent is attached as Exhibit 3.\n\n 36. The \u2019867 Patent improves upon prior GPU acceleration technology by disclosing\n\nand claiming a novel hardware and firmware system for performing a numerical simulation that\n\npermits dynamic editing of the outputs that flow from intermediate \u201csteps\u201d of that simulation,\n\nbefore they become inputs to the next \u201cstep.\u201d In particular, the \u2019867 patent discloses a CPU tethered\n\nto a GPU-based accelerator, each with their own corresponding memories, and an accelerator\n\n\u201ccontroller\u201d that coordinates transfers of data between the CPU and the GPU-based accelerator,\n\nsuch that the intermediate results from one step can be transferred from the GPU-based accelerator\n\nto the CPU, reviewed and corrected by the CPU, and transferred back to the GPU-based accelerator\n\nbefore the next computational cycle begins.\n\n 37. The \u2019867 Patent explains that performing the numerical computation in this\n\nstepwise fashion enables the system to eliminate \u201crace conditions,\u201d i.e., conflicts that occur when\n\ntwo programmatic \u201cthreads\u201d attempt to change the same shared data at the same time, which would\n\notherwise occur when other system elements attempt to access intermediate outputs of the\n\nnumerical computation. (See \u2019867 Patent, 5:60-6:31.) This avoids the computational overhead\n\nprevalent in conventional GPU-based accelerator architectures when transferring data from the\n\naccelerator to the CPU.\n\n\n\n\n 10\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 12 11\n of 95\n of 94\n\n\n\n\n 38. By enabling such \u201ccontroller-driven data exchange\u201d between the GPU-based\n\naccelerator and the CPU, the system described in the \u2019867 Patent allows for an \u201cinput parser\u201d\n\nexecuting on a CPU core to \u201cchange input\u2026on the fly during the simulation,\u201d thus enabling\n\nautomatic review and dynamic error correction of the numerical simulations or neural networks\n\nthat are being executed on the claimed system. (See id., 9:8-17.) Such dynamic, in-execution\n\nreview and error correction of whether each intermediate \u201cstep\u201d of a simulation or neural network\n\nis generating correct results is essential to the performance and reliability of large language models,\n\nimage classification, and image generation models that have become prevalent today. Because of\n\nthe scale to which such simulations and models have grown, it is no longer feasible to \u201crestart\u201d\n\nthem from scratch, only to correct them as they execute.\n\n 39. The \u2019461 and \u2019438 Patents disclose hardware and firmware configurations similar\n\nto those of the \u2019867 Patent, but are directed to using those configurations to process the layers of\n\nan artificial neural network (\u201cANN\u201d). The \u2019461 Patent is directed to further interplay between the\n\nCPU and the GPU-based accelerator: separating the CPU and GPU-based accelerator into separate\n\n\u201cstreams,\u201d whereby the CPU executes a \u201cuser interaction stream\u201d (e.g., enabling the parsing and\n\ndynamic editing of intermediate outputs, or for the ANN to be paused and resumed), while the\n\naccelerator executes a \u201ccomputational stream\u201d that executes the layers of the artificial neural\n\nnetwork. When the ANN is initialized, control over the generation of outputs shifts to the\n\ncomputational stream. However, once a pre-defined layer of the ANN has completed execution,\n\nor is interrupted, control over the generation of outputs and feeding of inputs is shifted back to the\n\nCPU\u2019s user interaction stream.\n\n 40. The Asserted Patents describe this \u201cshift of priorities\u201d as \u201c[t]he crucial feature of\n\nthe interaction between the User Interaction Stream and the Computational Stream.\u201d (\u2019867 Patent,\n\n\n\n\n 11\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 13 12\n of 95\n of 94\n\n\n\n\n7:45-47.) Even though the computational stream is in control during the ANN computation,\n\npriority shifting enables \u201c[t]he user [to] retain[] the ability to interrupt the simulation, change the\n\ninput, or to change the display properties of the framework\u201d because the user\u2019s \u201cinteractions are\n\nqueued to be performed at times determined by the controller-driven data exchange to avoid\n\ncorruption of the data.\u201d (Id., 8:47-57.)\n\n ACCUSED PRODUCTS\n\n 41. Nvidia offers, sells, and uses several products that provide and implement GPU-\n\nacceleration hardware, software, platforms, and services for individuals and enterprises and\n\nincorporate Plaintiff\u2019s patented technologies. (See https://www.nvidia.com/en-\n\nus/solutions/ai/inference/; https://marketplace.nvidia.com/en-us/data-center/?page=4;\n\nhttps://marketplace.nvidia.com/en-us/laptops-workstations/?page=9;\n\nhttps://marketplace.nvidia.com/en-us/software/?page=3.)\n\n 42. The Accused Products include Nvidia\u2019s GPU accelerators and superchips. (See\n\nhttps://resources.nvidia.com/l/en-us-gpu.) Nvidia\u2019s GPU accelerators include Nvidia\u2019s GPUs with\n\nNvidia\u2019s \u201cHopper,\u201d \u201cAda Lovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d \u201cPascal,\u201d and \u201cMaxwell\u201d\n\nGPU architectures. (See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-\n\nmatrix/index.html.) These GPUs are specifically designed to run and implement GPU-based\n\nhardware acceleration using Nvidia\u2019s proprietary CUDA (Compute Unified Device Architecture)\n\nplatform and CUDA libraries for GPU acceleration. (See id. (Nvidia GPU architectures\n\nimplementing Nvidia\u2019s cuDNN (CUDA Deep Neural Network) library for GPU acceleration.);\n\nhttps://developer.nvidia.com/cuda-gpus.)\n\n 43. Nvidia\u2019s Hopper GPUs include the H100 and H200 GPUs. (See\n\nhttps://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/ (Hopper\n\n\n\n\n 12\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 14 13\n of 95\n of 94\n\n\n\n\narchitecture); https://www.nvidia.com/en-us/data-center/h100/ (H100);\n\nhttps://www.nvidia.com/en-us/data-center/h200/ (H200).) In addition, Nvidia\u2019s superchips that\n\nimplement GPU accelerators include the GH200, or Grace Hopper Superchip, which implements\n\nthe Hopper-GPU architecture. (See https://www.nvidia.com/en-us/data-center/grace-hopper-\n\nsuperchip/ (GH200).)\n\n 44. Nvidia\u2019s Ada Lovelace (or Lovelace) GPUs include Nvidia Data Center GPUs,\n\nincluding L40, L40S, and L4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including\n\nRTX Ada Generations series GPUs and Laptop GPUs; and GeForce RTX 40 series GPUs and\n\nLaptop GPUs. (See https://www.nvidia.com/en-us/technologies/ada-architecture/ (Ada Lovelace\n\narchitecture). See https://www.nvidia.com/en-us/data-center/l40/ (L40);\n\nhttps://www.nvidia.com/en-us/data-center/l40s/ (L40S); https://www.nvidia.com/en-us/data-\n\ncenter/l4/ (L4). See https://resources.nvidia.com/en-us-design-viz-stories-ep/l40-linecard (Nvidia\n\nProfessional GPUs); https://www.nvidia.com/en-us/ai-on-rtx/ (RTX GPUs featuring \u201cAccelerated\n\nDevelopment\u201d); https://www.nvidia.com/en-us/design-visualization/desktop-graphics/ (RTX Ada\n\nGeneration GPUs); https://www.nvidia.com/en-us/design-visualization/rtx-professional-\n\nlaptops/compare-table/ (RTX Ada Generation Laptop GPUs). See https://www.nvidia.com/en-\n\nus/geforce/graphics-cards/40-series/ (GeForce RTX 40 GPUs); https://www.nvidia.com/en-\n\nus/geforce/graphics-cards/compare/ (GeForce RTX 40 GPUs); https://www.nvidia.com/en-\n\nus/geforce/laptops/compare/ (GeForce RTX 40 Laptop GPUs).)\n\n 45. Nvidia\u2019s Ampere GPUs include Nvidia Data Center GPUs, including A100, A40,\n\nA30, A16, A10, and A2 GPUs; Nvidia Workstation and Professional Laptop GPUs, including\n\nRTX A series GPUs and Laptop GPUs; GeForce RTX 30 series GPUs and Laptop GPUs; and\n\nGeForce MX570 Laptop GPU. (See https://www.nvidia.com/en-us/data-center/ampere-\n\n\n\n\n 13\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 15 14\n of 95\n of 94\n\n\n\n\narchitecture/ (Ampere architecture). See https://www.nvidia.com/en-us/data-center/a100/ (A100);\n\nhttps://www.nvidia.com/en-us/data-center/a40/ (A40); https://www.nvidia.com/en-us/data-\n\ncenter/a30/ (A30); https://www.nvidia.com/en-us/data-center/a16/ (A16);\n\nhttps://www.nvidia.com/en-us/data-center/a10/ (A10); https://www.nvidia.com/en-us/data-\n\ncenter/a2/ (A2). See https://www.nvidia.com/en-us/design-visualization/desktop-graphics/ (RTX\n\nA GPUs); https://www.nvidia.com/en-us/design-visualization/rtx-professional-laptops/compare-\n\ntable/ (RTX A Laptop GPUs). See https://www.nvidia.com/en-us/geforce/graphics-cards/30-\n\nseries/ (GeForce RTX 30 GPUs); https://www.nvidia.com/en-us/geforce/graphics-cards/compare/\n\n(GeForce RTX 30 GPUs); https://www.nvidia.com/en-us/geforce/laptops/compare/30-series/\n\n(GeForce RTX 30 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n570/ (GeForce MX570 Laptop GPU).)\n\n 46. Nvidia\u2019s Turing GPUs include Nvidia Data Center GPUs, including Tesla T4 GPUs\n\nand Quadro RTX 8000 (passive) and Quadro RTX 6000 (passive) GPUs; Nvidia Workstation and\n\nProfessional Laptop GPUs, including T series GPUs and Laptop GPUs, Quadro T series Laptop\n\nGPUs, and Quadro RTX series GPUs and Laptop GPUs; Titan series Titan RTX GPU;\n\nGeForce RTX 20 series and GeForce GTX 16 series GPUs and Laptop GPUs; and GeForce\n\nMX550, MX450, and MX430 Laptop GPUs. (See https://www.nvidia.com/en-us/geforce/turing/\n\n(Turing architecture). See https://www.nvidia.com/en-us/data-center/tesla-t4/ (Tesla T4);\n\nhttps://www.nvidia.com/en-gb/design-visualization/quadro-data-center/ (Quadro RTX 8000\n\n(passive) and Quadro RTX 6000 (passive). See https://www.nvidia.com/en-us/design-\n\nvisualization/quadro/ (T series GPUs/Laptop GPUs, Quadro T series Laptop GPUs, and Quadro\n\nRTX GPUs/Laptop GPUs); https://www.nvidia.com/en-us/design-visualization/desktop-graphics\n\n(T series GPUs/Laptop GPUs); https://www.nvidia.com/content/dam/en-\n\n\n\n\n 14\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 16 15\n of 95\n of 94\n\n\n\n\nzz/Solutions/titan/documents/titan-rtx-for-creators-us-nvidia-1011126-r6-web.pdf (Titan RTX);\n\nhttps://www.nvidia.com/en-us/geforce/20-series/ (GeForce RTX 20 GPUs);\n\nhttps://www.nvidia.com/en-us/geforce/graphics-cards/compare/ (GeForce RTX 20 GPUs and\n\nGeForce GTX 16 GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/compare-20-\n\nseries/ (GeForce RTX 20 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-\n\nlaptops/compare-16-series/ (GeForce GTX 16 Laptop GPUs); https://www.nvidia.com/en-\n\nus/geforce/gaming-laptops/mx-550/ (GeForce MX550 Laptop GPU); https://www.nvidia.com/en-\n\nus/geforce/gaming-laptops/mx-450/ (GeForce MX450 Laptop GPU);\n\nhttps://wccftech.com/nvidia-geforce-mx450-turing-discrete-notebook-gpu-gddr6-pcie-4/\n\n(GeForce M Laptop GPUs).)\n\n 47. Nvidia\u2019s Volta GPUs include Nvidia Data Center GPUs, including the Tesla V100\n\nGPU; Nvidia Workstation GPUs, including Quadro GV100; and Titan series Titan V GPU. (See\n\nhttps://www.nvidia.com/en-us/data-center/volta-gpu-architecture/ (Volta architecture);\n\nhttps://www.nvidia.com/en-us/data-center/v100/ (Tesla V100);\n\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\nvisualization/productspage/quadro/quadro-desktop/quadro-volta-gv100-data-sheet-us-nvidia-\n\n704619-r3-web.pdf (Quadro GV100); https://nvidianews.nvidia.com/news/nvidia-titan-v-\n\ntransforms-the-pc-into-ai-supercomputer (Titan V).)\n\n 48. Nvidia\u2019s Pascal GPUs include Nvidia Data Center GPUs, including Tesla P100,\n\nP40, and P4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including the\n\nQuadro GP100 GPU and Quadro P series GPUs and Laptop GPUs; Titan series Titan Xp and Titan\n\nX GPUs; GeForce GTX 10 series GPUs and Laptop GPUs; and GeForce MX300 series, MX200\n\nseries, and MX150 Laptop GPUs. (See https://developer.nvidia.com/pascal;\n\n\n\n\n 15\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 17 16\n of 95\n of 94\n\n\n\n\nhttps://www.nvidia.com/en-us/data-center/pascal-gpu-architecture/ (Pascal architecture). See\n\nhttps://www.nvidia.com/en-us/data-center/tesla-p100 (Tesla P100);\n\nhttps://developer.nvidia.com/cuda-gpus (Tesla P40 and P4);\n\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\nvisualization/productspage/quadro/quadro-desktop/quadro-pascal-gp100-data-sheet-us-nv-\n\n704562-r1.pdf (Quadro GP100); https://www.nvidia.com/en-us/design-visualization/quadro/\n\n(Quadro P series GPUs/Laptop GPUs). See https://www.nvidia.com/content/geforce-\n\ngtx/NVIDIA_TITAN_X_USER_GUIDE_v02.pdf (Titan X);\n\nhttps://www.nvidia.com/content/geforce-gtx/NVIDIA_TITAN_Xp_USER_GUIDE_v02.pdf\n\n(Titan Xp); https://www.nvidia.com/en-us/geforce/10-series/ (GeForce GTX 10);\n\nhttps://www.nvidia.com/en-us/geforce/graphics-cards/compare/ (GeForce GTX 10 GPUs);\n\nhttps://www.nvidia.com/en-us/geforce/news/gfecnt/nvidia-geforce-gtx-10-series-laptops/\n\n(GeForce GTX 10 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n350/ (GeForce MX350 Laptop GPU); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n330/ (GeForce MX330 Laptop GPU); https://wccftech.com/nvidia-geforce-mx450-turing-\n\ndiscrete-notebook-gpu-gddr6-pcie-4/ (GeForce M Laptop GPUs).)\n\n 49. Nvidia\u2019s Maxwell GPUs include Nvidia Data Center GPUs, including Tesla M60,\n\nM40, and M10 GPUs; Nvidia Workstation and Professional Laptop GPUs, including Quadro M\n\nseries GPUs and Laptop GPUs, the NVS 810 GPU, and Tesla M6 series Laptop GPUs; Titan series\n\nGTX Titan X GPU; GeForce GTX 900 series and GeForce GTX 700 series GPUs and Laptop\n\nGPUs; and GeForce MX130 series and MX110 Laptop GPUs. (See\n\nhttps://developer.nvidia.com/blog/maxwell-most-advanced-cuda-gpu-ever-made/ (Maxwell\n\narchitecture); https://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\n\n\n\n 16\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 18 17\n of 95\n of 94\n\n\n\n\nvisualization/solutions/resources/documents1/nvidia-m60-datasheet.pdf (M60);\n\nhttps://images.nvidia.com/content/tesla/pdf/78071_Tesla_M40_24GB_Print_Datasheet_LR.PDF\n\n(M40); https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-\n\nm10/pdf/188359-Tesla-M10-DS-NV-Aug19-A4-fnl-Web.pdf (M10);\n\nhttps://www.nvidia.com/en-us/design-visualization/quadro/ (Quadro M GPUs/Laptop GPUs);\n\nhttps://www.nvidia.com/docs/IO/146527/nvs-810-datasheet.pdf (NVS 810);\n\nhttps://images.nvidia.com/content/tesla/pdf/188300-Tesla-M6-DS-Aug19-A4-fnl-Web.pdf (Tesla\n\nM6); https://www.nvidia.com/content/geforce-gtx/GTX_TITAN_X_User_Guide.pdf (GTX Titan\n\nX); https://developer.nvidia.com/maxwell-compute-architecture (GeForce GTX 900 and 700\n\nGPUs/Laptop GPUs); https://wccftech.com/nvidia-geforce-mx450-turing-discrete-notebook-gpu-\n\ngddr6-pcie-4/ (GeForce M Laptop GPUs).)\n\n 50. These GPUs and superchips implement, and are specifically designed for, GPU-\n\nacceleration for artificial intelligence and neural networks. Nvidia\u2019s proprietary CUDA platform\n\nfor parallel computing, which includes GPU-acceleration libraries such as cuDNN (CUDA Deep\n\nNeural Network), is implemented in the Nvidia Hopper, Ada Lovelace, Ampere, Turing, Volta,\n\nPascal, and Maxwell GPU architectures.\n\n\n\n\n 17\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 19 18\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 51. The Accused Products further include Nvidia\u2019s supercomputers and servers that\n\nimplement its GPU accelerators and superchips. These supercomputers and servers include: the\n\nEGX line of servers for data centers and edge devices, the HGX line of supercomputers, the DGX\n\nline of supercomputers, and the OVX line of supercomputers. (See https://www.nvidia.com/en-\n\nus/data-center/solutions/accelerated-computing/.)\n\n 52. Nvidia\u2019s \u201cEGX hardware portfolio\u201d includes \u201caccelerators [that] combine the\n\nperformance of NVIDIA Ampere GPUs.\u201d (See https://www.nvidia.com/en-us/data-\n\ncenter/products/egx/; see https://www.nvidia.com/en-us/design-visualization/egx-graphics/.)\n\nNvidia\u2019s HGX \u201cAI supercomputing platform brings together the full power of NVIDIA GPUs,\n\nNVIDIA NVLink\u2122, NVIDIA networking, and fully optimized AI and high-performance\n\ncomputing (HPC) software stacks.\u201d (See https://www.nvidia.com/en-us/data-center/hgx/;\n\nhttps://nvdam.widen.net/s/5kgbjq2v2t/hpc-hgx-h100-datasheet-nvidia-web.) One example\n\n\n 18\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 20 19\n of 95\n of 94\n\n\n\n\nconfiguration includes \u201cfour or eight H200 or H100 GPUs.\u201d (See id.; see\n\nhttps://nvdam.widen.net/s/5kgbjq2v2t/hpc-hgx-h100-datasheet-nvidia-web.) Nvidia\u2019s DGX\n\nsupercomputers include the DGX H200, DGX BasePOD, and DGX SuperPOD with DGX GB200.\n\n(See https://www.nvidia.com/en-us/data-center/dgx-platform/; see also\n\nhttps://www.nvidia.com/en-us/data-center/base-command/; https://resources.nvidia.com/en-us-\n\ndgx-software/nvidia-base-command (DGX Base Command operating system for DGX data\n\ncenters.) And Nvidia\u2019s OVX supercomputers implement \u201cL40S GPUs . . . for both complex AI\n\nand graphics-intensive workloads.\u201d (See https://www.nvidia.com/en-us/data-center/products/ovx/;\n\nsee https://resources.nvidia.com/en-us-ovx/ovx-datasheet.)\n\n 53. The Accused Products further include Nvidia\u2019s software, platforms, and services\n\nfor accelerated computing. These include CUDA, Nvidia AI Enterprise, the DGX Platform, Nvidia\n\nOmniverse, Nvidia Drive, Nvidia Isaac Sim, and Nvidia NGC.\n\n 54. CUDA is Nvidia\u2019s proprietary \u201cparallel computing platform and programming\n\nmodel.\u201d (See https://developer.nvidia.com/cuda-zone.) CUDA is designed to support Nvidia\u2019s\n\nGPU accelerators and superchips and includes software specifically for GPU-acceleration such as\n\nthe cuDNN \u201cGPU-accelerated library.\u201d (See id.; https://developer.nvidia.com/cudnn.) In addition,\n\nNvidia\u2019s CUDA-X, built on top of CUDA, is a collection of \u201cGPU-accelerated microservices and\n\nlibraries for AI.\u201d (See https://www.nvidia.com/en-us/technologies/cuda-x/.) Nvidia also offers the\n\nCUDA Toolkit and SDK Manager for developing GPU-accelerated applications. (See\n\nhttps://developer.nvidia.com/cuda-toolkit; https://developer.nvidia.com/sdk-manager.)\n\n 55. In addition, Nvidia AI Enterprise is Nvidia\u2019s \u201cend-to-end, cloud-native software\n\nplatform\u201d for \u201caccelerat[ing] data science pipelines . . . and other generative AI applications.\u201d (See\n\nhttps://www.nvidia.com/en-us/data-center/products/ai-enterprise/.) It is Nvidia\u2019s \u201c\u2018operating\n\n\n\n\n 19\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 21 20\n of 95\n of 94\n\n\n\n\nsystem\u2019 for enterprise AI.\u201d (See id.)\n\n 56. In addition, Nvidia\u2019s DGX platform is \u201cis a complete AI solution, and includes\n\nNVIDIA AI Enterprise software.\u201d (See https://www.nvidia.com/en-us/data-center/dgx-platform/.)\n\nNvidia DGX Cloud is \u201can AI-training-as-a-service platform which includes cloud-based\n\ninfrastructure and software for AI, customizable pretrained AI models, and access to NVIDIA\n\nexperts.\u201d (See https://d18rn0p25nwr6d.cloudfront.net/CIK-0001045810/1cbe8fe7-e08a-46e3-\n\n8dcc-b429fc06c1a4.pdf, Nvidia U.S. Securities and Exchange Commission Form 10-K for Fiscal\n\nYear Ended January 28, 2024 at 6.)\n\n 57. In addition, Nvidia Omniverse is \u201ca development platform and operating system\n\nfor building virtual world simulation applications, available as a software subscription.\u201d (See\n\nhttps://d18rn0p25nwr6d.cloudfront.net/CIK-0001045810/1cbe8fe7-e08a-46e3-8dcc-\n\nb429fc06c1a4.pdf, Nvidia U.S. Securities and Exchange Commission Form 10-K for Fiscal Year\n\nEnded January 28, 2024 at 6.) Nvidia Omniverse implements software and services \u201cinto existing\n\nsoftware tools and simulation workflows for building AI systems.\u201d (See\n\nhttps://www.nvidia.com/en-us/omniverse/.)\n\n 58. In addition, Nvidia Drive is a platform that \u201cconsists of both the AI infrastructure\n\nand in-vehicle hardware and software\u201d for autonomous vehicles. (See https://www.nvidia.com/en-\n\nus/self-driving-cars/.) \u201cNVIDIA DRIVE Infrastructure encompasses data center hardware,\n\nsoftware, and workflows\u2014both on premises and in NVIDIA DGX Cloud & Omniverse.\u201d (See id.)\n\n 59. In addition, Nvidia Isaac Sim is a platform that enables \u201cdevelopers to design,\n\nsimulate, test, and train AI-based robots and autonomous machines in a physically-based virtual\n\nenvironment.\u201d (See https://developer.nvidia.com/isaac/sim.) It is built on Nvidia Omniverse. (See\n\nid.)\n\n\n\n\n 20\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 22 21\n of 95\n of 94\n\n\n\n\n 60. In addition, Nvidia NGC is a collection of software services and tools that support\n\n\u201cend-to-end AI and digital twin workflows\u201d that runs on \u201cNVIDIA GPU-accelerated platforms.\u201d\n\n(See https://www.nvidia.com/en-us/gpu-cloud/.) NGC \u201coffers a collection of cloud services . . .\n\nfor generative AI, drug discovery, and speech AI solutions, and the NGC Private Registry for\n\nsecurely sharing proprietary AI software.\u201d (See id.)\n\n FIRST CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019867 PATENT)\n\n 61. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n 62. Defendant has infringed and continues to infringe one or more claims of the \u2019867 Patent\n\nin violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will continue to\n\ndo so. The Accused Products, including features of, e.g., the Grace Hopper Superchip (GH200), at least\n\nwhen used for their ordinary and customary purposes, practice each element of at least claim 16 of the\n\n\u2019867 Patent as demonstrated below.\n\n 63. For example, claim 16 of the \u2019867 Patent recites:\n\n 16. A method for performing a numerical simulation on input data\n in a computer system including a central processing unit and an\n accelerator, the method comprising:\n\n receiving, by an accelerator, first input data from the central\n processing unit;\n\n transferring, by an accelerator controller, the first input data into a\n first partition, referenced by first pointer, of an accelerator memory\n before a first computational cycle of the numerical simulation;\n\n performing, by at least one graphics processing unit during the first\n computational cycle, at least one calculation on the first portion of\n the input data as to generate first output data;\n\n\n\n\n 21\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 23 22\n of 95\n of 94\n\n\n\n\n storing, by the accelerator controller, the first output data into a\n second partition, referenced by a second pointer, of the accelerator\n memory; and\n\n swapping the first pointer with the second pointer at the end of the\n first computational cycle, such that the first output data becomes an\n input for a second computational cycle of the numerical simulation.\n\n 64. The Accused Products perform each step of the method of claim 16 of the \u2019867\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod for performing a numerical simulation on input data in a computer system including a\n\ncentral processing unit and an accelerator, as further explained below. For instance, the Grace\n\nHopper Superchip (GH200) \u201cbrings together the groundbreaking performance of the NVIDIA\n\nHopper GPU with the versatility of the NVIDIA Grace\u2122 CPU . . . in a single Superchip.\u201d\n\n\n\n\n 22\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 24 23\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 65. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 66. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA enables NVIDIA GPUs to be used for general purpose computing tasks. CUDA further\n\nincludes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a compiler, development\n\ntools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural Network library\n\n(cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d It \u201cprovides\n\nhighly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n 23\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 25 24\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 67. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 24\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 26 25\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 68. The Accused Products perform a method that includes receiving, by an accelerator,\n\nfirst input data from the central processing unit. For instance, the \u201cCUDA programming model\u201d\n\nimplements programming functions and instructions for CPUs and GPUs. \u201cThe host is the CPU\n\navailable in the system\u201d and \u201csystem memory associated with the CPU is called host memory.\u201d\n\n\u201cThe GPU is called a device and GPU memory likewise called device memory.\u201d As an example,\n\nthe first main CUDA program execution step is \u201c[c]opy[ing] the input data from host [CPU]\n\nmemory to device [GPU] memory, also known as host-to-device transfer.\u201d\n\n\n\n\n 25\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 27 26\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 69. The Accused Products practice a method that includes transferring, by an\n\naccelerator controller, the first input data into a first partition, referenced by first pointer, of an\n\naccelerator memory before a first computational cycle of the numerical simulation. For instance,\n\nthe GPU architecture of the Accused Products implements a controller. As an example, the\n\nHopper-GPU architecture implements \u201cHBM3 memory controllers\u201d including \u201c12 512-bit\n\nmemory controllers\u201d coupled GPU memory including \u201c6 HBM3 or HBM2e stacks,\u201d \u201c80 GB\n\nHBM3, 5 HBM3 stacks,\u201d and \u201c80 GB HBM2e, 5 HBM2e stacks.\u201d\n\n\n\n\n 26\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 28 27\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ (emphasis added).)\n\n 70. The Accused Products implement CUDA, Nvidia\u2019s parallel computing platform.\n\nCUDA enables NVIDIA GPUs to be used for general purpose computing tasks and includes\n\nspecialized GPU-acceleration libraries such as cuDNN. Examples of parameters used in CUDA\n\ninclude pointers \u201cdst\u201d (\u201cDestination memory address\u201d) and \u201csrc\u201d (\u201cSource memory address\u201d). For\n\ninstance, exemplary CUDA function \u201ccudaMemcpy\u201d copies \u201cbytes [data] from the memory area\n\npointed to by src [source memory address pointer] to the memory area pointed to by dst\n\n\n\n 27\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 29 28\n of 95\n of 94\n\n\n\n\n[destination memory address pointer], where kind [type of transfer] specifies the direction of the\n\ncopy.\u201d One of the destinations is \u201ccudaMemcpyHostToDevice,\u201d or host (CPU) to device (GPU).\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 71. The Accused Products practice a method that includes performing, by at least one\n\ngraphics processing unit during the first computational cycle, at least one calculation on the first\n\nportion of the input data as to generate first output data. For instance, CUDA uses \u201cstreams\u201d to\n\nexecute a sequence of commands in order. As shown below, an exemplary CUDA function\n\n\u201ccudaMemcpyAsync\u201d is used to copy data between a host (CPU) and a device (GPU). \u201cEach\n\nstream copies its portion of input array hostPtr [pointer for CPU] to array inputDevPtr in device\n\n[GPU] memory.\u201d The stream then \u201cprocesses inputDevPtr on the device [GPU] by calling\n\nMyKernel(), and copies the result outputDevPtr back to the same portion of hostPtr.\u201d\n\n\n\n\n 28\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 30 29\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#creation-and-\n\ndestruction-of-streams (emphasis added).)\n\n 72. The Accused Products practice a method that includes storing, by the accelerator\n\ncontroller, the first output data into a second partition, referenced by a second pointer, of the\n\naccelerator memory. For instance, exemplary excerpts of CUDA code shown below demonstrate\n\nCUDA being used to calculate a square sub-matrix Csub of matrix C using the function MatMul.\n\nAn exemplary CUDA stream \u201callocate[s] [matrix] C in device memory.\u201d After Matrices A and B\n\nare synchronized and multiplied, the exemplary CUDA stream \u201c[w]rite[s] Csub to device [GPU]\n\nmemory.\u201d\n\n\n\n\n 29\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 31 30\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory\n\n(emphasis added).)\n\n 73. In addition, exemplary CUDA function \u201ccudaMalloc\u201d is used to \u201callocate weight,\n\nwork, and reserve space buffer sizes in the GPU memory.\u201d \u201cThe work-space buffer is used for\n\ntemporary storage\u201d and the \u201c content can be discarded or modified after all GPU kernels launched\n\nby the corresponding API complete.\u201d The \u201creserve-space buffer\u201d used to transfer intermediate\n\nresults is used for transferring \u201cintermediate results\u201d as used in the cuDNN GPU-acceleration\n\nlibrary for CUDA.\n\n\n 30\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 32 31\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-\n\n893/api/index.html#cudnnGetMultiHeadAttnBuffers (emphasis added).)\n\n 74. The Accused Products practice a method that includes swapping the first pointer\n\nwith the second pointer at the end of the first computational cycle, such that the first output data\n\nbecomes an input for a second computational cycle of the numerical simulation. For example, the\n\ncuDNN GPU-acceleration library of the Accused Products implement operations that \u201ctake tensors\n\nas input and produce tensors as output.\u201d\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n\n\n\n 31\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 33 32\n of 95\n of 94\n\n\n\n\n 75. Nvidia confirms that CUDA implements pointer swapping for device (GPU) pointers.\n\n\n\n\n *****\n\n\n\n\n(See https://forums.developer.nvidia.com/t/swap-device-pointers/38964 (emphasis added).)\n\n\n\n\n 32\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 34 33\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/#creation-and-destruction-of-\n\nstreams (emphasis added).)\n\n 76. Each claim in the \u2019867 Patent recites an independent invention. Neither claim 16,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019867 Patent.\n\n 77. Defendant has been aware of the technology patented by the \u2019867 Patent since at\n\nleast 2007, when the inventors of the Asserted Patents first discussed their patented technologies\n\n\n\n 33\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 35 34\n of 95\n of 94\n\n\n\n\nwith Mr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited the application for the \u2019867 Patent in its own patent portfolio since at least June 28, 2010.\n\n\n\n\n *****\n\n\n\n\n(See https://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy (emphasis\n\nadded).)\n\n\n\n\n *****\n\n\n\n\n 34\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 36 35\n of 95\n of 94\n\n\n\n\n(See https://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf\n\n(Nvidia U.S. Patent No. 8,922,566) (emphasis added).)\n\n\n\n\n(See\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner)\n\n(emphasis added).)\n\n 78. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019867 Patent and its related patents and applications. These discussions included at\n\nleast Mr. Alvin Lin, an Nvidia Senior Director of Business Development, and Mr. Jeff Herbst, then\n\nan Nvidia Vice President of Business Development and head of Nvidia\u2019s Inception GPU Ventures,\n\nin or around September 6, 2016. In or around October 2016, Nvidia, through its representatives,\n\n\n 35\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 37 36\n of 95\n of 94\n\n\n\n\ninitiated discussions with the inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 79. The inventors also discussed their patented technology, in addition to the \u2019867\n\nPatent and its family, with Defendant\u2019s representatives at Nvidia\u2019s artificial intelligence\n\nconference in or around June 2017. On or about June 26, 2017, Defendant received materials from\n\nthe inventors, in lieu of a meeting on or about June 29, that identified the \u2019867 Patent and its family\n\nand described the technology in detail. Defendant had previously stated it was interested in the\n\ninventors\u2019 solutions. Defendant also featured the inventors on its website as members of\n\nDefendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n 80. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\n 36\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 38 37\n of 95\n of 94\n\n\n\n\n\u2019867 Patent.\n\n 81. Defendant directly infringes at least claim 16 of the \u2019867 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 82. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 16 of the \u2019867 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 83. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 16 of the \u2019867 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 16 of the \u2019867 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 16 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\ninstallation, support, and distribution of the Accused Products.\n\n 84. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\n\n\n\n 37\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 39 38\n of 95\n of 94\n\n\n\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 85. Defendant further encourages and induces its customers to infringe claim 16 of the\n\n\u2019867 Patent: 1) by making its accelerated computing and data center services available on its\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 86. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 87. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\n\n\n\n 38\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 40 39\n of 95\n of 94\n\n\n\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019867 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 88. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019867 Patent.\n\n 89. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019867 Patent.\n\n 90. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\n\n\n\n 39\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 41 40\n of 95\n of 94\n\n\n\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 16 of the \u2019867 Patent.\n\n 91. Plaintiff has suffered and continue to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019867 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 92. Defendant\u2019s infringement of the \u2019867 Patent is knowing and willful. Defendant had\n\nactual knowledge of the \u2019867 Patent application since at least 2010 and actual knowledge of the\n\n\u2019867 Patent, and its family, since at least 2017.\n\n 93. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents\n\nand Plaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and\n\nservices that it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019867 Patent\n\nwith knowledge of the \u2019867 Patent constitutes willful infringement.\n\n SECOND CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019438 PATENT)\n\n 94. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n 95. Defendant has infringed and continues to infringe one or more claims of the \u2019438 Patent\n\nin violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will continue to\n\ndo so. The Accused Products, including features of, e.g., the Grace Hopper Superchip (GH200), at least\n\n\n\n 40\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 42 41\n of 95\n of 94\n\n\n\n\nwhen used for their ordinary and customary purposes, practice each element of at least claim 21 of the\n\n\u2019438 Patent as demonstrated below.\n\n 96. For example, claim 21 of the \u2019438 Patent recites:\n\n 21. A method of performing a sequence of computations\n representing an artificial neural network, the method comprising:\n\n receiving, at a central processing unit (CPU), first input data\n acquired from an external system in real time;\n\n initializing, by a controller operably coupled to a graphics\n processing unit (GPU), textures and shaders in a memory operably\n coupled to the GPU;\n\n transferring the first input data received by the CPU to the memory\n operably coupled to the GPU;\n\n performing, by the graphics processing unit (GPU), a first\n computation in the sequence of computations on the first input data\n based on the textures and shaders to generate first output data,\n computations in the sequence of computations representing\n respective layers of neurons in the artificial neural network, an\n output of the first computation in the sequence of computations\n representing an output of a first neuron in a first layer in the artificial\n neural network;\n\n storing, in the memory operably coupled to the GPU, the first input\n data and the first output data; and\n\n transferring second input data acquired from the external system in\n real time into the memory operably coupled to the GPU after the\n GPU starts the first computation and before the GPU starts a second\n computation of the sequence of computations, an output of the\n second computation in the sequence of computations representing\n an output of a second neuron in a second layer in the artificial neural\n network.\n\n 97. The Accused Products perform each step of the method of claim 21 of the \u2019438\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod of performing a sequence of computations representing an artificial neural network, as\n\nfurther explained below. For instance, the Grace Hopper Superchip (GH200) \u201cbrings together the\n\n\n\n 41\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 43 42\n of 95\n of 94\n\n\n\n\ngroundbreaking performance of the NVIDIA Hopper GPU with the versatility of the NVIDIA\n\nGrace\u2122 CPU . . . in a single Superchip.\u201d It includes the cuDNN (CUDA Deep Neural Network)\n\nlibrary for \u201c[d]eep neural networks.\u201d\n\n\n\n\n *****\n\n\n\n\n 42\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 44 43\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 98. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 99. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA further includes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a\n\ncompiler, development tools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural\n\nNetwork library (cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d\n\n\n\n 43\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 45 44\n of 95\n of 94\n\n\n\n\nIt \u201cprovides highly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 100. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 44\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 46 45\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 101. The Accused Products perform a method that includes receiving, at a central\n\nprocessing unit (CPU), first input data acquired from an external system in real time. For instance,\n\nas illustrated below, a diagram describing the architecture of the Grace Hopper Superchip depicts\n\na CPU (\u201cGrace CPU\u201d) with \u201c[u]p to 72 cores\u201d that receives and sends input and output data via\n\n\u201cHigh-Speed IO\u201d (\u201cPCIe-5\u201d). The Grace CPU is further depicted as being coupled to memory\n\n\u201cCPU LPDDR5X\u201d up to 480GB via a link up to \u201c500 GB/s.\u201d\n\n\n\n\n 45\n\fCase\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 47 46\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n 46\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 48 47\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 102. In a related diagram example, \u201cCPU PHYSICAL MEMORY\u201d (LPDDR5X) is\n\nillustrated as being accessed by a CPU (\u201cCPU-resident access\u201d).\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 103. As previously stated, the Accused Products implement CUDA and specialized\n\nGPU-acceleration libraries such as cuDNN. \u201cCUDA\u00ae is a parallel computing platform and\n\n\n\n 47\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 49 48\n of 95\n of 94\n\n\n\n\nprogramming model developed by NVIDIA for general computing on graphical processing units\n\n(GPUs)\u201d including \u201cGPU-accelerated applications.\u201d In GPU-accelerated applications, \u201cthe\n\nsequential part of the workload runs on the CPU \u2013 which is optimized for single-threaded\n\nperformance \u2013 while the compute intensive portion of the application runs on thousands of GPU\n\ncores in parallel.\u201d\n\n\n\n\n(See https://developer.nvidia. com/cuda-zone (emphasis added).)\n\n 104. The \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs and GPUs. \u201cThe host is the CPU available in the system\u201d and \u201csystem\n\nmemory associated with the CPU is called host memory.\u201d As an example, the first main CUDA\n\nprogram execution step is \u201c[c]opy[ing] the input data from host [CPU] memory.\u201d\n\n\n\n\n 48\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 50 49\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 105. The Accused Products perform a method that includes initializing, by a controller\n\noperably coupled to a graphics processing unit (GPU), textures and shaders in a memory operably\n\ncoupled to the GPU. For instance, as illustrated below, a GPU (\u201cHopper GPU\u201d) for the Grace\n\nHopper Superchip is depicted as accessing both CPU and GPU memory using \u201cNVLINK\u201d for both\n\naccessing and storing data. Indeed, \u201cNVIDIA GH200 is designed to accelerate applications with\n\nexceptionally large memory footprints.\u201d As illustrated, the Hopper GPUs are illustrated coupled\n\nto \u201cGPU HBM3\u201d high bandwidth memory or \u201cGPU HBM3e\u201d high bandwidth memory.\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 106. The GPU architecture of the Accused Products implements a controller. As an\n\nexample, the Hopper-GPU architecture includes \u201cGPU processing clusters\u201d and \u201ctexture\n\nprocessing clusters\u201d and implements \u201cHBM3 memory controllers\u201d including \u201c12 512-bit memory\n\ncontrollers\u201d coupled GPU memory including \u201c6 HBM3 or HBM2e stacks,\u201d \u201c80 GB HBM3, 5\n\nHBM3 stacks,\u201d and \u201c80 GB HBM2e, 5 HBM2e stacks.\u201d\n\n\n\n\n 49\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 51 50\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ (emphasis added).)\n\n 107. In addition, CUDA includes the exemplary NPP (Nvidia Performance Primitives)\n\nlibrary \u201cfor performing CUDA accelerated processing\u201d and \u201cperforming CUDA accelerated\n\nprocessing for 2D image and signal processing.\u201d\n\n\n\n\n 50\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 52 51\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/index.html (emphasis added).)\n\n\n\n\n(See https://docs.nvidia.com/cuda/npp/introduction.html (emphasis added).)\n\n 108. As an example, the exemplary CUDA NPP library passes image data using \u201c[a]\n\npointer to the image\u2019s underlying data type\u201d and \u201c[a] line step in bytes.\u201d In this example, the\n\npointer is passed \u201cto the underlying pixel data type\u201d and the pointer and line step are passed\n\nindividually for processing involving \u201cimage data.\u201d\n\n\n\n\n 51\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 53 52\n of 95\n of 94\n\n\n\n\n(See\n\nhttps://docs.nvidia.com/cuda/npp/introduction.html#nppi_conventions_lb_1passing_image_data\n\n(emphasis added).)\n\n 109. As another example, the exemplary CUDA NPP library implements function for\n\nimage color conversion. These functions \u201cmanipulat[e] an image\u2019s color model and sampling\n\nformat\u201d and \u201ccan be found in the nppicc [NVIDIA Performance Primitives Image Color\n\nConversion] library.\u201d As shown, these functions save \u201capplication load time\u201d and \u201cCUDA\n\nruntime.\u201d\n\n\n\n\n(See https://docs.nvidia.com/cuda/npp/image_color_conversion.html#image-color-model-\n\nconversion-functions (emphasis added).)\n\n 110. The Accused Products perform a method that includes transferring the first input\n\n\n 52\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 54 53\n of 95\n of 94\n\n\n\n\ndata received by the CPU to the memory operably coupled to the GPU. For instance, the \u201cCUDA\n\nprogramming model\u201d implements programming functions and instructions for CPUs and GPUs.\n\n\u201cThe host is the CPU available in the system\u201d and \u201csystem memory associated with the CPU is\n\ncalled host memory.\u201d \u201cThe GPU is called a device and GPU memory likewise called device\n\nmemory.\u201d As an example, the first main CUDA program execution step is \u201c[c]opy[ing] the input\n\ndata from host [CPU] memory to device [GPU] memory, also known as host-to-device transfer.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 111. The Accused Products perform a method that includes performing, by the graphics\n\nprocessing unit (GPU), a first computation in the sequence of computations on the first input data\n\nbased on the textures and shaders to generate first output data, computations in the sequence of\n\ncomputations representing respective layers of neurons in the artificial neural network, an output\n\nof the first computation in the sequence of computations representing an output of a first neuron\n\nin a first layer in the artificial neural network. For instance, the CUDA platform programming\n\nimplemented in the Accused Products utilizes the GPU and GPU memory. As an example, after\n\nthe \u201chost-to-device transfer\u201d (host (CPU) memory to device (GPU) memory), the second main\n\n\n 53\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 55 54\n of 95\n of 94\n\n\n\n\nstep is \u201c[l]oad the GPU program and execute, caching data on-chip for performance.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 112. In addition, as previously stated, cuDNN is a CUDA \u201cGPU-acceleration library of\n\nprimitives for deep neural networks.\u201d (See https://developer.nvidia.com/cudnn.) The exemplary\n\ncuDNN release notes below demonstrate computations implemented for RNNs and related data\n\nbeing transferred to GPU memory. As shown, users do \u201cnot need to transfer [an] array [from RNN\n\ndata descriptors] to device memory; the operation will be performed automatically by RNN APIs.\u201d\n\n\n\n\n *****\n\n\n\n\n 54\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 56 55\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-893/release-\n\nnotes/index.html#abstract (emphasis added).)\n\n 113. Relatedly, cuDNN operations below exemplify tensors being used as inputs and\n\noutputs (e.g., Tmp0). Exemplary \u201ccuDNN operations take tensors as input and produce tensors as\n\noutput.\u201d As part of CUDA, these cuDNN operations implement computer tasks performed by a\n\nGPU.\n\n\n\n\n 55\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 57 56\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n 114. The Accused Products perform a method that includes storing, in the memory\n\noperably coupled to the GPU, the first input data and the first output data. For instance, as\n\nillustrated below, a diagram describing the architecture of the Grace Hopper Superchip depicts a\n\nGPU (\u201cHopper GPU\u201d) in communication with GPU memory (\u201cGPUHBM3 or HBm3e\u201d high\n\nbandwidth memory).\n\n\n\n\n 56\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 58 57\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 115. As previously stated, the \u201cCUDA programming model\u201d implements programming\n\nfunctions and instructions for CPUs (host) and GPUs (device). For example, after the \u201chost-to-\n\ndevice transfer\u201d (CPU to GPU) first main step and \u201c[l]oad[ing] the GPU program and execut[ing]\u201d\n\n\n 57\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 59 58\n of 95\n of 94\n\n\n\n\nand \u201ccaching data on-chip for performance\u201d for the second main step, the \u201cresults\u201d are stored on\n\nGPU \u201cdevice memory.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded.)\n\n 116. The Accused Products perform a method that includes transferring second input\n\ndata acquired from the external system in real time into the memory operably coupled to the GPU\n\nafter the GPU starts the first computation and before the GPU starts a second computation of the\n\nsequence of computations, an output of the second computation in the sequence of computations\n\nrepresenting an output of a second neuron in a second layer in the artificial neural network. For\n\ninstance, as illustrated below, a diagram describing the architecture of the Grace Hopper Superchip\n\ndepicts a CPU (\u201cGrace CPU\u201d) in communication with a GPU (\u201cHopper GPU\u201d) via \u201cNVLink-C2C\u201d\n\n(chip-to-chip). \u201cHigh-Speed IO\u201d input and output data is received by the CPU via \u201cPCIe-5.\u201d\n\n\n\n\n 58\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 60 59\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 117. Furthermore, the Accused Products, including the Grace Hopper Superchip,\n\nimplement libraries and SDKs designed for neural networks that \u201care created from large numbers\n\nof identical neurons [that] are highly parallel by nature.\u201d The Accused Products implement\n\n\n 59\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 61 60\n of 95\n of 94\n\n\n\n\ncuDNN, a library \u201cmakes it easy to obtain state-of-the-art performance with Deep Neural\n\nNetworks,\u201d and TensorRT, a platform accelerator and runtime for optimizing, validating, and\n\ndeploying neural networks for inference (e.g., applying knowledge from a trained neural network\n\nmodel and inferring a result).\n\n\n\n\n(See https://developer.nvidia.com/discover/artificial-neural-network (emphasis added).)\n\n 118. An example below illustrates an exemplary neural network the Accused Products\n\nare designed to accelerate using parallel computations. \u201cInput\u201d (four) and \u201cOutput\u201d (eight) neurons\n\nare depicted below in a full-connected or linear layer structure in which all of the input neurons\n\ndepicted in a first layer are connected to all of the output neurons depicted in a second layer.\n\nComputations for the neural network are performed using, for example, \u201cNVIDIA Matrix\n\nMultiplication.\u201d Examples of inputs and outputs for forward propagation, activation gradient\n\ncomputation, and weight gradient computation (as matrix by matrix multiplications) are shown\n\nbelow.\n\n\n\n\n 60\n\fCase\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 62 61\n of 95\n of 94\n\n\n\n\n 61\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 63 62\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/performance/dl-performance-fully-\n\nconnected/index.html#performance (annotations added).)\n\n\n 62\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 64 63\n of 95\n of 94\n\n\n\n\n 119. Indeed, the cuDNN GPU-acceleration library of the Accused Products implement\n\noperations that \u201ctake tensors as input and produce tensors as output.\u201d\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n 120. Each claim in the \u2019438 Patent recites an independent invention. Neither claim 21,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019438 Patent.\n\n 121. Defendant has been aware of the \u2019438 Patent since at least the filing of the original\n\nComplaint. Defendant has been aware of the technology patented by the \u2019438 Patent since at least\n\n2007, when the inventors of the Asserted Patents first discussed their patented technologies with\n\nMr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited an ancestor of the \u2019438 Patent in its own patent portfolio since at least June 28, 2010 (See\n\nhttps://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy;\n\nhttps://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf;\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner).)\n\n 122. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019438 Patent family. These discussions included at least Mr. Alvin Lin, an Nvidia\n\nSenior Director of Business Development, and Mr. Jeff Herbst, then an Nvidia Vice President of\n\nBusiness Development and head of Nvidia\u2019s Inception GPU Ventures, in or around September 6,\n\n\n 63\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 65 64\n of 95\n of 94\n\n\n\n\n2016. In or around October 2016, Nvidia, through its representatives, initiated discussions with\n\nthe inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 123. The inventors also discussed their patented technology, including the underlying\n\ntechnology and family to the \u2019438 Patent (including U.S. Patent No. 9,189,828, the patent the \u2019438\n\nPatent reissued from), with Defendant\u2019s representatives at Nvidia\u2019s artificial intelligence\n\nconference in or around June 2017. On or about June 26, 2017, Defendant received materials from\n\nthe inventors, in lieu of a meeting on or about June 29, that identified patents related to the \u2019438\n\nPatent and described the technology in detail. Defendant had previously stated it was interested in\n\nthe inventors\u2019 solutions. Defendant also featured the inventors on its website as members of\n\nDefendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n\n 64\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 66 65\n of 95\n of 94\n\n\n\n\n 124. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\u2019438 Patent.\n\n 125. Defendant directly infringes at least claim 21 of the \u2019438 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 126. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 21 of the \u2019438 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 127. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 21 of the \u2019438 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 21 of the \u2019438 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 21 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\n\n\n\n 65\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 67 66\n of 95\n of 94\n\n\n\n\ninstallation, support, and distribution of the Accused Products.\n\n 128. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 129. Defendant further encourages and induces its customers to infringe claim 21 of the\n\n\u2019438 Patent: 1) by making its accelerated computing and data center services available on its\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 130. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 131. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\n\n\n\n 66\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 68 67\n of 95\n of 94\n\n\n\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019438 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 132. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019438 Patent.\n\n 133. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019438 Patent.\n\n\n\n\n 67\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 69 68\n of 95\n of 94\n\n\n\n\n 134. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 21 of the \u2019438 Patent.\n\n 135. Plaintiff has suffered and continues to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019438 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 136. Defendant\u2019s infringement of the \u2019438 Patent is knowing and willful. Defendant\n\nacquired actual knowledge of the patent that the \u2019438 Patent reissued from, and its family, since at\n\nleast 2017 and has acquired additional knowledge of the \u2019438 Patent since at least the filing of this\n\nlawsuit.\n\n 137. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents and\n\nPlaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and services\n\nthat it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019438 Patent with\n\nknowledge of the \u2019438 Patent constitutes willful infringement.\n\n THIRD CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019461 PATENT)\n\n 138. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n\n\n 68\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 70 69\n of 95\n of 94\n\n\n\n\n 139. Defendant has infringed and continues to infringe one or more claims of the \u2019461\n\nPatent in violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will\n\ncontinue to do so. The Accused Products, including features of, e.g., the Grace Hopper Superchip\n\n(GH200), at least when used for their ordinary and customary purposes, practice each element of\n\nat least claim 21 of the \u2019461 Patent as demonstrated below.\n\n 140. For example, claim 21 of the \u2019461 Patent recites:\n\n 21. A method of executing computations representing an artificial\n neural network on a computer system comprising at least one central\n processing unit (CPU), a processing unit, a first memory partition,\n and a second memory partition, the method comprising:\n\n executing, by the at least one CPU, a user interaction stream, the\n user interaction stream controlling transfer of inputs to the artificial\n neural network to the first memory partition and the second memory\n partition;\n\n executing, by the processing unit, a computational stream, the\n computational stream controlling data exchange between the user\n interaction stream and the computational stream during execution of\n the computations representing the artificial neural network;\n\n shifting control of a data exchange between the user interaction\n stream and the computational stream to the computational stream in\n response to starting execution of the computations representing the\n artificial neural network;\n\n shifting control of the data exchange between the user interaction\n stream and the computational stream to the user interaction stream\n in response to completion or interruption of the computations\n representing the artificial neural network;\n\n queueing a user command received by the user interaction stream\n during execution of the computations representing the artificial\n neural network; and\n\n executing the user command during execution of the computations\n representing the artificial neural network at times determined by the\n computational stream.\n\n 141. The Accused Products perform each step of the method of claim 21 of the \u2019461\n\n\n\n 69\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 71 70\n of 95\n of 94\n\n\n\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod of executing computations representing an artificial neural network on a computer system\n\ncomprising at least one central processing unit (CPU), a processing unit, a first memory partition,\n\nand a second memory partition, as further explained below. For instance, the Grace Hopper\n\nSuperchip (GH200) \u201cbrings together the groundbreaking performance of the NVIDIA Hopper\n\nGPU with the versatility of the NVIDIA Grace\u2122 CPU . . . in a single Superchip.\u201d It includes the\n\ncuDNN (CUDA Deep Neural Network) library for \u201c[d]eep neural networks.\u201d\n\n\n\n\n *****\n\n\n\n\n 70\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 72 71\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 142. As illustrated below, a diagram describing the architecture of the Grace Hopper\n\nSuperchip depicts a CPU (\u201cGrace CPU\u201d) with \u201c[u]p to 72 cores\u201d and CPU memory (\u201cCPU\n\nLPDDR5X\u201d) and a GPU (\u201cHopper GPU\u201d) and GPU memory (\u201cGPUHBM3 or HBm3e\u201d).\n\n\n\n\n 71\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 73 72\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 143. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 144. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA further includes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a\n\ncompiler, development tools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural\n\n\n\n 72\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 74 73\n of 95\n of 94\n\n\n\n\nNetwork library (cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d\n\nIt \u201cprovides highly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 145. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 73\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 75 74\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 146. The Accused Products perform a method that includes executing, by the at least\n\none CPU, a user interaction stream, the user interaction stream controlling transfer of inputs to\n\nthe artificial neural network to the first memory partition and the second memory partition. For\n\ninstance, as shown in the Grace Hopper Superchip architecture diagram below, the Grace Hopper\n\nSuperchip is illustrated below with a CPU (\u201cGRACE CPU\u201d). The CPU \u201cshare[s] a single per-\n\nprocess page table\u201d with a GPU (\u201cHopper GPU\u201d), \u201cenabling all CPU and GPU threads to access\n\nall system-allocated memory.\u201d The CPU is depicted as coupled to the GPU via \u201cNVLINK C2C\n\n[chip-to-chip],\u201d and can access the \u201cSystem Page Table\u201d and \u201cCPU PHYSICAL MEMORY\u201d via\n\n\u201cCPU-resident access\u201d and \u201cGPU PHYSICAL MEMORY\u201d via \u201c[r]emote access\u201d and \u201cPTE [page\n\ntable entry] B.\u201d The GPU can also access the System Page Table, and it can access \u201cGPU\n\nPHYSICAL MEMORY\u201d via \u201cGPU-resident access\u201d and \u201cCPU PHYSICAL MEMORY\u201d via\n\n\u201c[r]emote access\u201d and \u201cPTE A.\u201d Moreover, the \u201cSystem Page Table\u201d \u201c[t]ranslates CPU malloc()\n\n\n 74\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 76 75\n of 95\n of 94\n\n\n\n\n[memory allocation] to CPU or GPU.\u201d \u201cThe CPU heap, CPU thread stack, global variables\n\nmemory-mapped files, and inter-process memory are accessible to all CPU and GPU threads.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 147. The \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs and GPUs. \u201cThe host is the CPU available in the system\u201d and \u201csystem\n\nmemory associated with the CPU is called host memory.\u201d \u201cThe GPU is called a device and GPU\n\nmemory likewise called device memory.\u201d As an example, the first main CUDA program execution\n\nstep is \u201c[c]opy[ing] the input data from host [CPU] memory to device [GPU] memory, also known\n\nas host-to-device transfer.\u201d\n\n\n\n\n 75\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 77 76\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 148. The Accused Products perform a method that includes executing, by the processing\n\nunit, a computational stream, the computational stream controlling data exchange between the\n\nuser interaction stream and the computational stream during execution of the computations\n\nrepresenting the artificial neural network. For instance, the \u201cCUDA programming model\u201d\n\nimplements programming functions and instructions for CPUs and GPUs. As previously stated,\n\nthe host is the CPU and the device is the GPU. After \u201c[c]opy[ing] the input data from host [CPU]\n\nmemory to device [GPU] memory,\u201d the second main CUDA program execution step is\n\n\u201c[l]oad[ing] the GPU program and execut[ing].\u201d\n\n\n\n\n 76\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 78 77\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 149. Indeed, the Grace Hopper Superchip \u201cis designed to accelerate applications\u201d using\n\n\u201cExtended GPU Memory.\u201d As depicted in the gram of the Grace Hopper architecture below, a\n\nGPU (\u201cHOPPER GPU\u201d) can access \u201cLocal CPU,\u201d \u201cPeer CPU,\u201d and \u201cPeer GPU\u201d memory via\n\n\u201cNVLink.\u201d\n\n\n\n\n 77\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 79 78\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 150. For instance, exemplary CUDA library cuDNN function\n\n\u201ccudnnSetRNNDescriptor_v8\u201d \u201cinitializes a previously created RNN [recurrent neural network]\n\ndescriptor object.\u201d This function \u201cstore[s] all information needed to compute the total number of\n\nadjustable weights/biases in the RNN model.\u201d In addition, the parameters \u201cdirMode,\u201d\n\n\u201cinputMode,\u201d and \u201cdatatype\u201d confirm the exchange of calculations and values between the hidden\n\nlayers of an RNN.\n\n\n\n\n *****\n\n\n\n\n 78\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 80 79\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-891/pdf/cuDNN-API.pdf\n\n(emphasis added).)\n\n 151. The Accused Products perform a method that includes shifting control of a data\n\nexchange between the user interaction stream and the computational stream to the computational\n\nstream in response to starting execution of the computations representing the artificial neural\n\nnetwork. For instance, the \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs (host) and GPUs (device). As an example, the \u201chost-to-device transfer\u201d\n\n(CPU to GPU) first main step, the second main step is \u201c[l]oad the GPU program and execute\u201d and\n\nthe third main step is \u201c[c]opy the results from device [GPU] memory to host [CPU] memory, also\n\nknown as device-to-host transfer\u201d (GPU to CPU).\n\n\n\n\n 79\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 81 80\n of 95\n of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded.)\n\n 152. As previously stated, the Grace Hopper Superchip \u201cis designed to accelerate\n\napplications\u201d using \u201cExtended GPU Memory\u201d and the GPU can access local/peer CPU and peer\n\nGPU memory via \u201cNVLink.\u201d The Grace Hopper Superchip\u2019s Extended GPU Memory feature\n\n\u201cenables GPUs to access all the system memory efficiently\u201d and \u201cphysical memory in the system\n\ncan be allocated to be accessible from any GPU thread.\u201d\n\n\n\n\n 80\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 82 81\n of 95\n of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 153. The Accused Products perform a method that includes shifting control of the data\n\nexchange between the user interaction stream and the computational stream to the user interaction\n\nstream in response to completion or interruption of the computations representing the artificial\n\nneural network. For instance, after the \u201cCUDA programming model\u201d \u201chost-to-device transfer\u201d\n\n(CPU to GPU) and GPU program load and execution steps, the third main step is \u201c[c]opy the\n\nresults from device [GPU] memory to host [CPU] memory, also known as device-to-host transfer\u201d\n\n(GPU to CPU). The \u201chost-to-device transfer\u201d (CPU to GPU) first main step can be reintroduced\n\nfor additional computations.\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\n\n 81\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 83 82\n of 95\n of 94\n\n\n\n\nadded.)\n\n 154. In addition, as shown in the Grace Hopper Superchip architecture diagram below,\n\nthe CPU (\u201cGRACE CPU\u201d) \u201cshare[s] a single per-process page table\u201d with a GPU (\u201cHopper\n\nGPU\u201d), \u201cenabling all CPU and GPU threads to access all system-allocated memory.\u201d \u201cThe CPU\n\nheap, CPU thread stack, global variables memory-mapped files, and inter-process memory are\n\naccessible to all CPU and GPU threads.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 155. The Accused Products perform a method that includes queueing a user command\n\nreceived by the user interaction stream during execution of the computations representing the\n\nartificial neural network. For instance, as shown by publicly available CUDA toolkit\n\n\n\n 82\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 84 83\n of 95\n of 94\n\n\n\n\ndocumentation, CUDA implements exemplary \u201cmemory management functions\u201d that \u201c[c]op[y]\n\ndata between host [CPU] and device [GPU].\u201d This includes CUDA functions \u201ccudaMemcpy\u201d and\n\n\u201ccudaMemcpyAsync.\u201d\n\n\n\n\n *****\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 156. As an example, exemplary CUDA memory management function\n\n\u201ccudaMemcpyAsync\u201d \u201c[c]opies count bytes [data] from the memory area pointed to by src [source\n\nmemory address pointer] to the memory area pointed to by dst [destination memory address\n\npointer], where kind [type of transfer] specifies the direction of the copy.\u201d Destinations includes\n\n\u201ccudaMemcpyHostToDevice [CPU to device GPU], cudaMemcpyDeviceToHost [GPU to CPU],\n\ncudaMemcpyDeviceToDevice [GPU to GPU]. Because the function \u201ccudaMemcpyAsync() is\n\nasynchronous with respect to the host, [] the call may return before the copy is complete. The copy\n\ncan optionally be associated to a stream [identified stream] by passing a non-zero stream\n\nargument.\u201d\n\n\n\n\n 83\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 85 84\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 157. The Accused Products perform a method that includes executing the user command\n\nduring execution of the computations representing the artificial neural network at times\n\ndetermined by the computational stream. For instance, as shown by exemplary and publicly\n\navailable CUDA toolkit documentation, CUDA implements \u201cmemory management functions\u201d that\n\n\u201c[c]op[y] data between host [CPU] and device [GPU].\u201d\n\n\n\n\n 84\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 86 85\n of 95\n of 94\n\n\n\n\n *****\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 158. As an example, exemplary CUDA memory management function\n\n\u201ccudaMemcpyAsync\u201d \u201c[c]opies count bytes [data] from the memory area pointed to by src [source\n\nmemory address pointer] to the memory area pointed to by dst [destination memory address\n\npointer], where kind [type of transfer] specifies the direction of the copy.\u201d Because the function\n\n\u201ccudaMemcpyAsync() is asynchronous with respect to the host, [] the call may return before the\n\ncopy is complete. The copy can optionally be associated to a stream [identified stream].\u201d\n\n\n\n\n 85\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 87 86\n of 95\n of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 159. In another example, CUDA implements \u201cCUDA-specific memory APIs [that]\n\nprovide users with guarantees about where the memory resides, which threads can access it,\n\nwhether it is migratable, and many other features that enable users to extract all the performance\n\nthe hardware has to offer.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n\n 86\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 88 87\n of 95\n of 94\n\n\n\n\n 160. Each claim in the \u2019461 Patent recites an independent invention. Neither claim 21,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019461 Patent.\n\n 161. Defendant has been aware of the \u2019461 Patent since at least the filing of the original\n\nComplaint. Defendant has been aware of the technology patented by the \u2019461 Patent since at least\n\n2007, when the inventors of the Asserted Patents first discussed their patented technologies with\n\nMr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited an ancestor of the \u2019461 Patent in its own patent portfolio since at least June 28, 2010 (See\n\nhttps://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy;\n\nhttps://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf;\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner).)\n\n 162. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019461 Patent family. These discussions included at least Mr. Alvin Lin, an Nvidia\n\nSenior Director of Business Development, and Mr. Jeff Herbst, then an Nvidia Vice President of\n\nBusiness Development and head of Nvidia\u2019s Inception GPU Ventures, in or around September 6,\n\n2016. In or around October 2016, Nvidia, through its representatives, initiated discussions with\n\nthe inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 163. The inventors also discussed their patented technology, including the underlying\n\ntechnology and family to the \u2019461 Patent, with Defendant\u2019s representatives at Nvidia\u2019s artificial\n\nintelligence conference in or around June 2017. On or about June 26, 2017, Defendant received\n\n\n\n\n 87\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 89 88\n of 95\n of 94\n\n\n\n\nmaterials from the inventors, in lieu of a meeting on or about June 29, that identified patents related\n\nto the \u2019461 Patent and described the technology in detail. Defendant had previously stated it was\n\ninterested in the inventors\u2019 solutions. Defendant also featured the inventors on its website as\n\nmembers of Defendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n 164. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\u2019461 Patent.\n\n 165. Defendant directly infringes at least claim 21 of the \u2019461 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\n\n 88\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 90 89\n of 95\n of 94\n\n\n\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 166. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 21 of the \u2019461 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 167. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 21 of the \u2019461 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 21 of the \u2019461 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 21 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\ninstallation, support, and distribution of the Accused Products.\n\n 168. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 169. Defendant further encourages and induces its customers to infringe claim 21 of the\n\n\u2019461 Patent: 1) by making its accelerated computing and data center services available on its\n\n\n\n\n 89\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 91 90\n of 95\n of 94\n\n\n\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 170. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 171. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\n\n\n\n 90\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 92 91\n of 95\n of 94\n\n\n\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019461 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 172. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019461 Patent.\n\n 173. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019461 Patent.\n\n 174. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\n\n\n\n 91\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 93 92\n of 95\n of 94\n\n\n\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 21 of the \u2019461 Patent.\n\n 175. Plaintiff has suffered and continues to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019461 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 176. Defendant\u2019s infringement of the \u2019461 Patent is knowing and willful. Defendant\n\nacquired actual knowledge of the family of the \u2019461 Patent since at least 2017 and has acquired\n\nadditional knowledge of the \u2019461 Patent since at least the filing of this lawsuit.\n\n 177. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents and\n\nPlaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and services\n\nthat it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019461 Patent with\n\nknowledge of the \u2019461 Patent constitutes willful infringement.\n\n\n\n\n 92\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 94 93\n of 95\n of 94\n\n\n\n\n PRAYER FOR RELIEF\n\nWHEREFORE, Plaintiff respectfully requests the following relief:\n\n a) That this Court adjudge and decree that Defendant has been, and is currently,\n\n infringing each of the Asserted Patents;\n\n b) That this Court award Plaintiff damages to compensate for Defendant\u2019s past and\n\n future infringement of the Asserted Patents, through the life of the Asserted Patents;\n\n c) That this Court award Plaintiff pre- and post-judgment interest on such;\n\n d) That this Court order an accounting of damages incurred by Plaintiff from six years\n\n prior to the date this lawsuit was filed through entry of a final, non-appealable\n\n judgment;\n\n e) That this Court determine that this patent infringement case is exceptional and\n\n award Plaintiff its costs and attorneys\u2019 fees incurred in this action;\n\n f) That this Court award increased damages under 35 U.S.C. \u00a7 284; and\n\n g) That this Court award such other relief as the Court deems just and proper.\n\n DEMAND FOR JURY TRIAL\n\n Plaintiff respectfully requests a trial by jury on all issues triable thereby.\n\n\n\n\n 93\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-2 30 Filed\n Filed\n 08/18/26\n 12/12/24 Page\n Page\n 95 94\n of 95\n of 94\n\n\n\nDATED: December 12, 2024\n By:/s/ Mark D. Siegmund\n Mark D. Siegmund\n Texas Bar No. 24117055\n CHERRY JOHNSON SIEGMUND JAMES\n PLLC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n Telephone: (254) 732-2242\n Facsimile: (866) 627-3509\n msiegmund@cjsjlaw.com\n\n Christopher C. Campbell\n KING & SPALDING LLP\n 1700 Pennsylvania Avenue, NW\n Suite 900\n Washington, DC 20006\n Telephone: (202) 626-5578\n Facsimile: (202) 626-3737\n ccampbell@kslaw.com\n\n Britton F. Davis\n Brian Eutermoser (pro hac vice to be filed)\n KING & SPALDING LLP\n 1401 Lawrence Street\n Suite 1900\n Denver, CO 80202\n Telephone: (720) 535-2300\n Facsimile: (720) 535-2400\n bfdavis@kslaw.com\n beutermoser@kslaw.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that a true and correct copy of the foregoing document\n\nwas served on all counsel of record via the Court\u2019s electronic filing system on this 12th day of\n\nDecember 2024.\n\n By:/s/ Mark D. Siegmund\n Mark D. Siegmund\n\n\n\n 94\n\f","ocr_status":2,"date_upload":"2026-08-20T15:37:41.680357-07:00","document_number":"6","attachment_number":2,"pacer_doc_id":"181037220269","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 1","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554422/","id":490554422,"tags":[],"absolute_url":"/docket/74659430/6/3/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.177861-07:00","date_modified":"2026-08-23T09:39:45.058891-07:00","sha1":"f980a76a81e0a19c763fa55b5a22e7da1c5541ae","page_count":15,"file_size":1311511,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.3.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.3.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 1 of 15\n\n\n\n\n EXHIBIT\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 2 of 15\n\n USOO8648867B2\n\n\n(12) United States Patent (10) Patent No.: US 8,648,867 B2\n Gorchetchnikov et al. (45) Date of Patent: Feb. 11, 2014\n(54) GRAPHIC PROCESSORBASED (56) References Cited\n ACCELERATOR SYSTEMAND METHOD\n U.S. PATENT DOCUMENTS\n(75) Inventors: Anatoli Gorchetchnikov, Belmont, MA 5,388,206 A * 2/1995 Poulton et al. ................ 345,505\n (US); Heather Marie Ames, South 2005, 0166042 A1* 7, 2005 Evans ............ T13,150\n Boston, MA (US); Massimiliano 2007/0052713 A1* 3/2007 Chung et al. ... 34.5/5O1\n Versace, South Boston, MA (US); 2007/0279429 A1* 12/2007 Ganzer ......................... 345,582\n Fabrizio Santini, Jamaica Plain, MA\n (US) * cited by examiner\n(73) Assignee: Neurala LLC, Boston, MA (US) Primary Examiner \u2014 Maurice L. McDowell, Jr.\n(*) Notice: Subject to any disclaimer, the term of this (57) ABSTRACT\n patent is extended or adjusted under 35 An accelerator system is implemented on an expansion card\n U.S.C. 154(b) by 1030 days. comprising a printed circuit board having (a) one or more\n graphics processing units (GPU), (b) two or more associated\n(21) Appl. No.: 11/860,254 memory banks (logically or physically partitioned), (c) a\n(22) Filed: Sep. 24, 2007 specialized controller, and (d) a local bus providing signal\n coupling compatible with the PCI industry standards (this\n(65) Prior Publication Data includes but is not limited to PCI-Express, PCI-X, USB 2.0,\n or functionally similar technologies). The controller handles\n US 2008/O 117220 A1 May 22, 2008 most of the primitive operations needed to set up and control\n Related U.S. Application Data GPU computation. As a result, the computer's central pro\n cessing unit (CPU) is freed from this function and is dedicated\n(60) Provisional application No. 60/826,892, filed on Sep. to other tasks. In this case a few controls (simulation start and\n 25, 2006. stop signals from the CPU and the simulation completion\n signal back to CPU), GPU programs and input/output data are\n(51) Int. C. the information exchanged between CPU and the expansion\n G06F 5/00 (2006.01) card. Moreover, since on every time step of the simulation the\n(52) U.S. C. results from the previous time step are used but not changed,\n USPC ........................................... 345/501; 34.5/503 the results are preferably transferred back to CPU in parallel\n(58) Field of Classification Search with the computation.\n USPC .................................................. 345/501,503\n See application file for complete search history. 19 Claims, 5 Drawing Sheets\n\n 140 200\n 8O 250\n -4\n\n\n SHADERE.K.\n MEMORY\n\n 50\n\n O EXRE\n - icon TROLLER w NEMORY\n\n C\n -\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 3 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 1 of 5 US 8,648,867 B2\n\n\n\n\n 7.\n\n\n\n\n $3.\n\n\n\n\n s\n\n\n\n\n FG,\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 4 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 2 of 5 US 8,648,867 B2\n\n\n\n\n g\n i.\n :\n i\n : r\n s\n al e3esiasti Od\n\n\n s\n\n 3\n\n\n s\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 5 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 3 of 5 US 8,648,867 B2\n\n\n\n\n 300\n CPU 120 Start\n 304\n Expansion Cardiso :\n iaia fict iss: Risitio: &::::::::8:::::::::\n Sires 301 Se::303\n 320\n Disk:O Graphic User Controller\n Interface initialization\n initialization Initialization\n\n\n\n\n 325\n User input Parser External input\n Texture textures from RAM to\n interaction Generator texture memory bank\n 326 :\n Population Parser Population shader\n Shader Generato binaries from RAM to\n and Compiler shader memory bank\n\n\n\n\n Simulation\n Initialization\n\n\n 330\n\n\n Progress Input Parser\n monitor Texture\n Generator Computation\n (see Fig. 4)\n Data output\n to Disk\n Output Data Data\n ACCumulation issils\n in RAM\n\n Last\n iteration\n 318 :\n\n Yes\n 350\n\n\n\n\n 3\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 6 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 4 of 5 US 8,648,867 B2\n\n\n Expansion Card 180\n\n Simulation\n start\n\n iDesi: {34}{ut : {iciplisiii) 2. iaia iiii:\n Si:Sires: Siii:Sire:: Silisirai?\n 402 403 404\n Input textures from\n texture memory\n bank to GPU\n\n\n\n Shaders from\n shader memory\n bank to GPU\n\n\n ... ...........\n\n\n\n\n Shader\n execution\n\n\n\n\n Output texture\n upload to texture\n memory bank\n\n New external input Data\n textures from RAM told -----------\n\n\n\n\n Wait for swap Swap input Wait for swap\n of input/output and output of input/output\n texture pointers texture pointers texture pointers\n\n 475\n Input textures from NO Last\n texture memory iteration?\n bank to RAM\n Yes\n\n\n\n iteration?\n Yes 490\n Wait for all three\n streams of\n execution to finish\n\n 499\n\n\n\n EIG 4\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 7 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 5 of 5 US 8,648,867 B2\n\n\n\n\n (p)\n\n\n\n\n *\u00b7,,?)&xaddO2OInputDependency(m gate1, O., O., 0.004, O., 0, 0);\nInteraction Stream, Data Output Stream, and Computational u->addFullDependency(m gate2, population());\nStream. Through TSimulator:timestep(), TSimulator:out 40\nfileInterval ( ), and TSimulator:outmode(), the application public: TCablePopRCF(): TPopulation(\u201ccompCPU RCF, w, h, true)\ncan set the time step of the simulation, the time step of disk { };\noutput, and the mode of the disk output. The external input ~TCablePopRCF() {if(m equation) delete m equation;\n if(m gate1) delete m gate1;\npattern should be packed into a TPattern object and bound to\nthe simulation object through TSimulator:resetInputs( ) 45 bool if(m gate2) delete m gate2:};\n fillElements(TSimulator sim);\nmethod. TSimulator::simLength( ) sets the length of the }:\nsimulation. bool TCablePopRCF::fillElements(TSimulatior sim)\n The second step is to create at least one population of { equation = new TEq RCF (this, m compet, m persist);\nequations (TPopulation object). Population holds one equa mcreateCatingStructure();\ntion object TEquation. This object contains only a formula 50 for(size ti = 0; i < x.Size(); ++i)\nand does not hold element-specific data, so all elements of the for(size tj = 0; j < ySize(); ++)\npopulation can share single TEquation. {\n The TEquation object is converted to a GPU program TElement u = new TCPUElement(this, m equation, i,j):\n sim->addUnit(u);\nbefore execution. GPU programs have to be executed within createUnitStructure(u);\na graphical context, which is stream specific. TSimulator 55 return true:\ncreates this context within a Computational Stream, therefore\nall programs and data arrays that are necessary for computa int\ntion have to be initialized within Computational Stream. Con main()\nstructor of TPopulation is called from User Interaction //{ Input pattern generation (309 in FIG. 3)\nStream, so no GPU-related objects can be initialized in this 60 uint32 t pat = new uint32 twh;\nCOnStructOr. TRandom-float randGen (O);\n TPopulation::fillElements( ) is a virtual method designed for(uint32 ti = 0; i < wh; ++i)\nto overcome this difficulty. It is called from within the Com pati = randGen.random ();\nputational Stream after TSimulator::networkCreate( ) is TPattern p = new TPattern (pat, w, h);\n if Setting up the simulation\ncalled in the User Interaction Stream. A user has to override 65 TSimulator cableSim = new TSimulator(\u201cdata'); fi(308 and 320 in\nTPopulation::fillElements( ) to create TEquation and other FIG. 3)\ncomputation related objects both element independent and\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 13 of 15\n\n\n US 8,648,867 B2\n 11 12\n -continued need to perform the computations and initiates the upload 435\n of them onto the GPU 240. The GPU 240 can communicate\ncableSim->timestep(0.05); fi(320 in FIG. 3)\ncableSim->resetInputs(p); (325 in FIG. 3) directly with the texture memory bank 250 to upload the\ncableSim->OutfileInterval(0.1); (308 in FIG. 3) appropriate texture to perform the computations. The control\ncableSim->Outmode(SANNDRA::timefunc); (308 in FIG. 3)\ncableSim->simLength(60.0); (320 in FIG. 3)\n ler 220 also pulls the first shader (known by the stored order)\nif Preparing the population from the shader memory bank 210 and uploads 450 it onto the\nTPopulation* cablePop = new TCablePopRCF(); //(310 in FIG. 3) GPU 240.\ncableSim->networkCreate(); //(326 in FIG. 3) The GPU 240 executes the following operations in this\nuint16 t user = 1; order: performs the computation (execution of the shader)\nwhile(user) 10\n 470; tells the controller 220 that it is done with the computa\nif(cableSim->simulationStart(true, 1)) (311 in FIG. 3) tions for the current shader; and after all shaders for this\nexit(1): particular equation are executed sends 480 the output textures\nstd::cout-\u201cRepeat?\\n: //(305 in FIG. 3)\nstd::cin>user; //(305 in FIG. 3) to the output portion of the texture memory bank 250. This\nif(user == 1) 15 cycle continues through all of the equations based on the\ncableSim->networkReset(); //(305 in FIG. 3) branching step 482.\nif cableSim) An example shader that performs fourth order Runge\ndelete cableSim; Also deletes cablePop and its internals Kutta numerical integration is shown in Listing 2 using GLSL\nexit(0); notation;\nListing 1.\n\n FIG. 4 is a detailed flow diagram illustrating a part of an uniform sampler2DRect Variable;\n uniform float integration step;\nexemplary implementation of the bottom level system and float halfstep = integration step*0.5;\nmethod performed during the computation on the GPU accel 25 float fl. 6 step = integration stepf 6.0:\nerator of the expansion card 180 and is a more detailed view vec4 output = texture2DRect(Variable, gl TexCoord O.st);\n if define equation() here\nof the computational box 330 in FIG. 3. FIG. 4 is a represen vec4 rungekutta4(vec4 x)\ntation of one of several ways in which a system and method\nfor processing numerical techniques can be implemented. const vecA k1 = equation(x);\n const vec4 k2 = equation(x + halfstep*k1);\n With systems of equations that have complex interdepen 30 const vec4 k3 = equation(x + halfstep*k2);\ndencies it is likely that the variable in Some equation from a const vecA k4 = equation(x + integration Step*k3);\nprevious time step has to be used by some other equation after return fl. 6step*(k1 + 2.0* (k2 + k3) + k4);\nthe new values of this variable are already computed for new\ntime step. To avoid data confusion, the new values of variables void main (void)\n {\nshould be rendered in a separate texture. After the time step is 35 output += rungekutta4(output);\ncompleted for all equations, these new values should be cop gl FragColor = output;\nied over old values so that they are used as input during the Listing 2.\nnext time step. Copying textures is an expensive operation,\ncomputationally, but since the textures are referred to by\ntexture IDs (pointers), Swapping these pointers for input and 40 The shader in Listing 2 can be executed on conventional\noutput textures after each time step achieves the same resultat video card. Using the controller 220 this code can be further\na much lesser cost. optimized, however. Since the integration step does not\n In the hardware solution suggested herein, ID Swapping is change during the simulation, the step itself as well as the\nequivalent to Swapping the base memory address for two halfstep and /6 of the step can be computed once per simula\npartitions of the texture memory bank 250. They are swapped 45 tion, and updated in all shaders by a shader update procedures\n485 during synchronization (485, 430, and 455) so that data\ntransfer 445 and the computation 435-487 proceeds immedi 310,326 discussed above.\nately and in parallel with data transfer as shown in FIG. 4. A computed theofmain\n After all the equations in the computational cycle are\nhardware solution allows this parallelism through access of 220 can switch 485execution the\n substream 403 on the controller\n reference pointers of the input and\nthe controller 220 to the onboard texture memory bank 250. 50 output portions of the texture memory bank 250.\n The main computation and data exchange are executed by The two other substreams of execution on the controller\nthe controller 220. It runs three parallel substreams of execu\ntion: Computational Substream 403, Data Output Substream 220 are waiting (blocks 430 and 455, respectively) for this\n402, and Data Input Substream 404. These streams are syn switch to begin their execution. The Data Input Substream\nchronized with each other during the swap of pointers 485 to 55 404 is controlling 440 the input of additional data from the\nthe input and output texture memory partitions of the texture CPU 120. This is necessary in cases where the simulation is\nmemory bank 250 and the check for the last iteration 487. monitoring the changing input, for example input from a\nAlgorithmically, these two operations are a single atomic video camera or other recording device in the real time. This\noperation, but the block diagram shows them as two separate substream uploads new external input from the CPU 120 to\nblocks for clarity. 60 the texture memory bank 250 so it can be used by the main\n The Computational Substream 403 performs a computa computational Substream 403 on the next computational step\ntional cycle including a sequential execution of all shaders and waits for the next iteration 475. The Data Output Sub\nthat were stored in the shader memory bank 210 using the stream 445 controls the output of simulation results to the\nappropriate input and output textures. To begin the simulation CPU 120 if requested by the user. This substream uploads the\nthe controller 220 initializes three execution substreams 403, 65 results of the previous step to the main RAM 130 so that the\n402, and 404. On every simulation step, the Computational CPU 120 can save them on disk 140 or show them on the\nSubstream 403 determines which textures the GPU 240 will results display 313 and waits for the next iteration 460.\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 14 of 15\n\n\n US 8,648,867 B2\n 13 14\n Since the Computational Substream 403 determines the While this invention has been particularly shown and\ntiming of input 440 and output 445 data transfers, these data described with references to preferred embodiments thereof,\ntransfers are driven by the controller 220. To further reduce it will be understood by those skilled in the art that various\nthe data transfer overhead (and disk 140 overhead also) the changes in form and details may be made therein without\ncontroller 220 initiates transfer only after selected computa departing from the scope of the invention encompassed by the\ntional steps. For example, if the experimental data that is appended claims.\nsimulated was recorded every 10 milliseconds (msec) and the\nsimulation for better precision was computed every 1 mSec, What is claimed is:\nthen only every tenth result has to be transferred to match the 1. A computer system for performing a numerical simula\nexperimental frequency. 10 tion over a plurality of computational cycles including at least\n This solution stores two copies of output data, one in the a first computational cycle and a second computational cycle,\nexpansion card texture memory bank 250 and another in the the computer system comprising:\nsystem RAM 130. The copy in the system RAM 130 is a central processing unit;\naccessed twice: for disk I/O and screen visualization 313. An a main memory, operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a 15 unit, to store input data to be accessed by the central\ndirect read access to the onboard texture memory bank 250 by processing unit in performing the numerical simulation;\nmapping the memory of the hardware onto a global memory a video system, operably coupled to the central processing\nspace. The alternative solution will double the communica unit, to drive a video monitor to display an indication of\ntion through the local bus 190. Since the goal discussed herein the numerical simulation in response to the computer\nis reducing the information transfer through the local bus 190, system performing the numerical simulation;\nthe former solution is favored. an accelerator, operably coupled to the central processing\n The main stream substream 403 determines if this is the last unit, to receive at least a portion of the input data from\niteration 487. If it is the last iteration, the controller 220 waits the central processing unit and to provide first output\nfor the all of the execution substreams to finish 490 and then data generated during the first computational cycle to the\nreturns the control to the CPU 120, otherwise it begins the 25 central processing unit after a conclusion of the first\nnext computational cycle. computational cycle, the accelerator comprising:\n This repeats through all of the computational cycles of the at least one graphics processing unit to generate second\nsimulation. output data, during the second computational cycle,\n Conclusion by performing at least one calculation on the first\n This GPU accelerator system offers the following potential 30 output data; and\nadvantages: an accelerator memory, operably coupled to the at least\n 1. Limited computations on the CPU 120. The CPU 120 is one graphic processing unit, the accelerator memory\nonly used for user input, sending information to the controller comprising:\n220, receiving output after each computational cycle (or less a first partition, referenced by a first pointer, to store\nfrequently as defined by the user), writing this output to disk 35 the first output data during the second computa\n140, and displaying this output on the monitor 170. This frees tional cycle; and\nthe CPU 120 to execute other applications and allows the a second partition, referenced by a second pointer, to\nexpansion card to run at its full capacity without being slowed store the second output data generated during the\ndown by extensive interactions with the CPU 120. second computational cycle; and\n 2. Minimizing data transfer between the expansion card 40 an accelerator controller, operably coupled to the accelera\n180 and the system bus 200. All of the information needed to tor memory and the central processing unit, to transfer\nperform the simulations will be stored on the expansion card the at least the portion of the input data into the accel\n180 and all simulations will take place on it. Furthermore, erator memory before the first computational cycle, to\nwhatever data transfer remains necessary will take place in transfer the first output data from the accelerator\nparallel with the computation, thus reducing the impact of this 45 memory to the main memory during the second compu\ntransfer on the performance. tational cycle, to direct the second output data into the\n 3. New way to execute GPU programs (shaders). Previ second partition during the second computational cycle,\nously, the CPU 120 had full control over the order of shaders and to Swap the first pointer and the second pointer at the\nexecution and was required to produce specific commands on conclusion of the second computational cycle Such that\nevery cycle to tell the GPU 240 which shader to use. With the 50 the second output data becomes an input for a third\ninvention disclosed herein, shaders will initially be stored on computational cycle of the plurality of computational\nthe shader memory bank 210 on the expansion card 180 and cycles.\nwill be sent to the GPU 240 for execution by the general 2. The computer system as claimed in claim 1, wherein the\npurpose controller 220 located on the expansion card. accelerator controller is configured to dictate an order of\n 4. Multiple parallelisms. The GPU 240 is inherently paral 55 execution of instructions to the at least one graphics process\nlel and is well suited to perform parallel computations. In ing unit.\nparallel with the GPU 240 performing the next calculation, 3. The computer system as claimed in claim 1, wherein the\nthe controller 220 is uploading the data from the previous accelerator controller comprises an interface controller to\ncalculation into main memory 130. Furthermore, the CPU communicate with the central processing unit over a bus of\n120 at the same time uses uploaded previous results to save 60 the computer system.\nthem onto disk 140 and to display them on the screen through 4. The computer system as claimed in claim 1, wherein the\nthe system bus 200. accelerator memory comprises a texture memory bank to\n 5. Reuse of existing and affordable technology. All hard store the at least the portion of the input data and the first\nware used in the invention and mentioned here-in are based on output data and a shader memory bank to store instructions\ncurrently available and reliable components. Further advance 65 for performing a set of operations to be performed on the at\nof these components will provide straightforward improve least the portion of the input data by the at least one graphic\nments of the invention. processing unit.\n\f Case 7:26-mc-00318-LS Document 6-3 Filed 08/18/26 Page 15 of 15\n\n\n US 8,648,867 B2\n 15 16\n 5. The computer system as claimed in claim 4, wherein the bank to store instructions for processing operations to be\ntexture memory is partitioned into the first partition, the sec performed on the input data by the at least one graphic pro\nond partition, a third partition to store internal variables, a cessing unit.\nfourth partition to store data textures used as input at a par 13. The accelerator system as claimed in claim 12, wherein\nticular computation cycle of the plurality of computational 5 the texture memory is partitioned into a first partition to store\ncycles. the input data, a second partition to store internal variables, a\n 6. The computer system as claimed in claim 1, wherein the third partition to store data textures used as input at a particu\naccelerator controller inputs the at least the portion of the lar computation cycle of the numerical simulation, and a\ninput data and a series of instructions into the at least one fourth partition to store the output data.\ngraphic processing unit, wherein the at least one graphics 10\nprocessing unit then executes the instructions on the at least the14. The accelerator system as claimed in claim 9, wherein\n accelerator controller is configured to perform successive\nthe portion of the input data. computational cycles of the numerical simulation by feeding\n 7. The computer system as claimed in claim 1, wherein the the output data generated\naccelerator controller comprises a set of instructions stored ing unit from a previous by the at least one graphic process\n computational cycle and the input\non a memory. 15\n 8. The computer system as claimed in claim 1, wherein the data for a next computational cycle into the at least one\nat least the portion of the input data represents an initial graphic processing unit.\ncondition of the numerical simulation. 15. The accelerator system as claimed in claim 9, wherein\n 9. An accelerator system for a computer system performing the accelerator controller comprises a set of instructions\na numerical simulation, the accelerator system comprising: 20 stored in a memory.\n 16. A method for performing a numerical simulation on\n at least one graphics processing unit to generate output data input data in a computer system including a central process\n by performing at least one computation during a first\n computational cycle of the numerical simulation; ing unit and an accelerator, the method comprising:\n an accelerator memory, operably coupled to the at least one receiving, by an accelerator, first input data from the central\n processing unit;\n graphics processing unit, to store data used to perform 25 transferring,\n the at least one computation; and by an accelerator controller, the first input\n an accelerator controller, operably coupled to the accelera data into a first partition, referenced by first pointer. ofan\n tor memory and the at least one graphics processing unit, accelerator memory before a first computational cycle of\n to execute: the numerical simulation;\n (i) a computational stream controlling performance of 30 performing, by at least one graphics processing unit during\n the first computational cycle, at least one calculation on\n the at least one computation by the at least one graph the first portion of the input data as to generate first\n ics processing unit; output data;\n (ii) an output stream controlling transfer of the output storing, by the accelerator controller, the first output data\n data from the at least one graphics processing unit to into a second partition, referenced by a second pointer,\n the accelerator memory during the first computational 35 of the accelerator memory; and\n cycle; and\n (iii) an input stream controlling transfer of input data to Swapping the first pointer with the second pointer at the end\n the accelerator memory for use by the at least one of the first computational cycle, such that the first output\n graphics processing unit during a second computa data becomes an input for a second computational cycle\n of the numerical simulation.\n tional cycle of the numerical simulation. 40\n 10. The accelerator system as claimed in claim 9, wherein 17. The method as claimed inclaim 16, further comprising:\nthe accelerator controller is configured to dictate an order of sending, by the accelerator controller, instructions for per\nexecution of instructions to the at least one graphics process forming the at least one calculation to the at least one\ning unit. graphics processing unit.\n 11. The accelerator system as claimed in claim 9, wherein 45 18. The method as claimed in claim 16, further comprising:\nthe accelerator controller is configured to send instructions to partitioning the accelerator memory into the first partition,\nthe at least one graphics processing unit, wherein the at least the second partition, a third partition to store internal\none graphics processing unit is configured to execute the Variables, and a fourth partition to store data used as\ninstructions on the input data, and wherein the accelerator input at a particular computation cycle of the numerical\n simulation.\ncontroller is configured to transfer the output data from the 50 19. The method of claim 16, further comprising:\naccelerator memory to a main memory during the execution transferring, by the accelerator controller, the first output\nof the instructions by the at least one graphics processing unit. data to the main memory during the second computa\n 12. The accelerator system as claimed in claim 9, wherein tional cycle.\nthe accelerator memory comprises a texture memory bank to\nStore the input data and the output data and a shader memory\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:43.744082-07:00","document_number":"6","attachment_number":3,"pacer_doc_id":"181037220270","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 2","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554423/","id":490554423,"tags":[],"absolute_url":"/docket/74659430/6/4/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.193402-07:00","date_modified":"2026-08-23T04:37:19.092869-07:00","sha1":"b72d019f4eff2fbe3c07d14d8e38794b2d8b3bd1","page_count":22,"file_size":2170190,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.4.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.4.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 1 of 22\n\n\n\n\n EXHIBIT\n\n 3\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 2 of 22\n\n USOORE48438E\n(( 1912)) United\n United States\n Reissued Patent ( 10 ) Patent Number : US RE48,438 E\n Gorchetchnikov et al . (45 ) Date of Reissued Patent : * Feb . 16, 2021\n( 54 ) GRAPHIC PROCESSOR BASED ( 56 ) References Cited\n ACCELERATOR SYSTEM AND METHOD\n U.S. PATENT DOCUMENTS\n( 71 ) Applicant: Neurala, Inc. , Boston, MA (US) 5,063,603 A 11/1991 Burt\n( 72 ) Inventors: Anatoli Gorchetchnikov , Belmont, MA 5,136,687 A 8/1992 Edelman et al .\n (US ) ; Heather Marie Ames , Milton , (Continued )\n MA (US ) ; Massimiliano Versace, FOREIGN PATENT DOCUMENTS\n Milton, MA (US ); Fabrizio Santini ,\n Jamaica Plain, MA (US) EP 1 224 622 B1 7/2002\n WO WO 2014/190208 11/2014\n( 73 ) Assignee : Neurala, Inc. , Boston, MA (US) ( Continued )\n( * ) Notice: This patent is subject to a terminal dis\n claimer . OTHER PUBLICATIONS\n( 21 ) Appl. No .: 15 /808,201 Cornwall et al . , \u201c Automatically Translating a General Purpose C ++\n Image Processing Library for GPUs \u201d , IEEE , Jun . 2006 , 8 pages .\n(22 ) Filed : Nov. 9 , 2017 ( Year: 2006 ) . *\n Related U.S. Patent Documents ( Continued )\nReissue of:\n( 64) Patent No .: 9,189,828 Primary Examiner William H. Wood\n Issued: Nov. 17 , 2015 (74 ) Attorney, Agent, or Firm -Smith Baluch LLP\n Appl. No .: 14 /147,015 ( 57 ) ABSTRACT\n Filed : Jan. 3 , 2014\nU.S. Applications: An accelerator system is implemented on an expansion card\n( 63 ) Continuation of application No. 11 / 860,254 , filed on comprising a printed circuit board having ( a) one or more\n Sep. 24 , 2007 , now Pat . No. 8,648,867 . graphics processing units ( GPUs ) , ( b ) two or more associ\n (Continued ) ated memory banks ( logically or physically partitioned ), (c )\n a specialized controller, and ( d) a local bus providing signal\n( 51 ) Int. Ci. coupling compatible with the PCI industry standards. The\n GO6T 1/60 ( 2006.01 ) controller handles most of the primitive operations to set up\n G06F 9/50 ( 2006.01 ) and control GPU computation. Thus, the computer's central\n (Continued ) processing unit ( CPU) can be dedicated to other tasks . In this\n( 52 ) U.S. Ci . case a few controls ( simulation start and stop signals from\n CPC G06T 1/20 (2013.01 ) ; G06F 9/5027\n the CPU and the simulation completion signal back to CPU ),\n GPU programs and input/output data are exchanged between\n ( 2013.01 ) ; G06T 1/60 ( 2013.01 ) ; CPU and the expansion card . Moreover, since on every time\n ( Continued ) step of the simulation the results from the previous time step\n( 58 ) Field of Classification Search are used but not changed, the results are preferably trans\n CPC ... GO6F 9/5027 ; G06F 2209/509 ; G06T 1/20 ; ferred back to CPU in parallel with the computation .\n GO6T 1/60 ; G06N 99/005 ; GO6N 37063\n See application file for complete search history . 56 Claims , 5 Drawing Sheets\n Expansion Cards 40\n\n 420\n ? e.com\n\n 403 435\n xxtes\n 19x VI 9XY\n SKM 2\n 450\n Swarum\n shokolaty\n bars CPU\n\n\n 19:43 Shader GPU .\n\n 180\n Outdoor\n yoles exi\n Teney bank\n 440\n 482 Now external\n texmex Tom FRAM\n UM\n WOX + vary but\n\n\n Wait for SWRP Svima i wait for swap\n otrputloutpan\n xure points\n ? output\n texture packs\n otinescioutput\n extere gointate\n plexulante\n 478\n ******** * herbimity\n Dank AM leration harasana\n\n\n en?\n 490\n Waxa ste\n tra\n axexkoren\n 499\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 3 of 22\n\n\n US RE48,438 E\n Page 2\n\n Related U.S. Application Data 2014/0032461 Al 1/2014 Weng\n 2014/0089232 A1 3/2014 Buibas et al .\n( 60 ) Provisional application No. 60 /826,892 , filed on Sep. 2014/0052679 Al\n 2015/0127149 Al\n 11/2014 Sinyavskiy et al .\n 5/2015 Sinyavskiy et al .\n 25 , 2006 . 2015/0134232 Al 5/2015 Robinson\n 2015/0224648 A1 8/2015 Lee et al .\n( 51 ) Int . Ci . 2016/0075017 A1 3/2016 Laurent et al.\n GOOT 1/20 ( 2006.01 ) 2016/0082597 A1 3/2016 Gorchetchnikov et al .\n GOON 37063 (2006.01 ) 2016/0096270 A1 4/2016 Gabardos et al .\n GOON 20/00 2016/0198000 Al 7/2016 Gorchetchnikov et al .\n ( 2019.01 ) 2017/0024877 A1 1/2017 Versace et al .\n( 52) U.S. Ci. 2017/0076194 A1 3/2017 Versace et al .\n CPC G06F 2209/509 (2013.01 ) ; GO6N 37063 2017/0193298 Al 7/2017 Versace et al .\n (2013.01 ) ; GOON 20/00 (2019.01 ) FOREIGN PATENT DOCUMENTS\n( 56) References Cited WO WO 2014/204615 12/2014\n U.S. PATENT DOCUMENTS WO WO 2015/143173 9/2015\n WO WO 2016/014137 1/2016\n 5,142,665 A * 8/1992 Bigus GOON 3/04\n 706/16 OTHER PUBLICATIONS\n 5,172,253 A 12/1992 Lynne\n 5,388,206 A 2/1995 Poulton et al . Montrym et al . , \u201c The GeForce 6800 \u201d , IEEE , 2005 , 11 pages. ( Year :\n 6,018,696 A 1/2000 Matsuoka et al . 2005) . *\n 6,336,051 B1 1/2002 Pangels et al. Perumalla , Kalyan S. , \u201c Discrete - event Execution Alternatives on\n 6,647,508 B2 11/2003 Zalewski et al .\n 7,119,810 B2 * 10/2006 Sumanaweera G06T 15/005 General Purpose Graphical Processing Units (GPGPUs)\", IEEE ,\n 345/506 \u201c Proceedings of the 20th Workshop on Principles of Advanced and\n 7,219,085 B2 * 5/2007 Buck GOON 3/08 Distributed Simulation ( PADS'06 ) \u201d , Mar. 2006 , 8 pages . ( Year:\n 706/12 2006 ) . *\n 7,477,256 B1 * 1/2009 Johnson G06F 3/14 Setoain et al . , \u201c Parallel Hyperspectral Image Processing on Com\n 345/501 modity Graphics Hardware \u201d , IEEE , \u201c Proceedings of the 2006\n 7,525,547 B1 * 4/2009 Diard G06T 1/20 International Conference on Parallel Processing Workshops\n 345/502\n 7,765,029 B2 7/2010 Fleischer et al . ( ICPPW'06 ) \" , Mar. 2006 , 8 pages . ( Year: 2006 ) . *\n 7,861,060 B1 * 12/2010 Nickolls et al . 712/22 Luo et al ., \u201c Artificial Neural Network Computation on Graphic\n 7,873,650 B1 1/2011 Chapman et Process Unit \u201d , IEEE , \u201c Proceedings of International Joint Confer\n 8,392,346 B2 3/2013 Ueda et al . ence on Neural Networks, Montreal, Canada, Jul. 31 - Aug. 4 , 2005 \u201d ,\n 8,510,244 B2 8/2013 Carson et al . Feb. 2005 , pp . 622-626 ( Year: 2005 ) . *\n 8,583,286 B2 11/2013 Fleischer et al . Non - Final Office Action dated Jan. 4 , 2018 from U.S. Appl. No.\n 8,648,867 B2 * 2/2014 Gorchetchnikov et al. .. 345/501 15/262,637 , 23 pages.\n 9,031,692 B2 5/2015 Zhu Wu, Yan & J. Cai , H. ( 2010 ) . A Simulation Study of Deep Belief\n 9,177,246 B2 11/2015 Buibas et al .\n 9,626,566 B2 4/2017 Versace et al. Network Combined with the Self -Organizing Mechanism of Adap\n 10,083,523 B2 9/2018 Versace et al. tive Resonance Theory. 10.1109 /CISE.2010.5677265 , 4 pages .\n 2001/0010034 A1 * 7/2001 Burton G06F 17/5022 Al-Kaysi, A. M. et al . , A Multichannel Deep Belief Network for the\n 703/17 Classification of EEG Data, from Ontology -based Information\n 2002/0046271 A1 4/2002 Huang Extraction for Residential Land Use Suitability: A Case Study of the\n 2002/0050518 A1 5/2002 Roustaei City of Regina, Canada, DOI 10.1007/978-3-319-26561-2_5 , 8\n 2002/0064314 Al 5/2002 Comaniciu et al . pages (Nov. 2015 ) .\n 2002/0168100 A1 11/2002 Woodall\n 2003/0026588 A1 2/2003 Elder et al . Apolloni, B. et al . , Training a network of mobile neurons, Proceed\n 2003/0078754 Al 4/2003 Hamza ings of International Joint Conference on Neural Networks, San\n 2004/0015334 Al 1/2004 Ditlow et al. Jose , CA , doi: 10.1109 /IJCNN.2011.6033427, pp . 1683-1691 ( Jul.\n 2005/0166042 Al 7/2005 Evans 31 -Aug. 5 , 2011 ) .\n 2006/0129506 A1 * 6/2006 Edelman GO5D 1/0088 Boddapati, V., Classifying Environmental Sounds with Image Net\n 706/12 works, Thesis , Faculty of Computing Blekinge Institute of Tech\n 2006/0184273 A1 8/2006 Sawada et al. nology, 37 pages . ( Feb. 2017 ) .\n 2007/0052713 A1 3/2007 Chung et al .\n 2007/0198222 A1 8/2007 Schuster et al . Khaligh -Razavi, S.-M. et al . , Deep Supervised, but Not Unsuper\n 2007/0279429 A1 12/2007 Ganzer vised, Models May Explain IT Cortical Represen on, PLoS\n 2008/0033897 A1 2/2008 Lloyd Computational Biology , vol . 10 , Issue 11 , 29 pages . (Nov. 2014 ) .\n 2008/0066065 Al 3/2008 Kim et al . Kim , S. , Novel approaches to clustering , biclustering and algo\n 2008/0258880 A1 10/2008 Smith et al . rithms based on adaptive resonance theory and intelligent control,\n 2009/0080695 Al 3/2009 Yang et al . Doctoral Dissertations, Missouri University of Science and Tech\n 2009/0089030 Al 4/2009 Sturrock et al .\n 2009/0116688 A1 5/2009 Monacos et al . nology, 125 pages . ( 2016 ) .\n 2010/0048242 A1 2/2010 Rhoads et al . Notice of Alllowance dated May 22 , 2018 from U.S. Appl. No.\n 2011/0004341 Al 1/2011 Sarvadevabhatla et al . 15/262,637 , 6 pages.\n 2011/0173015 A1 7/2011 Chapman et al . Ren, Y. et al . , Ensemble Classification and Regression Recent\n 2011/0279682 A1 11/2011 Li et al . Developments, Applications and Future Directions , in IEEE Com\n 2012/0072215 A1 3/2012 Yu et al . putational Intelligence Magazine , 10.1109 /MCI.2015.2471235 , 14\n 2012/0089552 A1 4/2012 Chang et al . pages (2016) .\n 2012/0197596 A1 8/2012 Comi\n 2012/0316786 A1 12/2012 Liu et al . Sun , Z. et al . , Recognition of SAR target based on multilayer\n 2013/0126703 A1 5/2013 Caulfield auto -encoder and SNN , International Journal of Innovative Com\n 2013/0131985 Al 5/2013 Weiland et al . puting, Information and Control, vol . 9 , No. 11 , pp . 4331-4341 ,\n 2014/0019392 A1 1/2014 Buibas et al. Nov. 2013 .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 4 of 22\n\n\n US RE48,438 E\n Page 3\n\n( 56 ) References Cited Coifman , R.R. , Lafon , S. , Lee , A.B. , Maggioni , M. , Nadler, B. ,\n Warner, F. , and Zucker, S.W. Geometric diffusions as a tool for\n OTHER PUBLICATIONS harmonic analysis and structure definition of data : Diffusion maps.\n Proceedings of the National Academy of Sciences of the United\nAdelson , E. H. , Anderson, C. H. , Bergen , J. R. , Burt, P. J. , & Ogden, States of America , 102 ( 21 ) : 7426 , 2005 .\nJ. M. ( 1984 ) . Pyramid methods in image processing. RCA engineer, Davis, C. E. 2005. Graphic Processing Unit Computation of Neural\n29(6) , 33-41 . Networks. Master's thesis, University of New Mexico , Albuquer\nAggarwal, Charu C , Hinneburg, Alexander, and Keim , Daniel A. On que, NM , 121 pages .\nthe surprising behavior of distance metrics in high dimensional Dosher, B.A. , and Lu , Z.L. ( 2010 ) . Mechanisms of perceptual\nspace . Springer, 2001 . attention in precuing of location . Vision Res . , 40 ( 10-12 ) . 1269\nAmes, H , Versace , M. , Gorchetchnikov, A. , Chandler, B. , Livitz , G. , 1292 .\nL\u00e9veill\u00e9 , J. , Mingolla, E. , Carter, D. , Abdalla, H. , and Snider, G. Ellias , S. A. , and Grossberg, S. ( 1975 ) . Pattern formation, contrast\n( 2012 ) Persuading computers to act more like brains. In Advances control and oscillations in the short term memory of shunting\nin Neuromorphic Memristor Science and Applications, Kozma , on -center off - surround networks. Biol Cybern 20 , pp . 69-98 .\nR.Pino , R . , and Pazienza , G. ( eds ) , Springer Verlag . Extended European Search Report and Written Opinion dated Jun .\nAmes, H. Mingolla, E. , Sohail , A. , Chandler, B. , Gorchetchnikov, 1 , 2017 from European Application No. 14813864.7 , 10 pages .\nA. , L\u00e9veill\u00e9 , J. , Livitz , G. and Versace, M. ( 2012 ) The Animat . IEEE Extended European Search Report and Written Opinion dated Oct.\nPulse , Feb. 2012 , 3 ( 1 ) , 47-50 . 12 , 2017 from European Application No. 14800348.6 , 12 pages .\nArtificial Intelligence As a Service. Invited talk, Defrag, Broomfield , Extended European Search Report and Written Opinion Oct. 23 ,\nCO , Nov. 4-6 ( 2013 ). 2017 from European Application No. 15765396.5 , 8 pages .\nAryananda, L. ( 2006 ) . Attending to learn and learning to attend for Fazl , A. , Grossberg , S. , and Mingolla , E. ( 2009 ) . View - invariant\na social robot . Humanoids 06 , pp . 618-623 . object category learning, recognition, and search : How spatial and\nBaraldi, A. and Alpaydin , E. ( 1998 ) . Simplified ART: A new class object attention are coordinated using surface -based attentional\nof ART algorithms. International Computer Science Institute, Berke shrouds . Cognitive Psychology 58 , 1-48 .\nley, CA , TR- 98-004 , 1998 . F\u00f6ldi\u00e1k , P. ( 1990 ) . Forming sparse representations by local anti\nBaraldi, A. and Alpaydin , E. ( 2002 ) . Constructive feedforward ART Hebbian learning, Biological Cybernetics, vol . 64 , pp . 165-170 .\nclustering networks \u2014 Part I. IEEE Transactions on Neural Net Friston K. , Adams R. , Perrinet L. , & Breakspear M. ( 2012 ) . Per\nworks 13 ( 3 ) , 645-661 . ceptions as hypotheses: saccades as experiments . Frontiers in Psy\nBaraldi, A. and Parmiggiani, F. ( 1997 ) . Fuzzy combination of chology, 3 ( 151 ), 1-20 .\nKohonen's and ART neural network models to detect statistical Galbraith , B.V, Guenther, F.H. , and Versace , M. ( 2015 ) A neural\nregularities in a random sequence of multi - valued input patterns. In network -based exploratory learning and motor planning system for\nInternational Conference on Neural Networks, IEEE . co - robots.Frontiers in Neuroscience, in press .\nBaraldi, Andrea and Alpaydin , Ethem . Constructive feedforward George , D. and Hawkins, J. ( 2009 ) . Towards a mathematical theory\nART clustering networks \u2014 part II . IEEE Transactions on Neural of cortical micro - circuits . PLoS Computational Biology 5 ( 10 ) , 1-26 .\nNetworks, 13 ( 3 ) : 662-677 , May 2002. ISSN 1045-9227 . doi: 10.1109/ Georgeii , J. , and Westermann , R. ( 2005 ) . Mass - spring systems on\ntnn.2002.1000131\n1000131 .\n . URL http://dx.doi.org/10.1109/tnn.2002 . the GPU . Simulation Modelling Practice and Theory 13 , pp . 693\n 702 .\nBengio , Y., Courville, A. , & Vincent, P. Representation learning: A Gorchetchnikov A. , Hasselmo M.E. ( 2005 ) . A biophysical imple\nreview and new perspectives , IEEE Transactions on Pattern Analy mentation of a bidirectional graph search algorithm to solve mul\nsis and Machine Intelligence , vol . 35 Issue 8 , Aug. 2013 , pp . tiple goal navigation tasks . Connection Science , 17 ( 1-2 ) , pp . 145\n1798-1828 . 166 .\nBerenson , D. et al ., A robot path planning framework that learns Gorchetchnikov A. , Hasselmo M.E. ( 2005 ) . A simple rule for\nfrom experience, 2012 International Conference on Robotics and spike -timing -dependent plasticity : local influence of AHP current.\nAutomation, 2012 , 9 pages [ retrieved from the internet] URL : http : // Neurocomputing, 65-66 , pp . 885-890 .\nusers.wpi.edu/-dberenson/lightning.pdf. Gorchetchnikov A. , Versace, M. , Hasselmo M.E. ( 2005 ) . A Model\nBernhard , F. , and Keriven, R. ( 2005 ) . Spiking Neurons on GPUs. of STDP Based on Spatially and Temporally Local Information :\nTech . Rep. 05-15 , Ecole Nationale des Ponts et Chauss\u2019es, 8 pages. Derivation and Combination with Gated Decay. Neural Networks,\nBesl , P. J. , & Jain , R. C. ( 1985 ) . Three - dimensional object recog 18 , pp . 458-466 .\nnition . ACM Computing Surveys (CSUR ), 17 ( 1 ) , 75-145 . Gorchetchnikov A. , Versace, M. , Hasselmo M.E. ( 2005 ) . Spatially\nBohn , C.-A. Kohonen . ( 1998 ) . Feature Mapping Through Graphics and temporally local spiketiming -dependent plasticity rule. In :\nHardware. In Proceedings of 3rd Int. Conference on Computational Proceedings of the International Joint Conference on Neural Net\nIntelligence and Neurosciences , 4 pages . works, No. 1568 in IEEE CD - ROM Catalog No. 05CH37662C , pp .\nBradski, G. , & Grossberg, S. ( 1995 ) . Fast -learning Viewnet archi 390-396 .\ntectures for recognizing three - dimensional objects from multiple Gorchetchnikov A. An Approach to a Biologically Realistic Simu\ntwo - dimensional views. Neural Networks, 8 ( 7-8 ) , 1053-1080 . lation of Natural Memory . Master's thesis , Middle Tennessee State\nCanny, J.A. ( 1986 ) . Computational Approach to Edge Detection, University, Murfreesboro, TN , 70 pages .\nIEEE Trans. Pattern Analysis and Machine Intelligence , 8 (6 ):679 Grossberg, S. ( 1973 ) . Contour enhancement, short -term memory ,\n698 . and constancies in reverberating neural networks. Studies in Applied\nCarpenter, G.A. and Grossberg, S. ( 1987 ) . A massively parallel Mathematics 52 , 213-257 .\narchitecture for a self -organizing neural pattern recognition machine . Grossberg , S. , and Huang, T.R. ( 2009 ) . Artscene : A neural system\nComputer Vision, Graphics, and Image Processing 37 , 54-115 . for natural scene classification . Journal of Vision, 9 (4 ) , 6.1-19 .\nCarpenter, G.A. , and Grossberg , S. ( 1995 ) . Adaptive resonance doi : 10.1167 /9.4.6 .\ntheory ( ART ). In M. Arbib ( Ed . ) , The handbook of brain theory and Grossberg, S. , and Versace , M. ( 2008 ) Spikes , synchrony, and\nneural networks. (pp . 79-82 ) . Cambridge , M.A .: MIT press . attentive learning by laminar thalamocortical circuits . Brain Research ,\nCarpenter, G.A. , Grossberg, S. and Rosen , D.B. ( 1991 ) . Fuzzy ART: 1218C , 278-312 [ Authors listed alphabetically ].\nFast stable learning and categorization of analog patterns by an Hagen , T.R. , Hjelmervik, J. , Lie , K.-A. , Natvig , J. and Ofstad\nadaptive resonance system . Neural Networks 4 , 759-771 . Henriksen , M. ( 2005 ) . Visual simulation of shallow -water waves .\nCarpenter, Gail A and Grossberg, Stephen . The art of adaptive Simulation Modelling Practice and Theory 13 , pp . 716-726 .\npattern recognition by a self -organizing neural network . Computer, Hasselt , Hado Van . Double q -learning. In Advances in Neural\n21 ( 3 ) : 77-88 , 1988 . Information Processing Systems , pp . 2613-2621 , 2010 .\nCoifman , R.R. and Maggioni , M. Diffusion wavelets . Applied and Hinton , G. E. , Osindero , S. , and Teh , Y. ( 2006 ) . A fast learning\nComputational Harmonic Analysis, 21 ( 1 ) : 53-94 , 2006 . algorithm for deep belief nets. Neural Computation, 18 , 1527-1554 .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 5 of 22\n\n\n US RE48,438 E\n Page 4\n\n( 56 ) References Cited L\u00e9veill\u00e9 , J. , Ames, H. , Chandler, B. , Gorchetchnikov, A. , Mingolla ,\n E. , Patrick , S. , and Versace, M. ( 2010 ) Learning in a distributed\n OTHER PUBLICATIONS software architecture for large -scale neural modeling. BIONET\n ICS10 , Boston , MA , USA .\nHodgkin , A.L. , and Huxley, A.F. ( 1952 ) . Quantitative description of Livitz , G. , Versace, M. , Gorchetchnikov, A. , Vasilkoski, Z. , Ames,\nmembrane current and its application to conduction and excitation H. , Chandler, B. , L\u00e9veill\u00e9 , J. , Mingolla , E. , Snider, G. , Amerson , R. ,\nin nerve. J Physiol 117 , pp . 500-544 . Carter, D. , Abdalla, H. , and Qureshi , S. ( 2011 ) Visually -Guided\nHopfield , J. ( 1982 ) . Neural networks and physical systems with Adaptive Robot ( ViGUAR ). Proceedings of the International Joint\nemergent collective computational abilities . In Proc Natl Acad Sci Conference on Neural Networks ( IJCNN) 2011 , San Jose , CA ,\nUSA , vol . 79 , pp . 2554-2558 . USA .\nIlie , A. ( 2002 ) . Optical character recognition on graphics hardware . Lowe , D.G. ( 2004 ) . Distinctive Image Features from Scale - Invariant\nTech . Rep . integrative paper, UNCCH , Department of Computer Keypoints. Journal International Journal of Computer Vision archive\nScience, 9 pages . vol . 60 , 2 , 91-110 .\nInternational Preliminary Report on Patentability in related PCT Lu , Z.L. , Liu , J. , and Dosher, B.A. ( 2010 ) Modeling mechanisms of\nApplication No. PCT/US2014 /039162 filed May 22 , 2014 , dated perceptual learning with augmented Hebbian re -weighting. Vision\nNov. 24 , 2015 , 7 pages. Research , 50 ( 4 ) . 375-390 .\nInternational Preliminary Report on Patentability in related PCT Mahadevan, S. Proto -value functions: Developmental reinforce\nApplication No. PCT/US2014 /039239 filed May 22 , 2014 , dated ment learning. In Proceedings of the 22nd international conference\nNov. 24 , 2015 , 8 pages. on Machine learning, pp . 553-560 . ACM , 2005 .\nInternational Preliminary Report on Patentability dated Nov. 8 , Meuth , J.R. and Wunsch , D.C. ( 2007 ) A Survey of Neural Compu\n2016 from International Application No. PCT/US2015 /029438, 7 tation on Graphics Processing Hardware. 22nd IEEE International\npages. Symposium on Intelligent Control, Part of IEEE Multi-conference\nInternational Search Report and Written Opinion dated Feb. 18 , on Systems and Control, Singapore, Oct. 1-3 , 2007 , 5 pages .\n2015 from International Application No. PCT /US2014 /039162, 12 Mishkin M , Ungerleider LG . ( 1982 ) . \" Contribution of striate inputs\npages. to the visuospatial functions of parieto -preoccipital cortex in mon\nInternational Search Report and Written Opinion dated Feb. 23 , keys,\u201d Behav Brain Res, 6 ( 1 ) : 57-77 .\n2016 from International Application No. PCT /US2015 /029438 , 11 Mnih , Volodymyr, Kavukcuoglu , Koray, Silver, David , Rusu , Andrei\npages . A , Veness, Joel , Bellemare, Marc G , Graves, Alex , Riedmiller,\nInternational Search Report and Written Opinion dated Jul. 6 , 2017 Martin , Fidjeland, Andreas K , Ostrovski, Georg , et al . Human -level\nfrom International Application No. PCT/US2017 /029866 , 12 pages . control through deep reinforcement learning. Nature, 518 (7540 ):529\nInternational Search Report and Written Opinion dated Nov. 26 , 533 , Feb. 25 , 2015 .\n2014 from International Application No. PCT /US2014 /039239, 14 Moore , Andrew W and Atkeson, Christopher G. Prioritized sweep\npages. ing : Reinforcement learning with less data and less time . Machine\nInternational Search Report and Written Opinion dated Sep. 15 , Learning, 13 ( 1 ) : 103-130 , 1993 .\n2015 from International Application No. PCT /US2015 /021492 , 9 Najemnik, J. , and Geisler, W. ( 2009 ) . Simple summation rule for\npages. optimal fixation selection in visual search . Vision Research . 49 ,\nItti, L. , and Koch , C. ( 2001 ) . Computational modelling of visual 1286-1294 .\nattention . Nature Reviews Neuroscience, 2 ( 3 ) , 194-203 . Notice of Allowance dated Jul. 27 , 2016 from U.S. Appl. No.\nItti, L. , Koch , C. , and Niebur, E. ( 1998 ) . A Model of Saliency - Based 14/662,657 .\nVisual Attention for Rapid Scene Analysis, 1-6 . Notice of Allowance dated Dec. 16 , 2016 from U.S. Appl. No.\nJarrett, K. , Kavukcuoglu , K. , Ranzato , M. A. , & LeCun , Y. ( Sep. 14/662,657 .\n2009 ) . What is the best multi - stage architecture for object recog Oh , K.-S. , and Jung, K. ( 2004 ) . GPU implementation of neural\nnition ?. In Computer Vision , 2009 IEEE 12th International Confer networks. Pattern Recognition 37 , pp . 1311-1314 .\nence on (pp . 2146-2153 ) . IEEE . Oja, E. ( 1982 ) . Simplified neuron model as a principal component\nKipfer, P., Segal , M. , and Westermann , R. ( 2004 ) . UberFlow : A analyzer. Journal of Mathematical Biology 15 ( 3 ) , 267-273 .\nGPU - Based Particle Engine . In Proceedings of the SIGGRAPH/ Partial Supplementary European Search Report dated Jul. 4 , 2017\nEurographics Workshop on Graphics Hardware 2004 , pp . 115-122 . from European Application No. 14800348.6 , 13 pages .\nKolb , A. , L. Latta , and C. Rezk - Salama. ( 2004 ) . \u201c Hardware - Based Raijmakers, M.E.J. , and Molenaar, P. ( 1997 ) . Exact ART: A com\nSimulation and Collision Detection for Large Particle Systems.\u201d in plete implementation of an ART network Neural networks 10 (4 ) ,\nProceedings of the SIGGRAPH /Eurographics Workshop on Graph 649-669 .\nics Hardware 2004 , pp . 123-131 . Ranzato , M. A. , Huang, F. J. , Boureau , Y. L. , & Lecun , Y. ( Jun .\nKompella, Varun Raj, Luciw , Matthew , and Schmidhuber, J\u00fcrgen. 2007 ) . Unsupervised learning of invariant feature hierarchies with\nIncremental slow feature analysis: Adaptive low - complexity slow applications to object recognition . In Computer Vision and Pattern\nfeature updating from high -dimensional input streams. Neural Com Recognition, 2007. CVPR\u201907 . IEEE Conference on (pp . 1-8 ) . IEEE .\nputation , 24 ( 11 ) : 2994-3024 , 2012 . Raudies, F. , Eldridge, S. , Joshi , A. , and Versace, M. ( Aug. 20 , 2014 ) .\nKowler, E. ( 2011 ) . Eye movements: The past 25years. Vision Learning to navigate in a virtual world using optic flow and stereo\nResearch , 51 ( 13 ) , 1457-1483 . doi : 10.1016 / j.visres.2010.12.014 . disparity signals. Artificial Life and Robotics, DOI 10.1007/ s10015\nLarochelle H. , & Hinton G. ( 2012 ) . Learning to combine foveal 014-0153-1 .\nglimpses with a third - order Boltzmann machine . NIPS 2010,1243 Riesenhuber, M. , & Poggio , T. ( 1999 ) . Hierarchical models of\n1251 . object recognition in cortex . Nature Neuroscience , 2 ( 11 ) , 1019\nLeCun , Y., Kavukcuoglu, K. , & Farabet, C. (May 2010 ) . Convolu 1025 .\ntional networks and applications in vision . In Circuits and Systems Riesenhuber, M. , & Poggio , T. ( 2000 ) . Models of object recogni\n( ISCAS ) , Proceedings of 2010 IEEE International Symposium on tion . Nature neuroscience , 3 , 1199-1204 .\n( pp . 253-256 ) . IEEE . Rolfes, T. ( 2004 ) . Artificial Neural Networks on Programmable\nLee , D. D. and Seung, H. S. ( 1999 ) . Learning the parts of objects Graphics Hardware . In Game Programming Gems 4 , A. Kirmse, Ed .\nby non -negative matrix factorization. Nature, 401 ( 6755 ) : 788-791 . Charles River Media , Hingham , MA , pp . 373-378 .\nLee , D. D. , and Seung , H. S. ( 1997 ) . \u201c Unsupervised learning by Rublee , E. , Rabaud, V., Konolige , K. , & Bradski, G. ( 2011 ) . ORB :\nconvex and conic coding .\u201d Advances in Neural Information Pro An efficient alternative to SIFT or SURF. In IEEE International\ncessing Systems , 9 . Conference on Computer Vision ( ICCV ) 2011 , 2564-2571 .\nLegenstein , R. , Wilbert, N. , and Wiskott, L. Reinforcement learning Ruesch, J. et al . ( 2008 ) . Multimodal Saliency -Based Bottom -Up\non slow features of high - dimensional input streams. PLoS Compu Attention a Framework for the Humanoid Robot iCub . 2008 IEEE\ntational Biology, 6 ( 8 ) , 2010. ISSN 1553-734X . International Conference on Robotics and Automation , pp . 962-965 .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 6 of 22\n\n\n US RE48,438 E\n Page 5\n\n( 56 ) References Cited Tong , F. , Ze -Nian Li , ( 1995 ) . Reciprocal -wedge transform for\n space - variant sensing, Pattern Analysis and Machine Intelligence,\n OTHER PUBLICATIONS IEEE Transactions on , vol . 17 , No. 5 , pp . 500-51 . doi: 10.1109 /\n 34.391393 .\nRumelhart D. , Hinton G. , and Williams, R. ( 1986 ) . Learning inter Torralba, A. , Oliva , A. , Castelhano, M.S. , Henderson, J.M. ( 2006 ) .\nnal representations by error propagation. In Parallel distributed Contextual guidance of eye movements and attention in real -world\nprocessing : explorations in the microstructure of cognition , vol . 1 , scenes: the role of global features in object search . Psychological\nMIT Press . Review , 113 ( 4 ) .766-786 .\nRumpf, M. and Strzodka, R. Graphics processor units : New pros Van Hasselt , Hado, Guez, Arthur, and Silver, David . Deep rein\npects for parallel computing. In Are Magnus Bruaset and Aslak forcement learning with double q - learning. arXiv preprint arXiv :\n 1509.06461 , Sep. 22 , 2015 .\nTveito , editors , Numerical Solution of Partial Differential Equations Versace, M. ( 2006 ) From spikes to interareal synchrony: how\non Parallel Computers, vol . 51 of Lecture Notes in Computational attentive matching and resonance control learning and information\nScience and Engineering , pp . 89-134 . Springer, 2005 . processing by laminar thalamocortical circuits . NSF Science of\nSchaul, Tom , Quan , John , Antonoglou, Ioannis, and Silver, David . Learning Centers PI Meeting, Washington , DC , USA .\nPrioritized experience replay. arXiv preprint arXiv: 1511.05952 , Versace, M. , ( 2010 ) Open - source software for computational neu\nNov. 18 , 2015 . roscience: Bridging the gap between models and behavior. In\nSchmidhuber, J. ( 2010 ) . Formal theory of creativity, fun , and Horizons in Computer Science Research ,vol. 3 .\nintrinsic motivation ( 1990-2010 ) . Autonomous Mental Develop Versace, M. , Ames, H. , L\u00e9veill\u00e9, J. , Fortenberry, B. , and Gorchetchnikov,\n A. ( 2008 ) KinNeSS : A modular framework for computational\nment, IEEE Transactions on , 2 ( 3 ) , 230-247 . neuroscience. Neuroinformatics, 2008 Winter ; 6 (4 ) : 291-309 . Epub\nSchmidhuber, J\u00fcrgen . Curious model -building control systems. In Aug 10 , 2008 .\nNeural Networks, 1991. 1991 IEEE International Joint Conference Versace, M. , and Chandler, B. ( 2010 ) MONETA : A Mind Made from\non , pp . 1458-1463 . IEEE , 1991 . Memristors. IEEE Spectrum , Dec. 2010 .\nSeibert, M. , & Waxman , A.M. ( 1992 ) . Adaptive 3 - D Object Rec Webster, Bachevalier, Ungerleider ( 1994 ) . Connections of IT areas\nognition from Multiple Views. IEEE Transactions on Pattern Analy TEO and TE with parietal and frontal cortex in macaque monkeys.\nsis and Machine Intelligence , 14 ( 2 ) , 107-124 . Cerebal Cortex, 4 ( 5 ) , 470-483 .\nSherbakov, L. and Versace, M. ( 2014 ) Computational principles for Wiskott, Laurenz and Sejnowski, Terrence . Slow feature analysis:\nan autonomous active vision system . Ph.D., Boston University,\n Unsupervised learning of invariances. Neural Computation , 14 (4 ):715\n 770 , 2002 .\nhttp://search.proquest.com/docview/1558856407. Livitz G. , Versace M. , Gorchetchnikov A. , Vasilkoski Z. , Ames H. ,\nSherbakov, L. , Livitz , G. , Sohail , A. , Gorchetchnikov, A. , Mingolla , Chandler B. , Leveille J. and Mingolla E. ( 2011 ) Adaptive, brain -like\nE. , Ames, H. , and Versace, M. ( 2013a) CogEye: An online active systems give robots complex behaviors, The Neuromorphic Engi\nvision system that disambiguates and recognizes objects. NeuComp neer, : 10.2417 / 1201101.003500 Feb. 2011. 3 pages .\n2013 . Salakhutdinov, R. , & Hinton, G. E. ( 2009 ) . Deep boltzmann machines.\nSherbakov, L. , Livitz , G. , Sohail , A. , Gorchetchnikov, A. , Mingolla , In International Conference on Artificial Intelligence and Statistics\nE. , Ames, H. , and Versace, M ( 2013b) A computational model of the (pp . 448-455 ).\nrole of eye -movements in object disambiguation . Cosyne, Feb. Sherbakov, L. et al . 2012. CogEye : from active vision to context\n28 -Mar. 3 , 2013. Salt Lake City, UT, USA . identification, youtube, retrieved from the Internet on Oct. 10 , 2017 :\nSmolensky, P. ( 1986 ) . Information processing in dynamical sys URL : //www.youtube.com/watch ? v = i5PQk962B1k, 1 page .\ntems : Foundations of harmony theory. In D. E. Sherbakov, L. et al. 2013. CogEye: system diagram module brain\nSpratling, M. W. ( 2008 ) . Predictive coding as a model of biased area function algorithm approx # neurons, retrieved from the\ncompetition in visual attention . Vision Research, 48 ( 12 ) : 1391-1408 . Internet on Oct. 12 , 2017 : URL : //http ://www-labsticc.univ-ubs.\nSpratling, M. W. ( 2012 ) . Unsupervised learning of generative and fr / ~ coussy /neucomp2013 / index_fichiers /material / posters /\ndiscriminative weights encoding elementary image components in a NeuComp2013_final56x36.pdf, 1 page .\npredictive coding model of cortical function . Neural Computation , Snider, Greg, et al . \u201c From synapses to circuitry : Using memristive\n24 ( 1 ) : 60-103 . memory to explore the electronic brain . \u201d IEEE computer, vol . 44 ( 2 ) .\nSpratling, M. W. , De Meyer, K. , and Kompass , R. ( 2009 ) . Unsu (2011 ) : 21-28 .\npervised learning of overlapping image components using divisive Versace, TEDx Fulbright, Invited talk, Washington DC , Apr. 5 ,\ninput modulation . Computational intelligence and neuroscience . 2014. 30 pages .\nSprekeler, H. On the relation of slow feature analysis and laplacian Versace , Brain - inspired computing. Invited keynote address, Bionet\neigenmaps. Neural Computation, pp . 1-16 , 2011 . ics 2010 , Boston , MA , USA . 1 page .\nSutton , Richard S and Barto , Andrew G. Reinforcement learning:\nAn introduction. MIT Press , 1998 . * cited by examiner\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 7 of 22\n\n\nU.S. Patent Feb.i6, 2021 Sheet 1 of 5 US RE48,438 E\n\n\n\n\n 14? *\n\n\n\n\n 130\n ????????????\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 8 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 2 of 5 US RE48,438 E\n\n\n\n\n I\n\n\n 220\n\n\n I\n\n\n\n\n RAM\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 9 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 3 of 5 US RE48,438 E\n\n\n\n 13:25 326 330 FIG\n 3\n .\n 320\n\n\n ECXPANRSIDON CONTROLER INITIALIZATION\n 180\n EIXNTEPRUNATL TFEXRTUOREMS RTOEXATUMRE MBEAMNORKY POPULATION SBIHNADRIERS FRAMTO SMHEAMDOERRY BANK COMPUTATION )\n 4\n .\n (\n SEE\n FIG\n\n\n\n\n COMPUTATIONAL 3ST0RE3AM\n 309\n DATA\n 310\n PSDAHRASDTEAR GAENERADTOR\n 316\n BATA\n 314\n DATA\n 350\n\n 304\n IPANRPSUETR TEXTURE GENERATOR POPULATION COMPILER IPANRPSUETR TGENXETRAUTROER ODUATPUAT ACUMULATION INRAM\n 315\n\n 306 LAST T?ERATION YES\n 311 312\n\n 300\n UINTSERAECTIRON 3ST0RE2AM 307 GUINRTSAEPREFHAIRCE 305 SIMULATION INITIALIZATION PROGRES MONITOR RESULTS DISPLAY NO\n 390\n START GURSAPEHIRC INTERFACE INITIALIZATION INTERACTION USER\n 313\n END\n\n\n 308 317 318\n CPU120\n DOUATTPUAT 3ST0RE1AM\n O\n /\n I\n DISK\n\n\n INITIALIZATION DOUATPTUAT TODISK\n NO\n LAST I?TERATION YES\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 10 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 4 of 5 US RE48,438 E\n\n\n Expansion Card 180\n\n\n\n Isoss foxxos ir\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 11 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 5 of 5 US RE48,438 E\n\n\n\n\n Blsthepofdiwfc.amnicpxklotuniolndwahrstg 5\n .\n FIG\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 12 of 22\n\n\n US RE48,438 E\n 1 2\n GRAPHIC PROCESSOR BASED for input /output should be designed so that it provides the\n ACCELERATOR SYSTEM AND METHOD synchronization with computation .\n In the case of GPGPU , the computation itself is performed\n outside\nMatter enclosed in heavy brackets [ ] appears in the 5 \u201c peripheral of the CPU , so the complete system comprises three\noriginal patent but forms no part of this reissue specifica hardware, and\u201d components: user interactive hardware , disk\ntion ; matter printed in italics indicates the additions ing unit ( CPUcomputational\n ) establishes\n hardware. The central process\n communication and synchroni\nmade by reissue ; a claim printed with strikethrough zation between peripherals. Each of the peripherals is pref\nindicates that the claim was canceled, disclaimed, or held erably controlled by a dedicated thread that is executed in\ninvalid by a prior post- patent action or proceeding . 10 parallel with minimal interactions and dependencies on the\n other threads .\n RELATED APPLICATIONS A GPU on a conventional video card is usually controlled\n through OpenGL , DirectX , or similar graphic application\n The present application is a broadening reissue applica programming\ntion of U.S. Pat. No. 9,189,828, filed Jan. 3 , 2014, which 15 context of graphic interfaces (APIs ). Such APIs establish the\nclaims a priority benefit, under 35 U.S.C. $ 120 , as a con GPU are made . This operations, within which all calls to the\ntinuation of U.S. application Ser. No. 11 / 860,254 , now U.S. within the same threadcontext\n of\n only works when initialized\n execution that uses it . As a result,\nPat . No. 8,648,867 B2 , filed Sep. 24 , 2007 , entitled \u201c Graphic in a preferred embodiment, the context is initialized within\nProcessor Based Accelerator System and Method , \u201d which in\nturn claims the priority benefit, under 35 U.S.C. 8119 (e ) , of 20 aevercomputational thread. This creates complications , how\n , in the interaction between the user interface thread that\nU.S. Application No. 60/ 826,892 , filed Sep. 25 , 2006. Each\nof the above - identified applications is incorporated herein by changes parameters of simulations and the computational\nreference in its entirety. More than one reissue application thread that uses these parameters.\nhas been filed for the reissue of U.S. Pat. No. 9,189,828 , A solution as proposed here is an implementation of the\nincluding this application and a reissue continuation appli- 25 computational stream of execution in hardware, so that\ncation filed Dec. 29, 2020. thread and context initialization are replaced by hardware\n initialization . This hardware implementation includes an\n BACKGROUND expansion card comprising a printed circuit board having ( a )\n one or more graphics processing units , (b ) two or more\n Graphics Processing Units (GPUs) are found in video 30 associated memory banks that are logically or physically\nadapters ( graphic cards) of most personal computers ( PCs ) , partitioned, (c ) a specialized controller, and (d) a local bus\nvideo game consoles , workstations, etc. and are considered providing signal coupling compatible with the PCI industry\nhighly parallel processors dedicated to fast computation of standards ( this includes but is not limited to PCI -Express ,\ngraphical content. With the advances of the computer and PCI - X , USB 2.0 , or functionally similar technologies ). The\nconsole gaming industries, the need for efficient manipula- 35 controller handles most of the primitive operations needed to\ntion and display of 3D graphics has accelerated the devel- set up and control GPU computation . As a result, the CPU\nopment of GPUs . is freed from this function and is dedicated to other tasks . In\n In addition , manufacturers of GPUs have included general this case a few controls ( simulation start and stop signals\npurpose programmability into the GPU architecture leading from the CPU and the simulation completion signal back to\nto the increased popularity of using GPUs for highly paral- 40 CPU) , GPU programs and input/output data are the infor\nlelizable and computationally expensive algorithms outside mation exchanged between CPU and the expansion card .\nof the computer graphics domain . When implemented on Moreover, since on every time step of the simulation the\nconventional video card architectures, these general purpose results from the previous time step are used but not changed ,\nGPU ( GPGPU) applications are not able to achieve optimal the results are preferably transferred back to CPU in parallel\nperformance , however. There is overhead for graphics- 45 with the computation .\nrelated features and algorithms that are not necessary for In general, according to one aspect , the invention features\nthese non-video applications. a computer system . This system comprises a central pro\n cessing unit , main memory accessed by the central process\n SUMMARY ing unit , and a video system for driving a video monitor in\n 50 response to the central processing unit as is common . The\n Numerical simulations, e.g. , finite element analysis , of computer system further comprises an accelerator that uses\nlarge systems of similar elements ( e.g. neural networks, input data from and provides output data to the central\ngenetic algorithms, particle systems, mechanical systems) processing unit. This accelerator comprises at least one\nare one example of an application that can benefit from graphics processing unit, accelerator memory for the graphic\nGPGPU computation . During numerical simulations, disk 55 processing unit, and an accelerator controller that moves the\nand user input /output can be performed independently of input data into the at least one graphics processing unit and\ncomputation because these two processes require interac- the accelerator memory to generate the output data .\ntions with peripheral hardware ( disk , screen , keyboard , In the preferred , the central processing unit transfers the\nmouse , etc ) and put relatively low load on the central input data for a simulation to the accelerator, after which the\nprocessing unit/system (CPU) . Complete independence is 60 accelerator executes simulation computations to generate\nnot desirable , however; user input might affect how the the output data, which is transferred to the central processing\ncomputation is performed and even interrupt it if necessary. unit. Preferably, the accelerator controller dictates an order\nFurthermore, the user output and the disk output are depen- of execution of instructions to the at least one graphics\ndent on the results of the computation. A reasonable solution processing unit . The use of the separate controller enables\nwould be to separate input/output into threads, so that it is 65 data transfer during execution such that the accelerator\ninteracting with hardware occurs in parallel with the com- controller transfers output data from the accelerator memory\nputation . In this case whatever CPU processing is required to main memory of the central processing unit .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 13 of 22\n\n\n US RE48,438 E\n 3 4\n In the preferred embodiment, the accelerator controller not limited to , workstations, server computers, supercom\ncomprises an interface controller that enables the accelerator puters, notebook computers, hand -held electronic devices\nto communicate over a bus of the computer system with the such as cell phones, mp3 players, or personal digital assis\ncentral processing unit. tants ( PDAs ) , multiprocessor systems, programmable con\n In general according to another aspect , the invention also 5 sumer electronics, networks of any of the above -mentioned\nfeatures an accelerator system for a computer system , which computing devices, and distributed computing environments\ncomprises at least one graphics processing unit , accelerator that including any of the above -mentioned computing\nmemory for the graphic processing unit and an accelerator devices.\ncontroller for moving data between the at least one graphics 10 In one implementation the GPU accelerator is imple\nprocessing unit and the accelerator memory .\n In general according to another aspect , the invention also mented as an expansion card 180 includes connections with\nfeatures a method for performing numerical simulations in a are installed along110with\n the motherboard , on which the one or more CPU's 120\n main , or system memory 130 and\ncomputer system . This method comprises a central process mass / non volatile data storage 140 , such as hard drive or\ning unit loading input data into an accelerator system from redundant array of independent drives (RAID ) array, for the\nmain memory of the central processing unit and an accel- 15 computer system 100. In the current example\nerator controller transferring the input data to a graphics card 180 communicates to the motherboard ,110 the expansion\nprocessing unit with instructions to be performed on the bus 190. This local bus 190 could be PCI , PCIviaExpress a local\ninput data . The accelerator controller then transfers output PCI - X , or any other functionally similar technology (de,\ndata generated by the graphic processing unit to the central 20 pending upon the availability on the motherboard 110 ) . An\nprocessing unit as output data .\n The above and other features of the invention including external version GPU accelerator is also a possible imple\nvarious novel details of construction and combinations of mentation . In this example, the external GPU accelerator is\nparts, and other advantages, will now be more particularly connected to the motherboard 110 through USB - 2.0 , IEEE\ndescribed with reference to the accompanying drawings and 1394 (Firewire ), or similar external /peripheral device inter\npointed out in the claims . It will be understood that the 25 face .\nparticular method and device embodying the invention are The CPU 120 and the system memory 130 on the moth\nshown by way of illustration and not as a limitation of the erboard 110 and the mass data storage system 140 are\ninvention . The principles and features of this invention may preferably independent of the expansion card 180 and only\nbe employed in various and numerous embodiments without communicate with each other and the expansion card 180\ndeparting from the scope of the invention . 30 through the system bus 200 located in the motherboard 110 .\n A system bus 200 in current generations of computers have\n BRIEF DESCRIPTION OF THE DRAWINGS bandwidths from 3.2 GB / s (Pentium 4 with AGTL + , Athlon\n XP with EVO ) to around 15 GB / s (Xeon Woodcrest with\n In the accompanying drawings, reference characters refer AGTL + , Athlon 64 /Opteron with Hypertransport), while the\nto the same parts throughout the different views. The draw- 35 local bus has maximal peak data transfer rates of 4 GB / s\nings are not necessarily to scale ; emphasis has instead been (PCI Express 16 ) or 2 GB / s ( PCI -X 2.0 ) . Thus the local bus\nplaced upon illustrating the principles of the invention . Of 190 becomes a bottleneck in the information exchange\nthe drawings: between the system bus 200 and the expansion card 180. The\n FIG . 1 is a schematic diagram illustrating a computer design of the expansion card and methods proposed herein\nsystem including the GPU accelerator according to an 40 minimizes the data transfer through the local bus 190 to\nembodiment of the present invention ; reduce the effect of this bottleneck .\n FIG . 2 is block diagram illustrating the architecture for the The system memory 130 is referred to as the main\nGPU accelerator according to an embodiment of the present random - access memory (RAM ) in the description herein .\ninvention; However, this is not intended to limit the system memory\n FIG . 3 is a block / flow diagram illustrating an exemplary 45 130 to only RAM technology. Other possible computer\nimplementation of the top level control of the GPU accel- storage media include , but are not limited to ROM ,\nerator system ; EEPROM , flash memory, or any other memory technology.\n FIG . 4 is a flow diagram illustrating an exemplary imple- In the illustrated example, the GPU accelerator system is\nmentation of the bottom level control of the GPU accelerator implemented on an expansion card 180 on which the one or\nsystem that is used to execute the target computation ; and 50 more GPU's 240 are mounted . It should be noted that the\n FIG . 5 is an example population of nine computational GPU accelerator system GPU 240 is separate from and\nelements arranged in a 3x3 square and a potential packing independent of any GPU on the standard video card 150 or\nscheme for texture pixels , according to an implementation of other video driving hardware such as integrated graphics\nthe present invention . systems. Thus the computations performed on the expansion\n 55 card 180 do not interfere with graphics display ( including\n DETAILED DESCRIPTION but not limited to manipulation and rendering of images ) .\n Various brand of GPU are relevant. Under current tech\n FIG . 1 shows a computer system 100 that has been nology, GPU's based on the GeForce series from NVIDIA\nconstructed according to the principles of the present inven- Corporation or the Catalyst series from ATI/ Advanced\ntion . 60 Micro Devices, Inc.\n In more detail, the computer system 100 in one example The output to a video monitor 170 is preferably through\nis a standard personal computer ( PC ) . However, this only the video card 150 and not the GPU accelerator system 180 .\nserves as an example environment as computing environ- The video card 150 is dedicated to the transfer of graphical\nment 100 does not necessarily depend on or require any information and connects to the motherboard 110 through a\ncombination of the components that are illustrated and 65 local bus 160 that is sometimes physically separate from the\ndescribed herein . In fact, there are many other suitable local bus 190 that connects the expansion card 180 to the\ncomputing environments for this invention, including, but motherboard 110 .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 14 of 22\n\n\n US RE48,438 E\n 5 6\n FIG . 2 is a block diagram illustrating the general archi- not require hardware implementation. Also the partitioning\ntecture of the GPU accelerator system and specifically the scheme is also altered based on new designs or needs of the\nexpansion card 180 in which at least one GPU 240 and algorithms being employed. The reason for this partitioning\nassociated memories 210 and 250 are mounted . Electrical is further explained in the Data Organization section, below .\n( signal) and mechanical coupling with a local bus 190 5 A local bus interface 230 on the controller 220 serves as\nprovides signal coupling compatible with the PCI industry a driver that allows the controller 220 to communicate\nstandards ( this includes but is not limited to PCI , PCI -X , PCI through the local bus 190 with the system bus 200 and thus\nExpress, or functionally similar technology ). the CPU 120 and RAM 130. This local bus interface 230 is\n The GPU accelerator further preferably comprises one not intended to be limited to PCI related technology. Other\nspecifically designed accelerator controller 220. Depending 10 drivers can be used to interface with comparable technology\nupon the implementation , the accelerator controller 220 is as a local bus 190 .\nfield programmable gate array ( FPGA ) logic , or custom built Data Organization\napplication - specific (ASIC ) chip mounted in the expansion Each computational element discussed above has output\ncard 180 , and in mechanical and signal coupling with the variables that affect the rest of the system . For example in\nGPU 240 and the associated memories 210 and 250. During 15 the case of a neural network it is the output of a neuron . A\ninitial design , a controller can be partially or even fully computational element also usually has several internal\nimplemented in software, in one example. variables that are used to compute output variables , but are\n The controller 220 commands the storage and retrieval of not exposed to the rest of the system , not even to other\narrays of data ( on a conventional video card the arrays of elements of the same population, typically. Each of these\ndata are represented as textures, hence the term ' texture' in 20 variables is represented as a texture . The important differ\nthis document refers to a data array unless specified other- ence between output variables and internal variables is their\nwise and each element of the texture is a pixel of color access .\ninformation ), execution of GPU programs (on a conven- Output variables are usually accessed by any element in\ntional video card these programs are called shaders , hence the system during every time step . The value of the output\nthe term \u201c shader ' in this document refers to a GPU program 25 variable that is accessed by other elements of the system\nunless specified otherwise ), and data transfer between the corresponds to the value computed on the previous, not the\nsystem bus 200 and the expansion card 180 through the local current, time step . This is realized by dedicating two textures\nbus 190 which allows communication between the main to output variables \u2014 one holds the value computed during\nCPU 120 , RAM 130 , and disk 140 . the previous time step and is accessible to all computational\n Two memory banks 210 and 250 are mounted on the 30 elements during the current time step , another is not acces\nexpansion card 180. In some example, these memory banks sible to other elements and is used to accumulate new values\nseparated in the hardware, as shown, or alternatively imple- for the variable computed during the current time step .\nmented as a single , logically partitioned memory compo- In -between time steps these tw te res are switched , so\nnent. that newly accumulated values serve as accessible input\n The reason to separate the memory into two partitions 210 35 during the next time step , while the old input is replaced with\n250 stems from the nature of the computations to which the new values of the variable. This switch is implemented by\nGPU accelerator system is applied . The elements of com- swapping the address pointers to respective textures as\nputation ( computational elements ) are characterized by a described in the System and Framework section .\nsingle output variable . Such computational elements often Internal variables are computed and used within the same\ninclude one or more equations. Computational elements are 40 computational element. There is no chance of a race con\nsame or similar within a large population and are computed dition in which the value is used before it is computed or\nin parallel. An example of such a population is a layer of after it has already changed on the next time step because\nneurons in an artificial neural network (ANN ), where all within an element the processing is sequential. Therefore, it\nneurons are described by the same equation. As a result, is possible to render the new value of internal variable into\nsome data and most of the algorithms are common to all 45 the same texture where the old was read from in the texture\ncomputational elements within population , while most of the memory bank . Rendering to more than one texture from a\ndata and some algorithms are specific for each equation . single shader is not implemented in current GPU architec\nThus, one memory , the shader memory bank 210 , is used to tures, so computational elements that track internal variables\nstore the shaders needed for the execution of the required would have to have one shader per variable . These shaders\ncomputations and the parameters that are common for all 50 can be executed in order with internal variables computed\ncomputational elements and is coupled with the controller first, followed by output variables .\n220 only. The second memory, the texture memory bank Further savings of texture memory is achieved through\n250 , is used to store all the necessary data that are specific using multiple color components per pixel ( texture element)\nfor every computational element (including, but not limited to hold data . Textures can have up to four color components\nto , input data , output data , intermediate results, and param- 55 that are all processed in parallel on a GPU . Thus, to\neters ) and is coupled with both the controller 220 and the maximize the use of GPU architecture it is desirable to pack\nGPU 240 . the data in such a way that all four components are used by\n The texture memory bank 250 is preferably further par- the algorithm . Even though each computational element can\ntitioned into four sections . The first partition 250 a is have multiple variables, designating one texture pixel per\ndesigned to hold the external input data patterns. The second 60 element is ineffective because internal variables require one\npartition 250b is designed to hold the data textures repre- texture and output variables require two textures . Further\nsenting internal variables. The third partition 250c is more , different element types have different numbers of\ndesigned to hold the data textures used as input at a variables and unless this number is precisely a multiple of\nparticular computation step on the GPU 240. The fourth four, texture memory can be wasted .\npartition 250d holds the data textures used to accommodate 65 A more reasonable packing scheme would be to pack four\nthe output of a particular computational step on the GPU computational elements into a pixel and have separate\n240. This partitioning scheme can be done logically , does textures for every variable associated with each computa\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 15 of 22\n\n\n US RE48,438 E\n 7 8\ntional element. In this case the packing scheme is identical The crucial feature of the interaction between the User\nfor all textures, and therefore can be accessed using the same Interaction Stream 302 and the Computational Stream 303 is\nalgorithm . Several ways to approach this packing scheme the shift of priorities. Outside of the simulation , the system\nare outlined here. An example population of nine computa- 100 is driven by the user input, thus the User Interaction\ntional elements arranged in a 3x3 square (FIG . 5a ) can be 5 Stream 302 has the priority and controls the data exchange\npacked by element (FIG . 5b ) , by row (FIG . 5c ) , or by square 304 between streams. After the user starts the simulation , the\n( FIG . 5d) . Computational Stream 303 takes the priority and controls\n Packing by element ( FIG . 5b ) means that elements 1,2,3,4 the data exchange between streams until the simulation is\n finished or interrupted 350 .\ngo into first pixel ; 5,6,7,8 go into second pixel ; 9 goes into\nthird pixel . This is the most compact scheme , but not 10 an The user starts 300 the framework through the means of\nconvenient because the geometrical relationship is not pre theoperating system and interacts with the software through\nserved during packing and its extraction depends on the size 306user interaction section 305 of the graphic user interface\n executed on the CPU 120. The start 300 of the imple\nof the population .\n Packing by row ( column; FIG . 5c ) means that elements 15 initializationbegins\n mentation with a user action that causes a GUI\n 307 , Disk input /output initialization 308 on the\n1,2,3 go into pixel ( 1,1 ) ; 3,4,5 go into pixel (2,1 ) , 7,8,9 go CPU 120 , and controller initialization 320 of the GPU\ninto pixel ( 3,1 ) . With this scheme the element\u2019s y coordinate accelerator on the expansion card 180. GUI initialization\nin the population is the pixel's y coordinate, while the includes opening of the main application window and setting\nelement\u2019s x coordinate in the population is the pixel's x the interface tools that allow the user to control the frame\ncoordinate times four plus the index of color component. 20 work . Disk I/O initialization can be performed at the start of\nFive by five populations in this case will use 2x5 texture, or the framework , or at the start of each individual simulation .\n10 pixels . Five of these pixels will only use one out of four The user interaction 305 controls the setting and editing of\ncomponents , so it wastes 37.5 % of this texture. 25x1 popu- the computational elements, parameters, and sources of\nlation will use 6x1 texture ( six pixels ) and will waste 12.5 % external inputs. It specifies which equations should have\nof it . 25 their output saved to disk and / or displayed on the screen . It\n Packing by square ( FIG . 5d ) means that elements 1,2,4,5 allows the user to start and stop the simulation . And it\ngo into pixel ( 1,1 ) ; 3,6 go into pixel ( 1,2 ) ; 7,8 go into pixel performs standard interface functions such as file loading\n( 2,1 ) , and 9 goes into pixel (2,2 ) . Both the row and the and saving , interactive help , general preferences and others.\ncolumn of the element are determined from the row (col- The user interaction 305 directs the CPU 120 to acquire\numn ) of the pixel times two plus the second ( first) bit of the 30 the new external input textures needed (this includes but is\ncolor component index . Five by five populations in this case not limited to loading from disk 140 or receiving them in\nwill use 3x3 texture , or 9 pixels . Four of these pixels will real time from a recording device ), parses them if necessary\nonly use out of four components, and one will only use 309 , and initializes their transfer the expansion card 180 ,\none component, so it wastes 34.4 % of this texture . This is where they are stored 325 in the texture memory bank 250\nmore advantageous than packing by row , since the texture is 35 by the controller 220. The user interaction 305 also directs\nsmaller and the waste is also lower. 25x1 population on the the CPU 120 to parse populations of elements that will be\nother hand will use 13x1 texture ( thirteen pixels ) and waste used in the simulation, convert them to GPU programs\n> 50 % of it , which is much worse than packing by row . ( shaders ) , compile them 310 , and initializes their transfer to\n In order to eliminate waste altogether the population the expansion card 180 , where they are stored 326 in the\nshould have even dimensions in the square packing, and it 40 shader memory bank 210 by the controller 220. This opera\nshould have a number of columns divisible by four in row tion is accompanied by the upload 309 of the initial data into\npacking. Theoretically, the chances are approximately the input partition of the texture memory bank 250 , and\nequivalent for both of these cases to occur, so the particular stores the shader order of execution in the controller 220 .\ntask and data sizes should determine which packing scheme The user can perform operations 309 and 310 as many times\nis preferable in each individual case . 45 as necessary prior to starting the simulation or between\n The System and Framework simulations .\n FIG . 3 shows an exemplary implementation of the top The editing of the system between simulations is difficult\nlevel system and method that is used to control the compu- to accomplish without the hardware implementation of the\ntation . It is a representation of one of several ways in which computational thread suggested herein . The system of equa\na system and method for processing numerical techniques 50 tions ( computational elements) is represented by textures\ncan be implemented in the invention described herein and so that track variables plus shaders that define processing\nthe implementation is not intended to be limited to the algorithms. As mentioned above , textures, shaders and other\nfollowing description and accompanying figure . graphics related constructs can only be initialized within the\n The method presented herein includes two execution rendering context, which is thread specific . Therefore tex\nstreams that run on the CPU 120 - User Interaction Stream 55 tures and shaders can only be initialized in the computa\n302 and Data Output Stream 301. These two streams pref- tional thread .\nerably do not interact directly, but depend on the same data Network editing is a user - interactive process, which\naccumulated during simulations . They can be implemented according to the scheme suggested above happens in the\nas separate threads with shared memory access and executed User Interaction Stream 302. The simulation software thus\non different CPUs in the case of multi -CPU computing 60 has to take the new parameters from the User Interaction\nenvironment. The third execution stream \u2014 Computational Stream 302 , communicate them to the Computational\nStream 303 runs on the GPU accelerator of the expansion Stream 303 and regenerate the necessary shaders and tex\ncard 180 and interacts with the User Interaction Stream 302 tures . This is hard to accomplish without a hardware imple\nthrough initialization routines and data exchange in between mentation of the Computational Stream 303. The Compu\nsimulations. The Computational Stream 303 interacts with 65 tational Stream 303 is forked from the User Interaction\nthe User Interaction Stream and the Data Output Stream Stream and it can access the memory of the parent thread,\nthrough synchronization procedures during simulations. but the reverse communication is harder to achieve. The\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 16 of 22\n\n\n US RE48,438 E\n 9 10\ncontroller 220 allows operations 309 and 310 to be per- timestep ( ) , TSimulator::outfileInterval ( ), and TSimulator::\nformed as many times as necessary by providing the nec- outmode ( ) , the application can set the time step of the\nessary communication to the User Interaction Stream 302 . simulation , the time step of disk output, and the mode of the\n After execution of the input parser texture generation 309 5 disk output. The external input pattern should be packed into\nand population parser shader generator and compiler 310 are a TPattern object and bound to the simulation object through\nperformed at least once , the user has the option to initialize TSimulator:: resetInputs( ) . method . TSimulator::\nthe simulation 311. During this initialization the main con simLength ( ) sets the length of the simulation .\ntrol of the framework is transferred to the GPU accelerator The second step is to create at least one population of\nsystem's accelerator controller 220 and computation 330 is equations ( Tpopulation object ). Population holds one equa\nstarted\ninterrupt(see\n the FIG . 4; 420, change\n simulation ). The the\n userinput\n retains\n , or the abilitythe\n to change to 10 tion object TEquation. This object contains only a formula\ndisplay properties of the framework , but these interactions the and does not hold element- specific data , so all elements of\nare queued to be performed at times determined by the population can share single TEquation .\ncontroller - driven data exchange 314 and 316 to avoid the before execution The TEquation object is converted to a GPU program\ncorruption of the data . 15 . GPU programs have to be executed within\n The progress monitor 312 is not necessary for perfor creates this context , within\n a graphical context which is stream specific . TSimulator\nmance , but adds convenience. It displays the percentage of fore all programs and dataa arrays Computational Stream , there\ncompleted time steps of the simulation and allows the user computation have to be initialized that within\n are necessary for\n Computational\nto plan the schedule using the estimates of the simulation Stream . Constructor of TPopulation is called from User\nwall clock times . Controller - driven data exchange 314 20 Interaction\nupdates the display of the results 313. Online screen output tialized in this constructor .\n Stream , so no GPU - related objects can be ini\nfor the user selected population allows the user to monitor\nthe activity and evaluate the qualitative behavior of the to TPopulation overcome\n :: fillElements ( ) is a virtual method designed\n this difficulty. It is called from within the\nnetwork . Simulations with unsatisfactory behavior can be Computational Stream after TSimulator ::user\n networkCreate ( ) is\nterminated early to change parameters and restart . Control- 25 called in the User Interaction Stream\nler - driven data exchange 314 also drives the output of the TPopulation :: fillElements ( ) to create TEquation and other\n . A has to override\nresults to disk 317. Data output to disk for convenience can computation related objects both element independent and\nbe done on an element per file basis . A suggested file format element-specific. Element independent objects include sub\nincludes a leftmost column that displays a simulated time for\neach of the simulation steps and subsequent columns that 30 components of TEquation and\n handle interdependencies objectsvariables\n between that describe how to\n implemented\ndisplay variable values during this time step in all elements through derivatives of TGate class .\nwith identical equations (e.g. all neurons in a layer of a Element - specific data is held in TElement objects. These\nneural network ).\n Controller -driven data exchange or input parser texture objects. Therereferences\n objects hold to TEquation and a set of TGate\n is one TElement per population, but the size\ngenerator 316 allows the user to change input that is gen- 35 of data arrays within this object corresponds to population\nerated on the fly during the simulation . This allows the size . All TElement objects have to be added to the TSimu\nframework monitoring of the input that is coming from a lator list of elements by calling TSimulator:: addUnit ( )\nrecording device ( video camera, microphone, cell recording method from TPopulation :: fillElements ( ).\nelectrode, etc ) in real time . Similar to the initial input parser Finally, TPopulation :: fillElements ( ) should contain a set\n309\ndata, itarray\n preprocesses\n suitable the\n for input\n textureintogeneration\n a universaland\n format of the 40 of TElement:: add* Dependency ( ) calls for each element.\n generates\ntextures . Unlike the initial parser 309 , here the textures are every Each of these calls sets a corresponding dependency for\ntransferred to hardware not whenever ready but upon the pendentTGatepart object. Here TGate object holds element inde\n of dependency and TElement::\nrequest of the controller 220 . add * Dependency sets element-specific details.\n The controller 220 also drives the conditional testing 315 45 System provided TPopulation handles the output of com\nand 318 informs the CPU - bound streams whether the simu putational elements, both when they need to exchange the\nlation is finished . If so , the control returns to the User data and when they need to output it to disk . User imple\nInteraction Stream . The user then can change parameters or\ninputs ( 309 and 310 ) , restart the simulation (311 ) or quit the mentation of TPopulation derivative can add screen output.\n Listing 1 is an example code of the user program that uses\nframework (390 ) . 50\n a recurrent competitive field (RCF ) equation:\n SANNDRA ( Synchronous Artificial Neuronal Network\nDistributed Runtime Algorithm ; http://www.kinness.net/\nDocs /SANNDRA /html) was developed to accelerate and Listing 1\noptimize processing of numerical integration of large non\nhomogenous systems of differential equations. This library 55 uint16_t static floatw m_compet\n = 3 , h = 3 ; = 0.5 ;\nis fully reworked in its version 2.x.x to support multiple static float m_persist = 1.0 ;\ncomputational backends including those based on multicore class TCablePopRCF : public TPopulation\nCPUs , GPUs and other processing systems . GPU based {TEq_RCF * m_equation ;\nbackend for SANNDRA -2.x.x can serve as an example TGate * m_gatel;\npractical software implementation of the method and archi- 60 TGate * m_gate2;\ntecture described above and pictorially represented in FIG . void createGatingStructure( )\n3. {\n To use SANNDRA, the application should create a m_gate2 m_gatel = new TGate ( 0 );\nTSimulator object either directly or through inheritance . } ; = new TGate ( 1 ) ;\nThis object will handle global simulation properties and 65 void createUnitStructure ( TBasicUnit* u )\ncontrol the User Interaction Stream , Data Output Stream , {\nand Computational Stream . Through TSimulator ::\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 17 of 22\n\n\n US RE48,438 E\n 11 12\n -continued expensive operation , computationally, but since the textures\n are referred to by texture IDs ( pointers ), swapping these\n Listing 1 pointers for input and output textures after each time step\nu-> addO20PInputDependency (m_gatel, O. , 0. , 0.004 , 0. , 0 , 0 ) ; achieves the same result at a much lesser cost .\n 5\nu- > addFullDependency (m_gate2, population ( ) ); In the hardware solution suggested herein , ID swapping is\n} equivalent to swapping the base memory address for two\npublic : TCablePopRCF ( ) : TPopulation ( \" compCPU RCF \u201d , w , h , true) { } ;\n-TCablePopRCFO ) { if (m_equation ) delete m_equation ; partitions of the texture memory bank 250. They are\n if (m_gatel) delete m_gatel;\n if (m_gate2) delete m_gate2 ; } ;\n swapped 485 during synchronization ( 485 , 430 , and 455 ) so\nbool fillElements( TSimulator * sim ) ; 10 that data transfer 445 and the computation 435-487 proceeds\n}; immediately and in parallel with data transfer as shown in\nbool TCablePopRCF :: fillElements ( TSimulatior* sim) FIG . 4. A hardware solution allows this parallelism through\n{ access of the controller 220 to the onboard texture memory\nm_equation = new TEQ_RCF (this, m_compet, m_persist );\ncreateGatingStructure ( ) ; bank 250 .\nfor( size_t i = 0 ; i < xSize ( ) ; ++ i ) 15\n The main computation and data exchange are executed by\n for(size_t j = 0 ; j < ySize ( ) ; ++ i)\n{ the controller 220. It runs three parallel substreams of\nTElement * u = new TCPUElement(this , m_equation, i , j ) ; execution : Computational Substream 403 , Data Output Sub\nsim-> addUnit( u ); stream 402 , and Data Input Substream 404. These streams\ncreateUnitStructure ( u );\n} 20 are synchronized with each other during the swap of pointers\nReturn true ; 485 to the input and output texture memory partitions of the\n} texture memory bank 250 and the check for the last iteration\nint\nmain ( ) 487. Algorithmically, these two operations are a single\n{ atomic operation, but the block diagram shows them as two\n// Input pattern generation ( 309 in FIG.3 ) 25 separate blocks for clarity.\nuint32_t * pat = new uint32_t [ w * h ];\nTRandom < float > randGen (0 ) ; The Computational Substream 403 performs a computa\nfor (uint32_t I = 0 ; I < w * h ; ++ i )\npat [i] = randGen.random ( ) ; tional cycle including a sequential execution of all shaders\nTpattern * p = new Tpattern (pat, w, h ) ; that were stored in the shader memory bank 210 using the\n// Setting up the simulation\n 30\n appropriate input and output textures. To begin the simula\nTSimulator * cableSim = new TSimulator ( \"data \" ) ; // ( 308 and 320 in tion the controller 220 initializes three execution substreams\nFIG. 3)\ncableSim- >timestep ( 0.05 ) ; // (320 in FIG . 3 ) 403 , 402 , and 404. On every simulation step , the Compu\ncableSim-> resetInputs (p ); // (325 in FIG . 3 ) tational Substream 403 determines which textures the GPU\ncableSim- > outfileInterval(0.1 ); // (308 in FIG . 3 ) 240 will need to perform the computations and initiates the\ncableSim- > outmode (SANNDRA ::timefunc ); // ( 308 in FIG . 3 )\ncableSim- > simLength (60.0 ); // (320 in FIG . 3 ) 35 upload 435 of them onto the GPU 240. The GPU 240 can\n// Preparing the population communicate directly with the texture memory bank 250 to\nTPopulation * cablePop new TCablePopRCFO ); // (310 in FIG . 3 ) upload the appropriate texture to perform the computations .\ncableSim- > networkCreate ( ); // (326 in FIG . 3 ) The controller 220 also pulls the first shader (known by the\nuint16_t user = 1 ;\nwhile (user) stored order ) from the shader memory bank 210 and uploads\n{ 40 450 it onto the GPU 240 .\nif (! cableSim-> simulationStart ( true, 1 ) ) // ( 311 in FIG . 3 )\nexit ( 1 ) ; The GPU 240 executes the following operations in this\nstd ::cout << \" Repeat ? \\ n \" ; // (305 in FIG . 3 )\nstd :: cin >> user; // ( 305 in FIG . 3 )\n order : performs the computation ( execution of the shader )\nif (user 1 ) 470 ; tells the controller 220 that it is done with the compu\ncableSim- > networkReset ( ); // ( 305 in FIG . 3 ) tations for the current shader; and after all shaders for this\n 45 particular equation are executed sends 480 the output tex\n{\nIf (cableSim ) tures to the output portion of the texture memory bank 250 .\nDelete cableSim ; // Also deletes cablePop and its internals This cycle continues through all of the equations based on\nexit (0 ) ;\n}; the branching step 482 .\n 50 An example shader that performs fourth order Runge\n FIG . 4 is a detailed flow diagram illustrating a part of an Kutta numerical integration is shown in Listing 2 using\nexemplary implementation of the bottom level system and GLSL notation ;\nmethod performed during the computation on the GPU\naccelerator of the expansion card 180 and is a more detailed Listing 2\nview of the computational box 330 in FIG . 3. FIG . 4 is a 55\nrepresentation of one of several ways in which a system and uniform sampler2DRect Variable ;\nmethod for processing numerical techniques can be imple uniform float integration_step ;\n float halfstep = integration_step * 0.5 ;\nmented .\n With systems of equations that have complex interdepen float fl_6step = integration_step / 6.0 ;\n vec4 output = texture2DRect (Variable, gl_TexCoord [0 ] .st );\ndencies it is likely that the variable in some equation from 60 // define equation here\na previous time step has to be used by some other equation vec4 rungekutta4 ( vec4 x )\nafter the new values of this variable are already computed {\n const vec4 kl = equation ( x );\nfor new time step . To avoid data confusion , the new values const vec4 k2 equation ( x + halfstep * kl ) ;\nof variables should be rendered in a separate texture. After const vec4 k3 equation ( x + halfstep * k2 );\nthe time step is completed for all equations, these new values 65 const vec4 k4 = equation ( x + integration step * k3 ) ;\nshould be copied over old values so that they are used as return fl_6step * (kl + 2.0 * (k2 + k3 ) + k4 );\ninput during the next time step . Copying textures is an\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 18 of 22\n\n\n US RE48,438 E\n 13 14\n -continued CONCLUSION\n Listing 2 This GPU accelerator system offers the following poten\n }\n tial advantages:\n Void main (void ) 5 1. Limited computations on the CPU 120. The CPU 120\n { is only used for user input, sending information to the\n output + = rungekutta4 (output );\n gl_FragColor = output;\n controller 220 , receiving output after each computational\n } cycle ( or less frequently as defined by the user ), writing this\n output to disk 140 , and displaying this output on the monitor\n 10 170. This frees the CPU 120 to execute other applications\n The shader in Listing 2 can be executed on conventional and allows the expansion card to run at its full capacity\nvideo card . Using the controller 220 this code can be further without being slowed down by extensive interactions with\noptimized , however. Since the integration step does not the CPU 120 .\nchange during the simulation , the step itself as well as the 2. Minimizing data transfer between the expansion card\nhalfstep and % of the step can be computed once per 15 180 and the system bus 200. All of the information needed\nsimulation , and updated in all shaders by a shader update to perform the simulations will be stored on the expansion\nprocedures 310 , 326 discussed above . card 180 and all simulations will take place on it . Further\n After all of the equations in the computational cycle are more , whatever data transfer remains necessary will take\ncomputed the main execution substream 403 on the control place in parallel with the computation , thus reducing the\nler 220 can switch 485 the reference pointers of the input and 20 impact\n 3. New of this\n waytransfer on the\n to execute GPUperformance\n programs (. shaders ). Previ\noutput portions of the texture memory bank 250 . ously, the CPU 120 had full control over the order of\n The two other substreams of execution on the controller shader's execution and was required to produce specific\n220 are waiting ( blocks 430 and 455 , respectively) for this commands on every cycle to tell the GPU 240 which shader\nswitch to begin their execution . The Data Input Substream 25 to use . With the invention disclosed herein , shaders will\n404 is controlling 440 the input of additional data from the initially be stored on the shader memory bank 210 on the\nCPU 120. This is necessary in cases where the simulation is expansion card 180 and will be sent to the GPU 240 for\nmonitoring the changing input, for example input from a execution by the general purpose controller 220 located on\nvideo camera or other recording device in the real time . This the expansion card .\nsubstream uploads new external input from the CPU 120 to 30 4. Multiple parallelisms . The GPU 240 is inherently\nthe texture memory bank 250 so it can be used by the main parallel and is well suited to perform parallel computations.\ncomputational substream 403 on the next computational step In parallel with the GPU 240 performing the next calcula\nand waits for the next iteration 475. The Data Output tion , the controller 220 is uploading the data from the\nSubstream 445 controls the output of simulation results to previous calculation into main memory 130. Furthermore,\nthe CPU 120 if requested by the user . This substream 35 the CPU 120 at the same time uses uploaded previous results\nuploads the results of the previous step to the main RAM to save them onto disk 140 and to display them on the screen\n130 so that the CPU 120 can save them on disk 140 or show through the system bus 200 .\nthem on the results display 313 and waits for the next 5. Reuse of existing and affordable technology. All hard\niteration 460 . ware used in the invention and mentioned here - in are based\n Since the Computational Substream 403 determines the 40 on currently available and reliable components. Further\ntiming of input 440 and output 445 data transfers, these data advance of these components will provide straightforward\ntransfers are driven by the controller 220. To further reduce improvements of the invention .\nthe data transfer overhead ( and disk 140 overhead also ) the While this invention has been particularly shown and\ncontroller 220 initiates transfer only after selected compu- described with references to preferred embodiments thereof,\ntational steps . For example, if the experimental data that is 45 it will be understood by those skilled in the art that various\nsimulated was recorded every 10 milliseconds (msec ) and changes in form and details may be made therein without\nthe simulation for better precision was computed every 1 departing from the scope of the invention encompassed by\nmsec , then only every tenth result has to be transferred to the appended claims .\nmatch the experimental frequency.\n This solution stores two copies of output data , one in the 50 What is claimed is :\nexpansion card texture memory bank 250 and another in the 1. \u00c0 computer system , comprising:\nsystem RAM 130. The copy in the system RAM 130 is a central processing unit to receive input data ;\naccessed twice : for disk I/O and screen visualization 313. An main memory , operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a unit via a bus , to store the input data received by the\ndirect read access to the onboard texture memory bank 250 55 central processing unit;\nby mapping the memory of the hardware onto a global an accelerator, operably coupled to the central processing\nmemory space . The alternative solution will double the unit and the [first] main memory via the bus , to receive\ncommunication through the local bus 190. Since the goal at least a portion of the input data from the main\ndiscussed herein is reducing the information transfer through memory , the accelerator comprising:\nthe local bus 190 , the former solution is favored . 60 at least one graphics processing unit to perform a\n The main substream 403 determines if this is the last sequence of computations on the at least a portion of\niteration 487. If it is the last iteration , the controller 220 the input data so as to generate output data , the\nwaits for the all of the execution substreams to finish 490 sequence of computations representing an artificial\nand then returns the control to the CPU 120 , otherwise it neural network, intermediate computations in the\nbegins the next computational cycle . 65 sequence of computations representing respective\n This repeats through all of the computational cycles of the layers of the artificial neural network and yielding\nsimulation . intermediate results; and\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 19 of 22\n\n\n US RE48,438 E\n 15 16\n accelerator memory , operably coupled to the [ graphic ] system comprising a central processing unit (CPU) , a main\n at least one graphics processing unit, to store the memory operably coupled to the central processing unit via\n results of the [ plurality of sequential] sequence of a bus , an accelerator operably coupled to the CPU and the\n computations; and main memory via the bus , the accelerator comprising a\n a controller, operably coupled to the at least one graphics 5 graphics processing unit (GPU) and an accelerator memory,\n processing unit and the accelerator memory, to initial- the method comprising:\n ize textures and shaders in the accelerator memory for ( A ) performing, by the GPU , the sequence of computa\n performing the sequence of computations, to control tions on a first portion of [ the] input data so as to\n performance of the sequence of computations by the at generate a first portion of [the] output data , the first\n least one graphics processing unit, to transfer the at 10 portion of the output data representing an output of a\n least a portion of the input data into the accelerator neuron in a first layer of the artificial neural network,\n memory during performance of the intermediate com intermediate computations in the sequence of compu\n putations in the sequence of computations by the at tations yielding intermediate results , wherein perform\n least one graphics processing unit, and to transfer at ing the sequence of computations on the first portion of\n least a portion of the output data from the accelerator 15 the input data comprises ( i ) assigning an output vari\n memory to the main memory during performance of the\n intermediate computations in the sequence of compu able to a first texture and a second texture , the output\n tations by the at least one [ graphic ] graphics processing variable being included in a first computational ele\n unit . ment of a plurality of computational elements, the\n 2. The computer system of claim 1 , wherein the central 20 plurality of computational elements representing the\nprocessing unit is configured to receive the input data in sequence of computations and ( ii ) accumulating a first\nresponse to a user interaction . value for the output variable in the first texture during\n 3. The computer system of claim 1 , wherein : a first time step ;\n the central processing unit is configured to receive the ( B ) in parallel with performing the sequence of compu\n input data at a first rate ; and 25 tations by the GPU in ( A ), transferring a second portion\n the at least one graphics processing unit is configured to of the input data from the main memory to the accel\n perform the sequence of computations at a second rate erator via the bus ; [ and]\n different than the first rate . ( C ) in parallel with performing the sequence of compu\n 4. The computer system of claim 1 , wherein the main tations by the GPU in ( A ), transferring a second portion\nmemory is configured to store a copy of the output data 30 of the output data from the accelerator memory to the\nstored in the accelerator memory . main memory via the bus, the second portion of the\n 5. The computer system of claim 1 , wherein an output of output data representing an output of a neuron in a\nat least one computation in the sequence of computations second layer in the artificial neural network ; and\nrepresents an output of at least one neuron in an artificial ( D ) performing, by the GPU , the sequence of computa\nneural network . 35 tions on the second portion of the input data , wherein\n 6. The computer system of claim 1 , wherein accelerator performing the sequence of computationson the second\nmemory comprises: portion of the input data comprises ( i ) accumulating a\n a first memory bank to store parameters common to all of second value for the output variable in the second\n the computations in the sequence of computations; and texture during a second time step and ( ii) making the\n a second memory bank to store data specific to at least one 40 first value of the output variable in the first texture\n computation in the sequence of computations. accessible to other computational elements in the plu\n 7. The computer system of claim 1 , wherein the controller rality of computational elements during the second\nis configured to transfer the output data from the accelerator time step.\nmemory to the main memory without transferring any of the 13. The method of claim 12 , further comprising:\nintermediate results from the accelerator memory to the 45 storing the input data in the main memory in response to\nmain memory so as to reduce data transfer via the bus . a user interaction .\n 8. The computer system of claim 1 , wherein the controller 14. The method of claim 12 , further comprising:\nis configured to transfer at least a portion of the output data receiving the input data at a first rate; and\nfrom the accelerator memory to the main memory after the wherein ( A ) comprises performing the sequence of com\nat least one graphics processing unit has begun to perform 50 putations at a second rate different than the first rate .\nanother sequence of computations. [ 15. The method of claim 12 , wherein ( A ) comprises:\n 9. The computer system of claim 8 , wherein the controller generating an output representative of an output of at least\nis configured to initiate transfer of the at least a portion of the one neuron in an artificial neural network .]\ninput data and to transfer the at least a portion of the output 16. The method of claim 12 , wherein (C ) comprises:\ndata in parallel with performance of at least one computation 55 transferring the second portion of the output data from the\nin the other sequence of computations by the at least one accelerator memory to the main memory without trans\ngraphics processing unit . ferring any of the intermediate results of the plurality of\n 10. The computer system of claim 1 , wherein the con sequential computations from the accelerator memory\ntroller is configured to control execution of the sequence of to the main memory so as to reduce data transfer via the\ncomputations by the at least one graphics processing unit . 60 bus .\n 11. The computer system of claim 1 , further comprising: 17. The method of claim 12 , wherein ( C ) comprises :\n at least one of a video camera, a microphone, or a cell transferring the second portion of the output data from the\n recording electrode, operably coupled to the central accelerator memory to the main memory after the GPU\n [ processor) processing unit , to acquire the input data in has begun to perform another sequence of computa\n real time . 65 tions.\n 12. A method of performing a sequence of computations 18. The method of claim 17 , wherein (C ) further com\nrepresenting an artificial neural network on a computer prises:\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 20 of 22\n\n\n US RE48,438 E\n 17 18\n initiating transfer of the second portion of the output data 27. The method of claim 26 , further comprising :\n in parallel with performance of at least one computa- storing, in a second memory partition of the memory, data\n tion in the other sequence of computations. specific to the first computation in the sequence of\n 19. The method of claim 12 , further comprising : computations.\n acquiring the input data in real time with at least one of 5 28. The method of claim 27, further comprising :\n a video camera , a microphone, or a cell recording storing, in the second memory partition , external input\n electrode operably coupled to the CPU . data patterns, representations of internal variables, an\n 20. The method of claim 12 , further comprising : input of the computation in the sequence of computa\n storing parameters common to all of the computations in 10 tions , and the output of the computation in the sequence\n the sequence of computations in a first memory bank in of computations.\n the accelerator memory ; and 29. The method of claim 21 , wherein storing the first\n storing data specific to at least one computation in the output data comprises :\n sequence of computations in a second memory bank in accumulating, in the memory, outputs of computational\n the accelerator memory. 15 elements executed by the GPU in performing the first\n 21. A method of performing a sequence of computations computation in the sequence of computations.\nrepresenting an artificial neural network, the method com- 30. The method of claim 21 , further comprising:\nprising : storing, in the memory, an output of a previous compu\n receiving, at a central processing unit ( CPU ), first input tation in the sequence of computations; and\n data acquired from an external system in real time; 20 accessing, by the GPU , the output of the previous com\n initializing, by a controller operably coupled to a graph- putation during performance of the computation in the\n ics processing unit (GPU ), textures and shaders in a sequence of computations.\n memory opera coupled to the GPU ; 31. The method of claim 21 , wherein performing the first\n transferring the first input data received by the CPU to the computation comprises executing a plurality of computa\n memory operably coupled to the GPU : 25 tional elements representing a layer of neurons in an arti\n performing, by the graphics processing unit (GPU ), a first ficial neural network.\n computation in the sequence of computations on the 32. The method of claim 31 , wherein all neurons in the\n first input data based on the textures and shaders to layer of neurons are described by the same equation .\n generate first output data , computations in the 30 33. The method of claim 21 , further comprising :\n acquiring the second input data with at least one of a\n sequence of computations representing respective lay\n ers of neurons in the artificial neural network , an video camera , a microphone, or a cell recording elec\n trode .\n output of the first computation in the sequence of 34. The method of claim 21 , further comprising :\n computations representing an output of a first neuron in loading the second input data from disk .\n a first layer in the artificial neural network ; 35 35. A system for performing a sequence of computations,\n storing , in the memory operably coupled to the GPU , the the system comprising:\n first input data and the first output data ; and a camera to generate input data in real time ;\n transferring second input data acquired from the external a first memory partition ;\n system in real time into the memory operably coupled a second memory partition operably coupled to the first\n to the GPU after the GPU starts the first computation 40 memory partition ; and\n and before the GPU starts a second computation of the a processing unit , operably coupled to the camera , the\n sequence of computations, an output of the second first memory partition , and the second memory parti\n computation in the sequence of computations repre tion , to perform the sequence of computations on a first\n senting an output of a second neuron in a second layer portion of the input data so as to generate a first\n in the artificial neural network . 45 portion of output data , intermediate computations in\n 22. The method of claim 21 , wherein transferring the the sequence of computations yielding intermediate\nsecond input data comprises transferring the second input results, the first portion of the output data representing\ndata via a bus operably coupled to the CPU . an output of an artificial neural network,\n 23. The method of claim 21 , further comprising: wherein the first memory partition is configured to trans\n transferring the first output data from the memory to 50 fer a second portion of the input data to the second\n another memory during the second computation in the memory partition in parallel with performance the\n sequence of computations. sequence of computations by the processing unit ,\n 24. The method of claim 23, further comprising: wherein the second memory partition is configured to\n storing intermediate results of the sequence of computa 55 transfer a second portion of the output data to the first\n tions in the memory, and memory partition in parallel with performance the\n wherein transferring the first output data from the sequence of computations by the processing unit, and\n wherein the sequence of computations represents the\n memory to the other memory occurs without transfer artificial neural network , each neuron in the artificial\n ring the intermediate results of the sequence of com neural network has an output variable assigned to a\n putations. 60 first texture and a second texture in the memory , the first\n 25. The method of claim 23 , wherein transferring the texture holds a first value of the output variable com\nsecond input data and transferring the first output data puted during a previous time step of the sequence of\noccurs in parallel . computations and accessible to other neurons in the\n 26. The method of claim 21 ,further comprising: neural network during a current time step of the\n storing , in a first memory partition of the memory, param- 65 sequence of computations and the second texture accu\n eters common to all of the computations in the mulates a second value of the output variable computed\n sequence of computations. during the current time step .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 21 of 22\n\n\n US RE48,438 E\n 19 20\n 36. The system of claim 35, wherein the first memory in the sequence of computations representing layers\npartition and the second memory partition are logical par of the neural network and yielding intermediate\ntitions . results ; and\n 37. The system of claim 35, wherein the processing unit is accelerator memory, operably coupled to the at least\ncomprises a graphics processing unit ( GPU ) . 5\n one processing unit, to store the results of the\n 38. The system of claim 35, wherein the processing unit is sequence of computations; and\nconfigured to receive the input data at a first rate and to a controller, operably coupled to the at least one\nperform the sequence of computations at a second rate is processing unit and the accelerator memory, to con\ndifferent than the first rate. trol transfer of the at least a portion of the input data\n 39. The system of claim 35, wherein the second memory 10 into the accelerator memory during performance of\npartition is configured to transfer the second portion of the the intermediate computations in the sequence of\noutput data to the first memory partition without transfer computations by the at least one processing unit, to\nring any of the intermediate results to the first memory control transfer at least a portion of the output data\npartition .\n 40. A system for executing an artificial neural network, 15 from the accelerator memory to the main memory\nthe system comprising: during performance of the intermediate computa\n a central processing unit (CPU ) to provide first input tions in the sequence of computations by the at least\n data ; one processing unit, and to control performance of\n a memory, operably coupled to the CPU , to store the first the sequence of computations by the at least one\n input data in a first partition, referenced by a first 20 processing unit.\n pointer, before computing a first layer of neurons of the 45. The computer system of claim 44, wherein the central\n artificial neural network ; processing unit is configured to receive the input data in\n a processing unit, operably coupled to the memory, to response to a user interaction .\n perform , during computation of the first layer of neu- 46. The computer system of claim 44, wherein :\n rons, at least one calculation on the first input data so 25 the central processing unit is configured to receive the\n as to generate first output data , the first output data input data at a first rate ; and\n representing an output of at least one neuron in the first the at least one processing unit is configured to perform\n layer of neurons ; and the sequence of computations at a second rate different\n a controller, operably coupled to the processing unit and than the first rate.\n the memory, to : 30 47. The computer system of claim 44, wherein the main\n store the first output data in a second partition of the memory is configured to store a copy of the output data\n memory, the second partition referenced by a second stored in the accelerator memory.\n pointer, and to swap the firstpointer with the second 48. The computer system of claim wherein an output\n pointer at the end of the computation of the first layer of at least one computation in the sequence of computations\n of neurons, such that the firstoutput data becomes an 35 represents an output of at least one neuron in an artificial\n input for a second layer of neurons of the artificial neural network .\n neural network, 49. The computer system of claim 44, wherein accelerator\n transfer the first output data to another memory during memory comprises:\n computation of the second layer of neurons, and a first memory partition to store parameters common to\n dictate an order of execution of instructions to the 40 all of the computations in the sequence of computa\n processing unit to perform the computation of the tions ; and\n first layer of neurons . a second memory partition to store data specific to at least\n 41. The system of claim 40 , wherein the processing unit one computation in the sequence of computations.\ncomprises a graphics processing unit . 50. The computer system of claim 44, wherein the con\n 42. The system of claim 40 , wherein the controller is 45 troller is configured to transfer the output data from the\nconfigured to send instructions for performing the at least accelerator memory to the main memory without transfer\none calculation to the processing unit. ring any of the intermediate results from the accelerator\n 43. The system of claim 40 , wherein the memory further memory to the main memory so as to reduce data transfer\ncomprises : via the bus.\n a third partition to store internal variables; and 50 51. The computer system of claim 44, wherein the con\n a fourth partition to store data used as input at a troller is configured to transfer at least a portion of the\n particular layer of neurons of the artificial neural output data from the accelerator memory to the main\n network. memory after the at least one processing unit has begun to\n 44. A computer system , comprising: perform another sequence of computations.\n a central processing unit to receive input data acquired 55 52. The computer system of claim 51 , wherein the con\n from an external system ; troller is configured to initiate transfer of the at least a\n main memory, operably coupled to the central processing portion of the input data and to transfer the at least a portion\n unit via a bus, to store the input data received by the of the output data in parallel with performance of at least\n central processing unit; one computation in the other sequence of computations by\n an accelerator, operably coupled to the central processing 60 the at least one processing unit .\n unit and the main memory via the bus, to receive at 53. The computer system of claim 44, wherein the con\n least a portion of the input data from the main memory , troller is configured to control execution of the sequence of\n the accelerator comprising : computations by the at least one processing unit .\n at least one processing unit to perform a sequence of 54. The computer system of claim 44, further comprising :\n computations representing an artificial neural net- 65 at least one of a video camera , a microphone, or a cell\n work on the at least a portion of the input data so as recording electrode, operably coupled to the central\n to generate output data, intermediate computations processing unit, to acquire the input data in real time .\n\f Case 7:26-mc-00318-LS Document 6-4 Filed 08/18/26 Page 22 of 22\n\n\n US RE48,438 E\n 21 22\n 55. The computer system of claim 1 , wherein the control\nler is configured to inform the central processing unit that\nthe sequence of computations is finished .\n 56. The computer system of claim 1 , wherein the control\nler is configured to reduce a processing load on the central 5\nprocessing unit.\n 57. The computer system of claim 1 , wherein the control\nler is configured to reduce interactions between the central\nprocessing unit and the accelerator.\n 10\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:43.057545-07:00","document_number":"6","attachment_number":4,"pacer_doc_id":"181037220271","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 3","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554425/","id":490554425,"tags":[],"absolute_url":"/docket/74659430/6/5/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.207930-07:00","date_modified":"2026-08-23T04:32:35.102145-07:00","sha1":"4b4bfc72562788dc4415954ce2a86bdd71ae5f7a","page_count":20,"file_size":1057929,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.5.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.5.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 1 of 20\n\n\n\n\n EXHIBIT\n\n 4\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 2 of 20\n I 1111111111111111\n 1111111111111111\n IIIIIIIIIIIII1111111111111111\n Illlllll llll\n US00RE49461E\nc19) United States\nc12) Reissued Patent (10) Patent Number: US RE49,461 E\n Gorchetchnikov et al. (45) Date of Reissued Patent: Mar. 14, 2023\n\n\n(54) GRAPHIC PROCESSOR BASED (56) References Cited\n ACCELERATOR SYSTEM AND METHOD\n U.S. PATENT DOCUMENTS\n(71) Applicant: Neurala, Inc., Boston, MA (US)\n 5,063,603 A 11/1991 Burt\n(72) Inventors: Anatoli Gorchetchnikov, Newton, MA 5,136,687 A 8/1992 Edelman et al.\n (US); Heather Marie Ames, Milton, (Continued)\n MA (US); Massimiliano Versace,\n Milton, MA (US); Fabrizio Santini, FOREIGN PATENT DOCUMENTS\n Jamaica Plain, MA (US) EP 1224622 Bl 11/2004\n WO 190208 11/2014\n(73) Assignee: Neurala, Inc., Boston, MA (US)\n (Continued)\n(21) Appl. No.: 17/136,343\n(22) Filed: Dec. 29, 2020 OTHER PUBLICATIONS\n Related U.S. Patent Documents Hodgkin, A. L., and Huxley, A. F. 1952. Quantitative description of\nReissue of: membrane current and its application to conduction and excitation\n(64) Patent No.: 9,189,828 m nerve. J Physiol 117, pp. 500-544.\n Issued: Nov. 17, 2015 (Continued)\n Appl. No.: 14/147,015\n Primary Examiner - William H. Wood\n Filed: Jan.3, 2014\n (74) Attorney, Agent, or Firm - Smith Baluch LLP\nU.S. Applications:\n(63) Continuation of application No. 15/808,201, filed on (57) ABSTRACT\n Nov. 9, 2017, now Pat. No. Re. 48,438, which is an An accelerator system is implemented on an expansion card\n (Continued) comprising a printed circuit board having (a) one or more\n graphics processing units (GPUs), (b) two or more associ-\n(51) Int. Cl. ated memory banks (logically or physically partitioned), (c)\n G06T 1160 (2006.01) a specialized controller, and (d) a local bus providing signal\n coupling compatible with the PCI industry standards. The\n G06F 9/50 (2006.01) controller handles most of the primitive operations to set up\n (Continued) and control GPU computation. Thus, the computer's central\n(52) U.S. Cl. processing unit (CPU) can be dedicated to other tasks. In this\n CPC .............. G06T 1120 (2013.01); G06F 9/5027 case a few controls (simulation start and stop signals from\n the CPU and the simulation completion signal back to CPU),\n (2013.01); G06T 1160 (2013.01); GPU programs and input/output data are exchanged between\n (Continued) CPU and the expansion card. Moreover, since on every time\n(58) Field of Classification Search step of the simulation the results from the previous time step\n CPC ... G06F 9/5027; G06F 2209/509; G06T 1/20; are used but not changed, the results are preferably trans-\n ferred back to CPU in parallel with the computation.\n G06T 1/60; G06N 3/00; G06N 3/02;\n (Continued) 21 Claims, 5 Drawing Sheets\n\n Expansion Card mo\n\n\n\n\n I ............\n\n\n 41 r:;-\n\n\n r~\u00b7--\u00b7 .tll)\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 3 of 20\n\n\n US RE49,461 E\n Page 2\n\n\n Related U.S. Application Data 2011/0004341 Al 1/2011 Sarvadevabhatla et al.\n 2011/0173015 Al 7/2011 Chapman et al.\n application for the reissue of Pat. No. 9,189,828, 2011/0279682 Al 11/2011 Li et al.\n which is a continuation of application No. 11/860, 2012/0072215 Al 3/2012 Yu et al.\n 2012/0089552 Al 4/2012 Chang et al.\n 254, filed on Sep. 24, 2007, now Pat. No. 8,648,867. 2012/0197596 Al 8/2012 Comi\n(60) Provisional application No. 60/826,892, filed on Sep. 2012/0316786 Al 12/2012 Liu et al.\n 2013/0126703 Al 5/2013 Caulfield\n 25, 2006. 2013/0131985 Al 5/2013 Weiland et al.\n 2014/0019392 Al 1/2014 Buibas et al.\n(51) Int. Cl. 2014/0032461 Al 1/2014 Weng\n G06T 1120 (2006.01) 2014/0052679 Al 2/2014 Sinyavskiy et al.\n G06N 3/063 (2023.01) 2014/0089232 Al 3/2014 Buibas et al.\n 2015/0127149 Al 5/2015 Sinyavskiy et al.\n G06N 20/00 (2019.01) 2015/0134232 Al 5/2015 Robinson\n(52) U.S. Cl. 2015/0224648 Al 8/2015 Lee et al.\n CPC ........ G06F 2209/509 (2013.01); G06N 3/063 2016/0075017 Al 3/2016 Laurent et al.\n (2013.01); G06N 20/00 (2019.01) 2016/0082597 Al 3/2016 Gorshechnikov et al.\n 2016/0096270 Al 4/2016 Gabardos et al.\n(58) Field of Classification Search 2016/0198000 Al 7/2016 Gorshechnikov et al.\n CPC ............ G06N 3/04; G06N 3/06; G06N 3/063; 2017 /0024877 Al 1/2017 Versace et al.\n G06N 3/08; G06N 3/1 O; G06N 5/00; 2017/0076194 Al 3/2017 Versace et al.\n G06N 7/00; G06N 7/02; G06N 7/04; 2017/0193298 Al 7/2017 Versace et al.\n G06N 7/046; G06N 7/06; G06N 20/00\n See application file for complete search history. FOREIGN PATENT DOCUMENTS\n\n WO 2014204615 A2 12/2014\n(56) References Cited WO 2015143173 A2 9/2015\n WO 2016014137 A2 1/2016\n U.S. PATENT DOCUMENTS\n\n 5,142,665\n A 8/ 1992 Bigus OTHER PUBLICATIONS\n 5,172,253\n A 12/1992 Lynne\n 5,388,206\n A 2/1995 Poulton et al. Hopfield, J. 1982. Neural networks and physical systems with\n 6,018,696\n A 1/2000 Matsuoka et al.\n emergent collective computational abilities. In Proc Natl Acad Sci\n 6,336,051\n Bl 1/2002 Pangels et al.\n 6,647,508\n B2 11/2003 Zalewski et al. USA, vol. 79, pp. 2554-2558.\n 7,119,810\n B2 10/2006 Sumanaweera et al. Ilie, A. 2002. Optical character recognition on graphics hardware.\n 7,219,085\n B2 * 5/2007 Buck . G06V 10/955 Tech. Rep. integrative paper, UNCCH, Department of Computer\n 706/12 Science, 9 pages.\n 7,477,256 Bl* 1/2009 Johnson .................... G06F 3/14 International Preliminary Report on Patentability in related PCT\n 345/506 Application No. PCT/US2014/039162 filed May 22, 2014, dated\n 7,525,547 Bl* 4/2009 Diard ........................ G06T 1/20\n Nov. 24, 2015, 7 pages.\n 345/522\n 7,765,029 B2 7/2010 Fleischer et al. International Preliminary Report on Patentability in related PCT\n 7,861,060 Bl* 12/2010 Nickolls et al. .. ... ... .... ... . 712/22 Application No. PCT/US2014/039239 filed May 22, 2014, dated\n 7,873,650 Bl 1/2011 Chapman et al. Nov. 24, 2015, 8 pages.\n 8,392,346 B2 3/2013 Ueda et al. International Preliminary Report on Patentability dated Nov. 8,\n 8,510,244 B2 8/2013 Carson et al. 2016 from International Application No. PCT/US2015/029438, 7\n 8,583,286 B2 11/2013 Fleischer et al. pages.\n 8,648,867 B2 * 2/2014 Gorchetchnikov et al ... 345/501\n International Search Report and Written Opinion dated Feb. 18,\n 9,031,692 B2 5/2015 Zhu\n 9,177,246 B2 11/2015 Bui bas et al. 2015 from International Application No. PCT/US2014/039162, 12\n 9,189,828 B2 11/2015 Gorchetchnikov et al. pages.\n 9,626,566 B2 4/2017 Versace et al. International Search Report and Written Opinion dated Feb. 23,\n 10,083,523 B2 9/2018 Versace et al. 2016 from International Application No. PCT/US2015/029438, 11\n RE48,438 E * 2/2021 Gorchetchnikov ... G06F 9/5027 pages.\n 2001/0010034 Al 7/2001 Burton International Search Report and Written Opinion dated Jul. 6, 2017\n 2002/0046271 Al 4/2002 Huang\n from International Application No.PCT/US2017/029866, 12 pages.\n 2002/0050518 Al 5/2002 Roustaei\n 2002/0064314 Al 5/2002 Comaniciu et al. International Search Report and Written Opinion dated Nov. 26,\n 2002/0168100 Al 11/2002 Woodall 2014 from International Application No. PCT/US2014/039239, 14\n 2003/0026588 Al 2/2003 Elder et al. pages.\n 2003/00787 54 Al 4/2003 Hamza International Search Report and Written Opinion dated Sep. 15,\n 2004/0015334 Al 1/2004 Ditlow et al. 2015 from International Application No. PCT/US2015/021492, 9\n 2005/0166042 Al 7/2005 Evans pages.\n 2006/0129506 Al* 6/2006 Edelman. G05D 1/0088\n Itti, L., and Koch, C. (2001). Computational modelling of visual\n 706/12\n 2006/0184273 Al 8/2006 Sawada et al. attention. Nature Reviews Neuroscience, 2 (3), 194-203.\n 2007/0052713 Al 3/2007 Chung et al. Itti, L., Koch, C., and Niebur, E. (1998). A Model of Saliency-Based\n 2007/0198222 Al 8/2007 Schuster et al. Visual Attention for Rapid Scene Analysis, 1-6.\n 2007/0279429 Al 12/2007 Ganzer Jarrett, K., Kavukcuoglu, K., Ranzato, M. A., & LeCun, Y. (Sep.\n 2008/0033897 Al 2/2008 Lloyd 2009). What is the best multi-stage architecture tor object recogni-\n 2008/0066065 Al 3/2008 Kim et al. tion?. In Computer Vision, 2009 IEEE 12th International Confer-\n 2008/0258880 Al 10/2008 Smith et al.\n ence on (pp. 2146-2153) IEEE.\n 2009/0080695 Al 3/2009 Yang\n 2009/0089030 Al 4/2009 Sturrock et al. Khaligh-Razavi, S.-M et al., Deep Supervised, but Not Unsuper-\n 2009/0116688 Al 5/2009 Monacos et al. vised, Models May Explain IT Cortical Representation, PLoS\n 2010/0048242 Al 2/2010 Rhoads et al. Computational Biology, vol. 10, Issue 11, 29 pages (Nov. 2014).\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 4 of 20\n\n\n US RE49,461 E\n Page 3\n\n\n(56) References Cited Minih, Volodymyr, Kavukcuoglu, Koray, Silver, David, Rusu, Andrei\n A, Veness, Joel, Bellemare, Marc G, Graves, Alex, Riedmiller,\n OTHER PUBLICATIONS Martin, Fidjeland, Andreas K, Ostrovski, Georg, et al. Human-level\n control through deep reinforcement learning. Nature, 518(7540):529-\nKim, S., Novel approaches to clustering, biclustering and algo- 533, Feb. 25, 2015.\nrithms based on adaptive resonance theory and Intelligent control, Montrym et al., The GeForce 6800, in IEEE Micro, vol. 25, No. 2,\nDoctoral Dissertations, Missouri University of Science and Tech- pp. 41-51, March-Apr. 2005.\nnology, 125 pages (2016). Moore, Andrew W and Atkeson, Christopher G. Prioritized sweep-\nKipfer, P., Segal, M., and Westermann, R. 2004. UberFlow: A ing: Reinforcement learning with less data and less time. Machine\nGPU-Based Particle Engine. In Proceedings of the SIGGRAPH/ Learning, 13(1): 103-130,1993.\nEurographics Workshop on Graphics Hardware 2004, pp. 115-122. Najernnik, J., and Geisler, W. (2009). Simple sununation rule for\nKolb, A., L. Latta, and C. RF7K-SALAMA. 2004. \"Hardware- optimal fixation selection in visual search. Vision Research. 49,\nBased Simulation and Collision Detection for Large Particle Sys- 1286-1294.\ntems.\" In Proceedings of the SIGGRAPH/Eurographics Workshop Non-Final Office Action dated Jan. 4, 2018 from U.S. Appl. No.\non Graphics Hardware 2004, pp. 123-131. 15/262,637, 23 pages.\nKompella, Varun Raj, Luciw, Matthew, and Schmidhuber, Jurgen. Non-Final Office Action dated May 31, 2018 from U.S. Appl. No.\nIncremental slow feature analysis: Adaptive low-complexity slow 14/947,516, 16 pages.\nfeature updating from high-dimensional input streams Neural Com- Notice of Alllowance dated May 22, 2018 from U.S. Appl. No.\nputation, 24( 11):2994-3024, 2012. 15/262,637, 6 pages.\nKowler, E. (2011). Eye movements: The past 25years. Vision Notice of Allowance dated Jul. 27, 2016 from U.S. Appl. No.\nResearch, 51(13), 1457-1483. doi:10.1016/j.visres.2010.12.014. 14/662,657.\nLarochelle H., & Hinton G. (2012). Learning to combine foveal Notice of Allowance dated Dec. 16, 2016 from U.S. Appl. No.\nglimpses with a third-order Boltzmann machine. NIPS 2010,1243- 14/662,657.\n1251. Oh, K.-S., and Jung, K. 2004. GPU implementation of neural\nLeCun, Y., Kavukcuoglu, K., & Farabet, C. (May 2010). Convolu- networks. Pattern Recognition 37, pp. 1311-1314.\ntional networks and applications in vision. In Circuits and Systems Oja, E. (1982). Simplified neuron model as a principal component\n(ISCAS), Proceedings of 2010 IEEE International Symposium on analyzer. Journal of Mathematical Biology 15(3), 267-273.\n(pp. 253-256). IEEE. Partial Supplementary European Search Report dated Jul. 4, 2017\nLee, D. D. and Seung, H. S. (1999). Learning the parts of objects from European Application No. 14800348.6, 13 pages.\nby non-negative matrix factorization. Nature, 401(6755):788-791. Perumalla, \"Discrete-event execution alternatives on general pur-\nLee, D. D., and Seung, H. S. (1997). \"Unsupervised learning by pose graphical processing units (GPGPU s).\" Proceedings of the\nconvex and conic coding.\" Advances in Neural Information Pro- 20th Workshop on Principles of Advanced and Distributed Simu-\ncessing Systems, 9. lation. IEEE Computer Society, 2006.8 pages.\nLegenstein, R., Wilbert, N., and Wiskott, L. Reinforcement learning Raijmakers, M.E.J., and Molenaar, P. (1997). Exact Art: A complete\non slow features of high-dimensional input streams. PLoS Compu- implementation of an ART network Neural networks 10 (4), 649-\ntational Biology, 6(8), 2010. ISSN 1553-734X. 13 pages. 669.\nLeveille, J., Ames, H., Chandler, B., Gorchetchnikov, A., Mingolla, Ranzato, M.A., Huang, F. J., Boureau, Y. L., & Lecun, Y. (2007,\nE., Patrick, S., and Versace, M. (2010) Learning in a distributed June). Unsupervised learning of invariant feature hierarchies with\nsoftware architecture for large-scale neural modeling. BIONET- applications to object recognition. In Computer Vision and Pattern\nICS 10, Boston, MA, USA. 8 pages. Recognition, 2007. CVPR'07. IEEE Conference on (pp. 1-8). IEEE.\nLivitz G., Versace M., Gorchetchnikov A., Vasilkoski Z., Ames H., Raudies, F., Eldridge, S., Joshi, A., and Versace, M. (Aug. 20, 2014).\nChandler B., Leveille J. andMingolla E. (2011) Adaptive, brain-like Learning to navigate in a virtual world using optic flow and stereo\nsystems give robots complex behaviors, The Neuromorphic Engi- disparity signals. Artificial Life and Robotics, DOI 10.1007/10015-\nneer,: 10.2417/1201101.003500 Feb. 2011. 3 pages. 014-0153-l. 15 pages.\nLivitz, G., Versace, M., Gorchetchnikov, A., Vasilkoski, Z., Ames, Adelson, E. H, Anderson, C. H, Bergen, JR., Burt, P. J, & Ogden,\nH., Chandler, B., Leveille, J., Mingolla, E., Snider, G., Amerson, R., J. M (1984) Pyramid methods in image processing. RCA engineer,\nCarter, D., Abdalla, H., and Qureshi, S. (2011) Visually-Guided 29(6), 33-41.\nAdaptive Robot (ViGuAR). Proceedings of the International Joint Aggarwal, Charu C, Hinneburg, Alexander, and Keim, Daniel A. On\nConference on Neural Networks (IJCNN) 2011, San Jose, CA, the surprising behavior of distance metrics in high dimensional\nUSA. 9 pages. space. Springer, 2001. 15 pages.\n Al-Kaysi, A. M. et al., A Multichannel Deep Belief Network for the\nLowe, D.G.(2004). Distinctive Image Features from Scale-Invariant\n Classification of EEG Data, from Ontology-based Information\nKeypoints. Journal International Journal of Computer Vision archive\n Extraction for Residential Land Use Suitability: A Case Study of the\nvol. 60, 2, 91-110.\n City of Regina, Canada, DOI 10.1007/978-3-319-26561-2_5, 8\nLu, Z.L., Liu, J., and Dosher, B.A.(2010) Modeling mechanisms of pages (Nov. 2015).\nperceptual learning with augmented Hebbian re-areighting Vision Ames, H, Versace, M., Gorchetchnikov, A., Chandler, B., Livitz, G.,\nResearch, 50(4). 375-390. Leveille, J., Mingolla, E., Carter, D., Abdalla, H., and Snider, G.\nLuo et al., \"Ailificial neural network computation on graphic (2012) Persuading computers to act more like brains. In Advances\nprocess unit.\" Neural Networks, 2005. IJCNN'05. Proceedings in Neuromorphic Mernristor Science and Applications, Kozma,\n2005 IEEE International Joint Conference on vol. 1 IEEE, 2005 pp. R.Pino,R., and Pazienza, G. (eds), Springer Verlag. 25 pages.\n622-626. Ames, H. Mingolla, E., Sohail, A., Chandler, B., Gorchetchnikov,\nMahadevan, S. Proto-value functions: Developmental reinforce- A., Leveille, J., Livitz, G. and Versace, M. (2012) The Animat. IEEE\nment learning. In Proceedings of the 22nd international conference Pulse, Feb. 2012, 3(1), 47-50.\non Machine learning, pp. 553-560. ACM, 2005. Apolloni, B. et al., Training a network of mobile neurons, Proceed-\nMeuth, J.R. and Wunsch, D.C. (2007) A Survey of Neural Compu- ings of International Joint Conference on Neural Networks, San\ntation On Graphics Processing Hardware. 22nd IEEE International Jose, CA, doi: 10.1109/IJCNN.2011.6033427, pp. 1683-1691 (Jul.\nSymposium on Intelligent Control, Part of IEEE Multi-conference 31-Aug. 5, 2011).\non Systems and Control, Singapore, Oct. 1-3, 2007, 5 pages. Artificial Intelligence as a Service. Invited talk, Defrag, Broomfield,\nMishkin M, Ungerleider LG. (1982). \"Contribution of striate inputs CO, Nov. 4-6, 2013. 22 pages.\nto the visuospatial functions of parieto-preoccipital cortex in mon- Aryananda, L. 2006. Attending to learn and learning to attend for a\nkeys,\" Behav Brain Res, 6 (1): 57-77. social robot. Humanoids 06, pp. 618-623.\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 5 of 20\n\n\n US RE49,461 E\n Page 4\n\n\n(56) References Cited Ellias, S. A., and Grossberg, S. 1975. Pattern formation, contrast\n control and oscillations in the short term memory of shunting\n OTHER PUBLICATIONS on-center off-surround networks Biol Cybern 20, pp. 69-98.\n Extended European Search Report and Written Opinion dated Jun.\nBaraldi, A. and Alpaydin, E. ( 1998). Simplified Art: A new class of 1, 2017 from European Application No. 14813864.7, 10 pages.\nArt algorithms. International Computer Science Institute, Berkeley, Extended European Search Report and Written Opinion dated Oct.\nCA, TR-98-004, 1998. 42 pages. 12, 2017 from European Application No. 14800348.6, 12 pages.\nBaraldi, A. and Alpaydin, E. (2002). Constructive feedforward Art Extended European Search Report and Written Opinion Oct. 23,\nclustering networks-Part I. IEEE Transactions on Neural Net- 2017 from European Application No. 15765396.5, 3 pages.\nworks 13(3), 645-661. Faz!, A., Grossberg, S., and Mingolla, E. (2009). View-invariant\nBaraldi, A. and Parmiggiani, F. (1997). Fuzzy combination of object category learning, recognition, and search: How spatial and\nKohonen's and ART neural network models to detect statistical object attention are coordinated using surface-based attentional\nregularities in a random sequence of multi-valued input patterns. In shrouds. Cognitive Psychology 58, 1-48.\nInternational Conference on Neural Networks, IEEE. 6 pages. Fiildiak, P. (1990). Forming sparse representations by local anti-\nBaraldi, Andrea and Alpaydin, Ethem. Constructive feedforward Hebbian learning, Biological Cybernetics, vol. 64, pp. 165-170.\nART clustering networks-part II IEEE Transactions on Neural Friston K., Adams R., Perrinet L., & Breakspear M. (2012). Per-\nNetworks, 13(3):662-677, May 2002. ISSN 1045-9227. doi: 10.1109/ ceptions as hypotheses: saccades as experiments. Frontiers in Psy-\ntnn.2002.1000131. URL http://dx.doi.org/l 0.1109/tnn.2002. chology, 3 (151), 1-20.\n1000131. Galbraith, B.V, Guenther, F.H., and Versace, M. (2015) A neural\nBengio, Y., Courville, A., & Vincent, P. Representation learning: A network-based exploratory learning and motor planning system for\nreview and new perspectives, IEEE Transactions on Pattern Analy- co-robots Frontiers in Neuroscience, in press. 10 pages.\nsis and Machine Intelligence, vol. 35 Issue 8, Aug. 2013. pp. George, D. and Hawkins, J. (2009). Towards a mathematical theory\n1798-1828. of cortical micro-circuits. PLoS Computational Biology 5( 10), 1-26.\nBerenson, D et al., A robot path planning framework that learns Georgii, J., and Westermann, R. 2005. Mass-spring systems on the\nfrom experience, 2012 International Conference on Robotics and GPU. Simulation Modelling Practice and Theory 13, pp. 693-702.\nAutomation, 2012, 9 pages [retrieved from the internet] URL:http:// Gorchetchnikov A., Hasselmo M. E. (2005). A biophysical imple-\nusers. wpi .edu/-dberenson/lightning. pdf. mentation of a bidirectional graph search algorithm to solve mul-\nBernhard, F., and Keriven, R. 2005. Spiking Neurons on GPUs. tiple goal navigation tasks. Connection Science, 17(1-2), pp. 145-\nTech. Rep. 05-15, Ecole Nationale des Ponts et Chauss'es, 8 pages. 166.\nBes!, P. J., & Jain, R. C. (1985). Three-dimensional object recog- Gorchetchnikov A., Hasselmo M. E. (2005). A simple rule for\nnition. ACM Computing Surveys (CSUR), 17(1), 75-145. spike-timing-dependent plasticity: local influence of AHP current.\nBoddapati, V., Classifying Environmental Sounds with Image Net- Neurocomputing, 65-66, pp. 885-890.\nworks, Thesis, Faculty of Computing Blekinge Institute of Tech- Gorchetchnikov A., Versace M., Hasselmo M. E. (2005). A Model\nnology, 37 pages (Feb. 2017). of STDP Based on Spatially and Temporally Local Information:\nBohn, C.-A. Kohonen. 1998. Feature Mapping Through Graphics Derivation and Combination with Gated Decay. Neural Networks,\nHardware. In Proceedings of 3rd Int. Conference on Computational 18, pp. 458-466.\nIntelligence and Neurosciences, 4 pages. Gorchetchnikov A., Versace M., Hasselmo M. E. (2005). Spatially\nBradski, G., & Grossberg, S. (1995). Fast-learning Viewnet archi- and temporally local spike-timing-dependent plasticity rule. In:\ntectures for recognizing three-dimensional objects from multiple Proceedings of the International Joint Conference on Neural Net-\ntwo-dimensional views. Neural Networks, 8 (7-8), 1053-1080. works, No. 1568 in IEEE CD-ROM Catalog No. 05CH37662C, pp.\nCanny, J.A. (1986). Computational Approach To Edge Detection, 390-396.\nIEEE Trans. Pattern Analysis and Machine Intelligence, 8(6):679- Gorcheichnikov, A. 2017. An Approach to a Biologically Realistic\n698. Simulation of Natural Memory. Master's thesis, Middle Tennessee\nCarpenter, G.A. and Grossberg, S. (1987). A massively parallel State University, Murfreesboro, TN, 70 pages.\narchitecture for a self-organizing neural pattern recognition machine. Grossberg, S. (1973). Contour enhancement, short-term memory,\nComputer Vision, Graphics, and Image Processing 37, 54-115. and constancies in reverberating neural networks. Studies in Applied\nCarpenter, G.A., and Grossberg, S. (1995). Adaptive resonance Mathematics 52, 213-257.\ntheory (ART). In M. Arbib (Ed.), The handbook of brain theory and Grossberg, S., and Huang, T.R. (2009). Artscene: A neural system\nneural networks, (pp. 79-82). Cambridge, M.A.: MIT press. for natural scene classification. Journal of Mision, 9 (4), 6.1-19.\nCarpenter, G.A., Grossberg, S. and Rosen, D.B. (1991). Fuzzy Art: doi: 10.1167/9.4.6.\nFast stable learning and categorization of analog patterns by an Grossberg, S., and Versace, M. (2008) Spikes, synchrony, and\nadaptive resonance system Neural Networks 4, 759-771. attentive learning by laminar thalamocortical circuits. Brain Research,\nCarpenter, Gail A and Grossberg, Stephen. The art of adaptive 1218C, 278-312 [Authors listed alphabetically].\npattern recognition by a self-organizing neural network. Computer, Hagen, T. R., Hjelmervik, J., Lie, K.-A., Natvig, J., and Ofstad\n21(3):77-88, 1988. Henriksen, M. 2005. Visual simulation of shallow-water waves.\nCoifman, R.R. and Maggioni, M. Diffusion wavelets. Applied and Simulation Modelling Practice and Theory 13, pp. 716-726.\nComputational Harmonic Analysis, 21(1):53-94, 2006. Hasselt, Hado Van. Double q-learning. In Advances in Neural\nCoifman, R.R., Lafon, S., Lee, A.B., Maggioni, M., Nadler, B., Information Processing Systems, pp. 2613-2621,2010.\nWarner, F., and Zucker, S.W. Geometric diffusions as a tool for Hinton, G. E., Osindero, S., and Teh, Y. (2006). A fast learning\nharmonic analysis and structure definition of data: Diffusion maps. algorithm for deep belief nets. Neural Computation, 18, 1527-1554.\nProceedings of the National Academy of Sciences of the United Ren, Y et al., Ensemble Classification and Regression-Recent\nStates of America, 102(21):7426, 2005. 21 pages. Developments, Applications and Future Directions, in IEEE Com-\nCornwall et al., Automatically translating a general purpose C++ putational Intelligence Magazine, 10.1109/MCI.2015.2471235, 14\nimage processing library for GPUs. Proceedings 20th IEEE Inter- pages (2016).\nnational Parallel & Distributed Processing Symposium, 2006, 8 Riesenhuber, M., & Poggio, T. (1999). Hierarchical models of\npages. object recognition in cortex. Nature Neuroscience, 2 (11), 1019-\nDavis, C. E. 2005. Graphic Processing Unit Computation of Neural 1025.\nNetworks. Master's thesis, University of New Mexico, Albuquer- Riesenhuber, M., & Poggio, T. (2000). Models of object recogni-\nque, NM, 121 pages. tion. Nature neuroscience, 3, 1199-1204.\nDosher, B.A., and Lu, Z.L. (2010). Mechanisms of perceptual Rolfes, T. 2004. Artificial Neural Networks on Programmable\nattention in preening of location. Vision Res., 40(10-12). 1269- Graphics Hardware. In Game Programming Gems 4, A. Kirmse, Ed.\n1292. Charles River Media, Hingham, MA, pp. 373-378.\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 6 of 20\n\n\n US RE49,461 E\n Page 5\n\n\n(56) References Cited Snider, Greg, et al. \"From synapses to circuitry: Using mernristive\n memory to explore the electronic brain.\" IEEE Computer, vol.\n OTHER PUBLICATIONS 44(2). (2011): 21-28.\n Spratling, M. W. (2008). Predictive coding as a model of biased\nRublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011). ORB: competition in visual attention. Vision Research, 48(12): 1391-1408.\nAn efficient alternative to SIFT or SURF. In IEEE International Spratling, M. W. (2012). Unsupervised learning of generative and\nConference on Computer Vision (ICCV) 2011, 2564-2571. discriminative weights encoding elementary image components in a\nRuesch, J et al. 2008. Multimodal Saliency-Based Bottom-Up predictive coding model of cortical function. Neural Computation,\nAttention a Framework for the Humanoid Robot iCub. 2008 IEEE 24(1):60-103.\nInternational Conference on Robotics and Automation, pp. 962-965. Spratling, M. W., De Meyer, K., and Kompass, R. (2009). Unsu-\nRumelhart D., Hinton G., and Williams, R. (1986). Learning inter- pervised learning of overlapping image components using divisive\nnal representations by error propagation. In Parallel distributed input modulation. Computational intelligence and neuroscience. 20\nprocessing: explorations in the microstructure of cognition, vol. 1, pages.\nMIT Press. 45 pages. Sprekeler, H. On the relation of slow feature analysis and laplacian\nRumpf, M. and Strzodka, R. Graphics processor units: New pros- eigenmaps. Neural Computation, pp. 1-16, 2011.\npects for parallel computing. In Are Magnus Bruaset and Aslak Sun, Z et al., Recognition of SAR target based on multilayer\nTveito, editors, Numerical Solution of Partial Differential Equations auto-encoder and SNN, International Journal of Innovative Com-\non Parallel Computers, vol. 51 of Lecture Notes in Computational puting, Information and Control, vol. 9, No. 11, pp. 4331-4341,\nScience and Engineering, pp. 89-134. Springer, 2005. Nov. 2013.\nSalakhutdinov, R., & Hinton, G. E. (2009). Deep boltzmann machines. Sutton, R. S., and Barto, A.G. (1998). Reinforcement learning: An\nIn International Conference on Artificial Intelligence and Statistics introduction(vol. 1, No. 1). Cambridge: MIT press. 10 pages.\n(pp. 448-455). Tong, F., Ze-Nian Li, (1995). Reciprocal-wedge transform for\nSchaul, Tom, Quan, John, Antonoglou, Ioannis, and Silver, David. space-variant sensing, Pattern Analysis and Machine Intelligence,\nPrioritized experience replay. arXiv preprint arXiv: 1511.05952, IEEE Transactions on , vol. 17, No. 5, pp. 500-551 doi: 10\nNov. 18, 2015. 21 pages. 1109/34.391393.\nSchmidhuber, J. (2010). Formal theory of creativity, fun, and Torralba, A., Oliva, A., Castelhano, M.S., Henderson, J.M. (2006).\nintrinsic motivation (1990-2010). Autonomous Mental Develop- Contextual guidance of eye movements and attention in real-world\nment, IEEE Transactions on, 2(3), 230-247. scenes: the role of global features in object search Psychological\nSchmidhuber, Jurgen. Curious model-building control systems. In Review, 113(4).766-786.\nNeural Networks, 1991. 1991 IEEE International Joint Conference Van Hasselt, Hado, Guez, Arthur, and Silver, David. Deep rein-\non, pp. 1458-1463. IEEE, 1991. forcement learning with double q-learning. arXiv preprint arXiv:\nSeibert, M., & Waxman, A.M. (1992). Adaptive 3-D Object Rec- 1509.06461, Sep. 22, 2015. 7 pages.\nognition from Multiple Views. IEEE Transactions on Pattern Analy- Versace, Brain-inspired computing. Invited keynote address, Bionet-\nsis and Machine Intelligence, 14 (2), 107-124. ics 2010, Boston, MA, USA. 1 page.\nSetoain et al., \"Parallel hyperspectral image processing on com- Versace, M. (2006) From spikes to interareal synchrony: how\nmodity graphics hardware.\" Parallel Processing Workshops, 2006. attentive matching and resonance control learning and Information\nICPP 2006 Workshops 2006 International Conference on. IEEE, processing by laminar thalamocortical circuits. NSF Science of\n2006. 8 pages. Learning Centers PI Meeting, Washington, DC, USA. 1 page.\nSherbakov, L. and Versace, M. (2014) Computational principles for Versace, M., (2010) Open-source software for computational neu-\nan autonomous active vision system. Ph.D., Boston University, roscience: Bridging the gap between models and behavior. In\nhttp://search.proquest.com/docview/l 558856407. 194 pages. Horizons in Computer Science Research, vol. 3 43 pages.\nSherbakov, L. et al. 2012. CogEye: from active vision to context Versace,M., Ames, H., Leveille, J., Fortenberry,B., and Gorchetchnikov,\nidentification, youtube, retrieved from the Internet an Oct. 10, 2017: A. (2008) KlnNeSS: A modular framework for computational\nURL://www.youtube.com/watch?v~i5PQk962Blk, 1 page. neuroscience Neuroinforrnatics, 2008 Winter; 6(4):291-309. Epub\nSherbakov, L. et al. 2013. CogEye: system diagram module brain Aug. 10, 2008.\narea function algorithm approx # neurons, retrieved from the Versace, M., and Chandler, B. (2010) MoNeta: A Mind Made from\nInternet on Oct. 12, 2017: URL://http://www-labsticc.univ-ubs.fr/ Mernristors. IEEE Spectrum, Dec. 2010. 8 pages.\n--coussy/neucomp2013/index_fichiers/material/posters/NeuCornp2013_ Versace, TEDx Fulbright, Invited talk, Washington DC, Apr. 5,\nfinal56x36.pdf, 1 page. 2014. 30 pages.\nSherbakov, L., Livitz, G., Sohail, A., Gorchetchnikov, A., Mingolla, Webster, Bachevalier, Ungerleider (1994). Connections of IT areas\nE., Ames, H., and Versace, M (2013b) A computational model of the TEO and TE with parietal and frontal cortex in macaque monkeys.\nrole of eye-movements in object disambiguation. Cosyne, Feb. Cerebal Cortex, 4(5), 470-483.\n28-Mar. 3, 2013. Salt Lake City, UT, USA. 2 pages. Wiskott, Laurenz and Sejnowski, Terrence. Slow feature analysis:\nSherbakov, L., Livitz, G., Sohail, A., Gorchetchnikov, A., Mingolla, Unsupervised learning ofinvariances. Neural Computation, 14(4):715-\nE., Ames, H., and Versace, M. (2013a) CogEye: An online active 770, 2002.\nvision system that disambiguates and recognizes objects NeuComp Wu, Yan & J. Cai, H. (2010). A Simulation Study of Deep Belief\n2013.2 pages. Network Combined with the Self-Organizing Mechanism of Adap-\nSmolensky, Paul. Information processing in dynamical systems: tive Resonance Theory. 10.1109/CISE.2010.56//265, 4 pages.\nFoundations of harmony theory. No. CU-CS-321-86. Colorado\nUniv At Boulder Dept of Computer Science, 1986. 88 pages. * cited by examiner\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 7 of 20\n\n\nU.S. Patent Mar.14,2023 Sheet 1 of 5 US RE49,461 E\n\n\n\n\n ,/\n ........... \u00b7\u00b7\u00b7\u00b7\u00b7t--\u00b7.c:\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 8 of 20\n\n\nU.S. Patent Mar.14,2023 Sheet 2 of 5 US RE49,461 E\n\n\n\n\n w\n rt: ~ X:\n \\,,,\n ;- 0\n :) 2;\n\n ..., X\n < ~\n ~\n d w w 00 :t, ..-. -0\n 0\n i,t:<\n N\n \\{)\n (~,\n ..1 r- :fil <.(}\n t>J\n ':\")\n ;-b\n (~\n\n\n\n /\n\n\n CJ\n \"<......\n N\n\n ~~ >-\n ~\n 0\n 0\n N\n .,\n <\".\\!\n\n\n -\n u !!\n w\n ..J\n V '?\n\n /\n z.U ,.J.\n 0 :)\n l.\n :Ez\n ~\n \u00b5 tt; I~ a..\n f.!: \u00b7'\u00a2(\n wm '(--,\n }\u00ab\n ~\n 'r-r {;)\u00b7\n 0 0\n >.::t \u00b7,AWMWM ,A~ \u2022M\u00b7,A \u2022, 'OW\n\n\n %:\n w - Mi2J.1ma11:)d\n _.,,,~..,\n ~\n 0\n (.'>)\n N\n '\"I ..,, 'T\n\n\n - -\n\n\n ,,.,\n 0\n -.:.:::-:\n ..... --\u00b7\n f~\n .......\n ~\n ,~\\;,j\u2022t: r.?-!.if:~/? ~/:f::\n \u00b7\u00b7, :.~.....'!'\u00b7\";\u2022,;-\u00b7.:\u00b7~:\u00b7:\u00b7~ .:-:;::~\n lh.~-:;_/~ ~/~//~\n 435\n 40;: 403\n !: ~rsp.H~ t~xr:1r:=:-~~-ff>i:=:.\n t>:::-:-:.~u:\n ::-:\u2022\n ~n>:::-~n:.ny\n I\n .j t;-;:u:}; ?:-;:GPU\n .\n Sl i~:~t1-:::-r~;:\n fi t)fn\n ~b:=:::!Jt:;::\u00b7\n r'f}f.::f::K::ry\n b::~nK:r:-GPU\n\n ;\u2666------------------\n :r\n\n\n\n\n 480,----~--~\n i Outp~:ttf:<_:,;\u00b7:sp<::,\n i vpk-.:;;:r!to ::';ix.~:Jrfs\n j r?>:3 r\":::(\u00b7!~).. (3;~t: k ,......................\n 440\n\n\n i::;:~~=~:,:~:;:::;t~:~::~\n f\n 4~}fj\n\n V'.J::i~t\n fr_;r:~w::tp i fr:=!\u00b7\n \\:\\'\u00b0::~~t f)W{)p\n r,..\u00b7\n \u2713 y5\n O{ lnpcl!i\u25catl!f,rn\n t:::\u2022:-:.b.Jr<::\n 44~=,\n ........\n p:_);ntr~f::'>\n 1............. t..l~~;~~;~r\u2022:::;~t',J\n N.:.'::..................\n L:~::-(t-.....\n .........,\n --------~~..........\n tt-~~ratK>11?,,.\"\"r\u00b7-.:--\n ..... ........_ ......... \u2713\n\n\n Y<::-s\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 11 of 20\n\n\n\n\n e\n \u2022\n 00\n \u2022\n ~\n ~\n ~\n ~\n\n =\n ~\n\n\n\n 3 .i 2 3 4 1 2 3 -\u00b7\u00b7.. -\u00b7\u00b7 ...\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\n .............-...... ....\n 1 2 3\n ~--\u2022-\u2022.\u2022\n -~\u00b7\u00b7\u00b7'\u00b7\u00b7\u00b7\u00b7'\n 6 5 6 7 8 4 5 6 4 5 6 ~\n ~\n\n7!8 9 \u00b7g 7 8 9 7 8 9 :-:\n ' \" ....\n ...\n ~\n N\n 0\n ::a) b) ........,.,.,\n ,, ...... ' .... -\u00b7\u00b7\n )1 \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7=\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7--\u2022\u2022'\u00b7\u2022\u00b7\u00b7\n N\n ~\n\n\n\n\n Bo1dlines shmv the packing of data into pixels ',Nlthfour color components,\n rJJ\n =-\n ('D\n ....\n ('D\n\n Ul\n\n ....\n 0\n\n Ul\n\n\n\n\n FIG\u00bb5 d\n rJl.\n\n ~\n ~\n \\0\n ~\n 0--,\n\n \"'\"\"'\n ~\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 12 of 20\n\n\n US RE49,461 E\n 1 2\n GRAPHIC PROCESSOR BASED processing unit/system (CPU). Complete independence is\n ACCELERATOR SYSTEM AND METHOD not desirable, however; user input might affect how the\n computation is performed and even interrupt it if necessary.\n Furthermore, the user output and the disk output are depen-\n Matter enclosed in heavy brackets [ ] appears in the 5 dent on the results of the computation. A reasonable solution\n original patent but forms no part of this reissue specifica- would be to separate input/output into threads, so that it is\n tion; matter printed in italics indicates the additions interacting with hardware occurs in parallel with the com-\n made by reissue; a claim printed with strikethrough putation. In this case whatever CPU processing is required\n indicates that the claim was canceled, disclaimed, or held for input/output should be designed so that it provides the\n invalid by a prior post-patent action or proceeding. 10 synchronization with computation.\n In the case of GPGPU, the computation itself is performed\n RELATED APPLICATIONS outside of the CPU, so the complete system comprises three\n \"peripheral\" components: user interactive hardware, disk\n The present application is a reissue continuation appli- hardware, and computational hardware. The central process-\n cation of U.S. application Ser. No. 15/808,201, which was 15 ing unit (CPU) establishes communication and synchroni-\nfiled on Nov. 9, 2017 as a broadening reissue application of zation between peripherals. Each of the peripherals is pref-\n U.S. Pat. No. 9,189,828, filed Jan. 3, 2014, which claims a erably controlled by a dedicated thread that is executed in\n priority benefit, under 35 U.S.C. \u00a7120, as a continuation of parallel with minimal interactions and dependencies on the\n U.S. application Ser. No. 11/860,254, now U.S. Pat. No. other threads.\n 8,648,867 B2, filed Sep. 24, 2007, entitled \"Graphic Pro- 20 A GPU on a conventional video card is usually controlled\n cessor Based Accelerator System and Method,\" which in through OpenGL, DirectX, or similar graphic application\n turn claims the priority benefit, under 35 U.S.C. \u00a7119(e), of programming interfaces (APis ). Such APis establish the\n U.S. Application No. 60/826,892, filed Sep. 25, 2006. The context of graphic operations, within which all calls to the\npresent application is also a broadening reissue application GPU are made. This context only works when initialized\n of U.S. Pat. No. 9,189,828, filed Jan. 3, 2014, which is a 25 within the same thread of execution that uses it. As a result,\n continuation of U.S. application Ser. No. 11/860,254, now in a preferred embodiment, the context is initialized within\n U.S. Pat. No. 8,648,867 B2, filed Sep. 24, 2007, which in a computational thread. This creates complications, how-\n turn claims the priority benefit, under 35 U.S. C. \u00a7 119(e), of ever, in the interaction between the user interface thread that\n U.S. Application No. 60/826,892, filed Sep. 25, 2006. Each changes parameters of simulations and the computational\n of the above-identified applications is incorporated herein by 30 thread that uses these parameters.\n reference in its entirety. More than one reissue application A solution as proposed here is an implementation of the\n has been filed for the reissue of U.S. Pat. No. 9,189,828, computational stream of execution in hardware, so that\n including this application and U.S. application Ser. No. thread and context initialization are replaced by hardware\n 15/808,201. initialization. This hardware implementation includes an\n 35 expansion card comprising a printed circuit board having (a)\n BACKGROUND one or more graphics processing units, (b) two or more\n associated memory banks that are logically or physically\n Graphics Processing Units (GPUs) are found in video partitioned, (c) a specialized controller, and (d) a local bus\n adapters (graphic cards) of most personal computers (PCs), providing signal coupling compatible with the PCI industry\n video game consoles, workstations, etc. and are considered 40 standards (this includes but is not limited to PCI-Express,\n highly parallel processors dedicated to fast computation of PCI-X, USB 2.0, or functionally similar technologies). The\n graphical content. With the advances of the computer and controller handles most of the primitive operations needed to\n console gaming industries, the need for efficient manipula- set up and control GPU computation. As a result, the CPU\n tion and display of 3D graphics has accelerated the devel- is freed from this function and is dedicated to other tasks. In\n opment of GPUs. 45 this case a few controls (simulation start and stop signals\n In addition, manufacturers of GPU shave included general from the CPU and the simulation completion signal back to\n purpose programmability into the GPU architecture leading CPU), GPU programs and input/output data are the infor-\n to the increased popularity of using GPU s for highly paral- mation exchanged between CPU and the expansion card.\n lelizable and computationally expensive algorithms outside Moreover, since on every time step of the simulation the\n of the computer graphics domain. When implemented on 50 results from the previous time step are used but not changed,\n conventional video card architectures, these general purpose the results are preferably transferred back to CPU in parallel\n GPU (GPGPU) applications are not able to achieve optimal with the computation.\n performance, however. There is overhead for graphics- In general, according to one aspect, the invention features\n related features and algorithms that are not necessary for a computer system. This system comprises a central pro-\n these non-video applications. 55 cessing unit, main memory accessed by the central process-\n ing unit, and a video system for driving a video monitor in\n SUMMARY response to the central processing unit as is common. The\n computer system further comprises an accelerator that uses\n Numerical simulations, e.g., finite element analysis, of input data from and provides output data to the central\n large systems of similar elements (e.g. neural networks, 60 processing unit. This accelerator comprises at least one\n genetic algorithms, particle systems, mechanical systems) graphics processing unit, accelerator memory for the graphic\n are one example of an application that can benefit from processing unit, and an accelerator controller that moves the\n GPGPU computation. During numerical simulations, disk input data into the at least one graphics processing unit and\n and user input/output can be performed independently of the accelerator memory to generate the output data.\n computation because these two processes require interac- 65 In the preferred, the central processing unit transfers the\n tions with peripheral hardware (disk, screen, keyboard, input data for a simulation to the accelerator, after which the\n mouse, etc) and put relatively low load on the central accelerator executes simulation computations to generate\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 13 of 20\n\n\n US RE49,461 E\n 3 4\nthe output data, which is transferred to the central processing In more detail, the computer system 100 in one example\nunit. Preferably, the accelerator controller dictates an order is a standard personal computer (PC). However, this only\nof execution of instructions to the at least one graphics serves as an example environment as computing environ-\nprocessing unit. The use of the separate controller enables ment 100 does not necessarily depend on or require any\ndata transfer during execution such that the accelerator 5 combination of the components that are illustrated and\ncontroller transfers output data from the accelerator memory described herein. In fact, there are many other suitable\nto main memory of the central processing unit. computing environments for this invention, including, but\n In the preferred embodiment, the accelerator controller not limited to, workstations, server computers, supercom-\ncomprises an interface controller that enables the accelerator puters, notebook computers, hand-held electronic devices\nto communicate over a bus of the computer system with the 10 such as cell phones, mp3 players, or personal digital assis-\ncentral processing unit. tants (PDAs), multiprocessor systems, progranimable con-\n In general according to another aspect, the invention also sumer electronics, networks of any of the above-mentioned\nfeatures an accelerator system for a computer system, which computing devices, and distributed computing environments\ncomprises at least one graphics processing unit, accelerator\n that including any of the above-mentioned computing\nmemory for the graphic processing unit and an accelerator 15\n devices.\ncontroller for moving data between the at least one graphics\n In one implementation the GPU accelerator is imple-\nprocessing unit and the accelerator memory.\n In general according to another aspect, the invention also mented as an expansion card 180 includes connections with\nfeatures a method for performing numerical simulations in a the motherboard 110, on which the one or more CPU's 120\ncomputer system. This method comprises a central process- 20 are installed along with main, or system memory 130 and\ning unit loading input data into an accelerator system from mass/non volatile data storage 140, such as hard drive or\nmain memory of the central processing unit and an accel- redundant array of independent drives (RAID) array, for the\nerator controller transferring the input data to a graphics computer system 100. In the current example, the expansion\nprocessing unit with instructions to be performed on the card 180 communicates to the motherboard 110 via a local\ninput data. The accelerator controller then transfers output 25 bus 190. This local bus 190 could be PCI, PCI Express,\ndata generated by the graphic processing unit to the central PCI-X, or any other functionally similar technology (de-\nprocessing unit as output data. pending upon the availability on the motherboard 110). An\n The above and other features of the invention including external version GPU accelerator is also a possible imple-\nvarious novel details of construction and combinations of mentation. In this example, the external GPU accelerator is\nparts, and other advantages, will now be more particularly 30 connected to the motherboard 110 through USB-2.0, IEEE\ndescribed with reference to the accompanying drawings and 1394 (Firewire), or similar external/peripheral device inter-\npointed out in the claims. It will be understood that the face.\nparticular method and device embodying the invention are The CPU 120 and the system memory 130 on the moth-\nshown by way of illustration and not as a limitation of the erboard 110 and the mass data storage system 140 are\ninvention. The principles and features of this invention may 35 preferably independent of the expansion card 180 and only\nbe employed in various and numerous embodiments without communicate with each other and the expansion card 180\ndeparting from the scope of the invention. through the system bus 200 located in the motherboard 110.\n A system bus 200 in current generations of computers have\n BRIEF DESCRIPTION OF THE DRAWINGS bandwidths from 3.2 GB/s (Pentium 4 withAGTL+, Athlon\n 40 XP with EV6) to around 15 GB/s (Xeon Woodcrest with\n In the accompanying drawings, reference characters refer AGTL+, Athlon 64/Opteron with Hypertransport), while the\nto the same parts throughout the different views. The draw- local bus has maximal peak data transfer rates of 4 GB/s\nings are not necessarily to scale; emphasis has instead been (PCI Express 16) or 2 GB/s (PCI-X 2.0). Thus the local bus\nplaced upon illustrating the principles of the invention. Of 190 becomes a bottleneck in the information exchange\nthe drawings: 45 between the system bus 200 and the expansion card 180. The\n FIG. 1 is a schematic diagram illustrating a computer design of the expansion card and methods proposed herein\nsystem including the GPU accelerator according to an minimizes the data transfer through the local bus 190 to\nembodiment of the present invention; reduce the effect of this bottleneck.\n FIG. 2 is block diagram illustrating the architecture for the The system memory 130 is referred to as the main\nGPU accelerator according to an embodiment of the present 50 random-access memory (RAM) in the description herein.\ninvention; However, this is not intended to limit the system memory\n FIG. 3 is a block/flow diagram illustrating an exemplary 130 to only RAM technology. Other possible computer\nimplementation of the top level control of the GPU accel- storage media include, but are not limited to ROM,\nerator system; EEPROM, flash memory, or any other memory technology.\n FIG. 4 is a flow diagram illustrating an exemplary imple- 55 In the illustrated example, the GPU accelerator system is\nmentation of the bottom level control of the GPU accelerator implemented on an expansion card 180 on which the one or\nsystem that is used to execute the target computation; and more GPU's 240 are mounted. It should be noted that the\n FIG. 5 is an example population of nine computational GPU accelerator system GPU 240 is separate from and\nelements arranged in a 3x3 square and a potential packing independent of any GPU on the standard video card 150 or\nscheme for texture pixels, according to an implementation of 60 other video driving hardware such as integrated graphics\nthe present invention. systems. Thus the computations performed on the expansion\n card 180 do not interfere with graphics display (including\n DETAILED DESCRIPTION but not limited to manipulation and rendering of images).\n Various brand of GPU are relevant. Under current tech-\n FIG. 1 shows a computer system 100 that has been 65 nology, GPU's based on the GeForce series from NVIDIA\nconstructed according to the principles of the present inven- Corporation or the Catalyst series from ATI/Advanced\ntion. Micro Devices, Inc.\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 14 of 20\n\n\n US RE49,461 E\n 5 6\n The output to a video monitor 170 is preferably through partition 250b is designed to hold the data textures repre-\nthe video card 150 and not the GPU accelerator system 180. senting internal variables. The third partition 250c is\nThe video card 150 is dedicated to the transfer of graphical designed to hold the data textures used as input at a\ninformation and connects to the motherboard 110 through a particular computation step on the GPU 240. The fourth\nlocal bus 160 that is sometimes physically separate from the 5 partition 250d holds the data textures used to accommodate\nlocal bus 190 that connects the expansion card 180 to the the output of a particular computational step on the GPU\nmotherboard 110. 240. This partitioning scheme can be done logically, does\n FIG. 2 is a block diagram illustrating the general archi- not require hardware implementation. Also the partitioning\ntecture of the GPU accelerator system and specifically the scheme is also altered based on new designs or needs of the\nexpansion card 180 in which at least one GPU 240 and 10 algorithms being employed. The reason for this partitioning\nassociated memories 210 and 250 are mounted. Electrical is further explained in the Data Organization section, below.\n(signal) and mechanical coupling with a local bus 190 A local bus interface 230 on the controller 220 serves as\nprovides signal coupling compatible with the PCI industry a driver that allows the controller 220 to communicate\nstandards (this includes but is not limited to PCI, PCI-X, PCI through the local bus 190 with the system bus 200 and thus\nExpress, or functionally similar technology). 15 the CPU 120 and RAM 130. This local bus interface 230 is\n The GPU accelerator further preferably comprises one not intended to be limited to PCI related technology. Other\nspecifically designed accelerator controller 220. Depending drivers can be used to interface with comparable technology\nupon the implementation, the accelerator controller 220 is as a local bus 190.\nfield programmable gate array (FPGA) logic, or custom built Data Organization\napplication-specific (ASIC) chip mounted in the expansion 20 Each computational element discussed above has output\ncard 180, and in mechanical and signal coupling with the variables that affect the rest of the system. For example in\nGPU 240 and the associated memories 210 and 250. During the case of a neural network it is the output of a neuron. A\ninitial design, a controller can be partially or even fully computational element also usually has several internal\nimplemented in software, in one example. variables that are used to compute output variables, but are\n The controller 220 commands the storage and retrieval of 25 not exposed to the rest of the system, not even to other\narrays of data (on a conventional video card the arrays of elements of the same population, typically. Each of these\ndata are represented as textures, hence the term 'texture' in variables is represented as a texture. The important differ-\nthis document refers to a data array unless specified other- ence between output variables and internal variables is their\nwise and each element of the texture is a pixel of color access.\ninformation), execution of GPU programs (on a conven- 30 Output variables are usually accessed by any element in\ntional video card these programs are called shaders, hence the system during every time step. The value of the output\nthe term 'shader' in this document refers to a GPU program variable that is accessed by other elements of the system\nunless specified otherwise), and data transfer between the corresponds to the value computed on the previous, not the\nsystem bus 200 and the expansion card 180 through the local current, time step. This is realized by dedicating two textures\nbus 190 which allows communication between the main 35 to output variables----one holds the value computed during\nCPU 120, RAM 130, and disk 140. the previous time step and is accessible to all computational\n Two memory banks 210 and 250 are mounted on the elements during the current time step, another is not acces-\nexpansion card 180. In some example, these memory banks sible to other elements and is used to accumulate new values\nseparated in the hardware, as shown, or alternatively imple- for the variable computed during the current time step.\nmented as a single, logically partitioned memory compo- 40 In-between time steps these two textures are switched, so\nnent. that newly accumulated values serve as accessible input\n The reason to separate the memory into two partitions 210 during the next time step, while the old input is replaced with\n250 stems from the nature of the computations to which the new values of the variable. This switch is implemented by\nGPU accelerator system is applied. The elements of com- swapping the address pointers to respective textures as\nputation (computational elements) are characterized by a 45 described in the System and Framework section.\nsingle output variable. Such computational elements often Internal variables are computed and used within the same\ninclude one or more equations. Computational elements are computational element. There is no chance of a race con-\nsame or similar within a large population and are computed dition in which the value is used before it is computed or\nin parallel. An example of such a population is a layer of after it has already changed on the next time step because\nneurons in an artificial neural network (ANN), where all 50 within an element the processing is sequential. Therefore, it\nneurons are described by the same equation. As a result, is possible to render the new value of internal variable into\nsome data and most of the algorithms are common to all the same texture where the old was read from in the texture\ncomputational elements within population, while most of the memory bank. Rendering to more than one texture from a\ndata and some algorithms are specific for each equation. single shader is not implemented in current GPU architec-\nThus, one memory, the shader memory bank 210, is used to 55 tures, so computational elements that track internal variables\nstore the shaders needed for the execution of the required would have to have one shader per variable. These shaders\ncomputations and the parameters that are common for all can be executed in order with internal variables computed\ncomputational elements and is coupled with the controller first, followed by output variables.\n220 only. The second memory, the texture memory bank Further savings of texture memory is achieved through\n250, is used to store all the necessary data that are specific 60 using multiple color components per pixel (texture element)\nfor every computational element (including, but not limited to hold data. Textures can have up to four color components\nto, input data, output data, intermediate results, and param- that are all processed in parallel on a GPU. Thus, to\neters) and is coupled with both the controller 220 and the maximize the use of GPU architecture it is desirable to pack\nGPU 240. the data in such a way that all four components are used by\n The texture memory bank 250 is preferably further par- 65 the algorithm. Even though each computational element can\ntitioned into four sections. The first partition 250 a is have multiple variables, designating one texture pixel per\ndesigned to hold the external input data patterns. The second element is ineffective because internal variables require one\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 15 of 20\n\n\n US RE49,461 E\n 7 8\ntexture and output variables require two textures. Further- Stream 303-runs on the GPU accelerator of the expansion\nmore, different element types have different numbers of card 180 and interacts with the User Interaction Stream 302\nvariables and unless this number is precisely a multiple of through initialization routines and data exchange in between\nfour, texture memory can be wasted. simulations. The Computational Stream 303 interacts with\n Amore reasonable packing scheme would be to pack four 5 the User Interaction Stream and the Data Output Stream\ncomputational elements into a pixel and have separate through synchronization procedures during simulations.\ntextures for every variable associated with each computa- The crucial feature of the interaction between the User\ntional element. In this case the packing scheme is identical Interaction Stream 302 and the Computational Stream 303 is\nfor all textures, and therefore can be accessed using the same the shift of priorities. Outside of the simulation, the system\nalgorithm. Several ways to approach this packing scheme 10 100 is driven by the user input, thus the User Interaction\nare outlined here. An example population of nine computa- Stream 302 has the priority and controls the data exchange\ntional elements arranged in a 3x3 square (FIG. Sa) can be 304 between streams. After the user starts the simulation, the\npacked by element (FIG. Sb), by row (FIG. Sc), or by square Computational Stream 303 takes the priority and controls\n(FIG. Sd). the data exchange between streams until the simulation is\n Packing by element (FIG. Sb) means that elements 1,2,3,4 15 finished or interrupted 350.\ngo into first pixel; 5,6,7,8 go into second pixel; 9 goes into The user starts 300 the framework through the means of\nthird pixel. This is the most compact scheme, but not an operating system and interacts with the software through\nconvenient because the geometrical relationship is not pre- the user interaction section 305 of the graphic user interface\nserved during packing and its extraction depends on the size 306 executed on the CPU 120. The start 300 of the imple-\nof the population. 20 mentation begins with a user action that causes a GUI\n Packing by row (colunm; FIG. Sc) means that elements initialization 307, Disk input/output initialization 308 on the\n1,2,3 go into pixel (1,1); 3,4,5 go into pixel (2,1), 7,8,9 go CPU 120, and controller initialization 320 of the GPU\ninto pixel (3,1). With this scheme the element's y coordinate accelerator on the expansion card 180. GUI initialization\nin the population is the pixel's y coordinate, while the includes opening of the main application window and setting\nelement's x coordinate in the population is the pixel's x 25 the interface tools that allow the user to control the frame-\ncoordinate times four plus the index of color component. work. Disk I/O initialization can be performed at the start of\nFive by five populations in this case will use 2x5 texture, or the framework, or at the start of each individual simulation.\n10 pixels. Five of these pixels will only use one out of four The user interaction 305 controls the setting and editing of\ncomponents, so it wastes 37.5% of this texture. 25xl popu- the computational elements, parameters, and sources of\nlation will use 6xl texture (six pixels) and will waste 12.5% 30 external inputs. It specifies which equations should have\nof it. their output saved to disk and/or displayed on the screen. It\n Packing by square (FIG. Sd) means that elements 1,2,4,5 allows the user to start and stop the simulation. And it\ngo into pixel (1,1); 3,6 go into pixel (1,2); 7,8 go into pixel performs standard interface functions such as file loading\n(2,1), and 9 goes into pixel (2,2). Both the row and the and saving, interactive help, general preferences and others.\ncolunm of the element are determined from the row (col- 35 The user interaction 305 directs the CPU 120 to acquire\nunm) of the pixel times two plus the second (first) bit of the the new external input textures needed (this includes but is\ncolor component index. Five by five populations in this case not limited to loading from disk 140 or receiving them in\nwill use 3x3 texture, or 9 pixels. Four of these pixels will real time from a recording device), parses them if necessary\nonly use two out of four components, and one will only use 309, and initializes their transfer to the expansion card 180,\none component, so it wastes 34.4% of this texture. This is 40 where they are stored 325 in the texture memory bank 250\nmore advantageous than packing by row, since the texture is by the controller 220. The user interaction 305 also directs\nsmaller and the waste is also lower. 25xl population on the the CPU 120 to parse populations of elements that will be\nother hand will use 13xl texture (thirteen pixels) and waste used in the simulation, convert them to GPU programs\n>50% of it, which is much worse than packing by row. (shade rs), compile them 310, and initializes their transfer to\n In order to eliminate waste altogether the population 45 the expansion card 180, where they are stored 326 in the\nshould have even dimensions in the square packing, and it shader memory bank 210 by the controller 220. This opera-\nshould have a number of columns divisible by four in row tion is accompanied by the upload 309 of the initial data into\npacking. Theoretically, the chances are approximately the input partition of the texture memory bank 250, and\nequivalent for both of these cases to occur, so the particular stores the shader order of execution in the controller 220.\ntask and data sizes should determine which packing scheme 50 The user can perform operations 309 and 310 as many times\nis preferable in each individual case. as necessary prior to starting the simulation or between\n The System and Framework simulations.\n FIG. 3 shows an exemplary implementation of the top The editing of the system between simulations is difficult\nlevel system and method that is used to control the compu- to accomplish without the hardware implementation of the\ntation. It is a representation of one of several ways in which 55 computational thread suggested herein. The system of equa-\na system and method for processing numerical techniques tions (computational elements) is represented by textures\ncan be implemented in the invention described herein and so that track variables plus shaders that define processing\nthe implementation is not intended to be limited to the algorithms. As mentioned above, textures, shaders and other\nfollowing description and accompanying figure. graphics related constructs can only be initialized within the\n The method presented herein includes two execution 60 rendering context, which is thread specific. Therefore tex-\nstreams that run on the CPU 120-User Interaction Stream tures and shaders can only be initialized in the computa-\n302 and Data Output Stream 301. These two streams pref- tional thread.\nerably do not interact directly, but depend on the same data Network editing is a user-interactive process, which\naccumulated during simulations. They can be implemented according to the scheme suggested above happens in the\nas separate threads with shared memory access and executed 65 User Interaction Stream 302. The simulation software thus\non different CPUs in the case of multi-CPU computing has to take the new parameters from the User Interaction\nenvironment. The third execution stream-Computational Stream 302, communicate them to the Computational\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 16 of 20\n\n\n US RE49,461 E\n 9 10\nStream 303 and regenerate the necessary shaders and tex- practical software implementation of the method and archi-\ntures. This is hard to accomplish without a hardware imple- tecture described above and pictorially represented in FIG.\nmentation of the Computational Stream 303. The Compu- 3.\ntational Stream 303 is forked from the User Interaction To use SANNDRA, the application should create a\nStream and it can access the memory of the parent thread, 5 TSimulator object either directly or through inheritance.\nbut the reverse communication is harder to achieve. The This object will handle global simulation properties and\ncontroller 220 allows operations 309 and 310 to be per- control the User Interaction Stream, Data Output Stream,\nformed as many times as necessary by providing the nec- and Computational Stream. Through TSimulator: :time-\nessary communication to the User Interaction Stream 302. step( ) TSimulator: :outfileinterval( ), and TSimulator: :out-\n After execution of the input parser texture generation 309 lO mode( ), the application can set the time step of the simu-\nand population parser shader generator and compiler 310 are lation, the time step of disk output, and the mode of the disk\nperformed at least once, the user has the option to initialize output. The external input pattern should be packed into a\nthe simulation 311. During this initialization the main con- TPattem object and bound to the simulation object through\ntrol of the framework is transferred to the GPU accelerator 15 TSimulator::resetinputs( ) method. TSimulator::sim-\nsystem's accelerator controller 220 and computation 330 is Length( ) sets the length of the simulation.\nstarted (see FIG. 4; 420). The user retains the ability to The second step is to create at least one population of\ninterrupt the simulation, change the input, or to change the equations (Tpopulation object). Population holds one equa-\n tion object TEquation. This object contains only a formula\ndisplay properties of the framework, but these interactions\n and does not hold element-specific data, so all elements of\nare queued to be performed at times determined by the 20\n the population can share single TEquation.\ncontroller-driven data exchange 314 and 316 to avoid the\n The TEquation object is converted to a GPU program\ncorruption of the data. before execution. GPU programs have to be executed within\n The progress monitor 312 is not necessary for perfor- a graphical context, which is stream specific. TSimulator\nmance, but adds convenience. It displays the percentage of creates this context within a Computational Stream, there-\ncompleted time steps of the simulation and allows the user 25 fore all programs and data arrays that are necessary for\nto plan the schedule using the estimates of the simulation computation have to be initialized within Computational\nwall clock times. Controller-driven data exchange 314 Stream. Constructor of TPopulation is called from User\nupdates the display of the results 313. Online screen output Interaction Stream, so no GPU-related objects can be ini-\nfor the user selected population allows the user to monitor tialized in this constructor.\nthe activity and evaluate the qualitative behavior of the 30 TPopulation: :fillElements( ) is a virtual method designed\nnetwork. Simulations with unsatisfactory behavior can be to overcome this difficulty. It is called from within the\nterminated early to change parameters and restart. Control- Computational Stream after TSimulator: :networkCreate( ) is\nler-driven data exchange 314 also drives the output of the called in the User Interaction Stream. A user has to override\nresults to disk 317. Data output to disk for convenience can 35 TPopulation::fillElements( ) to create TEquation and other\nbe done on an element per file basis. A suggested file format computation related objects both element independent and\nincludes a leftmost colunm that displays a simulated time for element-specific. Element independent objects include sub-\neach of the simulation steps and subsequent colunms that components of TEquation and objects that describe how to\ndisplay variable values during this time step in all elements handle interdependencies between variables implemented\n through derivatives of TGate class.\nwith identical equations (e.g. all neurons in a layer of a 40\n Element-specific data is held in TElement objects. These\nneural network).\n objects hold references to TEquation and a set of TGate\n Controller-driven data exchange or input parser texture\n objects. There is one TElement per population, but the size\ngenerator 316 allows the user to change input that is gen- of data arrays within this object corresponds to population\nerated on the fly during the simulation. This allows the size. All TElement objects have to be added to the TSimu-\nframework monitoring of the input that is coming from a 45 lator list of elements by calling TSimulator::addUnit( )\nrecording device (video camera, microphone, cell recording method from TPopulation: :fillElements( ).\nelectrode, etc) in real time. Similar to the initial input parser Finally, TPopulation::fillElements() should contain a set\n309, it preprocesses the input into a universal format of the of TElement::add*Dependency( ) calls for each element.\ndata array suitable for texture generation and generates Each of these calls sets a corresponding dependency for\ntextures. Unlike the initial parser 309, here the textures are 50 every TGate object. Here TGate object holds element inde-\ntransferred to hardware not whenever ready but upon the pendent part of dependency and TElement::\nrequest of the controller 220. add*Dependency( ) sets element-specific details.\n The controller 220 also drives the conditional testing 315 System provided TPopulation handles the output of com-\nand 318 informs the CPU-bound streams whether the simu-\n putational elements, both when they need to exchange the\nlation is finished. If so, the control returns to the User 55 data and when they need to output it to disk. User imple-\nInteraction Stream. The user then can change parameters or mentation of TPopulation derivative can add screen output.\ninputs (309 and 310), restart the simulation (311) or quit the Listing 1 is an example code of the user program that uses\nframework (390). a recurrent competitive field (RCF) equation:\n SANNDRA (Synchronous Artificial Neuronal Network\nDistributed Runtime Algorithm; http://www.kinness.net/ 60\n LISTING 1\nDocs/SANNDRA/html) was developed to accelerate and\noptimize processing of numerical integration of large non- uint16_t w - 3, h - 3;\nhomogenous systems of differential equations. This library static float m_compet = 0.5;\n static float m_persist = 1.0;\nis fully reworked in its version 2.x.x to support multiple class TCablePopRCF : public TPopulation\ncomputational backends including those based on multicore 65 {\nCPUs, GPUs and other processing systems. GPU based TEq_RCF* m_equation;\nbackend for SANNDRA-2.x.x can serve as an example\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 17 of 20\n\n\n US RE49,461 E\n 11 12\n LISTING I-continued for new time step. To avoid data confusion, the new values\n of variables should be rendered in a separate texture. After\nTGate* m_gatel; the time step is completed for all equations, these new values\nTGate* m_gate2;\nvoid createGatingStructure( )\n should be copied over old values so that they are used as\n{ 5 input during the next time step. Copying textures is an\nm_gatel - new TGate(0); expensive operation, computationally, but since the textures\nm_gate2 - new TGate(l); are referred to by texture IDs (pointers), swapping these\n}; pointers for input and output textures after each time step\nvoid createUnitStructure(TBasicUnit* u)\n achieves the same result at a much lesser cost.\n{\nu->addO2OPlnputDependency(m_gatel, 0., 0., 0.004, 0., 0, 0); 10\n In the hardware solution suggested herein, ID swapping is\nu->addFullDependency(m_gate2, population()); equivalent to swapping the base memory address for two\n} partitions of the texture memory bank 250. They are\npublic: TCablePopRCF() : TPopulation(\"compCPU RCF\", w, h, true) { }; swapped 485 during synchronization (485, 430, and 455) so\n~TCablePopRCF() {if(m_equation) delete m_equation;\n that data transfer 445 and the computation 435-487 proceeds\n if(m_gatel) delete m_gatel;\n if(m_gate2) delete m_gate2;}; immediately and in parallel with data transfer as shown in\n 15 FIG. 4. A hardware solution allows this parallelism through\nboo! fillElements(TSimulator* sim);\n}; access of the controller 220 to the onboard texture memory\nboo! TCablePopRCF::fillElements(TSimulatior* sim) bank 250.\n{ The main computation and data exchange are executed by\nm_equation - new TEq_RCF(this, m_compet, m_persist);\ncreateGatingStructure( );\n the controller 220. It runs three parallel substreams of\nfor(size_t i - 0; i < xSize( ); ++i) 20 execution: Computational Substream 403, Data Output Sub-\nfor(size_t j - 0; j < ySize( ); ++j) stream 402, and Data Input Substream 404. These streams\n{ are synchronized with each other during the swap of pointers\nTElement* u - new TCPUElement(this, m_equation, i, j); 485 to the input and output texture memory partitions of the\nsim->addUnit(u);\ncreate U nitStructure( u);\n texture memory bank 250 and the check for the last iteration\n} 25 487. Algorithmically, these two operations are a single\nReturn true; atomic operation, but the block diagram shows them as two\n} separate blocks for clarity.\nint The Computational Substream 403 performs a computa-\nmain() tional cycle including a sequential execution of all shaders\n{\n// Input pattern generation (309 in FIG.3) that were stored in the shader memory bank 210 using the\n 30\nuint32_t* pat - new uint32_t[w*h]; appropriate input and output textures. To begin the simula-\nTRandom randGen (0); tion the controller 220 initializes three execution sub streams\nfor(uint32_t I - 0; I < w*h; ++i) 403, 402, and 404. On every simulation step, the Compu-\npat[i] - randGen.random( ); tational Substream 403 determines which textures the GPU\nTpattern* p - new Tpattern(pat, w, h);\n// Setting up the simulation\n 240 will need to perform the computations and initiates the\n 35 upload 435 of them onto the GPU 240. The GPU 240 can\nTSimulator* cableSim - new TSimulator(\"data\"); //(308 and 320 in\nFIG. 3) communicate directly with the texture memory bank 250 to\ncableSim->timestep(0.05); //(320 in FIG. 3) upload the appropriate texture to perform the computations.\ncableSim->resetlnputs(p); //(325 in FIG. 3) The controller 220 also pulls the first shader (known by the\ncableSim->outfileinterval(0.1); //(308 in FIG. 3)\ncableSim->outmode(SANNDRA::timefunc); //(308 in FIG. 3)\n stored order) from the shader memory bank 210 and uploads\ncableSim->simLength(60.0); //(320 in FIG. 3) 40 450 it onto the GPU 240.\n// Preparing the population The GPU 240 executes the following operations in this\nTPopulation* cablePop - new TCablePopRCF( ); //(310 in FIG. 3) order: performs the computation (execution of the shader)\ncableSim->networkCreate( ); //(326 in FIG. 3) 470; tells the controller 220 that it is done with the compu-\nuintl 6_t user= 1;\n tations for the current shader; and after all shaders for this\nwhile(user)\n{ 45 particular equation are executed sends 480 the output tex-\nif(! cableSim->simulationStart(true, 1)) //(311 in FIG. 3) tures to the output portion of the texture memory bank 250.\nexit(!); This cycle continues through all of the equations based on\nstd::cout<<\"Repeat?ln\"; //(305 in FIG. 3) the branching step 482.\nstd::cin>>user; //(305 in FIG. 3) An example shader that performs fourth order Runge-\nif(user -- 1)\ncableSim->networkReset( ); //(305 in FIG. 3)\n Kutta numerical integration is shown in Listing 2 using\n 50\n{ GLSL notation;\nIf(cableSim)\nDelete cableSim; //Also deletes cablePop and its internals LISTING 2\nexit(0);\n}; uniform sarnpler2DRect Variable;\n 55 uniform float integration_step;\n float halfstep - integration_step*0.5;\n FIG. 4 is a detailed flow diagram illustrating a part of an float fl_6step - integration_step/6.0;\nexemplary implementation of the bottom level system and vec4 output - texture2DRect(Variable, gl_TexCoord[0].st);\n // define equation( ) here\nmethod performed during the computation on the GPU\n vec4 rungekutta4(vec4 x)\naccelerator of the expansion card 180 and is a more detailed {\nview of the computational box 330 in FIG. 3. FIG. 4 is a 60 canst vec4 kl - equation(x);\nrepresentation of one of several ways in which a system and canst vec4 k2 - equation(x + halfstep*kl);\nmethod for processing numerical techniques can be imple- canst vec4 k3 - equation(x + halfstep*k2);\n canst vec4 k4 - equation(x + integration step*k3);\nmented. return fl_6step*(kl + 2.0*(k2 + k3) + k4);\n With systems of equations that have complex interdepen- }\ndencies it is likely that the variable in some equation from 65 Void main(void)\na previous time step has to be used by some other equation {\nafter the new values of this variable are already computed\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 18 of 20\n\n\n US RE49,461 E\n 13 14\n LISTING 2-continued cycle (or less frequently as defined by the user), writing this\n output to disk 140, and displaying this output on the monitor\n output +- rungekutta4( output);\n 170. This frees the CPU 120 to execute other applications\n gl_FragColor - output;\n }\n and allows the expansion card to run at its full capacity\n 5 without being slowed down by extensive interactions with\n the CPU 120.\n The shader in Listing 2 can be executed on conventional 2. Minimizing data transfer between the expansion card\nvideo card. Using the controller 220 this code can be further 180 and the system bus 200. All of the information needed\noptimized, however. Since the integration step does not to perform the simulations will be stored on the expansion\nchange during the simulation, the step itself as well as the 10 card 180 and all simulations will take place on it. Further-\nhalfstep and 1/4 of the step can be computed once per more, whatever data transfer remains necessary will take\nsimulation, and updated in all shaders by a shader update place in parallel with the computation, thus reducing the\nprocedures 310, 326 discussed above. impact of this transfer on the performance.\n After all of the equations in the computational cycle are 3. New way to execute GPU programs (shaders). Previ-\ncomputed the main execution substream 403 on the control- 15 ously, the CPU 120 had full control over the order of\n!er 220 can switch 485 the reference pointers of the input and shader's execution and was required to produce specific\noutput portions of the texture memory bank 250. commands on every cycle to tell the GPU 240 which shader\n The two other substreams of execution on the controller to use. With the invention disclosed herein, shaders will\n220 are waiting (blocks 430 and 455, respectively) for this initially be stored on the shader memory bank 210 on the\nswitch to begin their execution. The Data Input Substream\n 20 expansion card 180 and will be sent to the GPU 240 for\n404 is controlling 440 the input of additional data from the execution by the general purpose controller 220 located on\nCPU 120. This is necessary in cases where the simulation is the expansion card.\nmonitoring the changing input, for example input from a 4. Multiple parallelisms. The GPU 240 is inherently\nvideo camera or other recording device in the real time. This\n parallel and is well suited to perform parallel computations.\nsubstream uploads new external input from the CPU 120 to\n In parallel with the GPU 240 performing the next calcula-\nthe texture memory bank 250 so it can be used by the main 25\ncomputational sub stream 403 on the next computational step tion, the controller 220 is uploading the data from the\nand waits for the next iteration 475. The Data Output previous calculation into main memory 130. Furthermore,\nSubstream 445 controls the output of simulation results to the CPU 120 at the same time uses uploaded previous results\nthe CPU 120 if requested by the user. This substream to save them onto disk 140 and to display them on the screen\nuploads the results of the previous step to the main RAM 30 through the system bus 200.\n130 so that the CPU 120 can save them on disk 140 or show 5. Reuse of existing and affordable technology. All hard-\nthem on the results display 313 and waits for the next ware used in the invention and mentioned here-in are based\niteration 460. on currently available and reliable components. Further\n Since the Computational Substream 403 determines the advance of these components will provide straightforward\ntiming of input 440 and output 445 data transfers, these data 35 improvements of the invention.\ntransfers are driven by the controller 220. To further reduce While this invention has been particularly shown and\nthe data transfer overhead (and disk 140 overhead also) the described with references to preferred embodiments thereof,\ncontroller 220 initiates transfer only after selected compu- it will be understood by those skilled in the art that various\ntational steps. For example, if the experimental data that is changes in form and details may be made therein without\nsimulated was recorded every 10 milliseconds (msec) and\n 40 departing from the scope of the invention encompassed by\nthe simulation for better precision was computed every 1 the appended claims.\nmsec, then only every tenth result has to be transferred to\nmatch the experimental frequency.\n This solution stores two copies of output data, one in the What is claimed is:\nexpansion card texture memory bank 250 and another in the [1. A computer system, comprising:\nsystem RAM 130. The copy in the system RAM 130 is 45 a central processing unit to receive input data;\naccessed twice: for disk I/O and screen visualization 313. An main memory, operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a unit via a bus, to store the input data received by the\ndirect read access to the onboard texture memory bank 250 central processing unit;\nby mapping the memory of the hardware onto a global an accelerator, operably coupled to the central processing\nmemory space. The alternative solution will double the 50 unit and the first memory via the bus, to receive at least\ncommunication through the local bus 190. Since the goal a portion of the input data from the main memory, the\ndiscussed herein is reducing the information transfer through accelerator comprising:\nthe local bus 190, the former solution is favored. at least one graphics processing unit to perform a\n The main substream 403 determines if this is the last sequence of computations on the at least a portion of\niteration 487. If it is the last iteration, the controller 220 55 the input data so as to generate output data, inter-\nwaits for the all of the execution substreams to finish 490 mediate computations in the sequence of computa-\nand then returns the control to the CPU 120, otherwise it tions yielding intermediate results; and\nbegins the next computational cycle.\n accelerator memory, operably coupled to the graphic\n This repeats through all of the computational cycles of the\n processing unit, to store the results of the plurality of\nsimulation.\n 60 sequential computations; and\n CONCLUSION a controller, operably coupled to the at least one graphics\n processing unit and the accelerator memory, to transfer\n This GPU accelerator system offers the following poten- the at least a portion of the input data into the accel-\ntial advantages: erator memory, and to transfer at least a portion of the\n 1. Limited computations on the CPU 120. The CPU 120 65 output data from the accelerator memory to the main\nis only used for user input, sending information to the memory during performance of the sequence of com-\ncontroller 220, receiving output after each computational putations by the at least one graphic processing unit.]\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 19 of 20\n\n\n US RE49,461 E\n 15 16\n [2. The computer system of claim 1, wherein the central [13. The method of claim 12, further comprising:\nprocessing unit is configured to receive the input data in storing the input data in the main memory in response to\nresponse to a user interaction.] a user interaction.]\n [3. The computer system of claim 1, wherein: [14. The method of claim 12, further comprising:\n the central processing unit is configured to receive the 5 receiving the input data at a first rate; and\n input data at a first rate; and wherein (A) comprises performing the sequence of com-\n the at least one graphics processing unit is configured to putations at a second rate different than the first rate.]\n perform the sequence of computations at a second rate [15. The method of claim 12, wherein (A) comprises:\n different than the first rate.] generating an output representative of an output of at least\n 10\n [4. The computer system of claim 1, wherein the main one neuron in an artificial neural network.]\nmemory is configured to store a copy of the output data [16. The method of claim 12, wherein (C) comprises:\nstored in the accelerator memory.] transferring the second portion of the output data from the\n [5. The computer system of claim 1, wherein an output of accelerator memory to the main memory without trans-\nat least one computation in the sequence of computations 15 ferring any of the intermediate results of the plurality of\nrepresents an output of at least one neuron in an artificial sequential computations from the accelerator memory\nneural network.] to the main memory so as to reduce data transfer via the\n [6. The computer system of claim 1, wherein accelerator bus.]\nmemory comprises: [17. The method of claim 12, wherein (C) comprises:\n a first memory bank to store parameters common to all of 20 transferring the second portion of the output data from the\n the computations in the sequence of computations; and accelerator memory to the main memory after the GPU\n a second memory bank to store data specific to at least one has begun to perform another sequence of computa-\n computation in the sequence of computations.] tions.]\n [7. The computer system of claim 1, wherein the control- [18. The method of claim 17, wherein (C) further com-\nler is configured to transfer the output data from the accel- 25 prises:\nerator memory to the main memory without transferring any initiating transfer of the second portion of the output data\nof the intermediate results from the accelerator memory to in parallel with performance of at least one computa-\nthe main memory so as to reduce data transfer via the bus.] tion in the other sequence of computations.]\n [8. The computer system of claim 1, wherein the control- [19. The method of claim 12, further comprising:\nler is configured to transfer at least a portion of the output 30 acquiring the input data in real time with at least one of\ndata from the accelerator memory to the main memory after a video camera, a microphone, or a cell recording\nthe at least one graphics processing unit has begun to electrode operably coupled to the CPU.]\nperform another sequence of computations.] [20. The method of claim 12, further comprising:\n [9. The computer system of claim 8, wherein the control- 35 storing parameters common to all of the computations in\n!er is configured to initiate transfer of the at least a portion the sequence of computations in a first memory bank in\nof the input data and to transfer the at least a portion of the the accelerator memory; and\noutput data in parallel with performance of at least one storing data specific to at least one computation in the\ncomputation in the other sequence of computations by the at sequence of computations in a second memory bank in\nleast one graphics processing unit.] 40 the accelerator memory.]\n [10. The computer system of claim 1, wherein the con- 21. A method of executing computations representing an\ntroller is configured to control execution of the sequence of artificial neural network on a computer system comprising\ncomputations by the at least one graphics processing unit.] at least one central processing unit (CPU), a processing\n [11. The computer system of claim 1, further comprising: unit, a first memory partition, and a second memory parti-\n at least one of a video camera, a microphone, or a cell 45 tion, the method comprising:\n recording electrode, operably coupled to the central executing, by the at least one CPU, a user interaction\n processor unit, to acquire the input data in real time.] stream, the user interaction stream controlling transfer\n [12. A method of performing a sequence of computations of inputs to the artificial neural network to the first\non a computer system comprising a central processing unit memory partition and the second memory partition;\n(CPU), a main memory operably coupled to the central 50 executing, by the processing unit, a computational stream,\nprocessing unit via a bus, an accelerator operably coupled to the computational stream controlling data exchange\nthe CPU and the main memory via the bus, the accelerator between the user interaction stream and the computa-\ncomprising a graphics processing unit (GPU) and an accel- tional stream during execution of the computations\nerator memory, the method comprising: representing the artificial neural network;\n (A) performing, by the GPU, the sequence of computa- 55 shifting control of a data exchange between the user\n tions on a first portion of the input data so as to generate interaction stream and the computational stream to the\n a first portion of the output data, intermediate compu- computational stream in response to starting execution\n tations in the sequence of computations yielding inter- of the computations representing the artificial neural\n mediate results; network;\n (B) in parallel with performing the sequence of compu- 60 shifting control of the data exchange between the user\n tations by the GPU in (A), transferring a second portion interaction stream and the computational stream to the\n of the input data from the main memory to the accel- user interaction stream in response to completion or\n erator via the bus; and interruption of the computations representing the arti-\n (C) in parallel with performing the sequence of compu- ficial neural network;\n tations by the GPU in (A), transferring a second portion 65 queueing a user command received by the user interaction\n of the output data from the accelerator memory to the stream during execution of the computations represent-\n main memory via the bus.] ing the artificial neural network; and\n\f Case 7:26-mc-00318-LS Document 6-5 Filed 08/18/26 Page 20 of 20\n\n\n US RE49,461 E\n 17 18\n executing the user command during execution of the a second memory partition;\n computations representing the artificial neural network at least one central processing unit (CPU), operably\n at times determined by the computational stream. coupled to the camera, the first memory partition, and\n 22. The method of claim 21, wherein the user interaction the second memory partition, to execute a user inter-\nstream controls the data exchange between the user inter- 5 action stream, the user interaction stream controlling\naction stream and the computational stream outside of transfer of the input data acquired by the camera to the\nexecution of the computations representing the artificial first memory partition and the second memory partition\nneural network. during execution of the computations representing the\n 23. The method of claim 21, wherein executing the user artificial neural network;\ninteraction stream comprises: 10\n a processing unit, operably coupled to the first memory\n controlling setting and editing of computational elements\n partition, the second memory partition, and the at least\n of the computations representing the artificial neural\n one CPU, to execute a computational stream, the\n network\n 24. The method of claim 21, wherein executing the user computational stream controlling transfer of the input\ninteraction stream comprises: 15\n data from the first memory partition and the second\n controlling setting and editing of parameters of the com- memory partition during execution of the computations\n putations representing the artificial neural network. representing the artificial neural network, the execution\n 25. The method of claim 21, wherein executing the user of the computations representing the artificial neural\ninteraction stream comprises: network occurring while the camera is acquiring the\n controlling setting and editing of parameters of the inputs 20 input data; and\n to the artificial neural network. a controller, operably coupled to the at least one CPU and\n 26. The method of claim 21, wherein executing the user the processing unit, to queue user interactions received\ninteraction stream comprises: by the user interaction stream during the execution of\n specifying an output to be saved to disk and/or displayed the computations representing the artificial neural net-\n on a screen. 25\n work for performance at times selected to avoid data\n 27. The method of claim 21, wherein executing the user corruption.\ninteraction stream comprises: 34. The system of claim 33, wherein the user interactions\n parsing elements to be used in the computations repre- cause interruption of the computations representing the\n senting the artificial neural network. artificial neural network.\n 28. The method of claim 27, wherein the processing unit 30 35. The system of claim 33, wherein the user interactions\ncomprises a graphics processing unit ( GPU) and executing cause a change in inputs to the artificial neural network.\nthe user interaction stream further comprises: 36. The system of claim 33, wherein the user interactions\n converting the elements into GPU programs. cause a change in display properties of an output of the\n 29. The method of claim 28, wherein executing the user computations representing the artificial neural network.\ninteraction stream comprises: 35\n 3 7. The system of claim 33, wherein the controller is\n compiling the GPU programs. configured to request the input data during the execution of\n 30. The method of claim 29, wherein executing the user the computations representing the artificial neural network.\ninteraction stream comprises: 38. The method of claim 21, wherein the user command\n transferring the GPU programs to the second memory causes interruption of the computations representing the\n partition. 40\n artificial neural network.\n 31. The method of claim 21, further comprising: 39. The method of claim 21, wherein the user command\n executing, by the at least one CPU, a data output stream, causes a change in the inputs to the artificial neural net-\n the data output stream controlling transfer of outputs of work.\n the computations representing the artificial neural net- 40. The method of claim 21, wherein the user command\n work to disk. 45\n causes a change in display properties of an output of the\n 32. The method of claim 21, further comprising: computations representing the artificial neural network.\n generating the inputs with a video camera during execu- 41. The method of claim 21, wherein, during execution of\n tion of the computations. the computations representing the artificial neural network,\n 33. A system for executing computations representing an the computational stream controls the data exchange\nartificial neural network, the system comprising: 50\n between the user interaction stream and the computational\n a camera to acquire input data for the artificial neural stream by requesting the inputs to the artificial neural\n network; network.\n a first memory partition; * * * * *\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:43.769809-07:00","document_number":"6","attachment_number":5,"pacer_doc_id":"181037220272","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 4","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554426/","id":490554426,"tags":[],"absolute_url":"/docket/74659430/6/6/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.228401-07:00","date_modified":"2026-08-23T09:39:48.897627-07:00","sha1":"85dc2f522459e775a4870bcfd7292d7c4490c0df","page_count":27,"file_size":3012734,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.6.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.6.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 1 of 27\n\n\n\n\n EXHIBIT\n\n 5\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 2 of 27\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n\n Plaintiff, Civil Action No. 7:24-cv-00221-ADA-DTG\n\n v. JURY TRIAL DEMANDED\n\nNVIDIA CORPORATION,\n\n Defendant.\n\n\n PLAINTIFF NEURAL AI, LLC\u2019S NOTICE OF SERVICE\n OF SUBPOENA TO TESLA, INC.\n\n PLEASE TAKE NOTICE that Plaintiff Neural AI, LLC (\u201cNeural AI\u201d) will serve (1) a\n\nSubpoena to Produce Documents, Information, or Objects on Tesla, Inc. (\u201cTesla\u201d), attached hereto\n\nas Attachment 1, and (2) a Subpoena for Testimony on Tesla, attached as Attachment 2.\n\nDated: June 24, 2026\n\n\n Respectfully submitted,\n\n /s/ Emily Portuguese\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n\fCase 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 3 of 27\n\n\n\n\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 5125174\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 4 of 27\n\n\n\n\n CERTIFICATE OF SERVICE\n\n\n I certify that on June 24, 2026 true and correct copies of Neural AI\u2019s Subpoena to\n\nProduce Documents, Information, or Objects and Subpoena for Testimony on Tesla has been\n\nserved all counsel of record electronically.\n\n\n /s/ Emily Portuguese\n Emily Portuguese\n\fCase 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 5 of 27\n\n\n\n\n Attachment 1\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 6 of 27\nAO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action\n\n\n UNITED STATES DISTRICT COURT\n for the\n __________\n Western District of __________\n Texas\n\n Neural Al, LLC )\n Plaintiff )\n v. ) Civil Action No. 7:24-cv-00221-ADA-DTG\n NVIDIA Corporation )\n ) (If the action is pending in another district, state where:\n Defendant ) __________ District of __________ )\n\n SUBPOENA TO PRODUCE DOCUMENTS, INFORMATION, OR OBJECTS\n OR TO PERMIT INSPECTION OF PREMISES IN A CIVIL ACTION\n Tesla,\nTo: Google Inc.\n LLC\n\n\n\n -\n c/o C T Corporation\n clo Corporation System, 1999\n Service Company BryanLawyers\n OBA CSC- St., Suite\n Inc,900, Dallas,\n 211 E. TX 75201\n 7th Street, Suite 620, Austin, TX 78701\n\n Production: YOU ARE COMMANDED to produce at the time, date, and place set forth below the following\ndocuments, electronically stored information, or objects, and permit their inspection, copying, testing, or sampling of the\nmaterial: See Exhibit A\n\n\n\n Place: Planet Depos- Downtown Austin clo Lexitas Legal, Date and Time:\n 100 Congress Ave, Ste. 2000, Austin, TX 78701\n 0711412026 11 :59 pm\n\n\n 0 Inspection of Premises: YOU ARE COMMANDED to permit entry onto the designated premises, land, or\nother property possessed or controlled by you at the time, date, and location set forth below, so that the requesting party\nmay inspect, measure, survey, photograph, test, or sample the property or any designated object or operation on it.\n\n Place: Date and Time:\n\n\n\n\n The provisions of Fed. R. Civ. P. 45(c), relating to your protection as a person subject to a subpoena, and Rule\n45 (d) and (e), relating to your duty to respond to this subpoena and the potential consequences of not doing so, are\nattached.\n\n\nDate: 06/24/2026\n\n CLERK OF COURT\n OR\n Isl Emily Portuguese\n Signature of Clerk or Deputy Clerk Attorney\u2019s signature\n\n\nThe name, address, e-mail, and telephone number of the attorney representing (name of party)\nNeural Al LLC , who issues or requests this subpoena, are:\nEmily Portuguese, Susman Godfrey LLP, One Manhattan West, 50th Floor, New York, New York 10001\neportuguese@susmangodfrey.com, 212-729-2082\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 7 of 27\nAO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action (Page 2)\n\nCivil Action No. 7:24-cv-00221-ADA-DTG\n\n PROOF OF SERVICE\n (This section should not be filed with the court unless required by Fed. R. Civ. P. 45.)\n\n This subpoena for (name of individual and title, if any)\nwas received by me on (date) .\n\n 0 I served the subpoena by delivering a copy to the named person as follows:\n\n\n on (date) ; or\n\n 0 I returned the subpoena unexecuted because:\n .\n\n Unless the subpoena was issued on behalf of the United States, or one of its officers or agents, I have also\n tendered to the witness fees for one day\u2019s attendance, and the mileage allowed by law, in the amount of\n $ .\n\nMy fees are $ for travel and $ for services, for a total of $ 0.00 .\n\n\n I declare under penalty of perjury that this information is true.\n\n\nDate:\n Server\u2019s signature\n\n\n\n Printed name and title\n\n\n\n\n Server\u2019s address\n\nAdditional information regarding attempted service, etc:\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 8 of 27\n AO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action(Page 3)\n\n\n\n\n Federal Rule of Civil Procedure 45 (c), (d), and (e) (Effective 12/1/07)\n(c) Protecting a Person Subject to a Subpoena. (d) Duties in Responding to a Subpoena.\n (1) Avoiding Undue Burden or Expense; Sanctions. A party or (1) Producing Documents or Electronically Stored Information.\nattorney responsible for issuing and serving a subpoena must take These procedures apply to producing documents or electronically\nreasonable steps to avoid imposing undue burden or expense on a stored information:\nperson subject to the subpoena. The issuing court must enforce this (A) Documents. A person responding to a subpoena to produce\nduty and impose an appropriate sanction \u2014 which may include lost documents must produce them as they are kept in the ordinary\nearnings and reasonable attorney\u2019s fees \u2014 on a party or attorney course of business or must organize and label them to correspond to\nwho fails to comply. the categories in the demand.\n (2) Command to Produce Materials or Permit Inspection. (B) Form for Producing Electronically Stored Information Not\n (A) Appearance Not Required. A person commanded to produce Specified. If a subpoena does not specify a form for producing\ndocuments, electronically stored information, or tangible things, or electronically stored information, the person responding must\nto permit the inspection of premises, need not appear in person at the produce it in a form or forms in which it is ordinarily maintained or\nplace of production or inspection unless also commanded to appear in a reasonably usable form or forms.\nfor a deposition, hearing, or trial. (C) Electronically Stored Information Produced in Only One\n (B) Objections. A person commanded to produce documents or Form. The person responding need not produce the same\ntangible things or to permit inspection may serve on the party or electronically stored information in more than one form.\nattorney designated in the subpoena a written objection to (D) Inaccessible Electronically Stored Information. The person\ninspecting, copying, testing or sampling any or all of the materials or responding need not provide discovery of electronically stored\nto inspecting the premises \u2014 or to producing electronically stored information from sources that the person identifies as not reasonably\ninformation in the form or forms requested. The objection must be accessible because of undue burden or cost. On motion to compel\nserved before the earlier of the time specified for compliance or 14 discovery or for a protective order, the person responding must show\ndays after the subpoena is served. If an objection is made, the that the information is not reasonably accessible because of undue\nfollowing rules apply: burden or cost. If that showing is made, the court may nonetheless\n (i) At any time, on notice to the commanded person, the serving order discovery from such sources if the requesting party shows\nparty may move the issuing court for an order compelling production good cause, considering the limitations of Rule 26(b)(2)(C). The\nor inspection. court may specify conditions for the discovery.\n (ii) These acts may be required only as directed in the order, and (2) Claiming Privilege or Protection.\nthe order must protect a person who is neither a party nor a party\u2019s (A) Information Withheld. A person withholding subpoenaed\nofficer from significant expense resulting from compliance. information under a claim that it is privileged or subject to\n (3) Quashing or Modifying a Subpoena. protection as trial-preparation material must:\n (A) When Required. On timely motion, the issuing court must (i) expressly make the claim; and\nquash or modify a subpoena that: (ii) describe the nature of the withheld documents,\n (i) fails to allow a reasonable time to comply; communications, or tangible things in a manner that, without\n (ii) requires a person who is neither a party nor a party\u2019s officer revealing information itself privileged or protected, will enable the\nto travel more than 100 miles from where that person resides, is parties to assess the claim.\nemployed, or regularly transacts business in person \u2014 except that, (B) Information Produced. If information produced in response to a\nsubject to Rule 45(c)(3)(B)(iii), the person may be commanded to subpoena is subject to a claim of privilege or of protection as trial-\nattend a trial by traveling from any such place within the state where preparation material, the person making the claim may notify any\nthe trial is held; party that received the information of the claim and the basis for it.\n (iii) requires disclosure of privileged or other protected matter, if After being notified, a party must promptly return, sequester, or\nno exception or waiver applies; or destroy the specified information and any copies it has; must not use\n (iv) subjects a person to undue burden. or disclose the information until the claim is resolved; must take\n (B) When Permitted. To protect a person subject to or affected by reasonable steps to retrieve the information if the party disclosed it\na subpoena, the issuing court may, on motion, quash or modify the before being notified; and may promptly present the information to\nsubpoena if it requires: the court under seal for a determination of the claim. The person\n (i) disclosing a trade secret or other confidential research, who produced the information must preserve the information until\ndevelopment, or commercial information; the claim is resolved.\n (ii) disclosing an unretained expert\u2019s opinion or information that\ndoes not describe specific occurrences in dispute and results from (e) Contempt. The issuing court may hold in contempt a person\nthe expert\u2019s study that was not requested by a party; or who, having been served, fails without adequate excuse to obey the\n (iii) a person who is neither a party nor a party\u2019s officer to incur subpoena. A nonparty\u2019s failure to obey must be excused if the\nsubstantial expense to travel more than 100 miles to attend trial. subpoena purports to require the nonparty to attend or produce at a\n (C) Specifying Conditions as an Alternative. In the circumstances place outside the limits of Rule 45(c)(3)(A)(ii).\ndescribed in Rule 45(c)(3)(B), the court may, instead of quashing or\nmodifying a subpoena, order appearance or production under\nspecified conditions if the serving party:\n (i) shows a substantial need for the testimony or material that\ncannot be otherwise met without undue hardship; and\n (ii) ensures that the subpoenaed person will be reasonably\ncompensated.\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 9 of 27\n\n\n\n\n EXHIBIT A\n\n DEFINITIONS AND INSTRUCTIONS\n\n 1. The term \u201cNVIDIA GPUs\u201d means the Hopper, Ada Lovelace, Ampere, Turing,\n\nVolta, Pascal, Maxwell, Jetson, and Blackwell architectures of NVIDIA graphics processing units.\n\nFor avoidance of doubt, those architectures include the following devices: DGX line of\n\nsupercomputers and servers (including at least DGX B300, DGX B200, DGX GB200, DGX\n\nGB300, DGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200,\n\nDGX H200, DGX H100, DGX BasePOD, DGX SuperPOD with H200, DGX A100), HGX line of\n\nsupercomputers and servers (including at least HGX B300, HGX B200, HGX H100, HGX H200,\n\nEos SuperPOD), OVX line of supercomputers and servers (including at least OVX L40S), EGX\n\nline of supercomputers and servers (including at least EGX Server with Quadro RTX A6000, EGX\n\nServer with A40, EGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000),\n\nGB300 NVL72, GB200 NVL72; Nvidia\u2019s GPU accelerators and superchips, including those with\n\nNVIDIA\u2019s Blackwell, Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, and Maxwell GPU\n\narchitectures, including at least, RTX PRO 6000 Server Edition, RTX PRO 6000 Workstation,\n\nRTX PRO 6000 Max-Q Workstation, RTX PRO 5000, RTX PRO 4500, RTX PRO 4000, RTX\n\nPRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, GB300, GB200, H100, H200,\n\nGH200, GH100, L40, L40S, L4, RTX 6000, RTX 6000 Ada, RTX 5000, Ada, RTX 4500 Ada,\n\nRTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3000, RTX 2000, RTX 1000, RTX 500, RTX\n\n4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060 Ti,\n\nand RTX 4060, GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce RTX\n\n4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU, A100,\n\nA40, A30, A16, A10, A2, A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 10 of 27\n\n\n\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500, GeForce RTX\n\n3090 Ti, GeForce RTX 3090, GeForce RTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti,\n\nGeForce RTX 3070, GeForce RTX 3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB),\n\nGeForce RTX 3050 (6 GB), GeForce RTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop\n\nGPU, GeForce RTX 3070 Ti Laptop GPU, GeForce RTX 3070 Laptop GPU, GeForce RTX 3060\n\nLaptop GPU, GeForce RTX 3050 Ti Laptop GPU, GeForce RTX 3050 Laptop GPU, GeForce\n\nMX570 Laptop GPU, Tesla T4 GPUs, Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 8000,\n\nQuadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000,\n\nT1000 8GB, T1200, Quadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB, Titan\n\nRTX, GeForce RTX 2080 Ti, GeForce RTX 2080, Super, GeForce RTX 2080, GeForce RTX 2070\n\nSuper, GeForce RTX 2070, GeForce RTX 2060 Super, GeForce RTX 2060, GeForce RTX 2500,\n\nGeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630, GeForce MX550, GeForce MX450, GeForce MX430, Tesla V100,\n\nQuadro GV100, Titan V GPU, Tesla P100, P40, P4, Quadro GP100, Quadro P6000, Quadro\n\nP5200, Quadro P5000, Quadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro\n\nP2200, Quadro P2000, Quadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500,\n\nQuadro P400, Titan Xp, Titan X, GeForce GTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070\n\nTi, GeForce GTX 1070, GeForce GTX 1060, GeForce GTX 1050 Ti, GeForce GTX 1050,\n\nGeForce MX300, GeForce MX200, GeForce MX150, Tesla M60, M40, M10, Quadro M6000\n\n24GB, Quadro M6000 (12GB), Quadro M5000, Quadro M5000M, Quadro M5500, Quadro\n\nM4000, Quadro M4000M, Quadro M3000M, Quadro M2200, Quadro M2000, Quadro M2000M,\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 11 of 27\n\n\n\n\nQuadro M1200, Quadro M1000M, Quadro M620, Quadro M600M, Quadro M520, Quadro\n\nM500M, NVS 810, Tesla M6, GTX Titan X, GeForce GTX 980Ti, GeForce GTX 980, GeForce\n\nGTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M,\n\nGeForce GTX 960M, GeForce GTX 950M, GeForce GTX 750 Ti, GeForce GTX 750, GeForce\n\nMX130, and GeForce MX110; and Jetson modules, including at least the Jetson Thor Series,\n\nJetson Thor, Jetson T5000, Jetson T4000, Jetson AGX Orin Series, Jetson AGX Orin Developer\n\nKit, Jetson AGX Orin 64GB, Jetson AGX Orin Industrial, Jetson AGX Orin 32GB, Jetson Orin\n\nNX Series, Jetson Orin NX 16GB, Jetson Orin NX 8GB, Jetson Orin Nano Series, Jetson Orin\n\nNano Super Developer Kit, Jetson Orin Nano 8GB, Jetson Orin Nano 4GB, Jetson AGX Xavier\n\nSeries, Jetson AGX Xavier Industrial, Jetson AGX Xavier 64GB, Jetson AGX Xavier 32GB,\n\nJetson Xavier NX Series, Jetson Xavier NX 16GB, Jetson Xavier NX 8GB, Jetson TX2 Series,\n\nJetson TX2i, Jetson TX2, Jetson TX2 4GB, Jetson TX2 NX, Jetson Nano, any and all variations\n\nof the aforementioned products (including at least products having different options for number of\n\nGPUs).\n\n 2. The terms \u201cand\u201d and \u201cor\u201d are not intended to be read disjunctively but rather\n\nconjunctively unless the context of a particular request clearly indicates otherwise. \u201cOr\u201d should be\n\nunderstood to include and encompass \u201cand\u201d; and \u201cand\u201d should be understood to include and\n\nencompass \u201cor.\u201d\n\n 3. The terms \u201cany\u201d or \u201ceach\u201d should be understood to include and encompass \u201call.\u201d\n\n 4. The terms \u201cconcerning,\u201d \u201crelated to\u201d or \u201crelating to\u201d, and \u201cregarding\u201d and any\n\nvariation of these terms mean analyzing, alluding to, concerning, considering, commenting on,\n\nconsulting, comprising, containing, contradicting, describing, dealing with, discussing,\n\nestablishing, evidencing, identifying, involving, noting, recording, reporting on, relating to,\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 12 of 27\n\n\n\n\nreflecting, referring to, regarding, stating, showing, studying, mentioning, memorializing, or\n\npertaining to, directly or indirectly, in whole or in part.\n\n 5. The term \u201cCPU(s)\u201d means Central Processing Unit(s).\n\n 6. The term \u201cDocument(s)\u201d shall have the broadest meaning possible under Federal\n\nRules 26 and 34 and shall include without limitation: documents, Electronically Stored\n\nInformation, communications in written, electronic, and recorded form, and tangible things. A\n\ndraft or non-identical copy of a document shall be considered a separate document within the\n\nmeaning of the term \u201cdocument.\u201d Any comment, notation, or other marking shall be sufficient to\n\ndistinguish documents that are otherwise similar in appearance and to make them separate\n\ndocuments for purposes of your response. Any preliminary form, intermediate form, superseded\n\nversion, or amendment of any document is to be considered a separate document.\n\n 7. The term \u201cGPU(s)\u201d means Graphics Processing Unit(s).\n\n 8. The terms \u201cinclude\u201d and \u201cincluding\u201d mean including without limitation.\n\n 9. The term \u201cNVIDIA\u201d means Defendant NVIDIA Corporation, its predecessors,\n\npresent and former directors, officers, accountants, affiliates, attorneys, partners, managers, agents,\n\nemployees, representatives, in-house and outside counsel, and any other person or entity acting on\n\nbehalf of or under control of Defendant NVIDIA Corporation.\n\n 10. The term \u201cperson(s)\u201d means and includes natural persons and formal or informal\n\nentities and organizations, including public and private corporations, partnerships, professional\n\ncorporations, limited liability companies, business trusts, banking institutions, associations, firms,\n\njoint ventures, commissions, bureaus, departments, and any other legal entity, including\n\nany divisions, subsidiaries, departments, and other units thereof.\n\n 11. The term \u201cSource Code\u201d means human-readable instructions written in a\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 13 of 27\n\n\n\n\nprogramming language, including all comments, annotations, declarations, functions, classes, and\n\nother components used to define the behavior of a software program. For purposes of these\n\nrequests, \u201cSource Code\u201d includes all associated files necessary to understand, compile, and execute\n\nthe code, such as scripts, header files, makefiles, configuration files, and documentation. Unless\n\notherwise stated, \u201cSource Code\u201d includes all versions and revisions relevant to the time periods\n\nand subject matter described in each interrogatory.\n\n 12. The terms \u201cYou\u201d or \u201cYour\u201d refer to Tesla, Inc., including but not limited to its\n\npredecessors, successors, parents, subsidiaries, divisions, affiliates, and all past or present\n\ndirectors, officers, partners, managers, employees, contractors, agents, representatives,\n\naccountants, consultants, in-house and outside counsel, and any other person or entity acting or\n\npurporting to act on its behalf or subject to its control. This definition expressly includes, without\n\nlimitation, any Tesla parent, subsidiary, affiliate, or other related entity that has used, licensed,\n\ndeployed, evaluated, or integrated NVIDIA GPUs or software.\n\n 13. The use of the singular form of any word includes the plural and vice versa.\n\n 14. These Requests seek the production of all documents, electronically stored\n\ninformation (\u201cESI\u201d), and tangible things in your possession, custody, or control, as that phrase is\n\nused in Federal Rule of Civil Procedure 34, as of the date of compliance with this subpoena and\n\nthat come into your possession, custody, or control at any time prior to production. This includes\n\nmaterials held by You directly, as well as by Your affiliates, subsidiaries, agents, representatives,\n\nor any other person or entity acting on Your behalf.\n\n 15. If You are aware of the existence (past or present) of any responsive documents,\n\nESI, or tangible items that are not in Your current possession, custody, or control, You must\n\nidentify such materials and provide:\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 14 of 27\n\n\n\n\n a. A description of the item(s);\n\n b. The name and contact information of the person or entity currently believed\n\n to have possession, custody, or control; and\n\n c. The reason You are unable to produce the material(s).\n\n 16. If You believe that no responsive documents, ESI, or tangible things exist in\n\nresponse to a particular request, You must state so in writing with respect to that request.\n\n 17. If You withhold any document, ESI, or portion thereof based on a claim of attorney-\n\nclient privilege, work product doctrine, or any other legal protection, You must produce a privilege\n\nlog that complies with Federal Rule of Civil Procedure 26(b)(5).\n\n 18. You must produce all documents and ESI: as they are kept in the usual course of\n\nbusiness or organized and labeled to correspond to the categories in this subpoena, as required\n\nunder FRCP 34(b)(2)(E); in their native electronic format (with original metadata intact) wherever\n\npossible, or as searchable, OCR-scanned PDFs with corresponding load files; with complete\n\nfamily groupings (e.g., attachments must be produced with their parent emails/documents); and in\n\nthe same folders or directories in which they were maintained, preserving original file structures\n\nand naming conventions.\n\n 19. You must maintain and produce a record of the source of each document or ESI\n\nitem produced, including:\n\n a. The file path or directory location;\n\n b. The name of the custodian (individual, team, or department) from whose\n files the document was collected;\n\n c. The system or platform (e.g., email server, shared drive, cloud service) from\n which the document was obtained.\n\nThis information may be provided in metadata load files, a source log, or other mutually agreed\n\nformat. This instruction is consistent with Federal Rule of Civil Procedure 34(b)(2)(E) and\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 15 of 27\n\n\n\n\nproportional discovery principles under Rule 26(b)(1). If any of the above information is not\n\nreasonably available or unduly burdensome to collect, You must so state and explain the basis for\n\nthat assertion.\n\n 20. For all ESI, You must preserve and produce standard metadata fields, including but\n\nnot limited to: filename, filepath, author, date created, date last modified, recipients, sender,\n\nsubject line (for emails), and document type. You must not degrade or alter metadata through\n\nprocessing or production.\n\n 21. You are under a continuing obligation to supplement or correct Your production if\n\nYou discover or obtain additional responsive materials prior to the close of discovery or the\n\nresolution of this matter.\n\n 22. Unless otherwise stated, the relevant period is from September 13, 2018 to the\n\npresent.\n\n 23. Unless otherwise stated, all requests herein are limited to documents, electronically\n\nstored information, and tangible things that relate to NVIDIA GPU-Acceleration Hardware or\n\nNVIDIA GPU-Acceleration Software that were:\n\n a. purchased, acquired, licensed, used, implemented, deployed, tested, or\n evaluated within the United States, or\n\n b. purchased, acquired, licensed, or used for the purpose of supporting,\n enabling, or operating any facility, system, team, data center, personnel,\n product, service, customer, or business activity located in or directed toward\n the United States.\n\nThis instruction is intended to encompass both U.S.-based activity and non-U.S. activity that\n\ndirectly supports or enables U.S. operations or usage.\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 16 of 27\n\n\n\n\n REQUESTS FOR PRODUCTION OF DOCUMENTS\n\n1. Documents sufficient to identify all software, frameworks, libraries, APIs, scripts, Source\n\n Code, configuration files, and custom code You use to perform computations on NVIDIA\n\n GPUs.\n\n2. Documents sufficient to show whether You use NVIDIA\u2019s Aerial, Clara Parabricks, cuBLAS,\n\n cuDNN, cuFFT, cuQuantum, cuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac,\n\n Isaac Lab, Maxine, Memory Map, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo,\n\n PyTorch, RAPIDS, Riva, Runtime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream),\n\n or any other NVIDIA software as part of computations You perform using NVIDIA GPUs.\n\n3. Documents sufficient to show whether You use sample Source Code provided by NVIDIA as\n\n part of computations You perform using NVIDIA GPUs.\n\n4. Documents sufficient to show whether and how any software You use to perform\n\n computations on NVIDIA GPUs calls, invokes, interfaces with, wraps, depends on, sits on top\n\n of, modifies, extends, or implements functionality provided by CUDA, cuDNN, TensorRT,\n\n CUDA libraries, CUDA drivers, CUDA runtime, CUDA applications or frameworks or any\n\n other NVIDIA software.\n\n5. Documents sufficient to show the architecture, design, data flow, control flow, and execution\n\n flow of any system in which You use NVIDIA GPUs to perform computations, including\n\n diagrams, technical specifications, design documents, Powerpoints, slide decks, internal and\n\n external presentations, Source Code, configuration files, build files, deployment files, runtime\n\n logs, and profiler traces.\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 17 of 27\n\n\n\n\n6. Documents sufficient to show whether computations You performed using NVIDIA GPUs\n\n involved artificial neural networks, neural-network computational layers or computations with\n\n outputs as inputs for other neurons or layers.\n\n7. Documents sufficient to show whether You use a pointer to data stored in memory (e.g.\n\n memory bank or partition), using as an input to a subsequent computational layer the pointer\n\n to output data from a GPU computation, using pointers in neural network computations,\n\n swapping an input pointer with the pointer to data output from a GPU computation, pointer\n\n swapping, pointer rotation, buffer swapping, ping-pong buffers, double or triple buffering,\n\n alternating input/output buffers, or any other technique in which output data from one\n\n computation, layer, iteration, time step, or cycle becomes input data for a later computation,\n\n layer, iteration, time step, or cycle.\n\n8. Documents sufficient to show whether You store input data, output data, intermediate results,\n\n tensors, activations, weights, parameters, internal variables, GPU programs, kernels, textures,\n\n shaders, or other GPU-computation-related data in separate, partitioned, logical, physical,\n\n first/second, input/output, texture, shader, shared, global, device, host, pinned, GPU RAM,\n\n GPU cache(s), or unified memory regions (shared by CPU and GPU) when performing\n\n computations using NVIDIA GPUs.\n\n9. Documents sufficient to show how input data is received, acquired, stored, transferred, copied,\n\n streamed, prefetched, staged, queued, or loaded from CPU memory, host memory, system\n\n memory, storage, sensors, cameras, or other input sources to NVIDIA GPU memory\u2014\n\n including GPU RAM (e.g. GPU HBM, GDDR) and/or GPU cache(s)\u2014before, during, or in\n\n parallel with computations You perform using NVIDIA GPUs.\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 18 of 27\n\n\n\n\n10. Documents sufficient to show how output data from a GPU computation(s), intermediate\n\n results of GPU computations, tensors, buffers, activations, variables, or other computation\n\n results are stored, transferred, copied, streamed, written back, returned, accumulated, reused,\n\n or made available including asynchronously from NVIDIA GPU memory to CPU memory,\n\n host memory, system memory, storage, display, network, or another memory location before,\n\n during, or in parallel with computations You perform using NVIDIA GPUs\u2014and also\n\n including in the opposite direction, copying data from CPU or host or other memory to a queue\n\n for GPU computation while other GPU computations are occurring.\n\n11. Documents sufficient to show how computations You perform using NVIDIA GPUs are\n\n scheduled, ordered, controlled, queued, synchronized, parallelized, launched, interrupted,\n\n resumed, or executed, including through kernels, CUDA streams, CUDA graphs, events,\n\n threads, controllers, schedulers, compilers, runtimes, inference engines, run lists, run engines,\n\n or custom software.\n\n12. Documents sufficient to show whether and how user inputs, user commands, configuration\n\n changes, parameter changes, model changes, computational-element changes, input changes,\n\n interruptions, or display/output changes affect computations You perform using NVIDIA\n\n GPUs and/or queue them for GPU computation.\n\fCase 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 19 of 27\n\n\n\n\n Attachment 2\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 20 of 27\n AO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action\n\n\n UNITED STATES DISTRICT COURT\n for the\n Western District of __________\n __________ Texas\n\n NeuralAl,\n Nerual AI,LLC\n LLC )\n Plaintiff )\n v. ) Civil Action No. 7:24-cv-00221-ADA-DTG\n NVIDIA Corporation )\n )\n Defendant )\n\n SUBPOENA TO TESTIFY AT A DEPOSITION IN A CIVIL ACTION\n\n To: Tesla, Inc. Google LLC\n c/o C\n c/o T Corporation\n Corporation System,\n Service 1999 Bryan\n Company St., Suite\n DBA CSC- 900, Inc,\n Lawyers 211 TX\n Dallas, 75201\n E. 7th Street, Suite 620, Austin, TX 78701\n (Name of person to whom this subpoena is directed)\n\n Testimony: YOU ARE COMMANDED to appear at the time, date, and place set forth below to testify at a\n deposition to be taken in this civil action. If you are an organization, you must designate one or more officers, directors,\n or managing agents, or designate other persons who consent to testify on your behalf about the following matters, or\n those set forth in an attachment:\n See Exhibit A.\n\n\n Place: Planet Depos - Downtown Austin c/o Lexitas Legal, Date and Time:\n 100 Congress Ave, Ste. 2000, Austin, TX 78701 07/21/2026 9:00 am\n\n\n The deposition will be recorded by this method: audio, video, and stenographic means\n\n 0 Production: You, or your representatives, must also bring with you to the deposition the following documents,\n electronically stored information, or objects, and must permit inspection, copying, testing, or sampling of the\n material:\n\n\n\n\n The following provisions of Fed. R. Civ. P. 45 are attached \u2013 Rule 45(c), relating to the place of compliance;\n Rule 45(d), relating to your protection as a person subject to a subpoena; and Rule 45(e) and (g), relating to your duty to\n respond to this subpoena and the potential consequences of not doing so.\n\n Date: 06/24/2026\n CLERK OF COURT\n OR\n /s/ Emily Portuguese\n Signature of Clerk or Deputy Clerk Attorney\u2019s signature\n\n The name, address, e-mail address, and telephone number of the attorney representing (name of party)\n Neural Al, LLC , who issues or requests this subpoena, are:\n Emily Portuguese, Susman Godfrey LLP, One Manhattan West, 50th Floor, New York, New York 10001\neportuguese\u00aesusmangodfi'ey.com, 212-729-2082\n Notice to the person who issues or requests this subpoena\n If this subpoena commands the production of documents, electronically stored information, or tangible things before\n trial, a notice and a copy of the subpoena must be served on each party in this case before it is served on the person to\n whom it is directed. Fed. R. Civ. P. 45(a)(4).\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 21 of 27\nAO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action (Page 2)\n\nCivil Action No. 7:24-cv-00221-ADA-DTG\n\n PROOF OF SERVICE\n (This section should not be filed with the court unless required by Fed. R. Civ. P. 45.)\n\n I received this subpoena for (name of individual and title, if any)\non (date) .\n\n 0 I served the subpoena by delivering a copy to the named individual as follows:\n\n\n on (date) ; or\n\n 0 I returned the subpoena unexecuted because:\n .\n\n Unless the subpoena was issued on behalf of the United States, or one of its officers or agents, I have also\n tendered to the witness the fees for one day\u2019s attendance, and the mileage allowed by law, in the amount of\n $ .\n\nMy fees are $ for travel and $ for services, for a total of $ 0.00 .\n\n\n I declare under penalty of perjury that this information is true.\n\n\nDate:\n Server\u2019s signature\n\n\n\n Printed name and title\n\n\n\n\n Server\u2019s address\n\nAdditional information regarding attempted service, etc.:\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 22 of 27\n\nAO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action (Page 3)\n\n Federal Rule of Civil Procedure 45 (c), (d), (e), and (g) (Effective 12/1/13)\n(c) Place of Compliance. (i) disclosing a trade secret or other confidential research, development,\n or commercial information; or\n (1) For a Trial, Hearing, or Deposition. A subpoena may command a (ii) disclosing an unretained expert\u2019s opinion or information that does\nperson to attend a trial, hearing, or deposition only as follows: not describe specific occurrences in dispute and results from the expert\u2019s\n (A) within 100 miles of where the person resides, is employed, or study that was not requested by a party.\nregularly transacts business in person; or (C) Specifying Conditions as an Alternative. In the circumstances\n (B) within the state where the person resides, is employed, or regularly described in Rule 45(d)(3)(B), the court may, instead of quashing or\ntransacts business in person, if the person modifying a subpoena, order appearance or production under specified\n (i) is a party or a party\u2019s officer; or conditions if the serving party:\n (ii) is commanded to attend a trial and would not incur substantial (i) shows a substantial need for the testimony or material that cannot be\nexpense. otherwise met without undue hardship; and\n (ii) ensures that the subpoenaed person will be reasonably compensated.\n (2) For Other Discovery. A subpoena may command:\n (A) production of documents, electronically stored information, or (e) Duties in Responding to a Subpoena.\ntangible things at a place within 100 miles of where the person resides, is\nemployed, or regularly transacts business in person; and (1) Producing Documents or Electronically Stored Information. These\n (B) inspection of premises at the premises to be inspected. procedures apply to producing documents or electronically stored\n information:\n(d) Protecting a Person Subject to a Subpoena; Enforcement. (A) Documents. A person responding to a subpoena to produce documents\n must produce them as they are kept in the ordinary course of business or\n (1) Avoiding Undue Burden or Expense; Sanctions. A party or attorney must organize and label them to correspond to the categories in the demand.\nresponsible for issuing and serving a subpoena must take reasonable steps (B) Form for Producing Electronically Stored Information Not Specified.\nto avoid imposing undue burden or expense on a person subject to the If a subpoena does not specify a form for producing electronically stored\nsubpoena. The court for the district where compliance is required must information, the person responding must produce it in a form or forms in\nenforce this duty and impose an appropriate sanction\u2014which may include which it is ordinarily maintained or in a reasonably usable form or forms.\nlost earnings and reasonable attorney\u2019s fees\u2014on a party or attorney who (C) Electronically Stored Information Produced in Only One Form. The\nfails to comply. person responding need not produce the same electronically stored\n information in more than one form.\n (2) Command to Produce Materials or Permit Inspection. (D) Inaccessible Electronically Stored Information. The person\n (A) Appearance Not Required. A person commanded to produce responding need not provide discovery of electronically stored information\ndocuments, electronically stored information, or tangible things, or to from sources that the person identifies as not reasonably accessible because\npermit the inspection of premises, need not appear in person at the place of of undue burden or cost. On motion to compel discovery or for a protective\nproduction or inspection unless also commanded to appear for a deposition, order, the person responding must show that the information is not\nhearing, or trial. reasonably accessible because of undue burden or cost. If that showing is\n (B) Objections. A person commanded to produce documents or tangible made, the court may nonetheless order discovery from such sources if the\nthings or to permit inspection may serve on the party or attorney designated requesting party shows good cause, considering the limitations of Rule\nin the subpoena a written objection to inspecting, copying, testing, or 26(b)(2)(C). The court may specify conditions for the discovery.\nsampling any or all of the materials or to inspecting the premises\u2014or to\nproducing electronically stored information in the form or forms requested. (2) Claiming Privilege or Protection.\nThe objection must be served before the earlier of the time specified for (A) Information Withheld. A person withholding subpoenaed information\ncompliance or 14 days after the subpoena is served. If an objection is made, under a claim that it is privileged or subject to protection as trial-preparation\nthe following rules apply: material must:\n (i) At any time, on notice to the commanded person, the serving party (i) expressly make the claim; and\nmay move the court for the district where compliance is required for an (ii) describe the nature of the withheld documents, communications, or\norder compelling production or inspection. tangible things in a manner that, without revealing information itself\n (ii) These acts may be required only as directed in the order, and the privileged or protected, will enable the parties to assess the claim.\norder must protect a person who is neither a party nor a party\u2019s officer from (B) Information Produced. If information produced in response to a\nsignificant expense resulting from compliance. subpoena is subject to a claim of privilege or of protection as\n trial-preparation material, the person making the claim may notify any party\n (3) Quashing or Modifying a Subpoena. that received the information of the claim and the basis for it. After being\n notified, a party must promptly return, sequester, or destroy the specified\n (A) When Required. On timely motion, the court for the district where information and any copies it has; must not use or disclose the information\ncompliance is required must quash or modify a subpoena that: until the claim is resolved; must take reasonable steps to retrieve the\n information if the party disclosed it before being notified; and may promptly\n (i) fails to allow a reasonable time to comply; present the information under seal to the court for the district where\n (ii) requires a person to comply beyond the geographical limits compliance is required for a determination of the claim. The person who\nspecified in Rule 45(c); produced the information must preserve the information until the claim is\n (iii) requires disclosure of privileged or other protected matter, if no resolved.\nexception or waiver applies; or\n (iv) subjects a person to undue burden. (g) Contempt.\n (B) When Permitted. To protect a person subject to or affected by a The court for the district where compliance is required\u2014and also, after a\nsubpoena, the court for the district where compliance is required may, on motion is transferred, the issuing court\u2014may hold in contempt a person\nmotion, quash or modify the subpoena if it requires: who, having been served, fails without adequate excuse to obey the\n subpoena or an order related to it.\n\n\n For access to subpoena materials, see Fed. R. Civ. P. 45(a) Committee Note (2013).\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 23 of 27\n\n\n\n\n EXHIBIT A\n\n DEFINITIONS AND INSTRUCTIONS\n\n 24. The term \u201cNVIDIA GPUs\u201d means the Hopper, Ada Lovelace, Ampere, Turing,\n\nVolta, Pascal, Maxwell, Jetson, and Blackwell architectures of NVIDIA graphics processing units.\n\nFor avoidance of doubt, those architectures include the following devices: DGX line of\n\nsupercomputers and servers (including at least DGX B300, DGX B200, DGX GB200, DGX\n\nGB300, DGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200,\n\nDGX H200, DGX H100, DGX BasePOD, DGX SuperPOD with H200, DGX A100), HGX line of\n\nsupercomputers and servers (including at least HGX B300, HGX B200, HGX H100, HGX H200,\n\nEos SuperPOD), OVX line of supercomputers and servers (including at least OVX L40S), EGX\n\nline of supercomputers and servers (including at least EGX Server with Quadro RTX A6000, EGX\n\nServer with A40, EGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000),\n\nGB300 NVL72, GB200 NVL72; Nvidia\u2019s GPU accelerators and superchips, including those with\n\nNVIDIA\u2019s Blackwell, Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, and Maxwell GPU\n\narchitectures, including at least, RTX PRO 6000 Server Edition, RTX PRO 6000 Workstation,\n\nRTX PRO 6000 Max-Q Workstation, RTX PRO 5000, RTX PRO 4500, RTX PRO 4000, RTX\n\nPRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, GB300, GB200, H100, H200,\n\nGH200, GH100, L40, L40S, L4, RTX 6000, RTX 6000 Ada, RTX 5000, Ada, RTX 4500 Ada,\n\nRTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3000, RTX 2000, RTX 1000, RTX 500, RTX\n\n4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060 Ti,\n\nand RTX 4060, GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce RTX\n\n4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU, A100,\n\nA40, A30, A16, A10, A2, A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 24 of 27\n\n\n\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500, GeForce RTX\n\n3090 Ti, GeForce RTX 3090, GeForce RTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti,\n\nGeForce RTX 3070, GeForce RTX 3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB),\n\nGeForce RTX 3050 (6 GB), GeForce RTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop\n\nGPU, GeForce RTX 3070 Ti Laptop GPU, GeForce RTX 3070 Laptop GPU, GeForce RTX 3060\n\nLaptop GPU, GeForce RTX 3050 Ti Laptop GPU, GeForce RTX 3050 Laptop GPU, GeForce\n\nMX570 Laptop GPU, Tesla T4 GPUs, Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 8000,\n\nQuadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000,\n\nT1000 8GB, T1200, Quadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB, Titan\n\nRTX, GeForce RTX 2080 Ti, GeForce RTX 2080, Super, GeForce RTX 2080, GeForce RTX 2070\n\nSuper, GeForce RTX 2070, GeForce RTX 2060 Super, GeForce RTX 2060, GeForce RTX 2500,\n\nGeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630, GeForce MX550, GeForce MX450, GeForce MX430, Tesla V100,\n\nQuadro GV100, Titan V GPU, Tesla P100, P40, P4, Quadro GP100, Quadro P6000, Quadro\n\nP5200, Quadro P5000, Quadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro\n\nP2200, Quadro P2000, Quadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500,\n\nQuadro P400, Titan Xp, Titan X, GeForce GTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070\n\nTi, GeForce GTX 1070, GeForce GTX 1060, GeForce GTX 1050 Ti, GeForce GTX 1050,\n\nGeForce MX300, GeForce MX200, GeForce MX150, Tesla M60, M40, M10, Quadro M6000\n\n24GB, Quadro M6000 (12GB), Quadro M5000, Quadro M5000M, Quadro M5500, Quadro\n\nM4000, Quadro M4000M, Quadro M3000M, Quadro M2200, Quadro M2000, Quadro M2000M,\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 25 of 27\n\n\n\n\nQuadro M1200, Quadro M1000M, Quadro M620, Quadro M600M, Quadro M520, Quadro\n\nM500M, NVS 810, Tesla M6, GTX Titan X, GeForce GTX 980Ti, GeForce GTX 980, GeForce\n\nGTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M,\n\nGeForce GTX 960M, GeForce GTX 950M, GeForce GTX 750 Ti, GeForce GTX 750, GeForce\n\nMX130, and GeForce MX110; and Jetson modules, including at least the Jetson Thor Series,\n\nJetson Thor, Jetson T5000, Jetson T4000, Jetson AGX Orin Series, Jetson AGX Orin Developer\n\nKit, Jetson AGX Orin 64GB, Jetson AGX Orin Industrial, Jetson AGX Orin 32GB, Jetson Orin\n\nNX Series, Jetson Orin NX 16GB, Jetson Orin NX 8GB, Jetson Orin Nano Series, Jetson Orin\n\nNano Super Developer Kit, Jetson Orin Nano 8GB, Jetson Orin Nano 4GB, Jetson AGX Xavier\n\nSeries, Jetson AGX Xavier Industrial, Jetson AGX Xavier 64GB, Jetson AGX Xavier 32GB,\n\nJetson Xavier NX Series, Jetson Xavier NX 16GB, Jetson Xavier NX 8GB, Jetson TX2 Series,\n\nJetson TX2i, Jetson TX2, Jetson TX2 4GB, Jetson TX2 NX, Jetson Nano, any and all variations\n\nof the aforementioned products (including at least products having different options for number of\n\nGPUs).\n\n 25. The terms \u201cand\u201d and \u201cor\u201d are not intended to be read disjunctively but rather\n\nconjunctively unless the context of a particular request clearly indicates otherwise. \u201cOr\u201d should be\n\nunderstood to include and encompass \u201cand\u201d; and \u201cand\u201d should be understood to include and\n\nencompass \u201cor.\u201d\n\n 26. The terms \u201cany\u201d or \u201ceach\u201d should be understood to include and encompass \u201call.\u201d\n\n 27. The terms \u201cconcerning,\u201d \u201crelated to\u201d or \u201crelating to\u201d, and \u201cregarding\u201d and any\n\nvariation of these terms mean analyzing, alluding to, concerning, considering, commenting on,\n\nconsulting, comprising, containing, contradicting, describing, dealing with, discussing,\n\nestablishing, evidencing, identifying, involving, noting, recording, reporting on, relating to,\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 26 of 27\n\n\n\n\nreflecting, referring to, regarding, stating, showing, studying, mentioning, memorializing, or\n\npertaining to, directly or indirectly, in whole or in part.\n\n 28. The term \u201cCPU(s)\u201d means Central Processing Unit(s).\n\n 29. The term \u201cGPU(s)\u201d means Graphics Processing Unit(s).\n\n 30. The terms \u201cinclude\u201d and \u201cincluding\u201d mean including without limitation.\n\n 31. The term \u201cNVIDIA\u201d means Defendant NVIDIA Corporation, its predecessors,\n\npresent and former directors, officers, accountants, affiliates, attorneys, partners, managers, agents,\n\nemployees, representatives, in-house and outside counsel, and any other person or entity acting on\n\nbehalf of or under control of Defendant NVIDIA Corporation.\n\n 32. The term \u201cperson(s)\u201d means and includes natural persons and formal or informal\n\nentities and organizations, including public and private corporations, partnerships, professional\n\ncorporations, limited liability companies, business trusts, banking institutions, associations, firms,\n\njoint ventures, commissions, bureaus, departments, and any other legal entity, including\n\nany divisions, subsidiaries, departments, and other units thereof.\n\n 33. The term \u201cSource Code\u201d means human-readable instructions written in a\n\nprogramming language, including all comments, annotations, declarations, functions, classes, and\n\nother components used to define the behavior of a software program. For purposes of these topics,\n\n\u201cSource Code\u201d includes all associated files necessary to understand, compile, and execute the\n\ncode, such as scripts, header files, makefiles, configuration files, and documentation. Unless\n\notherwise stated, \u201cSource Code\u201d includes all versions and revisions relevant to the time periods\n\nand subject matter described in each interrogatory.\n\n 34. The terms \u201cYou,\u201d or \u201cYour\u201d refer to Tesla, Inc., including but not limited to its\n\npredecessors, successors, parents, subsidiaries, divisions, affiliates, and all past or present\n\f Case 7:26-mc-00318-LS Document 6-6 Filed 08/18/26 Page 27 of 27\n\n\n\n\ndirectors, officers, partners, managers, employees, contractors, agents, representatives,\n\naccountants, consultants, in-house and outside counsel, and any other person or entity acting or\n\npurporting to act on its behalf or subject to its control. This definition expressly includes, without\n\nlimitation, any Tesla parent, subsidiary, affiliate, or other related entity that has used, licensed,\n\ndeployed, evaluated, or integrated NVIDIA Hardware and/or NVIDIA Software.\n\n 35. The use of the singular form of any word includes the plural and vice versa.\n\n DEPOSITION TOPICS\n\n1. The NVIDIA software and libraries You use to perform computations, including but not\n\n limited to NVIDIA\u2019s Aerial, Clara Parabricks, cuBLAS, cuDNN, cuFFT, cuQuantum,\n\n cuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac, Isaac Lab, Maxine, Memory\n\n Map, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo, PyTorch, RAPIDS, Riva,\n\n Runtime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream).\n\n2. The NVIDIA sample Source Code You use, in whole or in part, to conduct computations.\n\n3. Your customizations and/or data inputs to NVIDIA software that alter the way in which\n\n NVIDIA software performs computations and/or a description of the data input to NVIDIA\n\n software on which computations are run.\n\n4. Identification of Your software that uses NVIDIA GPUs to perform computations.\n\n5. Using Your software, the ways in which output data from a GPU computation(s), including\n\n intermediate results of GPU computations are stored, referenced by a pointer, transferred,\n\n copied, streamed, written back, returned, accumulated, reused, or made available including\n\n asynchronously from NVIDIA GPU memory to CPU memory, host memory, system memory,\n\n storage, display, network, or another memory location before, during, or in parallel with\n\n computations performed using NVIDIA GPUs.\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:44.920889-07:00","document_number":"6","attachment_number":6,"pacer_doc_id":"181037220273","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 5","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554428/","id":490554428,"tags":[],"absolute_url":"/docket/74659430/6/7/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.247142-07:00","date_modified":"2026-08-23T04:40:08.225950-07:00","sha1":"8931900d9b8f3f039ac018d0167555e2c35ca8c0","page_count":2,"file_size":67070,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.7.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.7.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-7 Filed 08/18/26 Page 1 of 2\n\n\n\n\n EXHIBIT\n\n 6\n\f Case 7:26-mc-00318-LS Document 6-7 Filed 08/18/26 Page 2 of 2\n\n\n\n\n AFFIDAVIT OF SERVICE\n UNITED STATES DISTRICT COURT\n Western District of Texas\nCase Number: 7:24-CV-00221-ADA-\nDTH\n\nPlaintiff:\nNeural Al, LLC\nVS.\n\nDefendant:\nNvidia Corporation\n\nFor:\nEmily Portuguese\n\nReceived by Anthony Collins on the 24th day of June, 2026 at 4:46 pm to be served on Tesla, Inc c/o CT\nCorporation System, 1999 Bryan St, Ste 900, Dallas, Dallas County, TX 75201.\n\nI, Anthony Collins, being duly sworn, depose and say that on the 25th day of June, 2026 at 2:55 pm, I:\n\nExecuted service by hand delivering a true copy of the Subpoena to: Al Johnson , an authorized\nacceptance agent employed by Registered Agent CT Corporation System, Inc., who is authorized to\naccept service of process for Tesla, Inc , at the address of: 1999 Bryan St, Ste 900, Dallas, Dallas\nCounty, TX 75201, and informed said person of the contents therein, in compliance with state statutes.\n\nDescription of Person Served: Age: 30s, Sex: M, Race/Skin Color: Black, Height: 5'11, Weight: 220, Hair:\nBald, Glasses: Y\n\n\"I certify that I am over the age of 18, have no interest in the above action. and authorized to serve\nprocess in the judicial circuit in which the process was served. I have personal knowledge of the facts set\nforth in this affidavit. I declare under the penalty of perjury that the fore ing i r e and correct.\n\n %\"`\"'\"'\u2022 ERIC JACOB HARRIS\n **\u2022:-g:allotary Public, State ofas\n :\u2022.\n COMM. Expires OS-12-2.7.2S\n \u00b0!;,* Notary ID 135527875\n Anthony Collins\n PSC-357 Expires 12/31/2027\nSubscribed and Sworn to before me on the 26th day\nof June, 2026 by the affiant who is personally known Lexitas\nto me. 1235 Broadway\n 2nd Floor\n New York, NY 10001\nNOTARY PUBLIC (718) 672-1117\n\n Our Job Serial Number: ONT-2026006077\n Ref: 27254654\n\n\n\n\n Copyright @ 1992-2026 DreamBuilt Software, LLC. - Process Server's Toolbox V9.Oe\n\n\n\n\n 1111111111111111111111111111111111\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:44.961979-07:00","document_number":"6","attachment_number":7,"pacer_doc_id":"181037220274","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 6","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554429/","id":490554429,"tags":[],"absolute_url":"/docket/74659430/6/8/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.263114-07:00","date_modified":"2026-08-23T09:39:46.763868-07:00","sha1":"5437012e59781003b1035e29123541e69ce1de0c","page_count":18,"file_size":220239,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.8.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.8.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 1 of 18\n\n\n\n\n EXHIBIT\n\n 7\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 2 of 18\n\n\n\n\n UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND-ODESSA DIVISION\nNEURAL AI, LLC )\n )\n )\n Plaintiff, )\nv. ) Civil Action No. 7:24-cv-00221\n )\nNVIDIA CORPORATION )\n ) JURY TRIAL DEMANDED\n )\n Defendant. )\n\n\n PLAINTIFF\u2019S AMENDED DISCLOSURE OF ASSERTED CLAIMS AND FINAL\nINFRINGEMENT CONTENTIONS PURSUANT TO THE COURT\u2019S STANDING ORDER\n GOVERNING PATENT PROCEEDINGS\n\n Pursuant to the Court\u2019s December 2, 2025 Order (Dkt. 132), the First Amended Scheduling\n\n Order (Dkt. 135), and the Parties December 29, 2025 Joint Stipulation (Dkt. 136), Plaintiff Neural\n\n AI, LLC (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) provides the following Disclosure of Asserted Claims and\n\n Final Infringement Contentions (\u201cDisclosure\u201d) as to U.S. Patent Nos. 8,648,867 (\u201cthe \u2019867\n\n Patent\u201d), RE49,461 (\u201cthe \u2019461 Patent\u201d), and RE48,438 (\u201cthe \u2019438 Patent\u201d) (collectively, the\n\n \u201cAsserted Patents\u201d or \u201cPatents-in-Suit\u201d) against Defendant Nvidia Corporation (\u201cNvidia\u201d or\n\n \u201cDefendant\u201d). This Disclosure is made solely for the purpose of this action.\n\n Pursuant to the Court\u2019s December 2, 2025 Order (Dkt. 132), Plaintiff identifies source code\n\n for each accused instrumentality based on Nvidia\u2019s productions prior to November 30, 2025.\n\n Nvidia continues to produce source code on a rolling basis. Plaintiff therefore expressly reserves\n\n the right to serve additional or supplemental claim chart exhibits identifying newly produced or\n\n previously unavailable source code that satisfies the asserted claim elements, including after\n\n service of this Disclosure. Moreover, Plaintiff\u2019s investigation regarding infringement and\n\n additional potential grounds of infringement is ongoing. Nvidia\u2019s has not produced all necessary\n\n 1\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 3 of 18\n\n\n\ntechnical documents or source code sufficient to show the operation of the Accused Products, and\n\nkey categories of technical materials, including source code, remain outstanding. See, e.g., Dkt.\n\n103, 125. This Disclosure is therefore based upon information that Plaintiff has been able to obtain\n\nand review to date, together with its good-faith beliefs regarding the Accused Products and their\n\noperation, and is made without prejudice to Plaintiff\u2019s right to supplement or amend its Disclosure\n\nas additional facts are ascertained, discovery is conducted, code is reviewed, analysis is done, and\n\nresearch is completed. Plaintiff reserves the right to amend and/or supplement its infringement\n\ncontentions as additional information becomes available.\n\n For each Asserted Patent, Plaintiff identifies the following Accused Products of which it is\n\ncurrently aware. The identification of Accused Products is based on Plaintiff\u2019s research and\n\nanalysis to date, without the benefit of full discovery. Indeed, Nvidia has not yet complied with\n\nits obligation under the OGP to produce sufficient \u201ctechnical documents, including software where\n\napplicable, sufficient to show the operation of the accused product(s).\u201d See, e.g., Dkt. 103, 125.\n\nNvidia has continued to belatedly produce source code well into the fact discovery period\u2014\n\nincluding most recently on January 13, 2026\u2014and has not produced related technical documents,\n\nand Plaintiff has been required to repeatedly press Nvidia to obtain these late and piecemeal\n\nproductions. As a result, Plaintiff\u2019s current identification of Accused Products is necessarily based\n\non publicly available and otherwise accessible information. Accordingly, Plaintiff\u2019s current\n\nidentification is based on publicly available and otherwise accessible information. Plaintiff\n\nexpressly reserves the right to amend or supplement these contentions\u2014including by identifying\n\nadditional Accused Products or asserting additional bases for infringement\u2014under the applicable\n\nrules and any Court orders, including to reflect future productions by Nvidia.\n\n Accused Products. The Accused Products, as described in the accompanying Exhibits 1\u2013\n\n599 and any later-served amended or supplemental Exhibits, comprise integrated combinations of\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 4 of 18\n\n\n\nNvidia\u2019s software and hardware that together implement GPU-accelerated computing. These\n\ninclude Nvidia\u2019s GPU accelerators and superchips; Nvidia\u2019s computers, supercomputers, data\n\ncenters, servers, and workstations incorporating those GPUs; and the full Nvidia software stack\n\nthat operates on and controls that hardware to enable accelerated execution.\n\n Based on its present understanding of Nvidia\u2019s infringing software architecture and how\n\nthe accused functionality is implemented across Nvidia\u2019s integrated hardware and software stack,\n\nNeural AI has organized its infringement charts by Nvidia application in Exhibits 100-199. These\n\napplication charts further reference sub-charts identifying hardware described in Exhibits 1-99 as\n\nwell as software and libraries described in Exhibits 200-599. The sub charts are broken out as\n\nfollows: Exhibits 1\u201399 include hardware such as Nvidia\u2019s GPU accelerators and superchips, and\n\nNvidia\u2019s computers, supercomputers, data centers, servers, workstations that implement its GPU\n\naccelerators and superchips, coupling of some of this hardware to CPUs, as well as information on\n\nNvidia\u2019s infringing software products and CUDA code. Exhibits 100-199 include Nvidia\u2019s\n\napplication frameworks and platforms. Application frameworks and platforms processed by the\n\nhardware, in turn, call on lower-level Nvidia software. Exhibits 200\u2013299 include Nvidia\u2019s\n\nmachine-learning frameworks and inference platforms used by these applications, including\n\nPyTorch and TensorRT. Exhibits 300\u2013399 include Nvidia\u2019s neural-network primitive libraries,\n\nacceleration libraries, and functions, including cuDNN and related components. Exhibits 400\u2013499\n\ninclude Nvidia\u2019s mathematical libraries. Exhibits 500\u2013599 include low-level CUDA runtime,\n\ndriver, memory-management components, memory maps, and sub-component code used in the\n\nCUDA stack.\n\n This organization reflects Neural AI\u2019s present understanding of Nvidia\u2019s software\n\narchitecture. Nvidia has not produced a complete source-code production for any bucket or\n\ncategory, and Neural AI therefore expressly reserves the right to chart additional software,\n\n 3\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 5 of 18\n\n\n\nlibraries, or components, and to modify or supplement its chart organization and infringement\n\ntheories as discovery continues. Moreover, the charts provided within each category are\n\nrepresentative of other similar Nvidia software, libraries, and components within that same\n\ncategory. Neural AI has charted exemplar implementations rather than every produced file, and\n\nreserves the right to rely on other Nvidia software and libraries within the same category as\n\nadditional accused instrumentalities.\n\n In further details, the Accused Products include, without limitation, the following: Nvidia\u2019s\n\nGPU accelerators and superchips, including those with Nvidia\u2019s \u201cBlackwell,\u201d \u201cHopper,\u201d \u201cAda\n\nLovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d \u201cPascal,\u201d and \u201cMaxwell\u201d GPU architectures. These\n\nGPUs and superchips implement, and are specifically designed for, GPU-acceleration for artificial\n\nintelligence and neural networks.\n\n Nvidia\u2019s Blackwell GPUs include RTX PRO 6000 Server Edition, RTX PRO 6000\n\nWorkstation, RTX PRO 6000 Max-Q Workstation, RTX PRO 6000, RTX PRO 5000, RTX PRO\n\n4500, RTX PRO 4000, RTX PRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, RTX\n\n5090, RTX 5090 D, RTX 5080, RTX 5070 Ti, RTX 5070, RTX 5060, RTX 5060 Ti, RTX 5050,\n\nRTX 5080 Laptop, RTX 5090 Laptop, RTX 5070 Ti Laptop, RTX 5060 Laptop, RTX 5070\n\nLaptop, and RTX 5050 Laptop. In addition, Nvidia\u2019s superchips that implement GPU accelerators\n\ninclude the GB300 and GB200.\n\n Nvidia\u2019s Hopper GPUs include the H100 and H200 GPUs, including by not limited to\n\nPCle, SXM, and NVL models. In addition, Nvidia\u2019s superchips that implement GPU accelerators\n\ninclude the GH200, or Grace Hopper Superchip, which implements the Hopper-GPU architecture.\n\n Nvidia\u2019s Ada Lovelace (or Lovelace) GPUs include Nvidia Data Center GPUs, including\n\nL40, L40S, and L4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including RTX\n\nAda Generations series GPUs and Laptop GPUs (including RTX 6000, RTX 6000 Ada, RTX 5000\n\n 4\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 6 of 18\n\n\n\nAda, RTX 4500 Ada, RTX 4050, RTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3050, RTX\n\n3000, RTX 2000, RTX 1000, RTX 500); and GeForce RTX 40 series GPUs and Laptop GPUs\n\n(RTX 4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060\n\nTi, and RTX 4060; GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce\n\nRTX 4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU).\n\n Nvidia\u2019s Ampere GPUs include Nvidia Data Center GPUs, including A100, A40, A30,\n\nA16, A10, and A2 GPUs; Nvidia Workstation and Professional Laptop GPUs, including RTX A\n\nseries GPUs and Laptop GPUs (A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500); GeForce\n\nRTX 30 series GPUs and Laptop GPUs (GeForce RTX 3090 Ti, GeForce RTX 3090, GeForce\n\nRTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti, GeForce RTX 3070, GeForce RTX\n\n3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB), GeForce RTX 3050 (6 GB), GeForce\n\nRTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop GPU, GeForce RTX 3070 Ti Laptop GPU,\n\nGeForce RTX 3070 Laptop GPU, GeForce RTX 3060 Laptop GPU, GeForce RTX 3050 Ti Laptop\n\nGPU, GeForce RTX 3050 Laptop GPU); and GeForce MX570 Laptop GPU.\n\n Nvidia\u2019s Turing GPUs include Nvidia Data Center GPUs, including Tesla T4 GPUs and\n\nQuadro RTX 8000 (passive) and Quadro RTX 6000 (passive) GPUs; Nvidia Workstation and\n\nProfessional Laptop GPUs, including T series GPUs and Laptop GPUs, Quadro T series Laptop\n\nGPUs, and Quadro RTX series GPUs and Laptop GPUs (Quadro RTX 8000, Quadro RTX 6000,\n\nQuadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000, T1000 8GB, T1200,\n\nQuadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB); Titan series Titan RTX GPU;\n\nGeForce RTX 20 series GPUs and Laptop GPUs (GeForce RTX 2080 Ti, GeForce RTX 2080\n\nSuper, GeForce RTX 2080, GeForce RTX 2070 Super, GeForce RTX 2070, GeForce RTX 2060\n\n 5\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 7 of 18\n\n\n\nSuper, GeForce RTX 2060, GeForce RTX 2500); GeForce GTX 16 series GPUs and Laptop GPUs\n\n(GeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630); and GeForce MX550, MX450, and MX430 Laptop GPUs.\n\n Nvidia\u2019s Volta GPUs include Nvidia Data Center GPUs, including the Tesla V100 GPU;\n\nNvidia Workstation GPUs, including Quadro GV100; and Titan series Titan V GPU.\n\n Nvidia\u2019s Pascal GPUs include Nvidia Data Center GPUs, including Tesla P100, P40, and\n\nP4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including the Quadro GP100 GPU\n\nand Quadro P series GPUs and Laptop GPUs (Quadro P6000, Quadro P5200, Quadro P5000,\n\nQuadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro P2200, Quadro P2000,\n\nQuadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500, Quadro P400); Titan\n\nseries Titan Xp and Titan X GPUs; GeForce GTX 10 series GPUs and Laptop GPUs (GeForce\n\nGTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070 Ti, GeForce GTX 1070, GeForce GTX\n\n1060, GeForce GTX 1050 Ti, GeForce GTX 1050); and GeForce MX300 series, MX200 series,\n\nand MX150 Laptop GPUs.\n\n Nvidia\u2019s Maxwell GPUs include Nvidia Data Center GPUs, including Tesla M60, M40,\n\nand M10 GPUs; Nvidia Workstation and Professional Laptop GPUs, including Quadro M series\n\nGPUs and Laptop GPUs (Quadro M6000 24GB, Quadro M6000 (12GB), Quadro M5000, Quadro\n\nM5000M, Quadro M5500, Quadro M4000, Quadro M4000M, Quadro M3000M, Quadro M2200,\n\nQuadro M2000, Quadro M2000M, Quadro M1200, Quadro M1000M, Quadro M620, Quadro\n\nM600M, Quadro M520, Quadro M500M), the NVS 810 GPU, and Tesla M6 series Laptop GPUs;\n\nTitan series GTX Titan X GPU; GeForce GTX 900 series GPUs and Laptop GPUs (GeForce GTX\n\n980Ti, GeForce GTX 980, GeForce GTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce\n\nGTX 970M, GeForce GTX 965M, GeForce GTX 960M, GeForce GTX 950M); GeForce GTX\n\n 6\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 8 of 18\n\n\n\n700 series GPUs and Laptop GPUs (GeForce GTX 750 Ti, GeForce GTX 750); and GeForce\n\nMX130 series and MX110 Laptop GPUs.\n\n The Accused Products further include Nvidia\u2019s computers, supercomputers, data centers,\n\nservers, and workstations that implement its GPU accelerators and superchips. These computer\n\nhardware systems include: the DGX line of supercomputers, the HGX line of supercomputers, the\n\nOVX line of supercomputers, and the EGX line of servers for data centers and edge devices.\n\nNvidia\u2019s DGX supercomputers include the DGX B300, DGX B200, DGX GB200, DGX GB300,\n\nDGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200, DGX\n\nH200, DGX BasePOD, DGX A100, and DGX SuperPOD with DGX GB200. Nvidia\u2019s HGX\n\nincludes least the HGX B300, HGX B200, HGX H100, HGX H200, and EoS SuperPOD. And\n\nNvidia\u2019s EGX includes at least EGX Server with Quadro RTX A6000, EGX Server with A40,\n\nEGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000), GB300 NVL72,\n\nGB200 NVL72.\n\n The Accused Products include Nvidia\u2019s software, platforms, libraries, and services for\n\naccelerated computing. These products include, without limitation, application frameworks,\n\nplatforms, and domains such as NVIDIA Drive, Isaac, Holoscan, RAPIDS, NVBlox, NeMo,\n\nMerlin, Modulus, MONAI, Morpheus, Riva, Maxine, Clara, Metropolis, Tokkio, Avatar, NIM\n\nMicroservices, Omniverse, Clara Train, TAO, DRIVE Sim, DLSS, PhysX, OptiX, Texture Tools,\n\nJetPack, DeepStream, DOCA, Magnum IO, Aerial, BioNeMo, CUDA-X HPC, Unified Compute\n\nFramework, AI Enterprise, the DGX Platform, NGC, and AI Foundation Models. Application\n\nframeworks and platforms processed by the hardware, in turn, call on lower-level Nvidia software.\n\nThe Accused Products further include machine-learning frameworks and inference platforms such\n\nas PyTorch, TensorRT, Triton, TensorFlow, Torch-TensorRT, JAX, and Spark. They also include\n\nneural-network, mathematical, and acceleration libraries such as cuDNN, cuFFT, cuDSS,\n\n 7\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 9 of 18\n\n\n\ncuSOLVER, cuRAND, CUTLASS, DALI, cuTensor, cuGraph, cuSPARSELt, NPP, NeuralVDB,\n\ncuNumeric, cuCIM, Sionna, cuBLAS, cuSPARSE, NCCL, Thrust, CUB, AmgX, and nvmath-\n\npython. The Accused Products further include low-level CUDA runtime, driver, and system\n\ncomponents, including CUDA, the CUDA Toolkit, CUDA Runtime and Driver components,\n\nCUDA Python, CUDA Quantum, Base Command, GPNVAPI, NVSHMEM, DCGM, Displaced\n\nMicro-Mesh, FLARE, GVDB Voxels, KickstartRT, Mesh Shading, Optical Flow, PTX, SASS,\n\nkernel implementations, firmware, scheduling logic, backend libraries, the CUDA Driver Internal\n\nLayer (CUI), CUDA Driver API, vGPU, GPUDirect, NVML, and other associated runtime, driver,\n\nand sub-component code. The Accused Products further include other Nvidia software, platforms,\n\nand services that implement similar accelerated computing functionality or operate using the same\n\nCUDA-based execution models, architectures, libraries, and runtime components, whether or not\n\nexpressly listed above. The Accused Products also encompass associated and underlying software\n\nlibraries and components that enable or support accelerated execution and any source code or sub-\n\ncomponent code necessary to understand or effect CUDA execution, whether or not separately\n\nanalyzed.\n\n The Accused Products further include (1) any additional products identified in the\n\naccompanying Exhibits 1\u2013599 attached hereto and any later-served amended or supplemental\n\nExhibits; (2) any products that include the same functionality or features described in the Exhibits;\n\nand (3) any prior or subsequent versions of the products identified in the Exhibits that include the\n\nsame features or functionality.\n\n The Accused Products infringe each of the Asserted Patents in a manner fully consistent\n\nwith the Court\u2019s Claim Construction Order (Dkt. 95). For example, as construed, the Accused\n\nProducts implement the claimed \u201caccelerator\u201d as hardware, software, or a combination thereof that\n\nneed not be physically separate from the CPU. The Accused Products likewise perform the claimed\n\n 8\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 10 of 18\n\n\n\nmethod steps in a pipelined and overlapping manner that satisfies the Court\u2019s ordering\n\nrequirements for the asserted claims of the \u2019867 and \u2019438 patents, including parallel execution\n\nwhere permitted and the specific sequencing constraints identified by the Court.\n\n On information and belief, the asserted claim elements charted for one chip architecture\n\nare evidenced by documentation and source code pertaining to the Accused Products for other chip\n\narchitectures, for which the Accused Products have capabilities and functionalities that are\n\nsubstantially the same for the asserted claim elements. Indeed, this is reflected by the different chip\n\narchitectures of the Accused Products sharing materially similar technical specifications and\n\noverlapping documentation as each Accused Product pertains to the accused claim elements.\n\nLikewise, the software implementations analyzed and charted are representative of other related\n\nNvidia software products that rely on the same CUDA-based execution models, libraries, runtime\n\ncomponents, and architectural design choices, and that therefore implement the asserted claim\n\nelements in substantially the same manner such as, for example, performing a math operation or\n\ncomputation in a neural network.\n\n The present infringement contentions also accuse Nvidia\u2019s newly announced Vera CPU\n\nand Rubin GPU products. Neural AI will supplement its infringement contentions as Nvidia\n\nproduces additional information and as these products become commercially available. The\n\npresent infringement contentions further accuse Nvidia\u2019s Jetson hardware products, which, based\n\non information obtained during ongoing fact and expert discovery, infringe one or more Asserted\n\nClaims. Neural AI will supplement its infringement contentions as Nvidia produces additional\n\ninformation regarding these products.\n\n These Final Infringement Contentions are based on public evidence and the limited\n\ndiscovery and source code made available by Nvidia to date, which largely consists of a single\n\nsoftware version for each accused application or library. Nvidia\u2019s has not produced all necessary\n\n 9\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 11 of 18\n\n\n\ntechnical documents or source code sufficient to show the operation of the Accused Products, and\n\nkey categories of technical materials and code remain outstanding. See, e.g., Dkt. 103, 125.\n\nDiscovery and expert analysis are ongoing and include, among other things, the multiple software\n\nversions and builds that may be used in combination, as well as additional libraries and software\n\nfunctions that infringe the asserted claims under the same theories reflected in the claim charts.\n\nThese include, for example, acceleration libraries and functions that infringe the claims in the same\n\nmanner as the cuDNN CTCLoss and RNNForward functions in the two cuDNN source-code\n\nversions produced by Nvidia. Accordingly, Plaintiff expressly reserves the right to amend,\n\nsupplement, or refine these contentions as additional facts are ascertained, discovery is conducted,\n\nanalysis is performed, and research is completed. Plaintiff further reserves the right to amend or\n\nsupplement its Disclosure under the applicable rules and any Court orders, including to reflect\n\nfuture productions by Nvidia.\n\n Defendant\u2019s Infringement. Based upon currently available information, Plaintiff\n\nidentifies the following asserted claims:\n\n \u2022 The \u2019867 Patent. Defendant has infringed and is infringing claims 16-19, literally and/or\n\n under the doctrine of equivalents. Defendant has infringed and is infringing these\n\n claims both directly and indirectly (by inducing infringement pursuant to 35 U.S.C. \u00a7\n\n 271(b) and/or by contributing to infringement pursuant to 35 U.S.C.\u00a7 271(c)).\n\n \u2022 The \u2019461 Patent. Defendant has infringed and is infringing claims 21-25, 27-30,\n\n literally and/or under the doctrine of equivalents. Defendant has infringed and is\n\n infringing these claims both directly and indirectly (by inducing infringement\n\n pursuant to 35 U.S.C. \u00a7 271(b) and/or by contributing to infringement pursuant to 35\n\n U.S.C.\u00a7 271(c)).\n\n \u2022 The \u2019438 Patent. Defendant has infringed and is infringing claims 1, 3-6, 8-9, 12, 14,\n 10\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 12 of 18\n\n\n\n 17-18, 21-23, 29-30, 32, 40, 43-44, 46, 48, 51-52, 55-57, literally and/or under the\n\n doctrine of equivalents. Defendant has infringed and is infringing these claims both\n\n directly and indirectly (by inducing infringement pursuant to 35 U.S.C. \u00a7 271(b) and/or\n\n by contributing to infringement pursuant to 35 U.S.C. \u00a7 271(c)).\n\n Based upon currently available information, Plaintiff asserts that Defendant has infringed\n\nand/or continues to infringe the patents and claims as identified and described in the infringement\n\ncharts for the Accused Products attached as the accompanying Exhibits 1\u2013599 and any later-served\n\namended or supplemental Exhibits. These exhibits contain illustrative examples of Defendant\u2019s\n\npresently known infringement of the Asserted Claims by evidencing the correspondence between\n\n(i) elements of the Asserted Claims and (ii) corresponding structures and/or functions of the\n\nAccused Products. Such examples are illustrative and not exhaustive, additional materials may\n\nevidence infringement, and additional bases of infringement may be present and uncovered during\n\ndiscovery. Plaintiff reserves the right to amend or supplement its Disclosure, including the\n\nattached claim charts, upon Nvidia\u2019s compliance with its discovery obligations.\n\n Each element of each asserted claim is presently alleged to be literally present. However,\n\nto the extent Defendant argues that a limitation is not literally present in the Accused Products, then\n\nDefendant still infringes under the doctrine of equivalents. Any differences alleged to exist\n\nbetween any of the Asserted Claims and any of the Accused Products are insubstantial, and\n\ntherefore each Accused Product also meets each limitation under the doctrine of equivalents, as the\n\nidentified features of the Accused Product perform substantially the same function in substantially\n\nthe same way to achieve substantially the same result as the corresponding claim limitations.\n\nPlaintiff reserves the right to supplement this Disclosure as discovery is conducted, Defendant\n\nprovides any alleged non-infringement positions, and claim construction is completed.\n\n Defendant directly infringes each of the asserted claims under 35 U.S.C. \u00a7271(a) at least\n\n 11\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 13 of 18\n\n\n\nby using, operating, testing, advertising, making, installing, maintaining, distributing, supporting,\n\nproviding instructions for, offering to sell, selling, and/or otherwise providing services including\n\nthe Accused Products\u2014or systems incorporating the Accused Products\u2014within the United States\n\nand/or importing the Accused Products into the United States. Defendant also directly infringes\n\neach of the claims at least by performing, or being responsible for the performance of (e.g., the\n\nacts are attributable to it), each of the claimed steps as set forth in the accompanying charts.\n\nDefendant\u2019s acts of direct infringement are further set forth in the accompanying Exhibits and any\n\nlater-served amended or supplemental Exhibits.\n\n Defendant also indirectly infringes the Asserted Claims by inducing infringement pursuant\n\nto 35 U.S.C. \u00a7 271(b)). Defendant has had knowledge of each of the asserted patents and of the\n\nspecific manner by which the Accused Products infringe each patent since at least September 2024,\n\nwhen Plaintiff filed and served its original complaint. Defendant knowingly induced one or more\n\nthird parties (e.g., business partners, customers, or others), to infringe the Asserted Claims by\n\nmaking the Accused Products available on Defendant\u2019s website, widely advertising the Accused\n\nProducts, providing applications that allow partners and users to access the Accused Products,\n\nproviding instructions for installing the Accused Products, and providing technical support to users\n\nand/or engaging in activities that aid and abet infringement of the Asserted Patents by end users\n\nwithin the United States, with knowledge and intent that performance of such actions would\n\ninfringe the Asserted Claims. Defendant committed these acts with knowledge or willful blindness\n\nthat such induced acts would constitute infringement of the Asserted Claims at least as of the filing\n\nof the original complaint. Defendant also has had actual or constructive notice of the technology\n\nclaimed in the Asserted Patents since at least 2007, when the inventors of the Asserted Patents first\n\ndiscussed their patented technologies with Mr. Sanford Russell, then the CTO of Nvidia. In\n\naddition, the inventors of the Asserted Patents held multiple discussions with Nvidia regarding a\n\n 12\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 14 of 18\n\n\n\npotential investment in or acquisition of their company, Neurala, Inc., and its assets, including the\n\npatent family that includes the Asserted Patents. Defendant knew or should have known that it\n\ninfringed the Asserted Patents based on its knowledge of the same. Alternatively or additionally,\n\nDefendant was willfully blind to the fact that it infringed the Asserted Patents despite its\n\nknowledge of the same, based on, for example, Defendant\u2019s having cited the application for the\n\n\u2019867 Patent on the face of its own patent since at least June 28, 2010, see Nvidia U.S. Patent No.\n\n8,922,566, and the similarity of the Accused Products to Plaintiff\u2019s patented technology.\n\nAdditional evidence of Defendant\u2019s inducement of infringement by others is set forth in each of\n\naccompanying Exhibits.\n\n Defendant also indirectly infringes the Asserted Claims by contributing to infringement\n\npursuant to 35 U.S.C. \u00a7 271(c). Each of the Accused Products is a material part of the claims, and\n\nDefendant knew each of the Accused Products is especially made or especially adapted for use in\n\nan infringement of the Asserted Claims. Further, the Accused Products have no substantial non-\n\ninfringing uses, as set forth in the example claim charts.\n\n Defendant also contributes to infringement by its customers and end users of the Accused\n\nProducts by offering to sell or selling within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice,\n\nmethods claimed in the Asserted Patents, constituting a material part of the inventions claimed,\n\nand not a staple article or commodity of commerce suitable for substantial non-infringing use.\n\nIndeed, the Accused Products and the exemplary functionality identified in the accompanying\n\nExhibits have no substantial non-infringing uses but instead are specifically designed to practice\n\nthe Asserted Patents. Additional evidence of Defendant\u2019s contributory infringement is set forth in\n\neach of the Exhibits and any later-served amended or supplemental Exhibits.\n\n Priority Dates. Plaintiff presently identifies the following priority dates for the Asserted\n\n 13\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 15 of 18\n\n\n\nPatents:\n\n \u2022 All Asserted Claims of the \u2019867 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019867 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed invention, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n See NAI_0007731-NAI_0007923.\n\n \u2022 All Asserted Claims of the \u2019461 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019461 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed inventions, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n See NAI_0007731-NAI_0007923.\n\n \u2022 All Asserted Claims of the \u2019438 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019438 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed inventions, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n 14\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 16 of 18\n\n\n\n See NAI_0007731-NAI_0007923.\n\n Plaintiff\u2019s investigation and analysis is ongoing, and Plaintiff reserves the right to assert and\n\nrely on an earlier invention date in the event Defendant identifies alleged prior art that dated earlier\n\nthan the identified priority date corresponding to a date of conception followed by diligence and\n\nreduction to practice of the claimed inventions.\n\n\n\n DATED: January 20, 2026 Respectfully submitted,\n\n /s/ Tanner Laiche\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n 15\n\fCase 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 17 of 18\n\n\n\n\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\n\n\n\n 16\n\f Case 7:26-mc-00318-LS Document 6-8 Filed 08/18/26 Page 18 of 18\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned hereby certifies that a true and correct copy of the foregoing document has\n\nbeen served on January 20, 2026 to all counsel of record via electronic mail.\n\n\n\n /s/ Tanner Laiche\n Tanner Laiche\n\n\n\n\n 17\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:49.916369-07:00","document_number":"6","attachment_number":8,"pacer_doc_id":"181037220275","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 7","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554430/","id":490554430,"tags":[],"absolute_url":"/docket/74659430/6/9/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.280181-07:00","date_modified":"2026-08-23T04:40:00.615921-07:00","sha1":"09d8e3c2df6f3827a138d56fd345dc5a9788e865","page_count":47,"file_size":458868,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.9.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.9.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-9 Filed 08/18/26 Page 1 of 47\n\n\n\n\n EXHIBIT\n\n 8\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 2 of147\n of 46\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n PUBLIC VERSION\nNEURAL AI, LLC\n Civil Action No. 7:24-cv-00221\n Plaintiff,\n\n v. JURY TRIAL DEMANDED\n\nNVIDIA CORPORATION.,\n\n Defendant.\n\n\n DEFENDANT\u2019S AMENDED ANSWER TO PLAINTIFF\u2019S AMENDED COMPLAINT\n\n Defendant NVIDIA Corporation (\u201cNVIDIA\u201d) hereby provides its amended answer to\n\nPlaintiff Neural AI, LLC\u2019s (\u201cPlaintiff\u201d) Amended Complaint for Patent Infringement (Dkt. 30)\n\n(\u201cComplaint\u201d). The headings and subheadings in Defendant\u2019s Answer are used solely for purposes\n\nof convenience and organization to mirror those appearing in the Complaint; to the extent that any\n\nheadings or other non-numbered statements in the Complaint contain or imply any allegations,\n\nDefendant denies each and every allegation therein. Except as expressly admitted, all allegations\n\nin the Complaint are denied.\n\n 1. Defendant admits that graphics processor units may be used for artificial\n\nintelligence, machine learning, or complex numerical simulation applications. Defendant admits\n\nthat GPU computing powers many of the most advanced and powerful forms of artificial\n\nintelligence over the past decade. Defendant denies the remaining allegations of Paragraph 1.\n\n 2. Defendant admits that graphics processor units may be used for complex numerical\n\nsimulation, machine learning, and training complex models. Defendant further admits that\n\ngraphics processor units typically have more computational processors than central processing\n\nunits and are capable of parallel processing. Defendant does not have knowledge or information\n\n\n\n\n 1\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 3 of247\n of 46\n\n\n\n\nsufficient to form a belief as to the truth of the remaining allegations contained in Paragraph 2 and\n\non that basis denies them.\n\n 3. Defendant denies the allegations in Paragraph 3 of the Complaint.\n\n 4. Defendant denies the allegations in Paragraph 4 of the Complaint.\n\n NATURE OF THE CASE\n\n 5. Defendant admits that Plaintiff has asserted claims for patent infringement arising\n\nunder 35 U.S.C. \u00a7 1, et seq., but denies all of Plaintiff\u2019s allegations of infringement. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 5 of the Complaint.\n\n 6. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 6 and on that basis denies them.\n\n 7. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 7 and on that basis denies them.\n\n 8. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 8 and on that basis denies them.\n\n 9. Defendant admits that Defendant is a Delaware corporation with its headquarters\n\nin Santa Clara, California. Defendant further admits that it is registered to conduct business in\n\nTexas. Defendant further admits that it has an office in Austin, Texas. Except as expressly\n\nadmitted, Defendant denies the remaining allegations in Paragraph 9 of the Complaint.\n\n JURISDICTION & VENUE\n\n 10. Defendant admits that Plaintiff purports to assert claims for patent infringement\n\narising under 35 U.S.C. \u00a7 1, et seq., but denies all claims of infringement by Defendant. Defendant\n\nadmits that the Court has subject matter jurisdiction pursuant to 28 U.S.C. \u00a7\u00a7 1331 and 1338(a).\n\n\n\n\n 2\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 4 of347\n of 46\n\n\n\n\nExcept as expressly admitted, Defendant denies the remaining allegations in Paragraph 10 of the\n\nComplaint.\n\n 11. Defendant admits that this Court has personal jurisdiction for the purpose of this\n\nparticular action and that it does business in Texas and in this District. Except as expressly\n\nadmitted, Defendant denies the remaining allegations in Paragraph 11 of the Complaint.\n\n 12. Defendant admits that Defendant has conducted business in within this District, but\n\ndenies all allegations of infringement. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 12 of the Complaint.\n\n 13. Defendant admits that this Court has personal jurisdiction for the purpose of this\n\nparticular action. To the extent the allegations of Paragraph 13 purport to quote from or\n\ncharacterize the contents of written documents, those documents speak for themselves. Defendant\n\ndenies the remaining allegations in Paragraph 13 of the complaint.\n\n 14. Defendant admits that venue is proper for this case but denies that it is a convenient\n\nforum for NVIDIA and its witnesses. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 14 of the complaint.\n\n 15. Defendant Nvidia Corporation is a registered business in Texas and has regular and\n\nestablished places of business Defendant admits that it is a registered business in Texas and\n\nmaintains an office located at 11001 Lakeline Blvd, Suite 100 Bldg. 2, Austin, Texas 78717. To\n\nthe extent the allegations of Paragraph 15 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 15 of the complaint.\n\n\n\n\n 3\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 5 of447\n of 46\n\n\n\n\n 16. Defendant admits that it has hundreds of employees in this District. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the remaining\n\nallegations contained in Paragraph 16 and on that basis denies them.\n\n 17. Defendant admits that it has open job postings for jobs that may be filled in a\n\nnumber of locations, including in this District. Defendant does not have knowledge or information\n\nsufficient to form a belief as to the truth of the remaining allegations contained in Paragraph 17\n\nand on that basis denies them.\n\n 18. Defendant admits that it engages it engages in business in this District. Defendant\n\nadmits that it has customer-facing personnel and operations in this District. Defendant admits that\n\nit provides technical support to partners and customers for its products in this District. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 18 of the Complaint.\n\n 19. Defendant denies the allegations in Paragraph 19 of the Complaint.\n\n 20. Defendant admits that it sells products and provides services in the State of Texas,\n\nbut denies that those products or services infringe the Asserted Patents. To the extent the\n\nallegations in Paragraph 20 of the Complaint relate to the knowledge or actions of third parties,\n\nDefendant does not have knowledge or information sufficient to form a belief as to the truth of\n\nthose allegations and on that basis denies them. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 20 of the Complaint.\n\n 21. Defendant admits that it sells products and provides services in the State of Texas,\n\nbut denies that the use of those products or services infringes the Asserted Patents. To the extent\n\nthe allegations in Paragraph 21 of the Complaint relate to the knowledge or actions of third parties,\n\nDefendant does not have knowledge or information sufficient to form a belief as to the truth of\n\n\n\n\n 4\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 6 of547\n of 46\n\n\n\n\nthose allegations and on that basis denies them. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 21 of the Complaint.\n\n 22. Defendant admits that it partners with resellers and managed service providers for\n\nthe sale or installation of certain NVIDIA products. To the extent the allegations of Paragraph 22\n\npurport to quote from or characterize the contents of websites, those documents speak for\n\nthemselves. Except as expressly admitted, Defendant denies the remaining allegations in\n\nParagraph 22 of the Complaint.\n\n 23. Defendant admits that it partners with data center providers. To the extent the\n\nallegations of Paragraph 23 purport to quote from or characterize the contents of written\n\ndocuments, those documents speak for themselves. To the extent the allegations in Paragraph 23\n\nof the Complaint relate to the knowledge or actions of third parties, Defendant does not have\n\nknowledge or information sufficient to form a belief as to the truth of those allegations and on that\n\nbasis denies them. Except as expressly admitted, Defendant denies the remaining allegations in\n\nParagraph 23 of the Complaint.\n\n 24. Defendant denies the allegations in Paragraph 24 of the Complaint.\n\n 25. To the extent the allegations of Paragraph 25 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 25 of the Complaint.\n\n 26. To the extent the allegations of Paragraph 26 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 26 of the Complaint.\n\n\n\n\n 5\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 7 of647\n of 46\n\n\n\n\n 27. To the extent the allegations of Paragraph 27 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 27 of the Complaint.\n\n 28. Defendant denies the allegations in Paragraph 28 of the Complaint.\n\n 29. To the extent the allegations of Paragraph 29 purport to quote from or characterize\n\nthe contents of written documents, those documents speak for themselves. Defendant denies the\n\nremaining allegations in Paragraph 29 of the Complaint.\n\n 30. Defendant denies the allegations in Paragraph 30 of the Complaint.\n\n PLAINTIFF\u2019S PATENTED INNOVATIONS\n\n 31. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 31 and on that basis denies them.\n\n The GPU-Based Acceleration Patents\n U.S. Patent Nos. 8,648,867, RE49,461, and RE48,438\n\n 32. Defendant admits that the \u2019461 Patent purports to be a continuation of the \u2019438\n\nPatent, which purports to be an application for reissue of U.S. Patent No. 9,189,828, which purports\n\nto be a continuation of the \u2019867 Patent. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 32 of the Complaint.\n\n 33. Defendant admits that Exhibit 1 to the Complaint appears to be a copy of the \u2019867\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nSeptember 24, 2007, and was issued on February 11, 2014. Defendant further admits that the \u2019867\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019867\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 33 of the Complaint.\n\n\n\n\n 6\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 8 of747\n of 46\n\n\n\n\n 34. Defendant admits that Exhibit 2 to the Complaint appears to be a copy of the \u2019438\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nNovember 9, 2017, and was issued on February 16, 2021. Defendant further admits that the \u2019438\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019438\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 34 of the Complaint.\n\n 35. Defendant admits that Exhibit 3 to the Complaint appears to be a copy of the \u2019461\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nDecember 29, 2020, and was issued on March 14, 2023. Defendant further admits that the \u2019461\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019461\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 35 of the Complaint.\n\n 36. To the extent the allegations of Paragraph 36 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 36 of the Complaint.\n\n 37. To the extent the allegations of Paragraph 37 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 37 of the Complaint.\n\n 38. To the extent the allegations of Paragraph 38 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 38 of the Complaint.\n\n\n\n\n 7\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 9 of847\n of 46\n\n\n\n\n 39. To the extent the allegations of Paragraph 39 purport to quote from or characterize\n\nthe contents of the \u2019461 and \u2019438 Patents, those documents speak for themselves. Defendant denies\n\nthe remaining allegations in Paragraph 39 of the Complaint.\n\n 40. To the extent the allegations of Paragraph 40 purport to quote from or characterize\n\nthe contents of the Asserted Patents, those documents speak for themselves. Defendant denies the\n\nremaining allegations in Paragraph 40 of the Complaint.\n\n ACCUSED PRODUCTS\n\n 41. Defendant admits that it offers GPUs and various hardware and software products,\n\nbut specifically denies that those products infringe the Asserted Patents. To the extent the\n\nallegations of Paragraph 41 purport to quote from or characterize the contents of websites, those\n\ndocuments speak for themselves. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 41 of the Complaint.\n\n 42. Defendant admits that \u201cHopper,\u201d \u201cAda Lovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d\n\n\u201cPascal,\u201d and \u201cMaxwell\u201d are architectures of Defendant\u2019s GPUs but specifically denies that those\n\nproducts infringe the Asserted Patents. To the extent the allegations of Paragraph 42 purport to\n\nquote from or characterize the contents of websites, those documents speak for themselves. Except\n\nas expressly admitted, Defendant denies the remaining allegations in Paragraph 42 of the\n\nComplaint.\n\n 43. Defendant admits that its products include the H100 and H200 GPUs, but\n\nspecifically denies that those products infringe the Asserted Patents. Defendant further admits that\n\nit offers the GH200 \u201cGrace Hopper Superchip,\u201d but likewise specifically denies that this product\n\ninfringes the Asserted Patents. To the extent the allegations of Paragraph 43 purport to quote from\n\n\n\n\n 8\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9 130FiledFiled\n 08/18/26\n 11/25/25PagePage\n 10 of9 47\n of 46\n\n\n\n\nor characterize the contents of websites, those documents speak for themselves. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 43 of the Complaint.\n\n 44. Defendant admits that its products include GPUs with the Ada Lovelace\n\narchitecture, but specifically denies that those products infringe the Asserted Patents. To the extent\n\nthe allegations of Paragraph 44 purport to quote from or characterize the contents of websites,\n\nthose documents speak for themselves. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 44 of the Complaint.\n\n 45. Defendant admits that its products include GPUs with the Ampere architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 45 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 45 of the Complaint.\n\n 46. Defendant admits that its products include GPUs with the Turing architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 46 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 46 of the Complaint.\n\n 47. Defendant admits that its products include GPUs with the Volta architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 47 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 47 of the Complaint.\n\n\n\n\n 9\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 11 of\n 1047\n of 46\n\n\n\n\n 48. Defendant admits that its products include GPUs with the Pascal architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 48 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 48 of the Complaint.\n\n 49. Defendant admits that its products include GPUs with the Maxwell architecture,\n\nbut specifically denies that those products infringe the Asserted Patents. To the extent the\n\nallegations of Paragraph 49 purport to quote from or characterize the contents of websites, those\n\ndocuments speak for themselves. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 49 of the Complaint.\n\n 50. Defendant admits that certain of Defendant\u2019s GPU architectures support the CUDA\n\nplatform. To the extent the allegations of Paragraph 50 purport to quote from or characterize the\n\ncontents of websites, those documents speak for themselves. Except as expressly admitted,\n\nDefendant denies the remaining allegations in Paragraph 50 of the Complaint.\n\n 51. Defendant admits that it has marketed products under the EGX, HGX, DGX, and\n\nOVX product names, but specifically denies that those products infringe the Asserted Patents. To\n\nthe extent the allegations of Paragraph 51 purport to quote from or characterize the contents of\n\nwebsites, those documents speak for themselves. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 51 of the Complaint.\n\n 52. To the extent the allegations of Paragraph 52 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 52 of the Complaint.\n\n\n\n\n 10\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 12 of\n 1147\n of 46\n\n\n\n\n 53. Paragraph 53 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 54. To the extent the allegations of Paragraph 54 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 54 of the Complaint.\n\n 55. To the extent the allegations of Paragraph 55 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 55 of the Complaint.\n\n 56. To the extent the allegations of Paragraph 56 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 56 of the Complaint.\n\n 57. To the extent the allegations of Paragraph 57 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 57 of the Complaint.\n\n 58. To the extent the allegations of Paragraph 58 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 58 of the Complaint.\n\n 59. To the extent the allegations of Paragraph 59 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 59 of the Complaint.\n\n 60. To the extent the allegations of Paragraph 60 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 60 of the Complaint.\n\n\n\n\n 11\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 13 of\n 1247\n of 46\n\n\n\n\n FIRST CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019867 PATENT)\n\n 61. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 62. Defendant denies the allegations in Paragraph 62 of the Complaint.\n\n 63. Paragraph 63 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 64. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 64 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 64 of the Complaint.\n\n 65. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 65 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 65 of the Complaint.\n\n 66. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 66 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 66 of the Complaint.\n\n 67. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 67 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 67 of the Complaint.\n\n\n\n\n 12\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 14 of\n 1347\n of 46\n\n\n\n\n 68. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 68 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 68 of the Complaint.\n\n 69. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 69 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 69 of the Complaint.\n\n 70. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 70 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 70 of the Complaint.\n\n 71. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 71 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 71 of the Complaint.\n\n 72. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 72 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 72 of the Complaint.\n\n 73. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 73 purport to quote from or characterize the contents of\n\n\n\n\n 13\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 15 of\n 1447\n of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 73 of the Complaint.\n\n 74. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 74 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 74 of the Complaint.\n\n 75. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 75 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 75 of the Complaint.\n\n 76. Defendant denies the allegations in Paragraph 76 of the Complaint.\n\n 77. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 77 of the Complaint.\n\n 78. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 78 of the Complaint.\n\n\n\n\n 14\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 16 of\n 1547\n of 46\n\n\n\n\n 79. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 79 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 79 of the Complaint.\n\n 80. Defendant denies the allegations in Paragraph 80 of the Complaint.\n\n 81. Defendant denies the allegations in Paragraph 81 of the Complaint.\n\n 82. Defendant denies the allegations in Paragraph 82 of the Complaint.\n\n 83. Defendant denies the allegations in Paragraph 83 of the Complaint.\n\n 84. Defendant denies the allegations in Paragraph 84 of the Complaint.\n\n 85. Defendant denies the allegations in Paragraph 85 of the Complaint.\n\n 86. Defendant denies the allegations in Paragraph 86 of the Complaint.\n\n 87. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n87 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 87 of the Complaint.\n\n 88. Defendant denies the allegations in Paragraph 88 of the Complaint.\n\n 89. Defendant denies the allegations in Paragraph 89 of the Complaint.\n\n 90. Defendant denies the allegations in Paragraph 90 of the Complaint.\n\n 91. Defendant denies the allegations in Paragraph 91 of the Complaint.\n\n 92. Defendant denies the allegations in Paragraph 92 of the Complaint.\n\n 93. Defendant denies the allegations in Paragraph 93 of the Complaint.\n\n\n\n\n 15\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 17 of\n 1647\n of 46\n\n\n\n\n SECOND CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019438 PATENT)\n\n 94. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 95. Defendant denies the allegations in Paragraph 95 of the Complaint.\n\n 96. Paragraph 96 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 97. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 97 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 97 of the Complaint.\n\n 98. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 98 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 98 of the Complaint.\n\n 99. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 99 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 99 of the Complaint.\n\n 100. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 100 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 100 of the Complaint.\n\n\n\n\n 16\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 18 of\n 1747\n of 46\n\n\n\n\n 101. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 101 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 101 of the Complaint.\n\n 102. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 102 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 102 of the Complaint.\n\n 103. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 103 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 103 of the Complaint.\n\n 104. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 104 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 104 of the Complaint.\n\n 105. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 105 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 105 of the Complaint.\n\n 106. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 106 purport to quote from or characterize the contents of\n\n\n\n\n 17\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 19 of\n 1847\n of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 106 of the Complaint.\n\n 107. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 107 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 107 of the Complaint.\n\n 108. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 108 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 108 of the Complaint.\n\n 109. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 109 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 109 of the Complaint.\n\n 110. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 110 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 110 of the Complaint.\n\n 111. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 111 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 111 of the Complaint.\n\n\n\n\n 18\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 20 of\n 1947\n of 46\n\n\n\n\n 112. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 112 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 112 of the Complaint.\n\n 113. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 113 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 113 of the Complaint.\n\n 114. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 114 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 114 of the Complaint.\n\n 115. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 115 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 115 of the Complaint.\n\n 116. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 116 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 116 of the Complaint.\n\n 117. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 117 purport to quote from or characterize the contents of\n\n\n\n\n 19\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 21 of\n 2047\n of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 117 of the Complaint.\n\n 118. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 118 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 118 of the Complaint.\n\n 119. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 119 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 119 of the Complaint.\n\n 120. Defendant denies the allegations in Paragraph 120 of the Complaint.\n\n 121. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Defendant admits that it became aware of the \u2019438 Patent since at\n\nleast the filing of this Complaint. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 121 of the Complaint.\n\n 122. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 122 of the Complaint.\n\n\n\n\n 20\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 22 of\n 2147\n of 46\n\n\n\n\n 123. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 123 purport to quote from or characterize the contents\n\nof websites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 123 of the Complaint.\n\n 124. Defendant denies the allegations in Paragraph 124 of the Complaint.\n\n 125. Defendant denies the allegations in Paragraph 125 of the Complaint.\n\n 126. Defendant denies the allegations in Paragraph 126 of the Complaint.\n\n 127. Defendant denies the allegations in Paragraph 127 of the Complaint.\n\n 128. Defendant denies the allegations in Paragraph 128 of the Complaint.\n\n 129. Defendant denies the allegations in Paragraph 129 of the Complaint.\n\n 130. Defendant denies the allegations in Paragraph 130 of the Complaint.\n\n 131. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n131 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 131 of the Complaint.\n\n 132. Defendant denies the allegations in Paragraph 132 of the Complaint.\n\n 133. Defendant denies the allegations in Paragraph 133 of the Complaint.\n\n 134. Defendant denies the allegations in Paragraph 134 of the Complaint.\n\n 135. Defendant denies the allegations in Paragraph 135 of the Complaint.\n\n 136. Defendant denies the allegations in Paragraph 136 of the Complaint.\n\n 137. Defendant denies the allegations in Paragraph 137 of the Complaint.\n\n\n\n\n 21\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 23 of\n 2247\n of 46\n\n\n\n\n THIRD CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019461 PATENT)\n\n 138. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 139. Defendant denies the allegations in Paragraph 139 of the Complaint.\n\n 140. Paragraph 140 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 141. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 141 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 141 of the Complaint.\n\n 142. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 142 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 142 of the Complaint.\n\n 143. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 143 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 143 of the Complaint.\n\n 144. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 144 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 144 of the Complaint.\n\n\n\n\n 22\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 24 of\n 2347\n of 46\n\n\n\n\n 145. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 145 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 145 of the Complaint.\n\n 146. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 146 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 146 of the Complaint.\n\n 147. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 147 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 147 of the Complaint.\n\n 148. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 148 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 148 of the Complaint.\n\n 149. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 149 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 149 of the Complaint.\n\n 150. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 150 purport to quote from or characterize the contents of\n\n\n\n\n 23\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 25 of\n 2447\n of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 150 of the Complaint.\n\n 151. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 151 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 151 of the Complaint.\n\n 152. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 152 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 152 of the Complaint.\n\n 153. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 153 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 153 of the Complaint.\n\n 154. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 154 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 154 of the Complaint.\n\n 155. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 155 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 155 of the Complaint.\n\n\n\n\n 24\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 26 of\n 2547\n of 46\n\n\n\n\n 156. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 156 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 156 of the Complaint.\n\n 157. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 157 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 157 of the Complaint.\n\n 158. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 158 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 158 of the Complaint.\n\n 159. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 159 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 159 of the Complaint.\n\n 160. Defendant denies the allegations in Paragraph 160 of the Complaint.\n\n 161. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Defendant admits that it became aware of the \u2019461 Patent since at\n\n\n\n\n 25\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 27 of\n 2647\n of 46\n\n\n\n\nleast the filing of this Complaint. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 161 of the Complaint.\n\n 162. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 162 of the Complaint.\n\n 163. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 161 purport to quote from or characterize the contents\n\nof websites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 163 of the Complaint.\n\n 164. Defendant denies the allegations in Paragraph 164 of the Complaint.\n\n 165. Defendant denies the allegations in Paragraph 165 of the Complaint.\n\n 166. Defendant denies the allegations in Paragraph 166 of the Complaint.\n\n 167. Defendant denies the allegations in Paragraph 167 of the Complaint.\n\n 168. Defendant denies the allegations in Paragraph 168 of the Complaint.\n\n 169. Defendant denies the allegations in Paragraph 169 of the Complaint.\n\n 170. Defendant denies the allegations in Paragraph 170 of the Complaint.\n\n 171. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n171 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 171 of the Complaint.\n\n 172. Defendant denies the allegations in Paragraph 172 of the Complaint.\n\n\n\n\n 26\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 28 of\n 2747\n of 46\n\n\n\n\n 173. Defendant denies the allegations in Paragraph 173 of the Complaint.\n\n 174. Defendant denies the allegations in Paragraph 174 of the Complaint.\n\n 175. Defendant denies the allegations in Paragraph 175 of the Complaint.\n\n 176. Defendant denies the allegations in Paragraph 176 of the Complaint.\n\n 177. Defendant denies the allegations in Paragraph 177 of the Complaint.\n\n PRAYER FOR RELIEF\n\n Defendant denies any factual assertions contained in Plaintiff\u2019s Prayer for Relief.\n\nDefendant further denies that Plaintiff is entitled to any relief whatsoever, including but not limited\n\nto the relief sought in Paragraphs A-G of the Complaint.\n\n DEMAND FOR JURY TRIAL\n\n A response is not required to Plaintiff\u2019s demand for a jury trial.\n\n DEFENSES\n\n Defendant repeats and re-alleges the allegations of the preceding Paragraphs as if fully set\n\nforth herein. Defendant asserts the following defenses to Plaintiff\u2019s Complaint, without admitting\n\nor acknowledging that Defendant bears the burden of proof as to any of them or that any must be\n\npleaded as defenses. Defendant specifically reserves all rights to allege additional defenses that\n\nbecome known through the course of discovery.\n\n FIRST DEFENSE\n (Non-Infringement)\n\n Defendant has not and does not infringe, either literally or under the doctrine of\n\nequivalents, any valid and enforceable claim of any Asserted Patent, whether directly, indirectly,\n\ncontributorily, by inducement, individually, jointly, willfully, or otherwise. Additionally, with\n\nrespect to Plaintiff\u2019s allegations of indirect, joint, and willful infringement, Defendant lacks the\n\nrequisite mens rea.\n\n\n\n 27\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 29 of\n 2847\n of 46\n\n\n\n\n SECOND DEFENSE\n (Invalidity & Ineligibility)\n\n The Asserted Claims are invalid under at least 35 U.S.C. \u00a7\u00a7 101, 102, 103, and/or 112.\n\nDefendant incorporates by reference its forthcoming invalidity contentions and all amendments\n\nthereto.\n\n THIRD DEFENSE\n (No Willfulness, Enhanced Damages, or Attorneys\u2019 Fees)\n\n Plaintiff is not entitled to enhanced damages under 35 U.S.C. \u00a7 284, at least because\n\nPlaintiff has failed to show, and cannot show, that any infringement has been willful and/or\n\nknowing. Plaintiff is not entitled to an award of attorney\u2019s fees under 35 U.S.C. \u00a7 285, at least\n\nbecause Plaintiff has failed to show, and cannot show, that this case is \u201cexceptional\u201d in Plaintiff\u2019s\n\nfavor as would be required by the statute.\n\n FOURTH DEFENSE\n (Statutory Limitation on Damages)\n\n Plaintiff\u2019s claims for relief are statutorily limited in whole or in part by 35 U.S.C. \u00a7\u00a7 286\n\nand/or 287. In addition, to the extent Plaintiff seeks damages for allegedly infringing acts\n\ncommitted more than six years prior to the filing of the Complaint in this action, it is barred from\n\nrecovery of such damages.\n\n Additionally, to the extent Plaintiff or any licensee of the Asserted Patent failed to properly\n\nmark any of their relevant products as required by 35 U.S.C. \u00a7 287 or otherwise failed to give\n\nproper notice that Defendant\u2019s actions allegedly infringed any Asserted Claim, Defendant is not\n\nliable to Plaintiff for the acts alleged to have been performed before Defendant received actual\n\nnotice of infringement.\n\n FIFTH DEFENSE\n (License, Exhaustion, Waiver, and Estoppel)\n\n Plaintiff\u2019s claims are barred, in whole or in part, by license, exhaustion, and/or the\n\n\n 28\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 30 of\n 2947\n of 46\n\n\n\n\ndoctrines of waiver and/or equitable estoppel.\n\n SIXTH DEFENSE\n (Inexcusable Delay)\n\n Plaintiff is barred from enforcing the Asserted Patents due to inexcusable delay in reviving\n\nthe \u2019867 patent after it was abandoned.\n\n SEVENTH DEFENSE\n (Intervening Rights)\n\n Plaintiff\u2019s claims are barred by the doctrine of absolute intervening rights and the doctrine\n\nof equitable intervening rights with respect to any Accused Product or technology that predates\n\nthe date of the revival of the \u2019867 patent and/or the date of the reissue of the \u2019461 and \u2019438 patents.\n\n35 U.S.C. \u00a7 252.\n\n EIGHTH DEFENSE\n (Improper Reissue)\n\n The claims of the \u2019438 and \u2019461 patents are invalid pursuant to 35 U.S.C. \u00a7 251 because\n\nthey enlarge the scope of the claims of the original patent and/or because they improperly recapture\n\nsubject matter that the patentee intentionally surrendered to obtain a valid patent.\n\n NINTH DEFENSE\n (28 U.S.C. \u00a7 1498)\n\n On information and belief, Plaintiff\u2019s claims against NVIDIA for patent infringement are\n\nbarred, in whole or in part, by 28 U.S.C. \u00a7 1498.\n\n TENTH DEFENSE\n (Ensnarement)\n\n Plaintiff is barred by the doctrine of ensnarement from contending that any Asserted Claim\n\ncovers any product, service, or method practiced, manufactured, used, sold, or offered for sale by\n\nDefendant in any manner that would ensnare the prior art.\n\n\n\n\n 29\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 31 of\n 3047\n of 46\n\n\n\n\n ELEVENTH DEFENSE\n (Territoriality)\n\n Plaintiff is not entitled to damages arising from any purported indirect infringement by\n\nDefendant that is premised on direct infringement by end-users occurring outside of the United\n\nStates under 35 U.S.C. \u00a7 271.\n\n TWELFTH DEFENSE\n (No Standing)\n\n Plaintiff\u2019s claims are barred because Plaintiff lacks standing to bring this suit. Specifically,\n\nPlaintiff cannot prove that it is the rightful owner of the Asserted Patents.\n\n THIRTEENTH DEFENSE\n (Failure to State a Claim)\n\n The Complaint fails to state a claim upon which relief may be granted.\n\n FOURTEENTH DEFENSE\n (Inconvenient Venue)\n\n For the convenience of parties and witnesses, venue for this action is not convenient\n\nin this district and would be more appropriate in another district. 28 U.S.C. \u00a7 1404.\n\n FIFTEENTH DEFENSE\n (Reservation of Defenses)\n\n Defendant reserves all affirmative defenses under Rule 8(c) of the Federal Rules of\n\nCivil Procedure, as well as any other defenses at law or in equity that may exist now or that may\n\nbe available in the future.\n\n SIXTEENTH DEFENSE\n (Unenforceability Due to Inequitable Conduct)\n\n 1. Each of the claims of the Asserted Patents is unenforceable due to inequitable\n\nconduct committed by prior assignee Neurala, one or more of the named inventors (Anatoli\n\nGorchetchnikov, Heather Marie Ames, Massimiliano Versace, and Fabrizio Santini), and/or\n\n\n\n\n 30\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 32 of\n 3147\n of 46\n\n\n\n\nprosecution counsel for prior assignee Neurala or current assignee NAI (including, but not limited\n\nto, Christopher Max Colice).\n\n 2. This case arises from a pattern of concealment and misrepresentation surrounding\n\na purported improvement to general-purpose computing on graphics processing units (\u201cGPGPU\u201d).\n\nLong before Neurala\u2019s initial patent application that led to the Asserted patents was filed, NVIDIA\n\nhad pioneered GPGPU computing, developing both the hardware and software foundations for\n\nexecuting general-purpose numerical computations on GPUs. NVIDIA\u2019s engineers\u2014through\n\ntechnologies such as BrookGPU and well-known books such as GPU Gems 2\u2014publicly disclosed\n\nthe very concepts later claimed by Neurala. These were not obscure academic papers; they were\n\nwell-known, widely cited works intended to teach the industry how to harness GPUs for scientific\n\ncomputing. Neurala and the named inventors were well aware of not only NVIDIA\u2019s role in\n\nGPGPU development, but of NVIDIA\u2019s specific teachings in GPU Gems 2 and other technologies.\n\n 3. Against this backdrop, the Asserted Patents claim a narrow and incremental\n\npurported improvement to GPGPU\u2014such as merely offloading certain setup and control functions\n\nfrom the host CPU to an \u201caccelerator controller.\u201d This supposed improvement did not create a\n\nnew GPGPU paradigm; it merely repeated well understood ideas from NVIDIA\u2019s prior work and\n\ncontributions to the field. Yet Neurala\u2019s inventors and attorneys withheld NVIDIA\u2019s key patents\n\nand publications\u2014including GPU Gems 2\u2014from the Patent Office while advancing their own\n\napplication.\n\n 4. There are two independent bases for an inequitable conduct finding, either of which,\n\nstanding alone, renders the Asserted Patents unenforceable. Together, they demonstrate a\n\ncoordinated pattern of misleading conduct intended to misdirect the Patent Office about the true\n\n\n\n\n 31\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 33 of\n 3247\n of 46\n\n\n\n\nstate of the art, the inventors\u2019 knowledge of it, and Neurala\u2019s desperate attempt to get patents on\n\ntechnology it did not invent.\n\n 5. Withholding of Material Prior Art. The claims of the Asserted Patents are\n\nunenforceable due to Neurala\u2019s and/or one or more of the named inventors\u2019 intentional withholding\n\nof material prior art references during prosecution of the Asserted Patents with an intent to deceive\n\nthe Patent Office.\n\n 6. False Declaration Regarding Abandonment. The claims of the Asserted Patents\n\nare also unenforceable due to Neurala\u2019s filing of a false declaration regarding abandonment of the\n\napplication that issued as the \u2019867 patent. Without that false declaration, none of the Asserted\n\nPatents would have issued.\n\n 7. Collectively, these acts form a coherent pattern of inequitable conduct\u2014a deliberate\n\neffort to obscure NVIDIA\u2019s pioneering role in GPGPU computing and to mislead the Patent Office\n\ninto granting patents on technology NVIDIA and others had already disclosed to the world.\n\nWithholding Material Prior Art\n\n 8. Neurala, each of the named inventors, and their counsel involved in the prosecution\n\nof the Asserted Patents had a duty of candor and good faith in dealing with the Patent Office, as\n\nrequired by 37 C.F.R. \u00a7 1.56. Their individual and collective failure to disclose known material\n\nprior art was done with specific intent to mislead or deceive the Patent Office into issuing each of\n\nthe Asserted Patents. As a result, all of the Asserted Patents are unenforceable due to inequitable\n\nconduct.\n\n\n\n\n 32\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 34 of\n 3347\n of 46\n\n\n\n\nKnowledge and Materiality of GPU Gems 2\n\n 9. Neurala, one or more of the named inventors, and/or Neurala\u2019s prosecution counsel\n\nintentionally withheld GPU Gems 2: Programming Techniques for High-Performance Graphics\n\nand General-Purpose Computation (\u201cGPU Gems 2\u201d) (March 2005) from the Patent Office.\n\n 10. NVIDIA published a series of books that provided practical guidance and\n\ntechniques for using GPUs in general-purpose applications, helping developers harness the parallel\n\nprocessing power of GPUs for a wide range of fields. One of those books was GPU Gems 2, which\n\nNVIDIA published on the Internet in March 2005 and made it available for download for free to\n\nanyone that wanted to download it. See Ex. 1 (April 1, 2025 Invalidity Contentions Ex. A11), Ex.\n\n2 (August 29, 2025 Supplemental Invalidity Contentions Supp. Ex. A11), Ex. 3 (April 1, 2025\n\nInvalidity Contentions App\u2019x B), Ex. 4 (April 1, 2025 Invalidity Contentions Ex. C11). At least\n\none of the named inventors\n\n\n\n . Despite extensive knowledge of GPU Gems 2 and its direct relevance for teaching\n\ntechniques for using GPUs in general-purpose applications, that inventor and Neurala withheld\n\nGPU Gems 2 from the Patent Office for all seven years that the \u2019867 patent was pending and every\n\nyear since.\n\n 11. not only was intimately familiar with GPU Gems 2 prior to the\n\nfiling of the provisional patent application from which the Asserted Patents claim priority but also\n\n\n\n\n 33\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 35 of\n 3447\n of 46\n\n\n\n\n 12. and one or more of the other named inventors met regularly to\n\ndiscuss the implementation of the project and the filing of the provisional patent application from\n\nwhich the Asserted Patents claim priority. See, e.g.,\n\nOn information and belief, the other named inventors and/or prosecution counsel were also aware\n\nof GPU Gems 2 due to the close nature of their working relationship with .\n\n 13. As NVIDIA detailed in its Invalidity Contentions served April 1, 2025, August 29,\n\n2025, and October 30, 2025, GPU Gems 2 is a prior art reference that is material to each of\n\nthe \u2019867, \u2019461, and \u2019438 patents. GPU Gems 2 anticipates the \u2019867 and \u2019438 patents and discloses\n\nkey elements of the \u2019461 patent claims. When combined with other unconsidered prior art,\n\nincluding NVIDIA\u2019s own patents, GPU Gems 2 renders obvious all of the Asserted Claims of the\n\nAsserted Patents. See Exs. 1, 2, 3, 4; see also Ex. 14 (April 1, 2025, Preliminary Invalidity\n\nContentions Cover Pleading), Ex. 15 (August 29, 2025 Supplemental Invalidity Contentions Cover\n\nPleading), Ex. 16 (October 30, 2025 Second Supplemental Invalidity Contentions Cover Pleading).\n\nThe Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but for the\n\nwithholding of GPU Gems 2. Therasense, Inc. v. Becton, Dickinson & Co., 649 F.3d 1276, 1290\u2013\n\n91 (Fed. Cir. 2011). For example, as NVIDIA details in its Invalidity Contentions with respect to\n\nclaim 16 of the \u2019867 patent, GPU Gems 2 teaches \u201can accelerator controller, operably coupled to\n\nthe accelerator memory and the central processing unit.\u201d See Ex. 1 at 15\u201344; Ex. 2; Ex. 17 (\u2019867\n\npatent file history) at 43. GPU Gems 2 teaches that the accelerator controller that \u201ctransfer[s] the\n\n\n\n\n 34\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 36 of\n 3547\n of 46\n\n\n\n\nat least the portion of the input data into the accelerator memory before the first computational\n\ncycle.\u201d See Ex. 1 at 27\u201344; Ex. 2; Ex. 17 at 43. GPU Gems 2 teaches that the accelerator controller\n\nthat \u201ctransfer[s] the first output data from the accelerator memory to the main memory during the\n\nsecond computational cycle\u201d and \u201cdirect[s] the second output data into the second partition during\n\nthe second computational cycle.\u201d See Ex. 1 at 54\u201369; Ex. 2; Ex. 17 at 43. GPU Gems 2 teaches\n\nthat the accelerator controller \u201cswap[s] the first pointer and the second pointer at the conclusion of\n\nthe second computational cycle such that the second output data becomes an input for a third\n\ncomputational cycle of the plurality of computational cycles.\u201d See Ex. 1 at 69\u201378; Ex. 2; Ex. 17\n\nat 43.\n\nKnowledge and Materiality of SANNDRA/KInNeSS\n\n 14. Mr. Gorchetchnikov and Mr. Massimiliano Versace\u2014named inventors of the\n\nAsserted Patents\u2014developed Synchronous Artificial Neuronal Networks Distributed Runtime\n\nAlgorithm (SANNDRA) and its implementation on KDE Integrated NeuroSimulation Software\n\n(KInNeSS) (\u201cSANNDRA/KInNeSS\u201d) (March 2005). SANNDRA version 1.1.x and KInNeSS\n\n0.3.3 (on which SANNDRA was implemented) were publicly available and in use by March 2005\n\nbased at least on the following information: KInNeSS: A new software environment for\n\nsimulations of neuronal activity; 9th International conference on Cognitive and Neural Systems\n\n(Boston, MA, 2004); https://web.archive.org/web/20051030032020/http://www.kinness.net/\n\n(KInNeSS documentation);\n\nhttps://web.archive.org/web/20080828055305fw /http://symphony.bu.edu/\n\nDocs/SANNDRA/html/index.html (SANNDRA API documentation); see also Ex. 5 (April 1, 2025\n\nInvalidity Contentions Ex. A7), Ex. 6 (April 1, 2025 Invalidity Contentions Ex. B7), Ex. 7 (April\n\n1, 2025 Invalidity Contentions Ex. C7).\n\n\n\n\n 35\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 37 of\n 3647\n of 46\n\n\n\n\n 15. Furthermore, the specification of the Asserted Patents acknowledges that\n\nSANNDRA \u201cwas developed to accelerate and optimize processing of numerical integration of\n\nlarge non-homogenous systems of differential equations.\u201d \u02bc867 patent at 9:25\u201235. And although\n\nthe Asserted Patents reference version 2.x.x of SANNDRA as \u201can example practical software\n\nimplementation of the method and architecture described above and pictorially represented in FIG.\n\n3,\u201d the applicant failed to identify previous, publicly available versions of SANNDRA or KInNeSS\n\nas relevant prior art to the Patent Office. Id.;\n\n\n\n Neurala, the named\n\ninventors, and their prosecution counsel further failed to fully disclose the relevance and\n\nmateriality of their own software to the Patent Office, despite being the ones in the best position\n\nto do so. NVIDIA expects further discovery, including complete production of the documents that\n\nNVIDIA requested from Neurala on July 30, 2025, to shed further light on Mr. Gorchetchnikov\u2019s\n\nand Mr. Versace\u2019s concealment of earlier versions of SANNDRA/KInNeSS.\n\n 16. SANNDRA 1.1.x and earlier versions, as implemented on KInNeSS, together with\n\nother undisclosed references (such as Nickolls and Kirk, among others) renders obvious all of the\n\nAsserted Claims of the Asserted Patents as shown by Exs. 5, 6, 7, and 16. The Patent Office would\n\nnot have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but for the withholding of\n\nSANNDRA/KInNeSS. Therasense, 649 F.3d at 1290\u201391.\n\nKnowledge and Materiality of Cg\n\n 17. The C for Graphics (Cg) language (2003) is a high-level shading language created\n\nby NVIDIA in collaboration with Microsoft to program graphics shaders on GPUs. Cg was made\n\navailable as an open-source release and in public use by 2003 based at least on the following\n\n\n\n\n 36\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 38 of\n 3747\n of 46\n\n\n\n\ninformation: The Cg Tutorial: The Definitive Guide to Programmable Real-Time Graphics (2003);\n\nhttps://web.archive.org/web/20041205090713/http://developer.nvidia.com:80/object/cg_toolkit.h\n\ntml (\u201cCg Toolkit\u201d); see also Ex. 8 (April 1, 2025 Invalidity Contentions Ex. A4), Ex. 9 (April 1,\n\n2025 Invalidity Contentions Ex. B4), Ex. 10 (April 1, 2025 Invalidity Contentions Ex. C4).\n\n 18.\n\n\n\n 19. The Cg language, together with other references (such as Nickolls and Kirk, among\n\nothers) renders obvious all of the Asserted Claims of the Asserted Patents as shown by Exs. 8, 9,\n\n10, and 16. The Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but\n\nfor the withholding of Cg. Therasense, 649 F.3d at 1290\u201391.\n\nKnowledge and Materiality of BrookGPU\n\n 20. The BrookGPU programming language (2004) is an early system developed at\n\nStanford University to enable general-purpose computing on graphics processing units (GPGPU).\n\nBrookGPU was publicly available and in use by 2004 based at least on the following information:\n\nBuck et al., Brook for GPUs: Stream Computing on Graphics Hardware, ACM, 2004 (\u201cBrook for\n\nGPU 2004\u201d); https://web.archive.org/web/20041205061111/http://graphics.stanford.edu/projects/\n\nbrookgpu/start.html (BrookGPU documentation); see also Ex. 11 (April 1, 2025 Invalidity\n\nContentions Ex. A5), Ex. 12 (April 1, 2025 Invalidity Contentions Ex. B5), Ex. 13 (April 1, 2025\n\nInvalidity Contentions Ex. C5).\n\n 21.\n\n\n\n 22. The BrookGPU language, together with other references (such as Nickolls and Kirk,\n\namong others) renders obvious all of the Asserted Claims of the Asserted Patents, as shown by Exs.\n\n\n\n\n 37\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 39 of\n 3847\n of 46\n\n\n\n\n11, 12, 13, and 16. The Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to\n\nissue but for the withholding of BrookGPU. Therasense, 649 F.3d at 1290\u201391.\n\nWithholding of Material References\n\n 23. Despite their awareness of numerous references relevant to the technology of the\n\nAsserted Patents, the named inventors did not provide any prior art to the Patent Office during\n\nprosecution of the \u02bc867 patent. Only four references are disclosed on the face of the \u2019867 patent\n\nas having been considered during prosecution, and all four were identified by the examiner in a\n\nNotice of References Cited. Although each of the four references cited by the examiner relates to\n\ngraphics rendering, none of the references provide the practical guidance and techniques for using\n\nGPUs in general-purpose applications that GPU Gems 2 does, or the relevant applied examples\n\nand implementations that system art, such as SANNDRA/KInNeSS, Cg, or BrookGPU provides.\n\nFor each of the two reissue patents, the applicant took the opposite approach and submitted\n\nhundreds of references, none of which was GPU Gems 2, SANNDRA/KInNeSS, Cg, or\n\nBrookGPU, and none of which provide the relevant applied examples and implementations that\n\nGPU Gems 2 does.\n\n 24. None of GPU Gems 2, SANNDRA/KInNeSS, Cg, or BrookGPU is cumulative of\n\nthe information already on record. Unlike the four graphics-rendering references cited by the\n\nexaminer, these materials disclose practical architectures, applied examples, and implementation-\n\nlevel guidance applicable to GPGPU\u2014the very subject matter of the Asserted Patents. But for\n\ntheir withholding, the Patent Office would not have allowed any of the \u2019867, \u2019461, or \u2019438 patents\n\nto issue. The deliberate withholding of these NVIDIA and other GPGPU-related references\n\ndeprived the examiner of the most relevant prior art and materially misled the Patent Office about\n\nthe true state of the art.\n\n\n\n\n 38\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 40 of\n 3947\n of 46\n\n\n\n\nIntent to Deceive\n\n 25. \u2014and by extension, Neurala\u2014knew of these material prior art\n\nreferences. On information and belief, the other named inventors and their counsel involved in\n\nprosecution were similarly aware of these references and knew of their materiality to the Asserted\n\nPatents. and others were aware of GPU Gems 2, SANNDRA/KInNeSS, Cg,\n\nor BrookGPU and made the conscious decision to withhold it from the Patent Office.\n\n 26. The inventors did not file their provisional application until September 25, 2006,\n\nover a year after first reviewed GPU Gems 2. And the inventors did not\n\ndisclose GPU Gems 2 to the Patent Office at any time during nearly seven years of prosecution of\n\nthe application that led to the \u2019867 patent.\n\n\n\n\n Each of these references was material and not\n\ncumulative of the bare record before the examiner during prosecution of the \u2019867 patent. It is\n\nsimply not credible that the named inventors did not think that any prior art was material to\n\nprosecution. These facts demonstrate an intent to deceive the Patent Office by not providing any\n\nprior art for its consideration. Thus, for the \u02bc867 patent, by withholding all known references, the\n\ninventors may have aimed to create a misleading impression of the uniqueness and inventiveness\n\nof their claims. For the two reissue patents, the applicant attempted to flood the Patent Office with\n\nreferences to distract from the key prior art that was omitted: GPU Gems 2, SANNDRA/KInNeSS,\n\nCg, and BrookGPU. These facts, and those yet to be ascertained through discovery demonstrate\n\nthat the most reasonable inference to draw is that the named inventors intended to deceive the\n\nPatent Office by withholding references during prosecution.\n\n\n\n\n 39\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 41 of\n 4047\n of 46\n\n\n\n\n 27. On information and belief, Neurala and the named inventors also intentionally\n\nchose not to disclose GPU Gems 2 to its attorneys responsible for prosecution of the patent\n\napplications at the Patent Office, knowing that such attorneys also owed a duty of candor to the\n\nPatent Office, and would disclose the reference to the Patent Office and the examiner if the\n\nattorneys were to become aware of it. In either case, Neurala and the named inventors violated\n\nthe duty of candor that each of them owed to the Patent Office.\n\n 28. A pattern of deceiving the Patent Office continues. For example, current assignee\n\nNeural AI recently paid the 11-year maintenance fee for the \u2019867 patent and did so with a\n\nrepresentation that it was entitled to small entity status, even though it knew it was no longer\n\nentitled to small entity status due to\n\n 29. These facts, when taken together with the evidence of intent presented for the other\n\nbasis of inequitable conduct, demonstrate a pattern of conduct that shows a continuing intent to\n\ndeceive the Patent Office.\n\nFiling a False Declaration Regarding Abandonment of an Application\n\n 30. In addition to failing to disclose material prior art references during the prosecution\n\nof the \u2019867 patent that, if cited, would have precluded the claims in that patent from issuing,\n\nNeurala and its prosecution counsel also affirmatively misled the Patent Office when it revived the\n\nabandoned application that issued as the \u2019867 patent. But for its misrepresentation, none of the\n\nAsserted Patents would have issued because the patent application from which all of those patents\n\nstem would have remained abandoned. This affirmative misrepresentation is part of Neurala\u2019s\n\ncontinued pattern of inequitable conduct in front of the Patent Office to obtain the Asserted Patents.\n\n 31. Specifically, Neurala and its prosecution counsel allowed the application that issued\n\nas the \u2019867 patent (U.S. Patent Appl. No. 11/860,254) (\u201cthe \u2019254 application\u201d) to become\n\n\n\n\n 40\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 42 of\n 4147\n of 46\n\n\n\n\nabandoned for more than two years before belatedly filing a false, generic declaration alleging that\n\nthe \u2019254 application had been unintentionally abandoned in an attempt to revive it. The claims of\n\neach of the Asserted Patents are therefore also unenforceable due to a false declaration regarding\n\nabandonment of the application that issued as the \u2019867 patent.\n\n 32. The application that issued as the \u2019867 patent (U.S. Patent Appl. No. 11/860,254)\n\n(\u201cthe \u2019254 application\u201d) was filed on September 24, 2007. The Patent Office issued a non-final\n\nOffice Action on September 16, 2010, with a three month non-statutory time period for reply. A\n\nresponse to the Office Action was due on December 16, 2010 without payment of extension fees,\n\nbut the applicant neither filed a response nor requested an extension of time under the provisions\n\nof 37 C.F.R. \u00a7 1.136(a), and the \u2019254 application became abandoned on December 17, 2010, the\n\nday after the expiration of the shortened non-statutory deadline established in the non-final Office\n\nAction.\n\n 33. On April 12, 2011, the Patent Office mailed a notice of abandonment to the\n\napplicant. It was not until July 31, 2013\u2014more than two years later\u2014that the applicant filed a\n\npetition to revive the \u2019254 application. The petition was signed by Christopher Max Colice of\n\nFoley & Lardner LLP and included the statement that \u201c[t]he entire delay in filing the required reply\n\nfrom the due date for the required reply until the filing of a grantable petition under 37 CFR 1.137(b)\n\nwas unintentional.\u201d Ex. 17 at 64\u201388. At the time, no power of attorney had been filed listing Mr.\n\nColice as Neurala\u2019s attorney of record. There is no indication in the statement what investigation\n\nMr. Colice undertook or how he determined that abandonment was \u201cunintentional.\u201d Furthermore,\n\nthe attorney advisor reviewing the petition noted that it was \u201cnot apparent whether the person\n\nsigning the statement of unintentional delay was in a position to have firsthand or direct knowledge\n\n\n\n\n 41\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 43 of\n 4247\n of 46\n\n\n\n\nof the facts and circumstances of the delay at issue,\u201d and that there was \u201cno indication that the\n\npetition [wa]s signed by a registered patent attorney or patent agent of record. Id. at 62.\n\n 34. Materiality. The material prong is met \u201c[w]hen the patentee has engaged in\n\naffirmative acts of egregious misconduct, such as the filing of an unmistakably false affidavit.\u201d\n\nTherasense, 649 F.3d at 1292; see also Rohm & Haas Co. v. Crystal Chem. Co., 722 F.3d 1556,\n\n1571 (Fed. Cir. 1983) (\u201cthere is no room to argue that submission of false affidavits is not\n\nmaterial\u201d); Intellect Wireless, Inc. v. HTC Corp., 732 F.3d 1339, 1342 (Fed. Cir. 2013); Apotex,\n\nInc. v. UCB, Inc., 763 F.3d 1354 (Fed. Cir. 2014). An affirmative act of egregious misconduct is\n\ninherently material. Therasense, 649 F.3d at 1292. The filing of a false revival petition under 37\n\nCFR \u00a7 1.137(a) is an affirmative act of egregious misconduct. In re Rembrandt Techs. LP Patent\n\nLitig., 899 F.3d 1254, 1272\u201374 (Fed. Cir. 2018). Because the \u02bc867 patent (which issued from the\n\n\u02bc254 application) was the first patent in the family, but for the false statement in the Petition for\n\nRevival, all Asserted Patents would not have issued and thus would no longer be in force. The\n\nfalse statement to the Patent Office is therefore material to patentability. See, e.g., 3D Med.\n\nImaging Sys. LLC v. Visage Imaging Inc., 228 F. Supp. 3d 1331, 1338\u201339 (N.D. Ga. 2017).\n\n 35. Intent. The intent of Mr. Colice and/or the named inventors to deceive the Patent\n\nOffice is evidenced at least by the length of time that elapsed between the dates of the Office\n\nAction (September 16, 2010), when the application became abandoned (December 18, 2010), the\n\nNotice of Abandonment (April 12, 2011), and the applicant\u2019s Petition for Revival (July 31, 2013).\n\nFor example, there is no explanation why it took almost three full years from when the Office\n\nAction was issued for the applicant to respond to the Office Action. Even after Neurala LLC was\n\nnotified of the abandonment, it took more than two years to file a Petition for Revival, accompanied\n\nby only a generic statement that the entire delay was unintentional. The most reasonable inference\n\n\n\n\n 42\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 44 of\n 4347\n of 46\n\n\n\n\nto draw is that there was an intent to allow the application to become abandoned, and that Mr.\n\nColice and/or the named inventors intended to deceive the Patent Office by submitting a false\n\ndeclaration stating that the abandonment of the \u2019254 application was unintentional.\n\n 36. The \u02bc254 application was the first application in the chain of applications for the\n\nAsserted Patents (it issued as the asserted \u02bc867 patent). If the \u02bc254 application had been deemed\n\nabandoned and no false declaration had been filed, no patents in the asserted patent family would\n\nhave issued. Therefore, all Asserted Patents should be rendered unenforceable due to the filing of\n\na false declaration in the prosecution of the \u02bc867 patent.\n\n 37. These facts, when taken together with the evidence of intent presented for the other\n\nbasis of inequitable conduct, demonstrate a pattern of conduct that shows a continuing intent to\n\ndeceive the Patent Office.\n\nInfectious Unenforceability\n\n 38. The \u02bc254 application\u2014from which the \u2019867 patent issued\u2014was the first application\n\nin the chain of applications for the Asserted Patents. Each of the \u2019461 and the \u2019438 patents is a\n\nchild of the \u2019867 patent, and the pattern of blatantly inequitable conduct that pervades the\n\nprosecution of this patent family renders the claims of each of the Asserted Patents unenforceable.\n\n 39. Here, but for the false declaration filed in support of the revival of the \u2019254\n\napplication, no patents in the asserted patent family would have ever issued. Lumenyte Intern.\n\nCorp. v. Cable Lite Corp., 96-1011, 1996 U.S. App. LEXIS 16400 (Fed. Cir. July 9, 1996) (a false\n\naffidavit filed to revive an abandoned patent results in the unenforceability of later-filed, related\n\npatents).\n\n 40. Additionally, but for the withholding of material prior art\u2014including GPU Gems\n\n2, Cg, SANNDRA/KInNeSS, and BrookGPU, the Patent Office would not have allowed any of\n\n\n\n\n 43\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 45 of\n 4447\n of 46\n\n\n\n\nthe Asserted Patents to issue. All of the Asserted Patents were procured through a pattern of\n\ninequitable conduct as discussed herein.\n\n 41. From withholding material prior art to falsely claiming unintentional abandonment\n\nto obtain a patent, Neurala, the named inventors, and prosecution counsel have exhibited a pattern\n\nof intentional deception of the Patent Office. The remedy for the repeated and egregious instances\n\nof inequitable conduct is to render each of the Asserted Patents unenforceable.\n\n PRAYER FOR RELIEF\n\n Wherefore, Defendant respectfully requests judgment in its favor with the following relief:\n\n a) The Complaint be dismissed with prejudice.\n\n b) Judgment that Defendant has not infringed and is not infringing, either directly or\n\n indirectly, any of the claims the \u2019867 Patent, the \u2019461 Patent, and the\u2019 438 Patent,\n\n in violation of 35 U.S.C. \u00a7 271.\n\n c) Judgment that the claims of the \u2019867 Patent, the \u2019461 Patent, and the\u2019 438 Patent\n\n are invalid.\n\n d) Judgment that the \u2019867 Patent, the \u2019461 Patent and the \u2019438 Patent, including all of\n\n their claims, are unenforceable due to inequitable conduct.\n\n e) Judgment and determination that this case is exceptional under 35 U.S.C. \u00a7 285 and\n\n that Defendant is entitled to its attorneys\u2019 fees, costs, and expenses in defending\n\n this action.\n\n f) Such other relief, including other monetary and equitable relief, as this Court deems\n\n just and proper.\n\n DEMAND FOR JURY TRIAL\n\n Defendant demands a jury trial on all issues so triable.\n\n\n\n\n 44\n\f Case\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 46 of\n 4547\n of 46\n\n\n\n\nDated: November 18, 2025\n\n /s/ L. Kieran Kieckhefer\n L. Kieran Kieckhefer (pro hac vice)\n Jaysen S. Chung (pro hac vice)\n GIBSON, DUNN & CRUTCHER LLP\n One Embarcadero Center, Suite 2600\n San Francisco, CA 94111\n (415) 393-8200\n kkieckhefer@gibsondunn.com\n jschung@gibsondunn.com\n\n Brian Rosenthal\n Ahmed ElDessouki (pro hac vice)\n GIBSON, DUNN & CRUTCHER LLP\n 200 Park Ave.\n New York, NY 10166\n (212) 351-4000\n brosenthal@gibsondunn.com\n aeldessouki@gibsondunn.com\n\n Lillian J. Mao (pro hac vice)\n GIBSON DUNN & CRUTCHER LLP\n 1881 Page Mill Road\n Palto Alto, CA 94301-1211\n (650) 849-5307\n lmao@gibsondunn.com\n\n Barry K. Shelton (Texas State Bar No. 24055029)\n SHELTON COBURN LLP\n 311 RR 620 S, Suite 205\n Austin, TX 78734\n (512) 263 2165\n bshelton@sheltoncoburn.com\n\n Counsel for Defendant NVIDIA Corporation\n\n\n\n\n 45\n\fCase\n Case\n 7:24-cv-00221-ADA-DTG\n 7:26-mc-00318-LS Document\n Document\n 6-9130Filed\n Filed\n 08/18/26\n 11/25/25Page\n Page\n 47 of\n 4647\n of 46\n\n\n\n\n CERTIFICATE OF SERVICE\n\n I hereby certify that all counsel of record are being served with a copy of the foregoing\n\n documents via electronic mail on November 18, 2025.\n\n /s/ L. Kieran Kieckhefer\n L. Kieran Kieckhefer\n\n\n\n\n 46\n\f","ocr_status":2,"date_upload":"2026-08-20T15:37:45.060512-07:00","document_number":"6","attachment_number":9,"pacer_doc_id":"181037220276","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 8","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554431/","id":490554431,"tags":[],"absolute_url":"/docket/74659430/6/10/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.310819-07:00","date_modified":"2026-08-23T09:39:39.875560-07:00","sha1":"2f784a1b64eb94b834c508d12c6d20d487bf6344","page_count":47,"file_size":264190,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.10.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.10.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 1 of 47\n\n\n\n\n EXHIBIT\n\n 9\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 2 of 47\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n Case No. 7:24-cv-00221-ADA-DTG\n Plaintiff,\n JURY TRIAL DEMANDED\n v.\n\nNVIDIA CORPORATION,\n\n Defendants.\n\n\n\n\n NON-PARTY TESLA, INC.\u2019S OBJECTIONS AND RESPONSES TO PLAINTIFF\n NEURAL AI, LLC\u2019S SUBPOENA\n Non-Party Tesla, Inc. (\u201cTesla\u201d) hereby serves the following objections and responses\n(\u201cResponses\u201d) to Plaintiff Neural AI, LLC\u2019s (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) (1) Subpoena to\nProduce Documents, Information, or Objects, and (2) Subpoena for Testimony.\n\n PRELIMINARY STATEMENT\n 1. Tesla\u2019s objections and responses to the Requests are made to the best of its\ncurrent knowledge, information, belief, and understanding of the Requests. Tesla reserves the\nright to supplement or amend any responses should future investigation indicate that such\nsupplementation or amendment is necessary.\n 2. Tesla\u2019s responses to the Requests are made solely for the purpose of and in\nrelation to the above-captioned action. Each response is given subject to all appropriate\nobjections (including, but not limited to, objections concerning privilege, competency,\nrelevancy, materiality, propriety, and admissibility). All objections are reserved and may be\ninterposed at any time.\n 3. Tesla\u2019s responses include only information that is within Tesla\u2019s possession,\ncustody, or control.\n\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 3 of 47\n\n\n\n\n 4. Tesla incorporates by reference each and every general objection set forth\nbelow into each and every specific response. From time to time, a specific response may repeat\na general objection for emphasis or some other reason. The failure to include any general\nobjection in any specific response shall not be interpreted as a waiver of any general objection\nto that response.\n 5. Nothing contained in these Responses and Objections or provided in response\nto the Requests consists of, or should be construed as, an admission relating to the accuracy,\nrelevance, existence, or nonexistence of any alleged facts or information referenced in any\nRequest.\n GENERAL OBJECTIONS\n 1. Tesla objects to the Subpoena to the extent it seeks the disclosure of Tesla's\nhighly confidential, proprietary, or trade secret technical and business information, including\nbut not limited to internal software architectures, source code, AI/ML model designs, GPU\ncomputing infrastructure, data flow and execution flow diagrams, and engineering\nspecifications. As a non-party to this litigation, the burden on Tesla to review and produce\nsuch highly sensitive competitive information outweighs the potential relevance to the\nunderlying action.\n 2. Tesla generally objects to each Request, including the Definitions and\nInstructions, on the grounds and to the extent that it purports to impose obligations beyond\nthose imposed or authorized by the Federal Rules, the Federal Rules of Evidence, the Local\nRules, the Court\u2019s standing orders, any other applicable federal or state law, and any\nagreements between the parties. Tesla will construe and respond to the Requests in accordance\nwith the requirements of the Federal Rules and other applicable rules or laws.\n 3. Tesla generally objects to Defendant\u2019s Requests on the grounds that they are\noverbroad, oppressive, unduly burdensome, and disproportionate to the needs of the case,\nparticularly given that Tesla is not a party to this litigation. Tesla objects that the overbreadth\n\n\n\n\n 2\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 4 of 47\n\n\n\n\nof this Subpoena subjects it to undue burden and expense in both searching for and producing\nthe documents called for by this Subpoena.\n 4. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent the Request seeks documents and information that are irrelevant to\nthe claims in, or defenses to, this action, is disproportionate to the needs of the case, and/or is\nof such marginal relevance that its probative value is outweighed by the burden imposed on\nTesla in having to provide such information, including any Request that seeks information for\nany time period outside that which is relevant to the claims and defenses asserted in this action\nparticularly given that Tesla is not a party to this litigation.\n 5. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent the Request is vague, ambiguous, unreasonably cumulative, or\nduplicative, including to the extent it seeks documents or communications that are otherwise\nresponsive to other specific Requests.\n 6. Tesla generally objects to the Requests, including the Definitions and\nInstructions, on the basis that they specify an overbroad and unduly burdensome time period,\nor a time period when the asserted patents in the underlying action were not in force, and seek\ndocuments and things outside of the time period relevant to the claims and defenses asserted\nin this action.\n 7. Tesla generally objects to the Requests, including the Definitions and\nInstructions, to the extent that they are overbroad and/or unduly burdensome, including to the\nextent that they call for the production of \u201cAny,\u201d \u201cany\u201d or \u201call\u201d documents or communications\nconcerning the subject matter referenced therein.\n 8. Tesla generally objects to each Request, including the Definitions and\nInstructions, on the grounds and to the extent the Request purports to request the identification\nand disclosure of any information, communication(s), or document(s) that were prepared in\nanticipation of litigation or in connection with any internal investigation conducted at the\ndirection of counsel, constitute attorney work product, reveal privileged attorney-client\n\n 3\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 5 of 47\n\n\n\n\ncommunications, are covered under the common interest privilege and/or joint defense\nprivilege, or are otherwise protected or immune from disclosure under any applicable\nprivilege(s), law(s), or rule(s). Tesla hereby asserts all such applicable privileges and\nprotections and excludes privileged and protected information from its responses to each\nRequest. See generally Fed. R. Evid. 502.\n 9. Tesla generally objects to any Request to the extent it seeks production of\ninformation and/or documents that comprise or contain confidential information of a third\nparty to whom Tesla believes it owes a duty of confidentiality or otherwise protected from\ndisclosure by agreements between Tesla and other parties.\n 10. Tesla generally objects to any Request to the extent that it seeks to require Tesla\nto provide any information beyond what is available to Tesla at the present time after\nreasonable search of its own records and a reasonable inquiry of its present employees. For\nexample, Telsa objects to any request that seeks information that is not within Tesla\u2019s\npossession, custody or control. Tesla also objects to any Request that seeks to impose a duty\non Tesla to create materials that Tesla does not create or maintain in the ordinary course of\nbusiness.\n 11. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent that it requests information that is confidential, proprietary, or\ncompetitively sensitive.\n 12. Tesla generally objects to each Request to the extent that the information sought\nis more appropriately pursued through another discovery tool.\n 13. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent it is argumentative, lacks foundation, or incorporates allegations and\nassertions that are disputed or erroneous. In furnishing the responses herein, Tesla does not\nconcede the truth of any factual assertion or implication contained in any Request, Definition,\nor Instruction.\n\n\n\n 4\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 6 of 47\n\n\n\n\n 14. Tesla generally objects to any Request to the extent that it seeks information or\nmaterials that are publicly available, already in Defendant\u2019s possession, custody, or control,\nor are equally available to Defendant from another less burdensome source, such as parties to\nthe litigation.\n 15. Tesla generally objects to any Request to the extent that it fails to describe the\ninformation requested with reasonable particularity, is indefinite as to time and scope, seeks\ninformation that is not relevant to the claims or defenses of the parties in this case, and/or is\nnot proportional to the needs of the case.\n 16. Tesla generally objects to any Request to the extent that it requires Tesla to\ndraw legal conclusions.\n 17. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent that it purports to impose an obligation to conduct anything beyond\na reasonable and diligent search of reasonably accessible files (including electronic files)\nwhere responsive documents reasonably would be expected to be found. Any Requests that\nseek to require Tesla to go beyond such a search are overbroad and unduly burdensome.\n 18. Tesla generally objects to the Subpoena to the extent it seeks production of\nTesla's proprietary source code, internal software, custom code, configuration files, build files,\ndeployment files, runtime logs, profiler traces, or other engineering artifacts. Such materials\nconstitute Tesla's core intellectual property and trade secrets, and their production to a non-\nparty in a dispute between Neural AI and NVIDIA is disproportionate, unduly burdensome,\nand risks competitive harm.\n 19. Tesla\u2019s willingness to provide any document or information in response to a\nRequest shall not be interpreted as an admission that such document or information exists, that\nit is relevant to a claim or defense in this action, or that it is admissible for any purpose. Tesla\ndoes not waive its right to object to the admissibility of any document or information produced\nby any party on any ground.\n\n\n\n 5\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 7 of 47\n\n\n\n\n OBJECTIONS TO THE DEFINITIONS AND INSTRUCTIONS\n 1. Tesla objects to Plaintiff's definition of \"NVIDIA GPUs\" as overbroad, unduly\nburdensome, and disproportionate to the needs of the case. The definition encompasses\nvirtually every NVIDIA GPU product ever manufactured across eight architecture generations,\nincluding hundreds of individual product SKUs spanning consumer, enterprise, data center,\nand embedded platforms. This sweeping definition, combined with the document requests,\nwould require Tesla to search for, collect, and review documents relating to any and all\nNVIDIA hardware it has ever used, regardless of relevance to the patents-in-suit. Further,\nsuch information can be obtained through other, less burdensome and more appropriate means,\nincluding from parties to the litigation.\n 2. Tesla objects to Plaintiff's definitions of \"and\" and \"or\" as overbroad, unduly\nburdensome, impermissibly vague, and not proportional to the needs of the case, to the extent\nthey purport to change the customary and usual meaning of these terms and alter the meaning\nof a phrase to impose requirements in excess of those under the Federal Rules of Civil\nProcedure and the Local Rules of this Court.\n 3. Tesla objects to Plaintiff's definitions of \"any\" and \"each\" as overbroad, unduly\nburdensome, impermissibly vague, and not proportional to the needs of the case, to the extent\nit purports to seek information that is unrelated to the present case, to the extent would capture\ndocuments of no evidentiary value and impose an undue burden on a non-party, and to the\nextent it exceeds the obligations imposed by the Federal Rules, the local rules of this Court,\nand any orders this Court entered in this case.\n 4. Tesla objects to Plaintiff's definitions of \"concerning,\" \"related to,\" \"relating\nto,\" and \"regarding\" as overbroad, vague, and disproportionate to the needs of the case. The\ndefinitions encompass over twenty verbs including \"alluding to,\" \"contradicting,\"\n\"mentioning,\" and \"memorializing,\" which would capture documents of no evidentiary value\nand impose an undue burden on a non-party.\n\n\n\n\n 6\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 8 of 47\n\n\n\n\n 5. Tesla objects to Plaintiff\u2019s definition of \u201cdocument(s)\u201d as overly broad and\nunduly burdensome, to the extent it purports to seek information that is unrelated to the present\ncase, and to the extent it exceeds the obligations imposed by the Federal Rules, the local rules\nof this Court, and any orders this Court entered in this case. Tesla recognizes that Plaintiff\u2019s\ndefinition of \u201cdocument\u201d does not include \u201cSource Code\u201d and Plaintiff has provided a separate\ndefinition for \u201cSource Code.\u201d\n 6. Tesla objects to Plaintiff's definition of \"persons\" as overbroad, unduly\nburdensome, vague, ambiguous, and not proportional to the needs of the case to the extent it\npurports to include \"formal or informal entities and organizations\" and extends to \"public and\nprivate corporations, partnerships, professional corporations, limited liability companies,\nbusiness trusts, banking institutions, associations, firms, joint ventures, commissions, bureaus,\ndepartments, and any other legal entity, including any divisions, subsidiaries, departments, and\nother units thereof\" regardless of relevance to any claim or defense in this case. Tesla further\nobjects to the inclusion of \"informal entities and organizations\" and \"any other legal entity\" as\nunbounded in scope and undefined, as these terms could be interpreted to encompass virtually\nany grouping of individuals or organizational unit without limitation. The breadth of this\ndefinition, when applied across the Requests, would impose an undue burden on Tesla\u2014a non-\nparty\u2014by requiring it to search for and identify documents involving an unlimited universe of\npersons and entities with no meaningful nexus to the claims or defenses in this action. Tesla\nwill interpret the term \"persons\" according to its customary usage in the context of a particular\nRequest and limit any response accordingly.\n 7. Tesla objects to Plaintiff's definition of \"Source Code\" as overbroad and unduly\nburdensome and as calling for information not relevant to the case. The definition encompasses\n\"all versions and revisions,\" \"all associated files,\" \"scripts, header files, makefiles,\nconfiguration files, and documentation\"\u2014effectively requiring Tesla to potentially produce\nentire software development repositories relating to any system that touches an NVIDIA GPU.\nTesla objects to this definition as it would require the disclosure of sensitive Tesla trade secrets\n\n 7\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 9 of 47\n\n\n\n\nand intellectual property in a dispute to which Tesla is not a party and without making any\nshowing of relevance or that the request is proportional to the needs of the case.\n 8. Tesla objects to Plaintiff\u2019s definition of \u201cYou,\u201d and \u201cYour\u201d and its inclusion of\n\u201cbut not limited to its predecessors, successors, parents, subsidiaries, divisions, affiliates, and\nall past or present directors, officers, partners, managers, employees, contractors, agents,\nrepresentatives, accountants, consultants, in-house and outside counsel\u201d as overbroad and\nunduly burdensome to the extent the terms are meant to include any individual(s), entit(ies),\nor any other person(s) over which Tesla exercises no control and to the extent Defendant\npurports to use the terms to impose obligations on Tesla that go beyond the requirements of\nthe Federal Rules and the Local Rules. Tesla further objects to these definitions to the extent\nthat Defendant purports to use these defined terms to seek documents that are not relevant to\nthe claims and defenses in this action, including seeking documents from Tesla\u2019s subsidiaries\nand with respect to products not at issue in this litigation. Tesla will construe the terms \u201cYou\u201d\nand \u201cYour\u201d so as to include only the following: Tesla, Inc. and its employees. Tesla further\nobjects to the definition and its inclusion of \u201cin-house and outside counsel\u201d to the extent that\nit seeks information protected by attorney-client or work product privilege.\n 9. Tesla objects to Plaintiff's Instructions to the extent they impose obligations that\ngo beyond the requirements of the Federal Rules and the Local Rules.\n 10. Tesla objects to Plaintiff's instruction regarding production format, metadata,\nnative format, and load file specifications as overbroad, unduly burdensome, and\ndisproportionate to the needs of this case.\n 11. Tesla objects to Plaintiff's instruction that Tesla identify materials not in its\npossession, custody, or control as overbroad, vague, and disproportionate to the needs of this\ncase to the extent it purports to impose interrogatory-style narrative obligations on Tesla to\nexplain the absence of documents or to speculate regarding the existence, destruction, or\nlocation of documents.\n\n\n\n 8\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 10 of 47\n\n\n\n\n 12. Tesla objects to Plaintiff's instruction regarding continuing obligations to\nsupplement to the extent it purports to impose obligations beyond those required by the Federal\nRules for non-parties.\n 13. Tesla objects to Plaintiff's Requests to the extent they fail to impose a\nmeaningful geographic scope limitation, and further objects to the extent it purports to\nencompass \"non-U.S. activity that directly supports or enables U.S. operations or usage.\" This\nlanguage is undefined, unbounded, and would require Tesla, a non-party, to make subjective\nlegal and factual determinations regarding the geographic nexus of its global computing\noperations. As written, the Requests could be construed to require Tesla to search for, collect,\nand produce documents from any facility, system, or operation worldwide that has any\narguable connection to the United States, imposing a burden and expense on a non-party that\nis grossly disproportionate to the needs of this case. Tesla will construe the Requests as limited\nto activity occurring within the United States.\n 14. Tesla objects to Plaintiff\u2019s \u201cInstructions\u201d to the extent they impose obligations\nthat go beyond the requirements of the Federal Rules and the Local Rules.\n 15. Tesla objects to Plaintiff\u2019s instruction regarding production format and load file\nspecifications as overbroad, unduly burdensome, and disproportionate to the needs of this case.\nTesla further objects to the extent that this instruction purports to impose obligations that go\nbeyond the requirements of the Federal Rules and the Local Rules, as applicable. Tesla will\nconstrue this instruction in accordance with the Federal Rules and the Local Rules.\n 16. Tesla objects to Plaintiff\u2019s instruction regarding the production of documents in\nnative format as overbroad, vague, and disproportionate to the needs of this case to the extent\nit purports to unilaterally define the criteria for native productions that differ from standard\nindustry practice.\n 17. Tesla objects to Plaintiff\u2019s instruction regarding metadata load file requirements\nas overbroad, unduly burdensome, and disproportionate to the needs of this case to the extent\nit purports to unilaterally dictate specific metadata that may not be reasonably available,\n\n 9\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 11 of 47\n\n\n\n\nautomatically generated, or maintained in the ordinary course of business without undue\nburden or expense. Tesla further objects to the extent it purports to unilaterally impose rigid\nmetadata reporting requirements, which as previously stated, may or may not be technically\nfeasible.\n 18. All Tesla\u2019s responses herein and/or related documents will be provided subject\nto any protective order entered in this case by the Court, or, if no protective order is entered,\nas HIGHLY CONFIDENTIAL ATTORNEYS\u2019 EYES ONLY.\n 19. The foregoing general reservations and objections are incorporated into each of\nthe responses and objections to the specific Request set forth below.\n\n\n\n\n 10\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 12 of 47\n\n\n\n\n OBJECTIONS AND RESPONSES TO DOCUMENT REQUESTS\nDOCUMENT REQUEST NO. 1:\n Documents sufficient to identify all software, frameworks, libraries, APIs, scripts,\nSource Code, configuration files, and custom code You use to perform computations on\nNVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 1:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla or third-party software, frameworks, libraries, APIs, scripts, Source Code, configuration\nfiles, and custom code.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case, particularly given that\nTesla is not a party to this litigation. The Request seeks identification of \"all software,\nframeworks, libraries, APIs, scripts, Source Code, configuration files, and custom code\" used\nto perform computations on NVIDIA GPUs, a scope that could potentially encompass virtually\nany software system Tesla operates. Tesla further objects to the terms \"all\" and \"Source Code\"\nas overbroad and unduly burdensome as they would require an exhaustive identification effort\nthat is disproportionate to the discovery need in a dispute to which Tesla is not a party.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\n\n 11\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 13 of 47\n\n\n\n\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla confidential trade secrets and proprietary technical\ninformation, internal software architectures, and custom-built AI/ML frameworks that\nconstitute core intellectual property and are not relevant to this case. Production of such\nmaterials, particularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to this Request as vague and ambiguous as to what\n\"computations\" are relevant to the underlying litigation, and as to the meaning and scope of\nthe terms \u201csoftware, frameworks, libraries, APIs, scripts, Source Code, configuration files, and\ncustom code.\u201d\n Tesla further objects to this Request to the extent it seeks information beyond the use\nof NVIDIA products.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to the extent this\nRequest seeks to impose a duty on Tesla to create materials or compile information that Tesla\ndoes not create or maintain in the ordinary course of business. Tesla further objects on the\ngrounds that information regarding NVIDIA's software, frameworks, and libraries is more\nreadily available from NVIDIA, the Defendant in this action.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that\nit calls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 12\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 14 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 2:\n Documents sufficient to show whether You use NVIDIA's Aerial, Clara Parabricks,\ncuBLAS, cuDNN, cuFFT, cuQuantum, cuSOLVER, cuSPARSE, Drive, DriveWorks,\nHoloscan, Isaac, Isaac Lab, Maxine, Memory Map, Merlin, Metropolis, Modulus, Monai,\nMorpheus, NeMo, PyTorch, RAPIDS, Riva, Runtime Driver, TensorFlow, TensorRT, Triton,\nVSS (Deepstream), or any other NVIDIA software as part of computations You perform using\nNVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 2:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla\u2019s use of software specifically listed in this Request or \u201cany other NVIDIA software.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The Request identifies over thirty\nspecific NVIDIA software products and then adds the catch-all phrase \"or any other NVIDIA\nsoftware,\" rendering the Request virtually unlimited in scope.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentationrepresenting\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's highly confidential and proprietary information\n\n\n 13\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 15 of 47\n\n\n\n\nregarding its internal software stack, computing infrastructure, and technology choices, which\nconstitute competitively sensitive business information. Disclosure of which specific NVIDIA\ntools Tesla does or does not use would reveal Tesla's internal technology strategy and\ncompetitive posture.\n Tesla further objects on the grounds that information regarding NVIDIA's software\nproducts and their usage by customers is more appropriately sought from NVIDIA, the\nDefendant in this action, which possesses licensing records, telemetry data, and customer\nusage information.\n Tesla objects that the catch-all phrase \"any other NVIDIA software\" renders the\nRequest vague, ambiguous, and boundless in scope, as it effectively extends the demand to\nevery NVIDIA product, tool, or library, without temporal or functional limitation.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects on the grounds\nthat the information sought is more readily obtainable from the parties to this litigation or from\nother less burdensome sources, thus disproportionate to the needs of the case.\n Tesla further objects to the extent that this Request is duplicative of Request No. 1.\nTesla objects to this Request on the grounds that it is unduly burdensome and oppressive to\nthe extent that it seeks information and documents that are equally available to the parties in\nthis litigation.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n 14\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 16 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 3:\n\n Documents sufficient to show whether You use sample Source Code provided by\nNVIDIA as part of computations You perform using NVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 3:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The term \"sample Source Code\" is\nvague and ambiguous, as it is unclear what specific sample code is being referenced, what\nuniverse of NVIDIA sample code is at issue, or how Tesla would be expected to identify\nwhether any portion of its codebase derives from, incorporates, or was inspired by NVIDIA\nsample code. Tesla objects to this Request to the extent that compliance would require Tesla\nto conduct an exhaustive audit of its entire codebase against an undefined body of NVIDIA\nsample code, a task that is extraordinarily burdensome and disproportionate for a non-party.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's proprietary\nsource code and internal software, which constitute trade secrets and core intellectual property.\nTesla further objects on the grounds that NVIDIA sample source code is publicly available\nand/or in NVIDIA's possession, custody, or control, and information regarding its distribution\nto customers is more appropriately sought from the Defendant. Tesla further objects to this\nRequest to the extent it seeks documents containing confidential, proprietary or trade secret\ninformation without making any showing of relevance or that the request is proportional to the\nneeds of the case.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case. Tesla objects to this Request to the extent it seeks\ninformation protected by the attorney-client privilege or the work-product doctrine. Tesla\n\n 15\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 17 of 47\n\n\n\n\nfurther objects to the extent this Request seeks to impose a duty on Tesla to create materials\nor compile analyses that Tesla does not create or maintain in the ordinary course of business.\nTesla further objects to the extent that this Request is duplicative of Requests Nos. 1 and 2.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 4:\n Documents sufficient to show whether and how any software You use to perform\ncomputations on NVIDIA GPUs calls, invokes, interfaces with, wraps, depends on, sits on top\nof, modifies, extends, or implements functionality provided by CUDA, cuDNN, TensorRT,\nCUDA libraries, CUDA drivers, CUDA runtime, CUDA applications or frameworks or any\nother NVIDIA software.\nRESPONSE TO DOCUMENT REQUEST NO. 4:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether and how Tesla software \u201cperform[s] computations on NVIDIA GPUs calls, invokes,\ninterfaces with, wraps, depends on, sits on top of, modifies, extends, or implements\nfunctionality provided by CUDA, cuDNN, TensorRT, CUDA libraries, CUDA drivers, CUDA\nruntime, CUDA applications or frameworks or any other NVIDIA software.\u201d\n\n 16\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 18 of 47\n\n\n\n\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to this\nRequest as being vague and ambiguous to the extent it seeks documents showing \"whether and\nhow\" Tesla's software \"calls, invokes, interfaces with, wraps, depends on, sits on top of,\nmodifies, extends, or implements\" NVIDIA functionality.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla objects to this Request to the extent it implicates a scope that would require Tesla\nto map and document every interaction between its proprietary software systems and\nNVIDIA's computing stack across its entire business. The enumeration of nine distinct verbs\ndescribing software interaction, combined with the catch-all phrase \"or any other NVIDIA\nsoftware,\" renders the Request virtually unlimited in scope.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary technical\ninformation, including internal software architectures, custom integrations, dependency\nstructures, and engineering designs that constitute core intellectual property.\n Tesla objects to this Request as compound and vague. The Request combines at least\nnine distinct NVIDIA technologies with at least nine distinct functional relationships, creating\na matrix of discrete inquiries posed as a single request.\n Tesla further objects to the terms \"sits on top of,\" \"wraps,\" and \"interfaces with\" as\nvague, ambiguous, and susceptible to multiple technical interpretations that render meaningful\ncompliance impossible without speculation. The catch-all phrase \"any other NVIDIA\n\n 17\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 19 of 47\n\n\n\n\nsoftware\" renders the Request boundless in scope, extending the demand to every NVIDIA\nproduct, tool, or library, without temporal or functional limitation.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as architectural mappings, dependency analyses, or software interaction\ndiagrams, that Tesla does not create or maintain in the ordinary course of business. Tesla\nobjects to this Request to the extent it seeks information protected by the attorney-client\nprivilege or the work-product doctrine. Tesla further objects to the extent that this Request is\nduplicative of Requests Nos. 1, 2, and 3.\n Tesla further objects on the grounds that information regarding NVIDIA's CUDA\nplatform, software libraries, and their interfaces is extensively documented in NVIDIA's\npublic developer resources and is more appropriately sought from the Defendant. Tesla further\nobjections to this Request to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 5:\n Documents sufficient to show the architecture, design, data flow, control flow, and\nexecution flow of any system in which You use NVIDIA GPUs to perform computations,\nincluding diagrams, technical specifications, design documents, Powerpoints, slide decks,\ninternal and external presentations, Source Code, configuration files, build files, deployment\nfiles, runtime logs, and profiler traces.\n\n\n 18\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 20 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 5:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201cthe architecture, design, data flow, control flow, and execution flow of any system in which\n[Tesla] use[s] NVIDIA GPUs to perform computations, including diagrams, technical\nspecifications, design documents, Powerpoints, slide decks, internal and external presentations,\nSource Code, configuration files, build files, deployment files, runtime logs, and profiler\ntraces.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. The Request\nseeks the \"architecture, design, data flow, control flow, and execution flow\" of every system\nin which Tesla uses NVIDIA GPUs, along with virtually every category of engineering artifact,\n\"diagrams, technical specifications, design documents, Powerpoints, slide decks, internal and\nexternal presentations, Source Code, configuration files, build files, deployment files, runtime\nlogs, and profiler traces.\" Tesla objects to this Request to the extent that this catch-all request\nwould potentially require Tesla to produce an unknowable quantum of engineering\ndocumentation and source code representing Tesla's valuable intellectual property. The\nRequest fails to identify with sufficient particularity what specific implementation or\napplication of these techniques is at issue in the underlying litigation, rendering meaningful\ncompliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's highly confidential trade secrets and proprietary\n\n 19\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 21 of 47\n\n\n\n\ntechnical information, including proprietary source code, internal software architectures, and\ncustom-built AI/ML frameworks that constitute core intellectual property. Production of such\nmaterials, particularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to this Request to the extent it seeks information beyond the use\nof NVIDIA products.\n Tesla further objects to the extent that the scope of this Request is overbroad and bears\nno reasonable relationship to the claims or defenses in the underlying action. Tesla further\nobjects to the extent this Request seeks to impose a duty on Tesla to create materials that Tesla\ndoes not create or maintain in the ordinary course of business. Tesla objects to this Request to\nthe extent it seeks information protected by the attorney-client privilege or the work-product\ndoctrine. Tesla further objects to the extent that this Request is duplicative of prior Requests,\nincluding Requests Nos. 1, 2, 3, and 4.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla\nwill not produce any documents in response to this Request.\n\n\nDOCUMENT REQUEST NO. 6:\n Documents sufficient to show whether computations You performed using NVIDIA\nGPUs involved artificial neural networks, neural-network computational layers or\ncomputations with outputs as inputs for other neurons or layers.\nRESPONSE TO DOCUMENT REQUEST NO. 6:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n\n 20\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 22 of 47\n\n\n\n\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether computations Tesla performed using NVIDIA GPUs involved artificial neural\nnetworks, neural-network computational layers or computations with outputs as inputs for\nother neurons or layers.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. Tesla further objects to this Request\nto the extent it seeks documents containing confidential, proprietary or trade secret information\nwithout making any showing of relevance or that the request is proportional to the needs of\nthe case.\n Tesla further objects to the terms \"artificial neural networks, neural-network\ncomputational layers or computations with outputs as inputs for other neurons or layers\" as\nvague and ambiguous to the extent they could encompass virtually any computation Tesla\nperforms on NVIDIA hardware.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to the extent this\nRequest seeks to impose a duty on Tesla to create materials that Tesla does not create or\nmaintain in the ordinary course of business. Tesla further objects on the grounds that\ninformation regarding NVIDIA GPUs' neural network capabilities is publicly available and\nmore appropriately sought from NVIDIA. Tesla further objects to the extent that this Request\nis duplicative of prior Requests, including Requests Nos. 1, 4, and 5.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\n\n\n\n 21\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 23 of 47\n\n\n\n\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 7:\n Documents sufficient to show whether You use a pointer to data stored in memory (e.g.\nmemory bank or partition), using as an input to a subsequent computational layer the pointer to\noutput data from a GPU computation, using pointers in neural network computations, swapping\nan input pointer with the pointer to data output from a GPU computation, pointer swapping,\npointer rotation, buffer swapping, ping-pong buffers, double or triple buffering, alternating\ninput/output buffers, or any other technique in which output data from one computation, layer,\niteration, time step, or cycle becomes input data for a later computation, layer, iteration, time\nstep, or cycle.\nRESPONSE TO DOCUMENT REQUEST NO. 7:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, whether\nTesla uses a pointer to data stored in memory.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to this\nRequest to the extent that it identifies a sweeping range of fundamental computing techniques,\npointer usage, pointer swapping, pointer rotation, buffer swapping, ping-pong buffers, double\nor triple buffering, alternating input/output buffers, and then adds the catch-all phrase \"or any\n\n\n 22\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 24 of 47\n\n\n\n\nother technique in which output data from one computation, layer, iteration, time step, or cycle\nbecomes input data for a later computation, layer, iteration, time step, or cycle.\"\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary technical information, including internal memory management\nstrategies, GPU optimization techniques, custom buffer management implementations, and\nlow-level engineering designs that constitute core intellectual property. Tesla further objects to\nthis Request to the extent it seeks documents containing confidential, proprietary or trade secret\ninformation without making any showing of relevance or that the request is proportional to the\nneeds of the case.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as memory architecture diagrams, pointer flow analyses, or buffer management\ndocumentation, that Tesla does not create or maintain in the ordinary course of business.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case.\n Tesla objects to this Request to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla objects to this Request to the extent it seeks\ninformation protected by the attorney-client privilege or the work-product doctrine.\n Tesla further objects on the grounds to the extent that the techniques described, pointer\nusage, buffer swapping, double buffering, ping-pong buffers, are well-known, standard\ncomputing techniques whose operation is extensively and publicly documented in computer\nscience literature, NVIDIA's developer guides, and CUDA programming documentation, and\nare more appropriately explored through NVIDIA's own materials.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\n\n 23\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 25 of 47\n\n\n\n\nlitigation, rendering meaningful compliance impossible without speculation. Tesla objects to\nthis Request to the extent it uses generic technical terminology to describe fundamental and\nubiquitous computing operations that are not unique to any particular proprietary technology,\npatented method, or party to this litigation. As drafted, this Request is overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla further objects to the extent that this Request is duplicative of prior Requests,\nincluding Requests Nos. 1, 4, 5, and 6.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 8:\n Documents sufficient to show whether You store input data, output data, intermediate\nresults, tensors, activations, weights, parameters, internal variables, GPU programs, kernels,\ntextures, shaders, or other GPU-computation-related data in separate, partitioned, logical,\nphysical, first/second, input/output, texture, shader, shared, global, device, host, pinned, GPU\nRAM, GPU cache(s), or unified memory regions (shared by CPU and GPU) when performing\ncomputations using NVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 8:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n\n 24\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 26 of 47\n\n\n\n\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether Tesla stores certain GPU-computation-related data in certain memory regions.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case to the extent that\nit would potentially require Tesla to produce an unknowable quantum of engineering\ndocumentation representing Tesla's valuable intellectual property. The Request fails to identify\nwith sufficient particularity what specific implementation or application of these techniques is\nat issue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary intellectual property, including GPU memory allocation\nstrategies, custom memory optimization techniques, tensor management implementations, and\ninternal computing architectures. Tesla further objects to the catch-all phrase \"or other GPU-\ncomputation-related data\" as vague, ambiguous, and unbounded.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any\nparticular proprietary technology, patented method, or party to this litigation. The Request\nfails to identify with sufficient particularity what specific implementation or application of\nthese techniques is at issue in the underlying litigation, rendering meaningful compliance\nimpossible without speculation. As drafted, this Request is therefore facially overbroad,\nunduly burdensome, and disproportionate to the needs of this case, particularly as directed to\na non-party.\n\n\n\n 25\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 27 of 47\n\n\n\n\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as memory allocation maps, data storage analyses, or memory partition\ndocumentation, that Tesla does not create or maintain in the ordinary course of business. Tesla\nobjects to this Request to the extent it seeks information protected by the attorney-client\nprivilege or the work-product doctrine.\n Tesla further objects on the grounds that NVIDIA's own documentation, developer\nguides, CUDA programming manuals, and hardware specifications describe in detail the\nmemory architecture, memory types, and memory management capabilities of NVIDIA GPUs\nand are publicly available and more appropriately sought from the Defendant. Tesla further\nobjects to the extent that this Request is duplicative of prior Requests, including Requests Nos.\n1, 4, 5, and 7.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 9:\n Documents sufficient to show how input data is received, acquired, stored, transferred,\ncopied, streamed, prefetched, staged, queued, or loaded from CPU memory, host memory,\nsystem memory, storage, sensors, cameras, or other input sources to NVIDIA GPU memory\nincluding GPU RAM (e.g. GPU HBM, GDDR) and/or GPU cache(s) before, during, or in\nparallel with computations You perform using NVIDIA GPUs.\n\n\n\n\n 26\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 28 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 9:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nhow Tesla handles input data related to GPU computations.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to the\nRequest to the extent that it seeks documents showing \"how\" input data is \"received, acquired,\nstored, transferred, copied, streamed, prefetched, staged, queued, or loaded\" from an extensive\nand open-ended list of sources, \"CPU memory, host memory, system memory, storage, sensors,\ncameras, or other input sources, \" to NVIDIA GPU memory, \"before, during, or in parallel with\"\ncomputations. This scope encompasses virtually every data pipeline and data ingestion pathway\nacross Tesla's entire computing infrastructure that touches any NVIDIA hardware.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary technical information, including internal data pipeline\narchitectures, sensor fusion systems, data preprocessing workflows, custom data loading and\nprefetching implementations, and GPU optimization strategies that constitute core intellectual\nproperty. Tesla specifically objects to the reference to \"sensors, cameras, or other input sources\"\nas a transparent attempt to compel disclosure of Tesla's most valuable and competitively\nsensitive proprietary intellectual property. Tesla further objects to this Request to the extent it\nseeks documents containing confidential, proprietary or trade secret information without\nmaking any showing of relevance or that the request is proportional to the needs of the case.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to identify\n\n 27\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 29 of 47\n\n\n\n\nwith sufficient particularity what specific implementation or application of these techniques is\nat issue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation. As drafted, this Request is therefore overbroad, unduly burdensome, and\ndisproportionate to the needs of this case, particularly as directed to a non-party.\n Tesla further objects to the catch-all phrase \"or other input sources\" as vague, ambiguous,\nand unbounded. Tesla further objects to the extent this Request seeks to impose a duty on Tesla\nto create materials, such as data flow diagrams, pipeline architecture documents, or data transfer\nanalyses, that Tesla does not create or maintain in the ordinary course of business. Tesla objects\nto this Request to the extent it seeks information protected by the attorney-client privilege or\nthe work-product doctrine.\n Tesla further objects on the grounds that the data transfer and memory management\nmechanisms described\u2014CPU-to-GPU transfers, memory staging, prefetching, streaming\u2014are\nstandard computing operations whose architecture is publicly documented by NVIDIA in its\nCUDA programming guides and developer documentation and are more appropriately sought\nfrom the Defendant. Tesla further objects to the extent that this Request is duplicative of prior\nRequests, including Requests Nos. 4, 5, 7, and 8.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 28\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 30 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 10:\n Documents sufficient to show how output data from a GPU computation(s),\nintermediate results of GPU computations, tensors, buffers, activations, variables, or other\ncomputation results are stored, transferred, copied, streamed, written back, returned,\naccumulated, reused, or made available including asynchronously from NVIDIA GPU\nmemory to CPU memory, host memory, system memory, storage, display, network, or\nanother memory location before, during, or in parallel with computations You perform using\nNVIDIA GPUs and also including in the opposite direction, copying data from CPU or host or\nother memory to a queue for GPU computation while other GPU computations are occurring.\nRESPONSE TO DOCUMENT REQUEST NO. 10:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nhow Tesla handles output data related to GPU computations.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. The Request\nseeks documents showing how output data, intermediate results, tensors, buffers, activations,\nvariables, and \"other computation results\" are \"stored, transferred, copied, streamed, written\nback, returned, accumulated, reused, or made available (including asynchronously)\" of\nunbounded scope, including different memory domains such as \"GPU memory to CPU memory,\nhost memory, system memory, storage, display, network, or another memory location\" and then\nextends the Request to encompass data movement \"in the opposite direction\" as well. Tesla\nobjects to the catch-all phrases \"or other computation results\" and \"or another memory location\"\nas vague, ambiguous, and unbounded, rendering the Request limitless in scope. Tesla objects\nto this Request to the extent that this catch-all request would potentially require Tesla to produce\nan unknowable quantum of engineering documentation representing Tesla's valuable\n\n 29\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 31 of 47\n\n\n\n\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\nproperty, including GPU-CPU data transfer strategies, asynchronous computing pipelines,\ncustom memory management implementations, and real-time inference data flow architectures.\nProduction of such materials in a dispute between Neural AI and NVIDIA would risk\ncompetitive harm to Tesla, a non-party.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any\nparticular proprietary technology, patented method, or party to this litigation. The Request\nfails to identify with sufficient particularity what specific implementation or application of\nthese techniques is at issue in the underlying litigation, rendering meaningful compliance\nimpossible without speculation. As drafted, this Request is therefore facially overbroad,\nunduly burdensome, and disproportionate to the needs of this case, particularly as directed to\na non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as bidirectional data flow diagrams, memory transfer analyses, or\nasynchronous pipeline documentation, that Tesla does not create or maintain in the ordinary\ncourse of business. Tesla objects to this Request to the extent it seeks information protected\nby the attorney-client privilege or the work-product doctrine.\n Tesla further objects on the grounds that the data transfer mechanisms described, GPU-\nto-CPU transfers, asynchronous memory operations, streaming, write-back, are standard\ncomputing operations extensively documented in NVIDIA's public CUDA programming\n\n 30\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 32 of 47\n\n\n\n\nguides and developer resources and are more appropriately sought from the Defendant. Tesla\nfurther objects to the extent that this Request is substantially duplicative of prior Requests,\nincluding Requests Nos. 4, 5, 7, 8, and 9.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 11:\n Documents sufficient to show how computations You perform using NVIDIA GPUs\nare scheduled, ordered, controlled, queued, synchronized, parallelized, launched, interrupted,\nresumed, or executed, including through kernels, CUDA streams, CUDA graphs, events,\nthreads, controllers, schedulers, compilers, runtimes, inference engines, run lists, run engines,\nor custom software.\nRESPONSE TO DOCUMENT REQUEST NO. 11:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201chow computations [Tesla] perform[s] using NVIDIA GPUs are scheduled, ordered,\ncontrolled, queued, synchronized, parallelized, launched, interrupted, resumed, or executed,\nincluding through kernels, CUDA streams, CUDA graphs, events, threads, controllers,\nschedulers, compilers, runtimes, inference engines, run lists, run engines, or custom software.\u201d\n\n\n\n 31\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 33 of 47\n\n\n\n\n Tesla objects to this Request on the grounds that it is facially overbroad, unduly\nburdensome, unnecessary, oppressive, and not proportional to the needs of this case. The\nRequest seeks documents showing how Tesla's GPU computations are \"scheduled, ordered,\ncontrolled, queued, synchronized, parallelized, launched, interrupted, resumed, or executed\"\nthrough an exhaustive and open-ended list of mechanisms, \"kernels, CUDA streams, CUDA\ngraphs, events, threads, controllers, schedulers, compilers, runtimes, inference engines, run lists,\nrun engines, or custom software.\" Tesla objects to the inclusion of \"custom software\" as a catch-\nall that would require disclosure of Tesla's proprietary computing systems in their entirety. Tesla\nobjects to this Request to the extent that this catch-all request would potentially require Tesla\nto produce an unknowable quantum of engineering documentation representing Tesla's valuable\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\nproperty, including custom GPU scheduling systems, proprietary inference engine designs,\ncompute orchestration strategies, compiler optimizations, and proprietary runtime\nenvironments that constitute core competitive advantages. Production of such materials in a\ndispute between Neural AI and NVIDIA would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as scheduling architecture analyses, execution flow documentation, or\norchestration diagrams, that Tesla does not create or maintain in the ordinary course of business.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to\n\n 32\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 34 of 47\n\n\n\n\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Request is therefore facially overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla objects to this Request to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects on the grounds that\ninformation regarding CUDA streams, CUDA graphs, kernel launching, GPU scheduling, and\ncompute execution is extensively documented in NVIDIA's public developer documentation,\nCUDA programming guides, and technical specifications and is more appropriately sought\nfrom the Defendant. Tesla further objects to the extent that this Request is substantially\nduplicative of prior Requests, including Requests Nos. 1, 4, 5, 7, and 8.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 12:\n Documents sufficient to show whether and how user inputs, user commands,\nconfiguration changes, parameter changes, model changes, computational-element changes,\ninput changes, interruptions, or display/output changes affect computations You perform\nusing NVIDIA GPUs and/or queue them for GPU computation.\n\n\n\n\n 33\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 35 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 12:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201cwhether and how user inputs, user commands, configuration changes, parameter changes,\nmodel changes, computational-element changes, input changes, interruptions, or\ndisplay/output changes affect computations [Tesla] perform[s] using NVIDIA GPUs and/or\nqueue them for GPU computation.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. The Request seeks\ndocuments showing \"whether and how\" an open-ended and effectively limitless list of changes,\n\"user inputs, user commands, configuration changes, parameter changes, model changes,\ncomputational-element changes, input changes, interruptions, or display/output changes,\" affect\nTesla's GPU computations or queue them for processing. This scope is virtually unlimited. Tesla\nobjects to this Request to the extent that this catch-all request would potentially require Tesla\nto produce an unknowable quantum of engineering documentation representing Tesla's valuable\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla objects to the terms \"computational-element changes,\" \"input changes,\" and\n\"display/output changes\" as vague, ambiguous, and undefined. It is unclear what constitutes a\n\"computational-element change\" or how broadly \"input changes\" is intended to sweep.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary technical\n\n 34\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 36 of 47\n\n\n\n\ninformation, including user interface architectures, real-time computing systems, dynamic\nGPU scheduling implementations, model update pipelines, interactive inference systems, and\nhuman-machine interface designs. Production of such materials in a dispute between Neural\nAI and NVIDIA would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as user interaction flow analyses, change-impact documentation, or input-to-\ncomputation mapping diagrams, that Tesla does not create or maintain in the ordinary course\nof business. Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Request is therefore facially overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla further objects on the grounds that the computing concepts described in this\nRequest, user input handling, parameter configuration, compute queuing, are standard GPU\ncomputing operations documented in NVIDIA's public developer resources and are more\nappropriately sought from the Defendant. Tesla further objects to the extent that this Request\nis substantially duplicative of prior Requests, including Requests Nos. 1, 4, 5, 7, and 11.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n\n\n\n 35\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 37 of 47\n\n\n\n\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 36\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 38 of 47\n\n\n\n\n OBJECTIONS AND RESPONSES TO DEPOSITION TOPICS\nDEPOSITON TOPIC NO. 1:\n The NVIDIA software and libraries You use to perform computations, including but\nnot limited to NVIDIA's Aerial, Clara Parabricks, cuBLAS, cuDNN, cuFFT, cuQuantum,\ncuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac, Isaac Lab, Maxine, Memory\nMap, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo, PyTorch, RAPIDS, Riva,\nRuntime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream).\nRESPONSE TO DEPOSITION TOPIC NO. 1:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including\nTesla\u2019s use of a long list of software and libraries. Tesla further objects to this Topic as vague\nand ambiguous as to what \"computations\" are relevant to the underlying litigation.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The Topic identifies almost thirty\nNVIDIA software and libraries, placing undue burden on Tesla, a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's highly confidential and proprietary information\nregarding its internal software stack, computing infrastructure, and technology choices, which\nconstitute competitively sensitive business information. Disclosure of which specific NVIDIA\ntools Tesla does or does not use would reveal Tesla's internal technology strategy and\ncompetitive posture.\n\n\n\n 37\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 39 of 47\n\n\n\n\n Tesla further objects to this Topic on the grounds that information regarding NVIDIA's\nsoftware products and their usage by customers is more appropriately sought from NVIDIA,\nthe Defendant in this action, which possesses licensing records, telemetry data, and customer\nusage information.\n Tesla further objects to this Topic on the grounds that the information sought is more\nreadily obtainable from the parties to this litigation or from other less burdensome sources,\nthus disproportionate to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITION TOPCI NO. 2:\n The NVIDIA sample Source Code You use, in whole or in part, to conduct\ncomputations.\nRESPONSE TO DEPOSITION TOPIC NO. 2:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action. Tesla further\nobjects to this Topic as vague and ambiguous as to what \"computations\" are relevant to the\nunderlying litigation.\n\n 38\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 40 of 47\n\n\n\n\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The term \"sample Source Code\" is\nvague and ambiguous, as it is unclear what specific sample code is being referenced, what\nuniverse of NVIDIA sample code is at issue, or how Tesla would be expected to identify\nwhether any portion of its codebase derives from, incorporates, or was inspired by NVIDIA\nsample code. Tesla objects to this Topic to the extent that compliance would require Tesla to\nconduct an exhaustive audit of its entire codebase against an undefined body of NVIDIA\nsample code, a task that is extraordinarily burdensome and disproportionate for a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's proprietary source code and internal software, which\nconstitute trade secrets and core intellectual property. Tesla further objects on the grounds that\nNVIDIA sample source code is publicly available and/or in NVIDIA's possession, custody, or\ncontrol, and information regarding its distribution to customers is more appropriately sought\nfrom the Defendant.\n Tesla further objects to this Topic on the grounds that the information sought is more\nreadily obtainable from the parties to this litigation or from other less burdensome sources,\nthus disproportionate to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n 39\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 41 of 47\n\n\n\n\nDEPOSITION TOPIC NO. 3:\n Your customizations and/or data inputs to NVIDIA software that alter the way in which\nNVIDIA software performs computations and/or a description of the data input to NVIDIA\nsoftware on which computations are run.\nRESPONSE TO DEPOSITON TOPCI NO. 3:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including how\nTesla handles input data related to GPU computations. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to the\nTopic to the extent that it seeks information showing Tesla\u2019s \u201ccustomizations\u201d or \u201calter the way\nin which NVIDIA software performs computations\u201d without providing specificity as to what\nscope of customizations or alterations are relevant to the underlying action. The Topic fails to\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Topic would extend to the entirety of Tesla's computing\noperations and is therefore overbroad, unduly burdensome, and disproportionate to the needs of\nthis case, particularly as directed to a non-party.\n\n\n\n\n 40\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 42 of 47\n\n\n\n\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITON TOPIC NO. 4:\n Identification of Your software that uses NVIDIA GPUs to perform computations.\nRESPONSE TO DEPOSITION TOPIC NO. 4:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic t as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla or third-party software that uses NVIDA GPUs. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case, particularly given that\nTesla is not a party to this litigation. The Topic seeks identification of software used to perform\n\n 41\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 43 of 47\n\n\n\n\ncomputations on NVIDIA GPUs without any specificity, a scope that could potentially\nencompass virtually any software system Tesla operates. The Topic fails to identify with\nsufficient particularity what specific implementation or application of these techniques is at\nissue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation. As drafted, this Topic would extend to the entirety of Tesla's computing operations\nand is therefore overbroad, unduly burdensome, and disproportionate to the needs of this case,\nparticularly as directed to a non-party.\n Tesla objects to this Topic to the extent it seeks Tesla confidential, proprietary or trade\nsecret information without making any showing of relevance or that the request is proportional\nto the needs of the case. Tesla objects to this Topic to the extent it seeks the disclosure of Tesla\nconfidential trade secrets and proprietary technical information, including proprietary source\ncode, internal software architectures, and custom-built AI/ML frameworks that constitute core\nintellectual property and are not relevant to this case. Disclosure of such information,\nparticularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Topic seeks to impose a duty on Tesla to create\nmaterials or compile analyses that Tesla does not create or maintain in the ordinary course of\nbusiness. Tesla objects to this Topic to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to this Request as\nNeural AI has failed to comply with the requirements of Fed. R. Civ. P. 45(d).\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITION TOPIC NO. 5:\n Using Your software, the ways in which output data from a GPU computation(s),\nincluding intermediate results of GPU computations are stored, referenced by a pointer,\ntransferred, copied, streamed, written back, returned, accumulated, reused, or made available\n\n 42\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 44 of 47\n\n\n\n\nincluding asynchronously from NVIDIA GPU memory to CPU memory, host memory, system\nmemory, storage, display, network, or another memory location before, during, or in parallel\nwith computations performed using NVIDIA GPUs\nRESPONSE TO DEPOSITON TOPIC NO. 5:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including how\nTesla handles output data related to GPU computations. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. Tesla objects\nto this Topic to the extent it uses generic technical terminology to describe fundamental and\nubiquitous computing operations that are not unique to any particular proprietary technology,\npatented method, or party to this litigation. The Topic fails to identify with sufficient\nparticularity what specific implementation or application of these techniques is at issue in the\nunderlying litigation, rendering meaningful compliance impossible without speculation. As\ndrafted, this Topic would extend to the entirety of Tesla's computing operations and is\ntherefore overbroad, unduly burdensome, and disproportionate to the needs of this case,\nparticularly as directed to a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\n\n 43\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 45 of 47\n\n\n\n\nproperty, including GPU-CPU data transfer strategies. Disclosure of such confidential\nmaterials in a dispute between Neural AI and NVIDIA would risk competitive harm to Tesla,\na non-party.\n Tesla further objects to the extent this Topic seeks to impose a duty on Tesla to create\nmaterials or compile analyses that Tesla does not create or maintain in the ordinary course of\nbusiness. Tesla objects to this Topic to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to this Request as\nNeural AI has failed to comply with the requirements of Fed. R. Civ. P. 45(d).\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\n\n\n 44\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 46 of 47\n\n\n\nDated: July 21, 2026 /s/ Jun Zheng _____________\n\n Jun Zheng\n TX Bar No. 24102681\n zhengjun@tesla.com\n Tesla, Inc.\n 1 Tesla Rd\n Austin, TX 78725\n (512) 417-3528\n\n Gina H. Cremona\n CA Bar No. 305392\n gcremona@tesla.com\n Tesla, Inc.\n 1501 Page Mill Rd.\n Palo Alto, CA 94304\n (650) 647-0015\n\n Ashraf Fawzy\n DC Bar No. 989132\n afawzy@tesla.com\n Tesla, Inc.\n 800 Connecticut Ave. NW\n Washington, DC 20006\n (202) 905-9221\n\n\n\n\n Attorneys for TESLA, INC.\n\n\n\n\n 45\n\f Case 7:26-mc-00318-LS Document 6-10 Filed 08/18/26 Page 47 of 47\n\n\n\n\n CERTIFICATE OF SERVICE\n\n I hereby certify that a true copy of the above document was served upon Plaintiff Neural\n\nAI\u2019s counsel of record via electronic mail on July 21, 2026.\n\n\n /s/ Jun Zheng\n Jun Zheng\n\n\n\n\n 46\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:48.604942-07:00","document_number":"6","attachment_number":10,"pacer_doc_id":"181037220277","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 9","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554432/","id":490554432,"tags":[],"absolute_url":"/docket/74659430/6/11/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.328217-07:00","date_modified":"2026-08-23T09:39:48.287361-07:00","sha1":"cf3be2d68c30e696a7362d8172fc2308221372e0","page_count":10,"file_size":213852,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.11.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.11.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 1 of 10\n\n\n\n\n EXHIBIT\n\n 10\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 2 of 10\n\n\n Friday, August 7, 2026 at 1:58:08 PM Paci\ufb01c Daylight Time\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\nDate: Tuesday, August 4, 2026 at 11:28:50 AM Paci\ufb01c Daylight Time\nFrom: Tanner Laiche\nTo: Rocco Magni, Gina Cremona\nCC: Jun Zheng, Emily Portuguese, Tamar Lusztig, Brian Melton, Max Tribble, Samuel Drezdzon, Richard\n Wojtczak, Rachel Hanna, Ashraf Fawzy\nAttachments: Neural AI, Third-Party Questions.docx, Neural AI, Draft Third-Party Declaration.docx\n\n\nCounsel,\n\nI am following up on Rocco\u2019s email.\n\nDespite the parties\u2019 prior meet-and-confers, document discovery in the underlying action\ncloses on August 11. Unless the parties can promptly reach a resolution, that deadline\nleaves Neural AI no practical alternative but to move to compel by the end of this week or,\nat the latest, August 10, to preserve its rights.\n\nTo reduce burden and potentially avoid motion practice, I am attaching a short set of\nquestions intended to guide your investigation and help identify the responsive\ninformation, and also recirculating the draft declaration we previously shared, and that\nTesla may revise to ensure its accuracy.\n\nIf Tesla commits to provide an executed declaration, Neural AI is willing to consider\naccepting the declaration in lieu of further document production and/or deposition\ntestimony, subject to resolving any material gaps. Otherwise, the discovery deadline will\nforce Neural AI to move to compel by or before August 10 to preserve its rights. Even if a\nmotion becomes necessary, we remain open to resolving the issues promptly and mooting\nor withdrawing the motion through compliance.\n\nWe are available this week to further meet and confer as necessary.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\nFrom: Rocco Magni \nDate: Sunday, August 2, 2026 at 6:57 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\n\n 1 of 9\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 3 of 10\n\n\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nOur discovery deadline is approaching soon. Please let us know when you can confer again\nthis upcoming week. Thanks.\n\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Rocco Magni \nDate: Tuesday, July 28, 2026 at 6:52 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nThanks for speaking today. Attached is a draft of the declaration I referred to on our call.\n\nBest,\n\nRocco\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\n\n 2 of 9\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 4 of 10\n\n\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Gina Cremona \nDate: Thursday, July 23, 2026 at 9:45 AM\nTo: Rocco Magni \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\n\nHi Rocco, we are not available today.\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 22, 2026 12:01 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche\n; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nWould tomorrow work?\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received this message in\n\n 3 of 9\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 5 of 10\n\n\nerror, please notify the sender and delete it immediately.\n\n\n\n On Jul 22, 2026, at 2:54 PM, Gina Cremona wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n We are not available on Friday, but can meet on Tuesday, July 28 between 8-10 am PT.\n\n Thanks,\n Gina\n\n\n From: Rocco Magni \n Sent: Wednesday, July 22, 2026 6:54 AM\n To: Jun Zheng ; Gina Cremona\n \n Cc: Tanner Laiche ; Emily Portuguese\n ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n ; Ashraf Fawzy ; Jun\n Zheng \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n Jun,\n\n Please provide times to meet and confer on Friday 7/24. Thanks.\n\n --\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you received this\n message in error, please notify the sender and delete it immediately.\n\n 4 of 9\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 6 of 10\n\n\nFrom: Jun Zheng \nDate: Tuesday, July 21, 2026 at 7:37 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nEXTERNAL Email\nRocco,\n\nWe understand from the email exchange below that the noticed date for\ntestimony subpoena is now Sept. 14, 2026. In the meantime, attached please \ufb01nd\nTesla's objections and responses to Neural AI's subpoena.\n\nThanks!\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n\nFrom: Gina Cremona \nSent: Monday, July 20, 2026 1:07 PM\nTo: Rocco Magni \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\n\n 5 of 9\n\f Case 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 7 of 10\n\n\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nHi Rocco,\n\nTesla cannot provide an agreed date for deposition until we have reviewed and\nresponded to the subpoenas. However, we can agree to Sept. 14 as the noticed date\nfor testimony subpoena for now, subject to modi\ufb01cation once we have responded to\nthe document subpoena.\n\nRegards,\nGina\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 15, 2026 12:58 PM\nTo: Gina Cremona \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nGina,\n\nWe\u2019ll agree to pull down the July 28 date once we have an agreed replacement date.\nLet us know what date works for you and we\u2019ll withdraw the current notice.\nThanks.\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received\nthis message in error, please notify the sender and delete it immediately.\n\n\n\n\n 6 of 9\n\fCase 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 8 of 10\n\n\n On Jul 15, 2026, at 3:54 PM, Gina Cremona \n wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n Understood. To con\ufb01rm, the deposition date of July 28 is ok calendar, and\n we will work on agreeing to a new date.\n\n Regards,\n Gina\n\n\n From: Rocco Magni \n Sent: Tuesday, July 14, 2026 5:28 PM\n To: Gina Cremona ; Tanner Laiche\n ; Emily Portuguese\n \n Cc: Tamar Lusztig ; Brian\n Melton ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Ms. Cremona,\n\n Discovery has not been extended; only depositions. Written\n discovery will still close on August 11.\n\n We can work with you on a deposition date between now and\n September 14. But we do not have \ufb02exibility on the timing of your\n RFP responses and document production beyond the 1 week\n extension we noted below.\n\n Best,\n\n Rocco\n\n --\n\n\n 7 of 9\n\fCase 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 9 of 10\n\n\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you\n received this message in error, please notify the sender and delete it\n immediately.\n From: Gina Cremona \n Date: Tuesday, July 14, 2026 at 8:19 PM\n To: Tanner Laiche ; Emily\n Portuguese \n Cc: Rocco Magni ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D.\n Tex.) - Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n It has come to our attention that the discovery deadline in this case has\n been extended. Due to summer vacations and people being out of the\n okice, we renew our request for an additional 2-week extension such that\n our deadlines will be Aug. 4 and Aug. 11.\n\n Regards,\n Gina\n\n\n From: Tanner Laiche \n Sent: Monday, July 6, 2026 5:20 PM\n To: Gina Cremona ; Emily Portuguese\n \n Cc: Rocco Magni ; Tamar\n Lusztig ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n\n 8 of 9\n\fCase 7:26-mc-00318-LS Document 6-11 Filed 08/18/26 Page 10 of 10\n\n\n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Counsel,\n\n Thanks for reaching out. Given the upcoming close of fact\n discovery, Neural AI is not able to agree to a three-week\n extension. That said, we can agree to a one-week extension for\n Tesla\u2019s written objections/responses to the subpoena(s). A copy\n of the Protective Order is attached.\n\n Regards,\n\n Tanner Laiche\n Susman Godfrey LLP\n 206.505.3816 | tlaiche@susmangodfrey.com\n 401 Union Street | Suite 3000 | Seattle, WA 98101\n HOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\n From: Gina Cremona \n Date: Thursday, July 2, 2026 at 7:21 PM\n To: Emily Portuguese \n Subject: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n We are in receipt of your Subpoena to Produce Documents and Subpoena\n for Testimony. Due to the holiday weekend and vacation schedules, we\n request a three-week extension to respond such that our deadlines will be\n Aug. 4 and Aug. 11.\n\n Additionally, please provide us with a copy of the protective order.\n\n Regards,\n Gina\n\n Gina H. Cremona\n Senior Counsel, IP Litigation\n 1501 Page Mill Rd., Palo Alto, CA 94304\n E. gcremona@tesla.com T. 650.647.0015\n\n\n \n\n\n\n\n 9 of 9\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:46.702098-07:00","document_number":"6","attachment_number":11,"pacer_doc_id":"181037220278","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 10","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554433/","id":490554433,"tags":[],"absolute_url":"/docket/74659430/6/12/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.342893-07:00","date_modified":"2026-08-23T04:28:50.192150-07:00","sha1":"e7710399b2030839f0e90c53e964106bc73f08e7","page_count":3,"file_size":89106,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.12.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.12.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-12 Filed 08/18/26 Page 1 of 3\n\n\n\n\n EXHIBIT\n\n 11\n\f Case 7:26-mc-00318-LS Document 6-12 Filed 08/18/26 Page 2 of 3\n\n\n\n\nGeneral Deployment\n \u2022 Do you have architecture diagrams for your hardware and software systems that use\n NVIDIA GPUs?\n \u2022 Do you use PyTorch or TensorRT?\n \u2022 Which of the following, if any, top-layer applications do you use: Modulus, Maxine,\n cuQuantum, Merlin, Aerial, Monai, Triton, Nemo, VSS Blueprint, Riva, Metropolis,\n Holoscan, Clara Parabricks, Rapids, Isaac, Isaac Lab, Drive, DriveWorks, and\n Morpheus?\n \u2022 For each NVIDIA software product identified, in the ordinary course, do you download\n and use it locally, use it via a cloud-based solution offered by NVIDIA, or through a\n third-party cloud provider?\n \u2022 Do you make modifications to the source code for PyTorch, TensorRT, or other\n NVIDIA-provided software when its deployed on NVIDIA GPUs?\n \u2022 In the ordinary course, approximately how frequently do you run computations utilizing\n PyTorch or TensorRT on NVIDIA GPUs (e.g., many times per day, every day, every\n week, or every month)?\n \u2022 Are the relevant systems operated in the United States or used to support U.S.-directed\n operations?\n\nInput Data Path\n \u2022 In the ordinary course, do you use GPUDirect Storage (GDS) to load input data directly\n from storage into GPU memory, bypassing CPU main memory? Or do you use CPU\n main memory?\n \u2022 In the ordinary course, do you use GPUDirect RDMA or any direct NIC-to-GPU memory\n path for receiving live input data?\n \u2022 In the ordinary course, do you use unified memory (cudaMallocManaged), Unified\n Virtual Memory (UVM), or a coherent CPU/GPU memory architecture (e.g., DGX Spark\n UMA, Grace Hopper coherent memory) for the input data path?\n \u2022 In the ordinary course, do you use the PyTorch function torch.cuda.gds.GdsFile or any\n GDS-enabled data loader (e.g., DALI GDS, KvikIO) to load data?\n\nOutput Data Path and Memory Transfers\n \u2022 In the ordinary course, are outputs of GPU computations copied back to CPU/main\n memory? If so, what data is copied (e.g., transcripts, generated tokens, logits,\n embeddings)?\n \u2022 In the ordinary course, do GPU-to-CPU (D2H) or CPU-to-GPU (H2D) data transfers\n occur during or in parallel with GPU computations, or only after each computation\n completes?\n \u2022 In the ordinary course, how are your memory partitions configured for GPU\n computations? Are there separate memory regions for input data, output data, workspace,\n and intermediate results?\n \u2022 In the ordinary course, are output buffers from one GPU computation reused as input\n buffers for a subsequent computation?\n \u2022 Do you use torch.compile, TorchDynamo, Triton, TensorRT-LLM compilation, NVRTC,\n or PTX JIT to generate GPU programs at runtime, or do you use only precompiled\n kernels?\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 6-12 Filed 08/18/26 Page 3 of 3\n\n\n\n\nTensorRT Usage\n \u2022 If you use TensorRT or TensorRT-LLM, how do you populate the input buffers and\n retrieve the outputs? Do you use CPU-side request processing and H2D/D2H copies?\n \u2022 Are TensorRT engines built by your organization, provided pre-built by NVIDIA, or\n obtained from another source?\n\n\n\n\n 2\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:53.706732-07:00","document_number":"6","attachment_number":12,"pacer_doc_id":"181037220279","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 11","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554434/","id":490554434,"tags":[],"absolute_url":"/docket/74659430/6/13/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.360679-07:00","date_modified":"2026-08-23T09:39:57.331587-07:00","sha1":"48046c7eead19b32fb28fbcc25e75da176a1d19d","page_count":4,"file_size":84539,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.13.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.13.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-13 Filed 08/18/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 12\n\f Case 7:26-mc-00318-LS Document 6-13 Filed 08/18/26 Page 2 of 4\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n NEURAL AI, LLC\n\n Plaintiff, Case No. 7:24-cv-00221-ADA-DTG\n v. JURY TRIAL DEMANDED\n NVIDIA CORPORATION\n\n Defendant.\n\n\n\n\n I, [DECLARANT NAME], hereby declare as follows:\n\n 1. I am [TITLE] of [COMPANY NAME]. I am over the age of eighteen and\n\ncompetent to make this declaration. I make this declaration based on my personal knowledge,\n\nincluding my general familiarity with [COMPANY NAME]\u2019s technical infrastructure, software\n\ndeployments, use of NVIDIA products, as well as the on the basis of a reasonably diligent\n\ninvestigation.\n\n 2. I understand that this declaration is submitted in connection with a subpoena served\n\non [INSERT DAT] on [COMPANY NAME] in the above-captioned action (the \u201cSubpoena\u201d).\n\n 3. In the ordinary course of business, [COMPANY NAME] uses NVIDIA GPUs (as\n\nthat term is defined in the Subpoena attached as Exhibit A) in its commercial operations.\n\nSpecifically, [COMPANY NAME] deploys GPUs of the [INSERT ARCHITECTURE]\n\narchitecture(s), including but not limited to [INSERT SPECIFIC GPU MODEL(S)] (collectively,\n\nthe \u201cDeployed NVIDIA GPUs\u201d).\n\n 4. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select among cuFFT,\n\f Case 7:26-mc-00318-LS Document 6-13 Filed 08/18/26 Page 3 of 4\n\n\n\n\ncuSparse, cuSolver, cuBlas, and cuDNN].\n\n 5. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select PyTorch and/or\n\nTensorRT].\n\n 6. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select among Modulus,\n\nMaxine, cuQuantum, Merlin, Ariel, Monai, Triton, Nemo, Riva, Metropolis, Holoscan, Clara\n\nParabricks, Rapids, Issac, Drive, and Morpheus].\n\n 7. In the ordinary course of business, [COMPANY NAME] uses the NVIDIA\n\nsoftware identified in Paragraphs 4-6 as provided by NVIDIA, without modification to the source\n\ncode.\n\n 8. To the best of [COMPANY NAME]\u2019s knowledge, when it uses the NVIDIA\n\nSoftware on the Deployed NVIDIA GPUs in the ordinary course of business, the hardware and\n\nNVIDIA software function together as designed and intended by NVIDIA.\n\n 9. In the ordinary course, [COMPANY NAME] uses one or more pretrained neural-\n\nnetwork models, model implementations, or model configurations distributed, made available, or\n\nrecommended by NVIDIA. In the ordinary course, [COMPANY NAME] deploys those models,\n\nimplementations, or configurations using PyTorch, TensorRT, or other NVIDIA software without\n\nmodifying their underlying network structure or low-level implementation code.\n\n 10. In the ordinary course of business, the Deployed NVIDIA GPUs are installed in\n\nsystems with separate CPU main memory and GPU memory, and the CPU and GPU in these\n\nsystems are connected via a bus.\n\n 11. In the ordinary course of business and for the operations described in this\n\f Case 7:26-mc-00318-LS Document 6-13 Filed 08/18/26 Page 4 of 4\n\n\n\n\ndeclaration, [COMPANY NAME] does not use Deployed NVIDIA GPUs with a unified\n\nCPU/GPU memory pool, GPUDirect Storage (\u201cGDS\u201d) to transfer input data directly from storage\n\nto GPU memory, NVIDIA Unified Virtual Memory (\u201cUVM\u201d), or CUDA managed memory to\n\nbypass the standard CPU-memory-to-GPU-memory data transfer path.\n\n 12. In [COMPANY NAME]\u2019s normal operations, when it uses the NVIDIA software\n\ndescribed in Paragraphs 4-6 on the Deployed NVIDIA GPUs, input data is received and processed\n\nby the CPU and stored in CPU main memory before being transferred to the GPU for computation.\n\nTo the best of my knowledge, this is the default and standard method by which data is loaded for\n\nprocessing on NVIDIA GPUs.\n\n 13. In the ordinary course of business and to the best of [COMPANY NAME]\u2019s\n\nknowledge, [COMPANY NAME] does not implement custom code or configurations that alter\n\nthe default data flow path provided by the NVIDIA Software with respect to how input data is\n\nreceived by the CPU, stored in main memory, and transferred to the GPU for computation.\n\n 14. To the best of [COMPANY NAME]\u2019s knowledge, the operations described in this\n\ndeclaration are performed in the United States using systems located in the United States or\n\nsystems that directly support [COMPANY NAME]\u2019s United States operations.\n\n 15. At least once per day, [COMPANY NAME] uses NVIDIA GPUs in conjunction\n\nwith at least one NVIDIA offering identified in each of Paragraphs 4-6.\n\n I declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct. Executed on [DATE], in [CITY, STATE].\n\n_______________________________________\n[DECLARANT NAME]\n[TITLE]\n[COMPANY NAME]\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:55.023045-07:00","document_number":"6","attachment_number":13,"pacer_doc_id":"181037220280","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 12","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554435/","id":490554435,"tags":[],"absolute_url":"/docket/74659430/6/14/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.375420-07:00","date_modified":"2026-08-23T09:32:09.076187-07:00","sha1":"45be58c439c8ca229c0b5f40e13766e5ad3f7b5a","page_count":4,"file_size":1280995,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.14.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.14.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-14 Filed 08/18/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 13\n\f Case 7:26-mc-00318-LS Document 6-14 Filed 08/18/26 Page 2 of 4\n Company Blog Artificial Intelligence AI Infrastructure Physical AI Gaming & Creating Industries Subscribe US \uf2bd Sign In\n\n\n\n\nTesla Unveils Top AV Training Supercomputer Powered by NVIDIA A100\nGPUs\n\u2018Incredible\u2019 GPU cluster powers AI development for Autopilot and full self-driving.\nJune 22, 2021 by Danny Shapiro\n\n\n 2 mins 0 Share\n\n\n\n\nTackling one of the largest computing challenges of this lifetime requires larger than life computing.\n\n\nAt CVPR this week, Andrej Karpathy, senior director of AI at Tesla, unveiled the in-house supercomputer the automaker is using to train\ndeep neural networks for Autopilot and self-driving capabilities. The cluster uses 720 nodes of 8x NVIDIA A100 Tensor Core GPUs (5,760\nGPUs total) to achieve an industry-leading 1.8 exaflops of performance.\n\n\n\u201cThis is a really incredible supercomputer,\u201d Karpathy said. \u201cI actually believe that in terms of flops, this is roughly the No. 5 supercomputer\nin the world.\u201d\n\nWith unprecedented levels of compute for the automotive industry at the center of its development cycle, Tesla is making it possible for NVIDIA GTC Berlin\nautonomous vehicle engineers to do their life\u2019s work efficiently and at the cutting edge. Registration Is Now Open\n\nNVIDIA A100 GPUs deliver acceleration at every scale to power the world\u2019s highest-performing data centers. Powered by the NVIDIA October 20-22\n\nAmpere Architecture, the A100 GPU provides up to 20x higher performance over the prior generation and can be partitioned into seven\n Register Now\nGPU instances to dynamically adjust to shifting demands.\n\n\n\n\n Recent News\n Gaming\n\n\n GeForce NOW Shakes Up August With 26\n New Games\n August 6, 2026\n\n\n\n AI\n\n\n Into the Omniverse: How Open World\n Models Push the Frontier of Physical AI\n August 6, 2026\n\n\n\n AI\n\n\n NVIDIA and Partners Build in America, for\n America\n August 5, 2026\n\n\n\n AI Infrastructure\n\n\nThe GPU cluster is part of Tesla\u2019s vertically integrated autonomous driving approach, which uses more than 1 million cars already driving NVIDIA Joins NSF State and Regional AI\non the road to refine and build new features for continuous improvement. Hubs Program to Expand AI Research and\n Education Across the US\n\f Case 7:26-mc-00318-LS Document 6-14 Filed 08/18/26 Page 3 of 4\n August 4, 2026\nFrom the Car to the\n Company Blog Data Center Subscribe US\n\n\nTesla\u2019s cyclical development begins in the car. A deep neural network running in \u201cshadow mode\u201d quietly perceives and makes predictions\n View All Recent News\nwhile the car is driving without actually controlling the vehicle.\n\nThese predictions are recorded, and any mistakes or misidentifications are logged. Tesla engineers then use these instances to create a\ntraining dataset of difficult and diverse scenarios to refine the DNN.\n\n\nThe result is a collection of roughly 1 million 10-second clips recorded at 36 frames per second, totaling a whopping 1.5 petabytes of\ndata. The DNN is then run through these scenarios in the data center over and over until it operates without a mistake. Finally, it\u2019s sent\nback to the vehicle and begins the process again.\n\n\nKarpathy said training a DNN in this manner and on such a large amount of data requires \u201ca huge amount of compute,\u201d which led Tesla to\nbuild and deploy the current generation supercomputer with high-performance A100 GPUs.\n\n\n\n\nContinuous Iteration\nIn addition to comprehensive training, Tesla\u2019s supercomputer gives autonomous vehicle engineers the performance needed to\nexperiment and iterate in the development process.\n\n\nKarpathy said the current DNN structure the automaker is deploying allows a team of 20 engineers to work on a single network at once,\nisolating different features for parallel development.\n\nThese DNNs can then be run through training datasets at speeds faster than what has been previously possible for rapid iteration.\n\n\n\u201cComputer vision is the bread and butter of what we do and enables Autopilot. For that to work, you need to train a massive neural\nnetwork and experiment a lot,\u201d Karpathy said. \u201cThat\u2019s why we\u2019ve invested a lot into the compute.\u201d\n\nWatch the full CVPR session.\n\n\nCategories: Driving\n\n\n\nTags: NVIDIA DGX Transportation\n\n\n\n\n Comments for this thread are now closed \u00d7\n\n\n0 Comments \ue603\n 1 Login\n\n\n\n \uf109 Share Best Newest Oldest\n\n\n\n\n This discussion has been closed.\n\n\n\n\n Subscribe Privacy Do Not Sell My Data\n\n\n\n\nRelated News\n\n\n\n\n AI AI AI AI\n\f Case 7:26-mc-00318-LS Document 6-14 Filed 08/18/26 Page 4 of 4\nInto the Omniverse: How Open\n Company Blog\n NVIDIA and Partners Build in AI Leaders Propose SAFE Industry Leaders Unite\n Subscribe US\n in Open\nWorld Models Push the America, for America Guidelines for Cybersecurity Secure AI Alliance for AI Safety\nFrontier of Physical AI Aug 5, 2026 Transparency and Security\nAug 6, 2026 Aug 4, 2026 Jul 27, 2026\n\n\n\n\n Follow Us \uf39e \uf16d \uf0e1 \ue61b \uf167\n\n\n United States\n\nPrivacy Policy Your Privacy Choices Terms of Service Accessibility Corporate Policies Product Security Contact\nCopyright \u00a9 2026 NVIDIA Corporation\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:51.677370-07:00","document_number":"6","attachment_number":14,"pacer_doc_id":"181037220281","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 13","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554436/","id":490554436,"tags":[],"absolute_url":"/docket/74659430/6/15/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.394170-07:00","date_modified":"2026-08-23T04:40:02.384369-07:00","sha1":"881cceb757b39e60c20045745fe2e473d32be3f8","page_count":4,"file_size":3964473,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.15.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.15.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-15 Filed 08/18/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 14\n\f Case 7:26-mc-00318-LS Document 6-15 Filed 08/18/26 Page 2 of 4\n US Edition RSS Sign in Search\n\n\n Best Picks CPUs GPUs PC Components News Laptops Desktops Software & AI Coupons Premium Forums\n\n\n\n\nTRENDING Tom's 30th Anniversary AMD Advancing AI RAM Shortage Vera Rubin AI Data Centers RAM Combo Deals TH Premium AMD Instinct MI455X DLSS\n\n\n\n PC Components > GPUs ADVERTISEMENT\n\n\n\n Tesla Brags About In-House\n Supercomputer, Now With 7,360 A100\n GPUs\n News By Mark Tyson Published August 17, 2022\n\n\n With a 28% increase in GPUs, it's now a top-7 supercomputer\n worldwide by GPU count\n\n\n\n\n (Image credit: Tesla)\n\n\n\n\n 9 Follow us Newsletter\n\n\n\n Tesla has boosted its in-house AI supercomputer with thousands of\n additional Nvidia A100 GPUs. The Tesla supercomputer had 5,760 A100\n GPUs about a year ago, and that count has since risen to 7,360 A100 GPUs\n \u2014 that's an additional 1,600 GPUs, or about a 28% increase.\n\n According to Tesla Engineering Manager Tim Zaman, this upgrade makes\n the firm's AI system a top-7 supercomputer worldwide by GPU count.\n\n An Nvidia A100 GPU is a powerful Ampere architecture solution aimed at\n data centers. Yes, it uses the same GPU architecture as GeForce RTX 30\n series GPUs, which are some of the best graphics cards currently\n ADVERTISEMENT\n available. However, there is no close consumer relation to the A100, which\n comes with 80GB of HBM2e memory on board, offers up to 2 TB/s\n bandwidth, and requires up to 400W of power. The architecture of the\n A100 has also been tweaked for accelerating tasks common in AI, data\n analytics, and high-performance computing (HPC) applications.\n\n\n\n Latest Videos From Tom's Hardware\n\n\n\n\n Ad 1 of 2.\n Watch full video here: How to get rid of Google's AI overviews\n\n\n\n The first system Nvidia showed wielding the A100 was the Nvidia DGX\n A100, which packed in eight A100 GPUs linked via six NVSwitch with 4.8\n TBps of bi-directional bandwidth for upJoin\n to 10 PetaOPS\n Tom\u2019s of INT8 today\n Hardware ADVERTISEMENT EXPLORE\n\f Case 7:26-mc-00318-LS Document 6-15 Filed 08/18/26 Page 3 of 4\nperformance, 5 PFLOPS of FP16, 2.5 TFLOPS of TF32, and 156 TFLOPS of\nFP64 in a single node.\n\nThat was eight A100 GPUs \u2014 Tesla's AI supercomputer now has 7,360 of\nthese. Tesla hasn't publicly benchmarked its AI supercomputer, but the\nsimilarly-equipped GPU-based NERSC Perlmutter, which has 6,144 Nvidia\nA100 GPUs, achieves 70.87 Linpack petaflops. Using this and data from\nother A100 GPU supercomputers as performance reference points, HPC\nWire estimates the Tesla AI supercomputer is capable of achieving about\n100 Linpack petaflops.\n\n YOU MAY LIKE\n\n\nChina bypasses US GPU bans with 1.54-exaflops 'LineShine'\nsupercomputer\n\n\n\nGoogle could build more AI accelerators than Nvidia sells in\n2028, analyst claims\n\n\n\nNvidia's memory costs soar 485%, latest AI systems now cost\n$7.8 million to build\n\n\nTesla doesn\u2019t intend to continue down the Nvidia GPU architecture path\nfor its in-house AI supercomputers long-term. This world\u2019s top-7 machine\nby GPU-count is merely a precursor to the upcoming Dojo supercomputer,\nwhich was first announced by Elon Musk back in 2020. A year ago we got a\nlook at the Tesla D1 Dojo chip, which are designed to supplant Nvidia's\nGPUs for \u201cmaximum performance, throughput and bandwidth at every\ngranularity.\u201d\n ADVERTISEMENT\n\n\n\n\n \uf125\n(Image credit: Tesla)\n\n\nThe Tesla Dojo D1 is a custom ASIC (application-specific integrated circuit)\ndesign, purposed for AI training, and it is one of the first ASICs in this field.\nCurrent D1 test chips are manufactured on TSMC N7 and pack in about 50\nmillion transistors.\n\n\n\n Stay On the Cutting Edge: Get the Tom's\n Hardware Newsletter\n Get Tom's Hardware's best news and in-depth reviews, straight to\n your inbox.\n\n\n Your Email Address SIGN ME UP\n\n\n\n By signing up, you agree to our Terms of services and acknowledge that you\n have read our Privacy Notice. You also agree to receive marketing emails from\n us that may include promotions from our trusted partners and sponsors, which\n you can unsubscribe from at any time.\n\n\n\n\nMore information about the Dojo D1 chip, and the Dojo system, might be\nrevealed at next week's Hot Chips Symposium \u2014 three Tesla\npresentations are schedule for next Tuesday, addressing Dojo D1 chip\narchitecture, Dojo and ML training, and enabling AI through system\nintegration.\n\n\n ADVERTISEMENT\nTOPICS\n\n Nvidia Science GeForce\n\n\n\n Ad 1 of 2.\n \uf123 SEE ALL COMMENTS (9)\n\n\n\n\n Mark Tyson News Editor\n Join Tom\u2019s Hardware today EXPLORE\n\f Case 7:26-mc-00318-LS Document 6-15 Filed 08/18/26 Page 4 of 4\nREPLY \uf105\n\n\n\ncirdecus\n\n Mandark said:\n\n Nothing like vertical integration. ALWAYS make your OWN stuff to\n control your destiny. Don\u2019t believe the fools that say it\u2019s cheaper to\n outsource it\u2019s not. Outsource companies need to make a profit too.\n And when you outsource you lose the ability to innovate\n\n\n I love to see Tesla vertically integrate. It\u2019s one of the main reasons for\n their success\n\n\n\n\nCompletely agree. It does, however, tend to condense power which\ncould mean less consumer choice, but vertical integration is where it's\nat. I never thought I'd see Amazon buying their own fleet of transport\nvehicles, planes and ships lol.\n\nREPLY \uf105\n\n\n\njtenorj\nNoticed a mistake in your article. You state that Tesla's D1 chip has 50\nmillion transistors when Telsa's slide for the chip clearly shows 50 billion.\nThat's a difference of 3 orders of magnitude. 50 billion is also much\nmore in line with the chip's size in mm squared as well as its 400w power\ndraw on a modern fairly compact process node.\nREPLY \uf105\n\n\n\n\n SHOW MORE COMMENTS \uf105\n\n\n\n\n Tom's Hardware is part of Future US Inc, an international media group and leading digital publisher. Visit our corporate site.\n\n\n\n\n Terms and conditions Contact Future's experts Privacy policy Cookies policy\n\n\n Accessibility Statement Advertise with us About us Coupons\n\n\n Careers Do not sell or share my personal\n information\n\n\n\n \u00a9 Future US, Inc. Full 7th Floor, 130 West 42nd Street, New York, NY 10036.\n\n\n\n\n Ad 1 of 2.\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:56.037188-07:00","document_number":"6","attachment_number":15,"pacer_doc_id":"181037220282","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 14","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554437/","id":490554437,"tags":[],"absolute_url":"/docket/74659430/6/16/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.407968-07:00","date_modified":"2026-08-23T09:40:00.638770-07:00","sha1":"442148fe8f0463d630d7e2cff0747cf3eaabc348","page_count":8,"file_size":9692108,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.16.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.16.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 1 of 8\n\n\n\n\n EXHIBIT\n\n 15\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 2 of 8\n 00:10:03:51 Get $400 off your Disrupt 2026 ticket: REGISTER NOW.\n\n\n Latest Startups Venture Apple Security AI Apps Disrupt 2026 Events Podcasts Newsletters\n\n\n\n\n TRANSPORTATION\n\n\n\n\n Tesla Dojo: The rise and fall of\n Elon Musk\u2019s AI supercomputer\n\n\n\n Rebecca Bellan 9:18 AM PDT \u00b7 September 2, 2025\n\n\n\n\n IMAGE CREDITS: BRYCE DURBIN/TECHCRUNCH\n\n\n\n\n For years, Elon Musk has spoken of the promise of Dojo, the AI supercomputer\n that was supposed to be the cornerstone of Tesla\u2019s AI ambitions. It was\n important enough to Musk that in July 2024, he said the company\u2019s AI team\n would \u201cdouble down\u201d on Dojo in the lead-up to Tesla\u2019s robotaxi reveal, which\n happened in October.\n\n After six years of hype, Tesla decided last month to shut down Dojo and\n disband the team behind the supercomputer in August 2025. Within weeks of\n projecting that Dojo 2, Tesla\u2019s second supercluster that was meant to be built\n on the company\u2019s in-house D2 chips, would reach scale by 2026, Musk\n reversed course, declaring it \u201can evolutionary dead end.\u201d\n\n Ad\n Manage your stocks and ET\n\n Fidelity Investments\n\n\n Learn more\n\n\n This article originally set out to explain what Dojo was and how it could help\n Tesla achieve full-self driving, autonomous humanoid robots, semiconductor\n autonomy, and more. Now, you can think of it more as an obituary of a project\n that convinced so many analysts and investors that Tesla wasn\u2019t just an\n automaker, it was an AI company.\n\n Dojo was Tesla\u2019s custom-built supercomputer that was designed to train its\n \u201cFull Self-Driving\u201d neural networks.\n\n Beefing up Dojo went hand-in-hand with Tesla\u2019s goal to reach full self-driving\n October 13 \u2013 15 San Francisco\n and bring a robotaxi to market. FSD (Supervised) is Tesla\u2019s advanced driver\n assistance system that\u2019s on hundreds of thousands of Tesla vehicles today Scale faster. Grow your portfolio. Gain\n practical expertise. No matter your goal,\n and can perform some automated driving tasks, but it still requires a human Disrupt can empower you.\n to be attentive behind the wheel. It\u2019s also the basis of similar technology\n Save up to $400 today!\n powering Tesla\u2019s limited robotaxi service that the company launched in Austin\n this June using Model Y SUVs. REGISTER NOW\n Ad : (0:03)\n Even as Dojo\u2019s raison d\u2019\u00eatre started to come to life, Tesla failed to attribute its\n self-driving successes \u2014 controversial as they were \u2014 to the supercomputer.\n In fact, Musk and Tesla had barely mentioned Dojo at all over the past year. In\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 3 of 8\n August 2024, Tesla began promoting Cortex, the company\u2019s \u201cgiant new AI\n training supercluster being built at Tesla HQ in Austin to solve real-world AI,\u201d\n which Musk has said would have \u201cmassive storage for video training of FSD\n Most\n and Optimus.\u201d Popular\n In Tesla\u2019s Q4 2024 shareholder deck, the company shared updates on Cortex, ChatGPT brings unlimited text\n chats to free users\n but nothing on Dojo. It\u2019s not clear whether Tesla\u2019s Dojo shutdown affects\n Cortex.\n Ford\u2019s new electric truck,\n \u2018Fathom,\u2019 starts at $28,350\n\n\n Bending Spoons to buy Airtable\n for $1.28B\n\n\n Influencers draw backlash for\n attending OpenAI\u2019s first luxury\n trip\n\n\n Sequoia\u2019s Shaun Maguire leads\n $1B round for nuclear startup\n Valar Atomics\n\n\n Malaysia is reportedly shutting\n down Balaji Srinivasan\u2019s\n Network School\n\n\n YouTuber Hank Green says his\n The response to Dojo\u2019s disbanding has been mixed. Some see it as another AI usage is \u2018not healthy\u2019\n example of Musk making promises he can\u2019t deliver on that comes at a time of\n falling EV sales and a lackluster robotaxi rollout. Others say the shutdown\n wasn\u2019t a failure, but a strategic pivot from a high-risk, self-reliant hardware to\n a streamlined path that relies on partners for chip development. Ad\n\n\n Dojo\u2019s story reveals what was on the line, where the project fell short, and\n what its shutdown signals for Tesla\u2019s future.\n\n\n\n A recap of Dojo\u2019s shutdown Manage your stocks and ET\n Fidelity Basket Portfolios is a faster and easier way to build a\n\n Tesla disbanded its Dojo team and shut down the project in mid-August 2025. basket of securities and\n\n\n Dojo\u2019s lead, Peter Bannon, left the company as well, following the departure of Fidelity Investments Learn more\n around 20 workers who left to start their own AI chip and infrastructure\n company called DensityAI.\n\n Analysts have pointed out that losing key talent can quickly derail a project,\n especially a highly specialized, internal tech project.\n Advertisement\n The shutdown came a couple of weeks after Tesla signed a $16.5 billion deal\n to get its next-generation AI6 chips from Samsung. The AI6 chip is Tesla\u2019s bet\n on a chip design that can scale from powering FSD and Tesla\u2019s Optimus\n humanoid robots to high-performance AI training in data centers.\n\n \u201cOnce it became clear that all paths converged to AI6, I had to shut down Dojo\n and make some tough personnel choices, as Dojo 2 was now an evolutionary\n dead end,\u201d Musk posted on X, the social media platform he owns. \u201cDojo 3\n arguably lives on in the form of a large number of AI6 [systems-on-a-chip] on\n a single board.\u201d\n\n\n\n Tesla\u2019s Dojo backstory\n\n\n\n\n IMAGE CREDITS:SUZANNE CORDEIRO / AFP / GETTY IMAGES\n\n\n\n Musk has insisted that Tesla isn\u2019t just an automaker, or even a purveyor of\n solar panels and energy storage systems. Instead, he has pitched Tesla as an\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 4 of 8\n AI company, one that has cracked the code to self-driving cars by mimicking\n human perception.\n\n Most other companies building autonomous vehicle technology rely on a\n combination of sensors to perceive the world \u2014 like lidar, radar and cameras\n \u2014 as well as high-definition maps to localize the vehicle. Tesla believes it can\n achieve fully autonomous driving by relying on cameras alone to capture\n visual data and then use advanced neural networks to process that data and\n make quick decisions about how the car should behave.\n\n The pitch has been that Dojo-trained AI software will eventually be pushed out\n to Tesla customers via over-the-air updates. The scale of FSD also means\n Tesla has been able to rake in millions of miles worth of video footage that it\n uses to train FSD. The idea there is that the more data Tesla can collect, the\n closer the automaker can get to actually achieving full self-driving.\n\n However, some industry experts say there might be a limit to the brute force\n approach of throwing more data at a model and expecting it to get smarter.\n\n \u201cFirst of all, there\u2019s an economic constraint, and soon it will just get too\n expensive to do that,\u201d Anand Raghunathan, Purdue University\u2019s Silicon Valley\n professor of electrical and computer engineering, told TechCrunch. Further,\n he said, \u201cSome people claim that we might actually run out of meaningful\n data to train the models on. More data doesn\u2019t necessarily mean more\n information, so it depends on whether that data has information that is useful\n to create a better model, and if the training process is able to actually distill\n that information into a better model.\u201d\n\n Raghunathan said despite these doubts, the trend of more data appears to be\n here for the short-term at least. And more data means more compute power\n needed to store and process it all to train Tesla\u2019s AI models. That was where\n Dojo, the supercomputer, came in.\n\n\n\n What is a supercomputer?\n Dojo was Tesla\u2019s supercomputer system that was designed to function as a\n training ground for AI, specifically FSD. The name is a nod to the space where\n martial arts are practiced.\n\n A supercomputer is made up of thousands of smaller computers called\n nodes. Each of those nodes has its own CPU (central processing unit) and\n GPU (graphics processing unit). The former handles overall management of\n the node, and the latter does the complex stuff, like splitting tasks into\n multiple parts and working on them simultaneously.\n\n GPUs are essential for machine learning operations like those that power FSD\n training in simulation. They also power large language models, which is why\n the rise of generative AI has made Nvidia the most valuable company on the\n planet.\n\n Even Tesla buys Nvidia GPUs to train its AI (more on that later).\n\n\n\n Why did Tesla need a\n supercomputer?\n Tesla\u2019s vision-only approach was the main reason Tesla needed a\n supercomputer. The neural networks behind FSD are trained on vast amounts\n of driving data to recognize and classify objects around the vehicle and then\n make driving decisions. That means that when FSD is engaged, the neural\n nets have to collect and process visual data continuously at speeds that\n match the depth and velocity recognition capabilities of a human.\n\n In other words, Tesla means to create a digital duplicate of the human visual\n cortex and brain function.\n\n To get there, Tesla needs to store and process all the video data collected\n from its cars around the world and run millions of simulations to train its\n model on the data.\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 5 of 8\n Elon Musk \u00b7 Jul 23, 2024\n @elonmusk \u00b7 Follow\n Replying to @elonmusk and @ajtourville\n And Dojo 1 will have roughly 8k H100-equivalent of training\n online by end of year.\n Not massive, but not trivial either.\n Elon Musk\n @elonmusk \u00b7 Follow\n Dojo pics\n\n\n\n\n 2:24 PM \u00b7 Jul 23, 2024\n 6.3K Reply Copy link\n Read 385 replies\n\n Tesla relied mainly on Nvidia to power its current Dojo training computer, but it\n didn\u2019t want to have all its eggs in one basket \u2014 not least because Nvidia chips\n are expensive. Tesla had hoped to make something better that increased\n bandwidth and decreased latencies. That\u2019s why the automaker\u2019s AI division\n decided to come up with its own custom hardware program that aimed to\n train AI models more efficiently than traditional systems.\n\n At that program\u2019s core was Tesla\u2019s proprietary D1 chips, which the company\n said were optimized for AI workloads.\n\n\n\n Tell me more about these chips\n\n\n\n\n GANESH VENKATARAMANAN, FORMER SENIOR DIRECTOR OF AUTOPILOT HARDWARE, PRESENTING THE\n D1 TRAINING TILE AT TESLA\u2019S 2021 AI DAY.\n\n IMAGE CREDITS:TESLA / SCREENSHOT\n\n\n\n Tesla, like Apple, thinks hardware and software should be designed to work\n together. That\u2019s why Tesla was working to move away from the standard GPU\n hardware and design its own chips to power Dojo.\n\n Tesla unveiled its D1 chip, a silicon square the size of a palm, on AI Day in 2021.\n The D1 chip entered into production around July 2023.\n\n The Taiwan Semiconductor Manufacturing Company (TSMC) manufactured\n the chips using 7 nanometer semiconductor nodes. The D1 has 50 billion\n transistors and a large die size of 645 millimeters squared, according to Tesla.\n This is all to say that the D1 promises to be extremely powerful and efficient\n and to handle complex tasks quickly.\n\n The D1 wasn\u2019t as powerful as Nvidia\u2019s A100 chip, though.\n\n Tesla had been working on a next-gen D2 chip that aimed to solve information\n flow bottlenecks. Instead of connecting the individual chips, the D2 would\n have put the entire Dojo tile onto a single wafer of silicon.\n\n Tesla never confirmed how many D1 chips it ordered or received. The\n company also never provided a timeline for how long it would have taken to\n get Dojo supercomputers running on D1 chips.\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 6 of 8\n What did Dojo mean for Tesla?\n\n\n\n\n VISITORS ARE VIEWING TESLA\u2019S HUMANOID ROBOT OPTIMUS PRIME II AT WAIC IN SHANGHAI,\n CHINA, ON JULY 7, 2024.\n\n IMAGE CREDITS:COSTFOTO / NURPHOTO / GETTY IMAGES\n\n\n\n Tesla\u2019s hope was that by taking control of its own chip production, it might\n one day be able to quickly add large amounts of compute power to AI training\n programs at a low cost.\n\n It also meant not having to rely on Nvidia\u2019s chips in the future, which are\n increasingly expensive and hard to secure. Now, Tesla is going all-in on\n partnerships \u2014 with Nvidia, AMD, and Samsung, which will build its next-gen\n AI6 chip.\n\n During Tesla\u2019s second-quarter 2024 earnings call, Musk said demand for\n Nvidia hardware was \u201cso high that it\u2019s often difficult to get the GPUs.\u201d He said\n he was \u201cquite concerned about actually being able to get steady GPUs when\n we want them, and I think this therefore requires that we put a lot more effort\n on Dojo in order to ensure that we\u2019ve got the training capability that we\n need.\u201d\n\n Dojo was a risky bet, one that Musk hedged several times by saying that Tesla\n might not succeed.\n\n In the long run, Tesla toyed with the idea of creating a new business model\n based on its AI division, with Musk even saying during a Q2 2024 earnings call\n that he saw \u201ca path to being competitive with Nvidia with Dojo.\u201d While D1 was\n more tailored for Tesla computer vision labeling and training \u2014 useful for FSD\n and Optimus training \u2014 it wouldn\u2019t have been useful for much else. Future\n versions would have to be more tailored to general-purpose AI training, Musk\n said.\n\n The problem that Tesla might have come up against is that almost all AI\n software out there has been written to work with GPUs. Using Dojo chips to\n train general-purpose AI models would have required rewriting the software.\n\n That is, unless Tesla rented out its compute, similar to how AWS and Azure\n rent out cloud computing capabilities \u2014 an idea that excited analysts. A\n September 2023 report from Morgan Stanley predicted that Dojo could add\n $500 billion to Tesla\u2019s market value by unlocking new revenue streams in the\n form of robotaxis and software services.\n\n In short, Dojo chips were an insurance policy for the automaker, but one that\n might have paid dividends.\n\n\n\n How far did Tesla Dojo get?\n\fCase 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 7 of 8\n\n\n\n\n NVIDIA CEO JENSEN HUANG AND TESLA CEO ELON MUSK\n\n IMAGE CREDITS:KIM KULISH / CORBIS / GETTY IMAGES\n\n\n\n Musk often provided progress reports, but many of his goals for Dojo were\n never reached.\n\n For instance, Musk suggested in June 2023 that Dojo had been online and\n running useful tasks for a few months.\u201d Around the same time, Tesla said it\n expected Dojo to be one of the top five most powerful supercomputers by\n February 2024 and had planned for total compute to reach 100 exaflops in\n October 2024, which would have required roughly 276,000 D1s, or around\n 320,500 Nvidia A100 GPUs.\n\n Tesla never provided an update or any information that would suggest it ever\n reached these goals.\n\n Tesla and Musk made numerous other pledges for Dojo, including financial\n ones. For instance, Tesla committed in January 2024 to spend $500 million to\n build a Dojo supercomputer at its gigafactory in Buffalo, New York, and has\n already spent $314 million of that, per a 2024 report.\n\n Just after Tesla\u2019s second-quarter 2024 earnings call, Musk posted photos of\n Dojo 1 on X, saying that it would have \u201croughly 8k H100-equivalent of training\n online by end of year. Not massive, but not trivial either.\u201d\n\n Despite all of this activity \u2014 particularly by Musk on X and in earnings calls \u2014\n mention of Dojo abruptly ended August 2024. And talk switched to Cortex.\n\n During the company\u2019s fourth-quarter 2024 earnings call, Tesla said it\n completed the deployment of Cortex, \u201ca ~50k H100 training cluster at\n Gigafactory Texas\u201d and that Cortex helped enable V13 of supervised FSD.\n\n In Q2 2025, Tesla noted it \u201cexpanded AI training compute with an additional\n 16k H200 GPUs at Gigafactory Texas, bringing Cortex to a total of 67k H100\n equivalents.\u201d During that same earnings call, Musk said he expected to have\n a second Dojo cluster operating \u201cat scale\u201d in 2026. He also hinted at potential\n redundancies.\n\n \u201cThinking about Dojo 3 and the AI6 inference chip, it seems like intuitively, we\n want to try to find convergence there, where it\u2019s basically the same chip,\u201d\n Musk said.\n\n A few weeks later, he reversed course and disbanded the Dojo team.\n\n TechCrunch confirmed in late August 2025 that Tesla still plans to commit\n $500 million to a supercomputer in Buffalo \u2014 it just won\u2019t be Dojo.\n\n This story originally published August 3, 2024. The article was updated for a\n final time September 2, 2025, with new information about Tesla\u2019s decision to\n shut down Dojo.\n\n\n Topics: AI Dojo Elon Musk evergreens Tesla Tesla FSD Transportation\n\n\n When you purchase through links in our articles, we may earn a small commission. This doesn\u2019t affect our\n editorial independence.\n\n\n\n\n Advertisement\n\f Case 7:26-mc-00318-LS Document 6-16 Filed 08/18/26 Page 8 of 8\n\n\n Rebecca Bellan\n Senior Repor ter |\n\n Rebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends\n shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and\u2026\n\n View Bio\n\n\n\n Loading the next article\n\n\n\n\n Advertisement\n\n\n\n\n TechCrunch Terms of Service OpenAI vs Apple\n\n Nous Research\n Staff Privacy Policy\n Space Data Centers\n Contact Us RSS Terms of Use\n Staya Nadella\n Advertise\n SpaceX Starship\n Crunchboard Jobs\n Tech Layoffs\n Site Map ChatGPT\n\n\n\u00a9 2026 TechCrunch Media LLC.\n\f","ocr_status":1,"date_upload":"2026-08-20T15:38:02.940636-07:00","document_number":"6","attachment_number":16,"pacer_doc_id":"181037220283","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 15","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554438/","id":490554438,"tags":[],"absolute_url":"/docket/74659430/6/17/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.423621-07:00","date_modified":"2026-08-23T09:39:59.301746-07:00","sha1":"ed4ab87a268d55f042c002ac836c2ba6f98640ba","page_count":6,"file_size":5977651,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.17.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.17.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 1 of 6\n\n\n\n\n EXHIBIT\n\n 16\n\f Case 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 2 of 6\n SCIENCE TECHNOLOGY ENVIRONMENT DIY GEAR MERCH NEWSLETTER\n\n\n\n\nTECHNOLOGY VEHICLES SELF DRIVING\n\n\nHow Tesla is using a supercomputer to train its self-driving tech\nTesla's approach to autonomy is controversial: It relies on just cameras to see and understand the roads.\nROB STUMPF / PUBLISHED JUN 26, 2021 5:00 AM EDT / ADD POPULAR SCIENCE\n\n ADVERTISEMENT\n\n\n\n\n Trending\n\n\n\n\n The supercomputer cluster has 5,760 GPUs\u2014processing power it needs to help power its self-driving aspirations. Tesla\n\n\n\n Get the Popular Science daily newsletter \ud83d\udca1 BIRDS\n\n Breakthroughs, discoveries, and DIY tips sent six days a week. Outside eagle spotted on Jackie and\n Shadow\u2019s nest\n Enter your email SIGN UP LAURA BAISAS\n\n By signing up, you confirm you are 16+, will receive newsletters and promotional content and agree to our Terms of Use and\n acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.\n\n\n You can\u2019t buy a fully self-driving car today, but automakers around the globe\n are racing to become the first company to place such a vehicle on dealer\n lots. No two companies are taking the same technological path to achieve\n this plan, either. Some make use of remote sensing methods like Light\n Detection and Ranging (LiDAR), while others rely on radar-based sensors to BEARS\n\n help pick out hard-to-see obstacles in the roadway. And typically, firms Why do mother bears kill cubs? The\n working on autonomous tech will use a combination of LiDAR, radar, and answer is complicated.\n cameras. JENNIFER BYRNE\n\n\n Then there\u2019s Tesla, which believes vision-based image recognition using only\n cameras is the key to affordable and reliable autonomy.\n\fCase 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 3 of 6\n ADVERTISEMENT\n\n\n\n\n Nomatic\n Nomatic\n\n\n\n But there\u2019s a catch to Tesla\u2019s method: perfecting vision-based autonomy is\n difficult. It requires the use of a continuously improving system that can\n quickly adapt to new and changing road conditions, and then it must be\n capable of sharing that information with other vehicles on the roadway. That\n kind of learning takes significantly more processing power than what is\n available in a single vehicle\u2014it takes a supercomputer.\n [Related: Everything self-driving cars calculate before changing lanes]\n During a talk at the International Joint Conference on Computer Vision and\n Pattern Recognition earlier this month, Tesla\u2019s senior director of AI, Andrej\n Karpathy, revealed that the automaker has been working on a project to do\n exactly that.\n Tesla\u2019s new supercomputer hasn\u2019t been named, at least not publicly. The\n cluster itself consists of 720 individual computers called nodes. Each node\n has eight Nvidia A100 80GB Graphics Processing Units (GPUs) capable of\n performing high-intensity floating point calculations with nearly 500 times as\n much power compared to a standard desktop processor.\n In total, the cluster has 5,760 GPUs, or enough hardware to achieve an\n insane 1.8 exaflops of processing power. Karpathy believes this makes\n Tesla\u2019s supercomputer the fifth most powerful computing environment in the\n entire world, at least on paper.\n Modern Teslas utilize an advanced driver assistance system called Autopilot.\n This suite of features allows the vehicle to make use of eight exterior-facing\n cameras to gather data about the vehicle\u2019s surroundings and, when engaged\n and where applicable, performs lateral (steering) and longitudinal\n (acceleration and braking) controls under driver supervision. While this\n shouldn\u2019t be confused with Waymo\u2019s advanced self-driving, it is an interim\n step that uses partial automation to bridge the gap between manual driving\n and fully autonomous control.\n ADVERTISEMENT\n\n\n\n\n [Related: How Waymo is teaching self-driving cars to deal with the chaos\n of parking lots]\n Autopilot uses information gathered from all Tesla vehicles on the road to\n improve its driving decisions. As a Tesla steers along the street, its exterior\n cameras are constantly gathering data on the outside environment.\n Computers within the car study this data and make predictions of how to\n behave in any given scenario without actually sending controls to the vehicle\n itself.\n This information is shared on a machine learning architecture called a neural\n network. The predictions are then recorded and sent back to Tesla to\n determine if the decision was correct or if any data was misidentified. If it\n was, then the data then continually runs through the supercomputer\n tweaking its behavior until it processes without a mistake, effectively training\n Tesla\u2019s ever-improving Autopilot model.\n [Related: Intel\u2019s new chip puts a teraflop in your desktop. Here\u2019s what that\n means]\n This method not only consumes a large amount of processing power, but it\n also requires significant storage in order to stockpile the one million 10-\n second clips used to make up the proprietary Tesla dataset training for\n\fCase 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 4 of 6\n Autopilot. These clips alone require 1.5 petabytes of storage, whereas the\n system itself is capable of hoarding approximately 10 petabytes of data on\n ultra-fast NVMe flash storage.\n Relatedly, Tesla CEO Elon Musk has previously teased \u201cProject Dojo,\u201d a\n supercomputer built on proprietary Tesla silicon specifically architectured for\n neural net model training. Musk noted that establishing high speed\n communication between components and efficient cooling was an ongoing\n challenge in late 2020, though the project was ongoing.\n ADVERTISEMENT\n\n\n\n\n Because Karpathy\u2019s cluster uses Nvidia-based GPUs, it doesn\u2019t appear to be\n affiliated with Project Dojo. However, it still plays an important role in Tesla\u2019s\n ultimate goal of being the first automaker capable of a fully self-driving\n vehicle on public roads.\n Tesla\u2019s rather ambitious goal has been met with quite a bit of skepticism by\n industry leaders and naysayers of vision-only vehicle autonomy, especially\n since the automaker rejected the use of ultra-precision LiDAR as part of its\n autonomy suite.\n [Related: This supercomputer will perform 1,000,000,000,000,000,000\n operations per second]\n No Tesla vehicle on the road today makes use of LiDAR. In fact, Elon Musk\n called LiDAR a \u201ccrutch\u201d in 2018, denouncing the technology in favor of\n Tesla\u2019s own vision-based system before doing away with supplemental\n radars earlier this year. That decision alone cost Tesla safety endorsements\n from the National Highway Traffic Safety Administration.\n Meanwhile, Volvo has chosen to implement LiDAR as a standard feature on\n the upcoming successor to its XC90 SUV.\n As for Tesla, its current-generation supercomputer will help to train its\n Autopilot model, and its upcoming Project Dojo likely even moreso. But only\n time will tell if its vision-based technology will prevail over competitors,\n meaning that it\u2019s a gambit that could make or break its position as a leader in\n the autonomy segment.\n\n\n\n 2026 Popular Science Home of the Future Awards\n\n 39 products that will improve your everyday life.\n\n SEE IT\n\n\n\n\n ROB STUMPF\n Contributor, Tech\n Rob has been covering emerging car tech for PopSci since 2021, and the automotive beat for its\n motoring-focused sibling publication, The Drive, since early 2017. He brings both a tech and\n automotive background to his work about what the future of mobility holds.\n\f Case 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 5 of 6\n\n\n\n\n More in Self Driving\n\n\n\n\nSELF DRIVING SELF DRIVING ELECTRIC VEHICLES SELF DRIVING\nTeslas keep hitting emergency vehicles, and The new Mission Master XT is a massive military We\u2019re getting closer to a world filled with self- What we know so far about the fatal Tesla crash\nnow the government is investigating self-driving pack mule parking cars in Paris\nROB VERGER KELSEY D. ATHERTON ROB STUMPF COLLEEN HAGERTY\n\n\n\n\nSELF DRIVING ELECTRIC VEHICLES SELF DRIVING SELF DRIVING\nThe 6 terms you need to know to understand The government is investigating why Tesla Tesla caps off a tumultuous year with a massive \u2018Autopilot\u2019 systems in cars are finally getting\nself-driving cars drivers can play solitaire at the wheel recall graded on safety\nDAN CARNEY COLLEEN HAGERTY CHARLOTTE HU ROB STUMPF\n\n\n\n SEE MORE\n\n\n\n\n More in Vehicles\n\n\n\n\nELECTRIC VEHICLES ELECTRIC VEHICLES ELECTRIC VEHICLES ELECTRIC VEHICLES\nVolvo\u2019s sleek electric car concept has big Ikea Tesla\u2019s new adapter will let other car companies GM is recalling all its Bolts, but there\u2019s no need to These are the most useless car tech features\nvibes use its Fast Charging stations panic about EV safety ROB STUMPF\nROB STUMPF ROB STUMPF ROB VERGER\n\f Case 7:26-mc-00318-LS Document 6-17 Filed 08/18/26 Page 6 of 6\n\n\n\n\nELECTRIC VEHICLES ELECTRIC VEHICLES SELF DRIVING ELECTRIC VEHICLES\nThe Tesla Model S \u2018Plaid\u2019 will go from 0-60 with Tesla\u2019s new battery tech promises a road to a Self-driving cars still have major perception This company plans to build a self-driving car\nrecord-breaking acceleration cheap self-driving electric car problems with a brain that runs on light\nROB VERGER STAN HORACZEK YULONG CAO & Z. MORLEY MAO/THE CONVERSATION ROB VERGER\n\n\n\n SEE MORE\n\n\n\n\n More in Technology\n\n\n\n\nAI DRONES TECHNOLOGY TECHNOLOGY\nTesla wants to make humanoid robots. Here\u2019s Elvis (the helicopter) is cheating death by Director Of FBI Addresses Congress About San Your McDonald\u2019s Happy Meal Box Is Now A Virtual\ntheir competition. becoming a drone Bernardino iPhone Reality Headset\nCHARLOTTE HU KELSEY D. ATHERTON XAVIER HARDING CARL FRANZEN\n\n\n\n\nCANCER TECHNOLOGY WEAPONS TECHNOLOGY\nHere\u2019s How Virtual Reality Could Help Doctors Feed Your 3D Printer Recycled Plastic Burn After Shooting: Army Employees Patent This Is The Defining Photo Of Virtual Reality (So\nTreat Cancer XAVIER HARDING Self-Destructing Bullets Far)\nALEXANDRA OSSOLA KELSEY D. ATHERTON CARL FRANZEN\n\n\n\n SEE MORE\n\n\n\n\n SCIENCE\n\n TECHNOLOGY\n\n ENVIRONMENT\nGet the Popular Science daily newsletter\n\ud83d\udca1 EDITORIAL STANDARDS HOW WE TEST AND REVIEW PRODUCTS ABOUT US CONTACT US\n DIY\n\nBreakthroughs, discoveries, and DIY tips sent six GEAR\n TV CHANNEL PRIVACY POLICY TERMS OF USE AFFILIATE DISCLOSURE SITEMAP\ndays a week.\n MERCH\n Enter your email\n SHOP\n\n SIGN UP DISCLAIMER(S)\n ARTICLES MAY CONTAIN AFFILIATE LINKS WHICH ENABLE US TO SHARE IN THE REVENUE OF ANY PURCHASES MADE. REGISTRATION\nBy signing up, you confirm you are ON OR USE OF THIS SITE CONSTITUTES ACCEPTANCE OF OUR TERMS OF SERVICE.\n16+, will receive newsletters and\npromotional content and agree to our \u00a9 2026 ZIFF DAVIS. ALL RIGHTS RESERVED.\nTerms of Use and acknowledge the\n WITHDRAW CONSENT / MANAGE PRIVACY PREFERENCES\ndata practices in our Privacy Policy.\n DO NOT SELL MY PERSONAL INFORMATION\nYou may unsubscribe at any time.\n\f","ocr_status":1,"date_upload":"2026-08-20T15:37:56.768385-07:00","document_number":"6","attachment_number":17,"pacer_doc_id":"181037220284","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 16","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554439/","id":490554439,"tags":[],"absolute_url":"/docket/74659430/6/18/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.436018-07:00","date_modified":"2026-08-23T09:40:05.653171-07:00","sha1":"e1d0cb2a4735e6a7eaa9f67b5464731957610021","page_count":2,"file_size":13695054,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.18.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.18.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-18 Filed 08/18/26 Page 1 of 2\n\n\n\n\n EXHIBIT\n\n 17\n\f Case 7:26-mc-00318-LS Document 6-18 Filed 08/18/26 Page 2 of 2\nCareers Explore Jobs Manufacturing AI Terafab Vehicle Software Internships About Us Profile US\n\n\n\n\n Build your Career at Tesla\n\n\n\n\n 14 Results Clear Filters (5)\n Field Service Technician, Utility\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n Search by role or keyword\n\n\n Job Category Field Service Technician, Crew Based, Commercial\n Learn More\n Energy - Solar & Storage\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n\n\n Job Type\n Licensed Journeyman Electrician, Commercial Self Perform\n Learn More\n - Select - Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n\n Region\n Construction Manager, Energy Projects\n Learn More\n North America\n Energy - Solar & Storage \u30fb Part-Time Austin, Texas\n\n\n Location\n Excavator Operator, Commercial Self Perform\n United States of America Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n State\n\n Texas Carpenter & Concrete Finisher, Commercial Self Perform\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n City Search by ZIP Code\n\n Austin Technical Support Engineer, Commercial Charging\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\f","ocr_status":1,"date_upload":"2026-08-20T15:38:03.762159-07:00","document_number":"6","attachment_number":18,"pacer_doc_id":"181037220285","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 17","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554440/","id":490554440,"tags":[],"absolute_url":"/docket/74659430/6/19/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.449226-07:00","date_modified":"2026-08-23T09:39:54.712480-07:00","sha1":"723f513471ba629e9d122be0a0d91cf610fb8acc","page_count":6,"file_size":7348032,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.19.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.19.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 1 of 6\n\n\n\n\n EXHIBIT\n\n 18\n\f Case 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 2 of 6\n\n\n\n\nBUSINESS\n\n\nTesla now employs 20,000 in Austin, could triple that as\nit ramps up Cybertruck production\nBy Eric Killelea , Staff writer\nUpdated Sep 25, 2023 3:54 p.m.\n\n\n\n\nEmployees at Tesla\u2019s factory in Austin pose in mid-July with a Cybertruck the company said was the first to be built at\nits Gigafactory Texas in Austin. A company official says the plant now employs about 20,000 people \u2014 and that number is\nexpected to triple.\nTesla/Courtesy\n\n\n\n\n Listen Now:\n Tesla now employs 20,000 in Austin, could triple that as it ramps up\n 1x\n Cybertruck production\n About 5 Minutes\n\n\n\n\n Tesla Inc. now employs 20,000 people at its plant near Austin and has the potential to\n triple that as it\u2019s quickly become one of the largest employers in Central Texas, the\n Austin Regional Manufacturers Association said.\n\n\n\n\n Want more Express-News?\n\n Add Preferred Source\n\n\n\n\n MORE NEWS\n Watch More\n San Antonio doctor\n loses 4 properties to\n foreclosure with bids\n totaling $26M\n\fCase 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 3 of 6\n The electric vehicle maker has added more than 7,700 employees since the start of the International\n homebuyers are\n year, when it had just shy of 12,300, according to the Texas-based company\u2019s annual flocking to Texas.\n compliance report to Travis County economic development officials. Tesla\u2019s growth in Where are they moving\n from?\n Texas has been rapid; it reported about 3,500 employees there in 2021, the year it\n began production. Home buyers need less\n income to buy starter\n homes, according to\n ADVERTISEMENT Redfin report\n Article continues below this ad\n\n Governor's aim to break\n up CPS rattles S.A.\n business community\n\n\n Alamo Heights\n neighbors push back on\n planned 5-story\n apartment complex\n\n\n\n\n Current employment was disclosed last week by the plant\u2019s director of manufacturing\n during a meeting of the manufacturers association.\n\n\n With this year\u2019s new hires, Elon Musk-led Tesla is now Austin\u2019s second-largest private\n employer, just behind San Antonio\u2019s H-E-B, which earlier this year reported 22,955\n employees in Austin, according to the Austin Business Journal. Dell Technologies\n employed 13,000 workers in the metro last year, according to filings with the U.S.\n Securities and Exchange Commission.\n\n\n RELATED: Heading toward Cybertruck, Tesla triples employment and builds 4,000\n Model Ys a week in Texas\n SALE: ONLY 25\u00a2 Sign in\n\n\n\n\n Want more Express-News?\n Make us a Preferred Source on Google to see more of us when you search.\n\n\n Add Preferred Source\n\n\n\n\n Tesla now is stepping up its hiring efforts to push employment to 60,000 at its sprawling\n Austin plant as it continues producing its Model Y mid-sized SUV and ramps up\n production of its long-delayed Cybertruck.\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n The disclosure reportedly was made by Jason Shawhan, Tesla\u2019s director of\n manufacturing, when he spoke at the manufacturers association\u2019s 2023 State of\n Manufacturing Conference and Expo last week.\n\n\n Neither Tesla nor the Austin Regional Manufacturers Association responded to requests\n for comment Monday.\n\n\n 845 openings Watch More\n\n\n\n\n As of Monday morning, Tesla had about 845 job openings listed in Texas on its website.\n Most of those were based in Austin.\n\n\n ADVERTISEMENT\n\fCase 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 4 of 6\n Article continues below this ad\n\n\n\n\n Among them are Austin metro area-based positions for manufacturing engineers,\n construction managers, supply chain parts advisors, environmental specialists, project\n designers for the company\u2019s solar and EV charging stations, attorneys \u2014 even\n an athletic trainer to focus on \u201csetting the right conditions of work \u2014 from\n manufacturing and maintenance to sales and service.\u201d\n\n\n RELATED: Tesla continues growth, will lease 1 million square feet in Hays County\n industrial park\n\n\n Tesla relocated to Texas from California in 2021 and began limited production of the\n Model Y at its sprawling factory later that year. By the end of 2022, the company said it\n had invested $5.81 billion into the sprawling facility.\n\n\n Development has continued in 2023. In January, Tesla said it was planning to spend\n $775 million on five construction projects covering 1.5 million square feet of its land off\n Texas 45 and U.S. 130 near Austin-Bergstrom International Airport. It also plans to open\n a new cathode production facility at the site this year. Cathodes are a key component\n for the batteries that power its electric vehicles.\n\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n Musk also told investors the company had installed Cybertruck production equipment at\n the Texas plant in anticipation of kicking off production this summer and ramping up\n manufacturing efforts by 2024.\n\n\n As Tesla moves toward production of the long-delayed pickup, the company on Monday\n had 65 Austin job openings specifically related to manufacture of the stainless-steel\n electric truck.\n\n\n Elsewhere in Texas\n The company also is building a $375 million lithium refinery in Robstown, near Corpus\n Christi. Musk has said it would begin producing enough of the material for batteries in\n about 1 million EVs a year by 2025.\n\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n Watch More\n\f Case 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 5 of 6\n As of Monday, Tesla had 39 job openings in Robstown.\n\n\n RELATED: Greg Abbott and Elon Musk celebrate construction of Tesla\u2019s lithium refinery\n in South Texas\n\n\n Tesla\u2019s hiring push comes as tech giants with offices in the Austin area have been laying\n off workers. Google, Microsoft, Amazon, Facebook parent Meta and Dell all have\n announced cuts. It also reflects Musk\u2019s expanding business footprint across the state.\n\n\n Hawthorne, Calif.-based SpaceX, where Musk is CEO, had about 1,700 employees at\n its Starbase facility in Boca Chica in April, according to former Brownsville Mayor Juan\n \u201cTrey\u201d Mendez III.\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n As of Monday, SpaceX was advertising 103 openings for its Starship program in Boca\n Chica and eight at its rocket testing site in McGregor, near Waco.\n\n\n Other Musk-owned businesses with Texas operations \u2014 tunneling firm The Boring Co.\n and brain implant company Neuralink \u2014 also were advertising dozens of job openings\n Monday.\n\n\n Sep 25, 2023 | Updated Sep 25, 2023 3:54 p.m.\n\n\n Eric Killelea\n\n\n\n\nAround the Web Powered by\n\n\n\n\nAfter 60, Leg Strength Comes From The Lawn Problem Most Neurologists Beg Seniors With\nOne Simple Daily Move Homeowners Notice but Can't Fix Neuropathy: Stop Doing This Now\nBy ApexLabs By Glosrity By Health Weekly\n\n\n\n\nSurgeons: This Simple Trick Will Honey: The Greatest Enemy of Women Are Obsessed With These\nEnd Knee Pain & Arthritis Quickly Memory Loss (See How to Use It) Beautiful Floral Caps\n(Try It) By Health Weekly\n By Peoasis\n\nBy Health Weekly\n\n\n\n\n Watch More\n\n\n\n\nA 78-Year-Old Master Craftsman Sciatica is Not From a Slipped 1 Simple Tip to Cut Your Electric\nMade This Hummingbird House. Disc. Meet The Real Enemy of Bill (Try Tonight)\nThen This Happened Sciatica (Stop This) By MadeInGenius\nBy Ribili By SmoothSpine\n\f Case 7:26-mc-00318-LS Document 6-19 Filed 08/18/26 Page 6 of 6\n\n\n\n\nDoctor Begs Seniors: Do This to Neuropathy is Not From Low Endocrinologist: If You Have\nStop Losing Muscle Vitamin B (Meet The Real Enemy) Diabetes, Read This Before It's\nBy ApexLabs By Health Weekly Removed!\n By Health Weekly\n\n\n\n\nMOST POPULAR\n SAN ANTONIO REAL ESTATE\n Alamo Heights neighbors push back on planned 5-\n story apartment complex\n Homeowners near the site of the proposed development want\n the city to enforce its heights restrictions; the developer says\n the project isn\u2019t feasible without keeping all five floors.\n\n\n\n\nNEWS FORECASTS\n\nBankruptcy court records offer first look at Camp Mystic\u2019s finances When is South Texas\u2019 next rain chance? Here\u2019s the\n latest update\nThe Texas Hill Country retreat\u2019s gross revenue dropped from $11 million in 2024 to $5.7 million in\n2025, the year that a flash flood killed 27 guests and the camp\u2019s owner.\n\n\nBUSINESS NEWS\nInternational\nNEWS homebuyers are Annelise\n BUSINESS Camp, 2-year-old at\nflocking to Texas. Where are center of Texas brain-death\n\u2018Your body belongs to me\u2019: Governor's aim to break up\nthey moving from? battle, dies\nFeds release incarcerated CPS rattles S.A. business\npastor's explicit texts community\n\n\n Return To Top\n\n About Contact Services Account\n\n Our Company Customer Service Archives Subscribe\n\n Careers Frequently Asked Questions Newspaper Archive Newsletters\n\n Our Use of AI Newsroom Contacts Advertising e-Edition\n\n Standards and Practices Copyright and Reprints Photo Store\n\n\n\n\n \u00a9 2026 Hearst Newspapers, LLC Terms of Use Privacy Notice DAA Industry Opt Out Your Privacy Choices (Opt Out of Sale/Targeted Ads)\n\n\n\n\n Watch More\n\f","ocr_status":1,"date_upload":"2026-08-20T15:38:06.168188-07:00","document_number":"6","attachment_number":19,"pacer_doc_id":"181037220286","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 18","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554441/","id":490554441,"tags":[],"absolute_url":"/docket/74659430/6/20/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.466276-07:00","date_modified":"2026-08-23T04:40:19.634151-07:00","sha1":"d3cd339448e819597a741ac2ca144fddd4929387","page_count":3,"file_size":11994367,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.20.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.20.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-20 Filed 08/18/26 Page 1 of 3\n\n\n\n\n EXHIBIT\n\n 19\n\f Case 7:26-mc-00318-LS Document 6-20 Filed 08/18/26 Page 2 of 3\nCareers Explore Jobs Manufacturing AI Terafab Vehicle Software Internships About Us Profile US\n\n\n\n\n Gigafactory\n Austin, Texas\n\fCase 7:26-mc-00318-LS Document 6-20 Filed 08/18/26 Page 3 of 3\n\n\n Join Us at Our Global Headquarters\n Covering 2,500 acres along the Colorado River with over 10 million square feet of factory\n floor, Gigafactory Texas is a U.S. manufacturing hub for Model Y and the home of Cybertruck.\n\n It doesn\u2019t matter where you come from, where you went to school or what industry you\u2019re in\u2014we\u2019re\n hiring individuals of all levels. If you\u2019ve done exceptional work, join us in solving the next generation of\n engineering, manufacturing and operational challenges.\n\n\n\n View Jobs\n\n\n\n\nManufacturing\nProducing vehicles quickly, efficiently and safely is one of the biggest challenges our Manufacturing team\nfaces. Implement manufacturing processes that evolve to meet rising global demand.\n\n\n\nSupervisor Technician Production Associate\nApply to lead a motivated team of individuals who are Apply to improve the devices, equipment and systems Apply to work across all areas of the vehicle\nresponsible for achieving our ambitious quality and that are critical to the efficiency of our production manufacturing process\u2014on-the-job training provided,\nproduction goals. lines. no prior experience required.\n\n\n\n\nEngineering From powerful battery systems to in-car technology, our Engineering team is developing new processes,\n equipment and tooling to help us shape the future of sustainable transportation. Create the next Cybertruck\n or continue to improve our current vehicle lineup.\n\n Learn more about Tesla battery cell development and apply for open positions to join our Cell team.\n\f","ocr_status":1,"date_upload":"2026-08-20T15:38:04.500907-07:00","document_number":"6","attachment_number":20,"pacer_doc_id":"181037220287","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 19","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554442/","id":490554442,"tags":[],"absolute_url":"/docket/74659430/6/21/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.478468-07:00","date_modified":"2026-08-23T04:40:45.193013-07:00","sha1":"721e2fc95fb860ecf5f0a34f13f0c87f1817bc21","page_count":3,"file_size":650609,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.21.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.21.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document6-21 Filed 08/18/26 Page 1 of 3\n\nEXHIBIT\n\n20\n\fCase 7:26-mc-00318-LS Document6-21 Filed 08/18/26 Page 2 of 3\n\nUNITED STATES DISTRICT COURT\nWESTERN DISTRICT OF TEXAS\nMIDLAND/ODESSA DIVISION\n\nNEURAL AI, LLC\n\nPlaintiff, Case No. 7:24-cv-00221-ADA-DTG\n\u2122 JURY TRIAL DEMANDED\nNVIDIA CORPORATION\n\nDefendant.\n\nI, Alon Daks, hereby declare as follows:\n\n1. I am a Senior Staff Software Engineer at Tesla Inc. (\u201cTesla\u201d). Iam over the age of\neighteen and competent to make this declaration. I make this declaration based on my personal\nknowledge, including my general familiarity with Tesla\u2019s technical infrastructure, software\ndeployments, use of NVIDIA products, as well as the on the basis of a reasonably diligent\ninvestigation.\n\n2. I understand that this declaration is submitted in connection with a subpoena Neural\nAl served on Tesla on June 25, 2026 in the above-captioned action (the \u201cSubpoena\u201d.\n\nq. In the ordinary course of business, Tesla deploys the following NVIDIA GPUs in\nthe United States: A100, H100, H200, GB300, V100, and A16.\n\n4. In the ordinary course of business, Tesla uses the following software and/or\nlibraries identified in the Subpoena (specifically, Request No. 2 for Production of Documents):\ncuBLAS, cuDNN, cuFFT, cuSOLVER, Isaac Lab, PyTorch, and Triton. I understand that\ncuBLAS, cuDNN, cuFFT, cuSOLVER, and Isaac Lab are provided by NVIDIA, but PyTorch is\n\nmanaged by the Linux Foundation and Triton is managed by https://triton-lang.org.\n\fCase 7:26-mc-00318-LS Document6-21 Filed 08/18/26 Page 3 of 3\n\n5. In the ordinary course of business, Tesla uses at least cuBLAS, cuDNN, cuFFT,\ncuSOLVER, and Triton as provided. I understand that Isaac Lab, PyTorch, and Triton are open\nsource software and/or libraries, but cuBLAS, cuDNN, cuFFT, and cuSOLVER are not open\n\nsource.\n\nI declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct.\n\nExecuted on August 10, 2026.\n\nAlon Daks\n\nAlon Daks (Aug 10, 2026 18:35:30 PDT)\n\nAlon Daks\n","ocr_status":1,"date_upload":"2026-08-20T15:40:12.325145-07:00","document_number":"6","attachment_number":21,"pacer_doc_id":"181037220288","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 20","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554443/","id":490554443,"tags":[],"absolute_url":"/docket/74659430/6/22/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.490478-07:00","date_modified":"2026-08-23T04:40:04.138964-07:00","sha1":"4a33afaf9f80ed17493e9fb48b47a951585cad5c","page_count":14,"file_size":242650,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.22.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.22.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 1 of 14\n\n\n\n\n EXHIBIT\n\n 21\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 2 of 14\n\n\n Monday, August 17, 2026 at 6:52:05 AM Paci\ufb01c Daylight Time\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\nDate: Tuesday, August 11, 2026 at 4:18:05 PM Paci\ufb01c Daylight Time\nFrom: Tanner Laiche\nTo: Jun Zheng, Rocco Magni, Gina Cremona\nCC: Emily Portuguese, Tamar Lusztig, Brian Melton, Max Tribble, Samuel Drezdzon, Richard Wojtczak, Rachel\n Hanna, Ashraf Fawzy\nAttachments: Outlook-6C509A71.png, Outlook-6C509A71.png\n\n\nCounsel,\n\nWe have reviewed the declaration. Neural AI does not consider this matter concluded. The\ndeclaration is materially insu`icient in several respects:\n\n a. The declaration addresses only GPU and partial software identi\ufb01cation. It omits\n entirely the paragraphs addressing unmodi\ufb01ed use of NVIDIA software (draft \u00b6 7),\n hardware/software functioning as designed by NVIDIA (draft \u00b6 8), use of NVIDIA-\n distributed pretrained models (draft \u00b6 9), CPU/GPU memory architecture (draft \u00b6\u00b6\n 10\u201311), standard data \ufb02ow from CPU memory to GPU memory including th use of\n GPUDirect (draft \u00b6\u00b6 12\u201313), U.S. operations (draft \u00b6 14), and frequency of use (draft \u00b6\n 15). These are key issues within the scope of the subpoena that you unilaterally\n refused to address.\n j. The draft declaration organized software into three categories (draft \u00b6\u00b6 4, 5, 6)\n re\ufb02ecting distinct layers of the accused stack. Tesla collapsed these into a single\n undi`erentiated list and did not address application-layer platforms.\n k. The declaration states that PyTorch and Triton are not NVIDIA-provided without\n con\ufb01rming whether Tesla uses them on NVIDIA GPUs. NVIDIA distributes and\n optimizes PyTorch for use within its GPU-accelerated ecosystem, and Neural AI\u2019s\n infringement contentions identify PyTorch as part of the accused software\n stack. Additionally, \"Triton\" refers to NVIDIA Triton Inference Server\u2014not the Triton\n compiler language\u2014which the declaration does not address.\n\nNeural AI's o`er to accept a declaration was expressly \"subject to resolving any material\ngaps.\" The gaps identi\ufb01ed above are material.\n\nGiven these de\ufb01ciencies, Neural AI intends to pursue (1) document production responsive\nto the outstanding requests, and (2) deposition testimony from a Tesla Rule 30(b)(6)\ndesignee. The testimony subpoena noticed for September 14, 2026 remains in e`ect.\n\nWe are available to meet and confer this week to resolve these issues short of motion\npractice. Please provide available times.\n\nRegards,\n\n 1 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 3 of 14\n\n\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\nThis e-mail contains privileged and con\ufb01dential information, which may be subject to the attorney-client\nprivilege and/or attorney work product protection. If you received this message in error, please notify the\nsender and delete it immediately.\n\n\nFrom: Jun Zheng \nDate: Monday, August 10, 2026 at 11:57 PM\nTo: Tanner Laiche ; Rocco Magni\n; Gina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\nCounsel,\n\nAs discussed during the meet-and-confers on July 28 and Aug. 7, and in view of Tesla\u2019s objections\nserved on July 21, Neural AI stated it would accept a declaration in lieu of document production and\ndeposition testimony. Accordingly, Tesla conducted a reasonable investigation into which NVIDIA\nGPUs and software identi\ufb01ed in Neural AI\u2019s subpoena Tesla uses and whether Tesla uses that\nsoftware o` the shelf. The attached declaration provides that information. Tesla considers this\nmatter concluded.\n\nRegards,\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\nFrom: Tanner Laiche \nSent: Thursday, August 6, 2026 12:14 PM\nTo: Jun Zheng ; Rocco Magni ;\nGina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n\n 2 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 4 of 14\n\n\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nCounsel,\n\nI am available tomorrow at 9:30 am CT. I will circulate a calendar invite.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\nFrom: Jun Zheng \nDate: Wednesday, August 5, 2026 at 1:31 PM\nTo: Tanner Laiche ; Rocco Magni\n; Gina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\nTanner,\n\nWe are available to meet and confer this Friday between 9-10 AM CT. Please let us\nknow if that works for your team.\n\nThanks,\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n 3 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 5 of 14\n\n\n\n\nFrom: Tanner Laiche \nSent: Tuesday, August 4, 2026 1:28 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Jun Zheng ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nCounsel,\n\nI am following up on Rocco\u2019s email.\n\nDespite the parties\u2019 prior meet-and-confers, document discovery in the underlying action\ncloses on August 11. Unless the parties can promptly reach a resolution, that deadline\nleaves Neural AI no practical alternative but to move to compel by the end of this week or,\nat the latest, August 10, to preserve its rights.\n\nTo reduce burden and potentially avoid motion practice, I am attaching a short set of\nquestions intended to guide your investigation and help identify the responsive\ninformation, and also recirculating the draft declaration we previously shared, and that\nTesla may revise to ensure its accuracy.\n\nIf Tesla commits to provide an executed declaration, Neural AI is willing to consider\naccepting the declaration in lieu of further document production and/or deposition\ntestimony, subject to resolving any material gaps. Otherwise, the discovery deadline will\nforce Neural AI to move to compel by or before August 10 to preserve its rights. Even if a\nmotion becomes necessary, we remain open to resolving the issues promptly and mooting\nor withdrawing the motion through compliance.\n\nWe are available this week to further meet and confer as necessary.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n 4 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 6 of 14\n\n\n\nFrom: Rocco Magni \nDate: Sunday, August 2, 2026 at 6:57 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nOur discovery deadline is approaching soon. Please let us know when you can confer again\nthis upcoming week. Thanks.\n\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Rocco Magni \nDate: Tuesday, July 28, 2026 at 6:52 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nThanks for speaking today. Attached is a draft of the declaration I referred to on our call.\n\nBest,\n\n\n 5 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 7 of 14\n\n\nRocco\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Gina Cremona \nDate: Thursday, July 23, 2026 at 9:45 AM\nTo: Rocco Magni \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\n\nHi Rocco, we are not available today.\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 22, 2026 12:01 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche\n; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nWould tomorrow work?\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\n\n 6 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 8 of 14\n\n\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\n\n\n\n On Jul 22, 2026, at 2:54 PM, Gina Cremona wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n We are not available on Friday, but can meet on Tuesday, July 28 between 8-10 am PT.\n\n Thanks,\n Gina\n\n\n From: Rocco Magni \n Sent: Wednesday, July 22, 2026 6:54 AM\n To: Jun Zheng ; Gina Cremona\n \n Cc: Tanner Laiche ; Emily Portuguese\n ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n ; Ashraf Fawzy ; Jun\n Zheng \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n Jun,\n\n Please provide times to meet and confer on Friday 7/24. Thanks.\n\n --\n Rocco F. Magni\n\n 7 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 9 of 14\n\n\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this\nmessage in error, please notify the sender and delete it immediately.\nFrom: Jun Zheng \nDate: Tuesday, July 21, 2026 at 7:37 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nEXTERNAL Email\nRocco,\n\nWe understand from the email exchange below that the noticed date for\ntestimony subpoena is now Sept. 14, 2026. In the meantime, attached please \ufb01nd\nTesla's objections and responses to Neural AI's subpoena.\n\nThanks!\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n\nFrom: Gina Cremona \nSent: Monday, July 20, 2026 1:07 PM\nTo: Rocco Magni \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n\n 8 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 10 of 14\n\n\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nHi Rocco,\n\nTesla cannot provide an agreed date for deposition until we have reviewed and\nresponded to the subpoenas. However, we can agree to Sept. 14 as the noticed date\nfor testimony subpoena for now, subject to modi\ufb01cation once we have responded to\nthe document subpoena.\n\nRegards,\nGina\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 15, 2026 12:58 PM\nTo: Gina Cremona \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nGina,\n\nWe\u2019ll agree to pull down the July 28 date once we have an agreed replacement date.\nLet us know what date works for you and we\u2019ll withdraw the current notice.\nThanks.\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\n 9 of 13\n\f Case 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 11 of 14\n\n\n\nThis e-mail may contain privileged and confidential information. If you received\nthis message in error, please notify the sender and delete it immediately.\n\n\n\n On Jul 15, 2026, at 3:54 PM, Gina Cremona \n wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n Understood. To con\ufb01rm, the deposition date of July 28 is o` calendar, and\n we will work on agreeing to a new date.\n\n Regards,\n Gina\n\n\n From: Rocco Magni \n Sent: Tuesday, July 14, 2026 5:28 PM\n To: Gina Cremona ; Tanner Laiche\n ; Emily Portuguese\n \n Cc: Tamar Lusztig ; Brian\n Melton ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Ms. Cremona,\n\n Discovery has not been extended; only depositions. Written\n discovery will still close on August 11.\n\n We can work with you on a deposition date between now and\n September 14. But we do not have \ufb02exibility on the timing of your\n RFP responses and document production beyond the 1 week\n extension we noted below.\n\n\n 10 of 13\n\fCase 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 12 of 14\n\n\n Best,\n\n Rocco\n\n --\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you\n received this message in error, please notify the sender and delete it\n immediately.\n From: Gina Cremona \n Date: Tuesday, July 14, 2026 at 8:19 PM\n To: Tanner Laiche ; Emily\n Portuguese \n Cc: Rocco Magni ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D.\n Tex.) - Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n It has come to our attention that the discovery deadline in this case has\n been extended. Due to summer vacations and people being out of the\n o`ice, we renew our request for an additional 2-week extension such that\n our deadlines will be Aug. 4 and Aug. 11.\n\n Regards,\n Gina\n\n\n From: Tanner Laiche \n Sent: Monday, July 6, 2026 5:20 PM\n To: Gina Cremona ; Emily Portuguese\n \n Cc: Rocco Magni ; Tamar\n Lusztig ; Brian Melton\n\n 11 of 13\n\fCase 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 13 of 14\n\n\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Counsel,\n\n Thanks for reaching out. Given the upcoming close of fact\n discovery, Neural AI is not able to agree to a three-week\n extension. That said, we can agree to a one-week extension for\n Tesla\u2019s written objections/responses to the subpoena(s). A copy\n of the Protective Order is attached.\n\n Regards,\n\n Tanner Laiche\n Susman Godfrey LLP\n 206.505.3816 | tlaiche@susmangodfrey.com\n 401 Union Street | Suite 3000 | Seattle, WA 98101\n HOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\n From: Gina Cremona \n Date: Thursday, July 2, 2026 at 7:21 PM\n To: Emily Portuguese \n Subject: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n We are in receipt of your Subpoena to Produce Documents and Subpoena\n for Testimony. Due to the holiday weekend and vacation schedules, we\n request a three-week extension to respond such that our deadlines will be\n Aug. 4 and Aug. 11.\n\n Additionally, please provide us with a copy of the protective order.\n\n Regards,\n Gina\n\n Gina H. Cremona\n Senior Counsel, IP Litigation\n 1501 Page Mill Rd., Palo Alto, CA 94304\n E. gcremona@tesla.com T. 650.647.0015\n\n\n 12 of 13\n\fCase 7:26-mc-00318-LS Document 6-22 Filed 08/18/26 Page 14 of 14\n\n\n\n \n\n\n\n\n 13 of 13\n\f","ocr_status":1,"date_upload":"2026-08-20T15:40:15.672503-07:00","document_number":"6","attachment_number":22,"pacer_doc_id":"181037220289","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 21","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490554444/","id":490554444,"tags":[],"absolute_url":"/docket/74659430/6/23/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-19T09:16:36.504761-07:00","date_modified":"2026-08-23T04:40:45.203345-07:00","sha1":"0890056a1f6a0505a0fa1f5757390f67fef61e97","page_count":2,"file_size":109895,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.23.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.6.23.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-LS Document 6-23 Filed 08/18/26 Page 1 of 2\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n NEURAL AI, LLC,\n Misc. Case No. 7:26-mc-00318-DC\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\n TESLA, INC.,\n\n Respondent.\n\n\n\n [PROPOSED] ORDER ON PETITIONER'S MOTION TO COMPEL\n COMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY TESLA, INC.\n\n Before the Court is Petitioner Neural AI, LLC\u2019s Motion to Compel Compliance with\n\nSubpoena Served on Third-Party Tesla, Inc. Having considered the Motion, the responses and\n\nobjections thereto, and the applicable law, the Court finds that the Motion should be and hereby is\n\nGRANTED.\n\n IT IS THEREFORE ORDERED that Third-Party Tesla, Inc. shall produce nonprivileged\n\ndocuments responsive to Requests for Production Nos. 1\u201312 on a rolling basis, with production to\n\nbegin within seven (7) days of the date of this Order and be completed within twenty-one (21)\n\ndays of this Order.\n\n IT IS FURTHER ORDERED that Third-Party Tesla, Inc. shall designate and produce a\n\nknowledgeable witness for deposition testimony on Deposition Topics 1\u20135 within thirty (30) days\n\nof the date of this Order.\n\n\n\n\n 1\n\f Case 7:26-mc-00318-LS Document 6-23 Filed 08/18/26 Page 2 of 2\n\n\n\n\nSO ORDERED: this _____ day of _____________________, 2026.\n\n\n\n\n United States District Judge\n\n\n\n\n 2\n\f","ocr_status":1,"date_upload":"2026-08-20T15:40:16.180401-07:00","document_number":"6","attachment_number":23,"pacer_doc_id":"181037220290","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Proposed Order","acms_document_guid":""}],"date_created":"2026-08-18T18:05:13.956434-07:00","date_modified":"2026-08-19T09:12:24.289593-07:00","date_filed":"2026-08-18","time_filed":"19:44:42","entry_number":6,"recap_sequence_number":"2026-08-18.006","pacer_sequence_number":25,"description":"CORRECTED MOTION to Compel Compliance With Subpoena Served on Third Party Tesla, Inc. by Neural AI, LLC. (Attachments: # 1 Affidavit Declaration of Tanner Laiche, # 2 Exhibit 1, # 3 Exhibit 2, # 4 Exhibit 3, # 5 Exhibit 4, # 6 Exhibit 5, # 7 Exhibit 6, # 8 Exhibit 7, # 9 Exhibit 8, # 10 Exhibit 9, # 11 Exhibit 10, # 12 Exhibit 11, # 13 Exhibit 12, # 14 Exhibit 13, # 15 Exhibit 14, # 16 Exhibit 15, # 17 Exhibit 16, # 18 Exhibit 17, # 19 Exhibit 18, # 20 Exhibit 19, # 21 Exhibit 20, # 22 Exhibit 21, # 23 Proposed Order)(Magni, Rocco) (Entered: 08/18/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474908234/","id":474908234,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490423344/","id":490423344,"tags":[],"absolute_url":"","date_created":"2026-08-18T12:42:49.118945-07:00","date_modified":"2026-08-18T12:42:49.118959-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"","attachment_number":null,"pacer_doc_id":"","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"","acms_document_guid":""}],"date_created":"2026-08-18T12:42:49.094724-07:00","date_modified":"2026-08-18T12:42:49.094744-07:00","date_filed":"2026-08-18","time_filed":null,"entry_number":null,"recap_sequence_number":"2026-08-18.005","pacer_sequence_number":null,"description":"DEFICIENCY NOTICE: re 1 MOTION to Compel Compliance with Subpoena Served on Third-Party Tesla, Inc. (Reason for deficiency, i.e. Header does not read Midland/Odessa Division). Please correct and refile. In the docket text, type Corrected Motion to Compel. This deficient document will not be forwarded to the assigned judge forconsideration. (ktm)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474900565/","id":474900565,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490415193/","id":490415193,"tags":[],"absolute_url":"","date_created":"2026-08-18T12:17:11.792896-07:00","date_modified":"2026-08-18T12:17:11.792915-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"","attachment_number":null,"pacer_doc_id":"","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Deficiency Notice","acms_document_guid":""}],"date_created":"2026-08-18T12:17:11.779640-07:00","date_modified":"2026-08-18T12:17:11.779656-07:00","date_filed":"2026-08-18","time_filed":"14:13:58","entry_number":null,"recap_sequence_number":"2026-08-18.001","pacer_sequence_number":null,"description":"","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474887156/","id":474887156,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490401458/","id":490401458,"tags":[],"absolute_url":"/docket/74659430/3/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-18T11:06:19.405294-07:00","date_modified":"2026-08-22T22:40:54.692970-07:00","sha1":"605e9dd7097194aeaafb7d24e26a1889a431bb06","page_count":1,"file_size":230681,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.3.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.3.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-DC Document 3 Filed 08/18/26 Page 1 of 1\n\n\n\n\n UNITED STATES DISTRICT COURT\nPHILIP J. DEVLIN WESTERN DISTRICT OF TEXAS ANNETTE FRENCH\n CLERK OF COURT 200 E. Wall Street, Suite 222 CHIEF DEPUTY\n Midland, TX 79701\n August 18, 2026\n\n Samuel Drezdzon\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana, Suite 5100\n Houston, TX 77002\n\n Re: 7:26-MC-00318 Neural AI, LLC v. Tesla Inc.\n\n Dear Counsel:\n\n Our records indicate that you are not admitted to practice in the Western District of Texas. Local\n Court Rule AT-1(f)(1) states:\n\n In General: An attorney who is licensed by the highest court of a state or another\n federal district court, but who is not admitted to practice before this court, may\n represent a party in this court pro hac vice only by permission of the judge\n presiding. Unless excused by the judge presiding, an attorney is ordinarily required\n to apply for admission to the bar of this court.\n\n Pursuant to Local Court Rule AT-2, if you are an attorney residing outside the Western District\n of Texas, the court may require you to designate local counsel in this case.\n\n Please submit a Motion to Appear Pro Hac Vice requesting the Court\u2019s permission to appear in\n the above-captioned cause, along with a $100 filing fee. Applications for permanent admission\n to the bar and Local Court Rules for the Western District of Texas are available on our website at\n www.txwd.uscourts.gov.\n\n Please be advised that the Clerk no longer automatically adds non-admitted attorneys to a\n case. Thus, non-admitted attorneys will not receive orders or notices (which may include\n deadlines, hearing dates, etc.) filed in their case. The Clerk will only add said attorneys to\n the case upon the granting of a Motion to Appear pro hac vice.\n\n Sincerely,\n\n KM\n Deputy Clerk\n\n KM\n\f","ocr_status":2,"date_upload":"2026-08-18T12:43:02.536070-07:00","document_number":"3","attachment_number":null,"pacer_doc_id":"181037214991","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Pro Hac Vice Letter","acms_document_guid":""}],"date_created":"2026-08-18T11:06:19.386607-07:00","date_modified":"2026-08-19T09:12:24.220935-07:00","date_filed":"2026-08-18","time_filed":"12:18:00","entry_number":3,"recap_sequence_number":"2026-08-18.002","pacer_sequence_number":16,"description":"Pro Hac Vice Letter for Attorney Samuel Drezdzon. (ktm) (Entered: 08/18/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474887125/","id":474887125,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490401427/","id":490401427,"tags":[],"absolute_url":"/docket/74659430/4/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-18T11:06:17.271784-07:00","date_modified":"2026-08-22T22:43:05.518120-07:00","sha1":"a1d1a90da135e77a04a1c2b61c39ae691e652ce4","page_count":1,"file_size":213054,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.4.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.4.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-DC Document 4 Filed 08/18/26 Page 1 of 1\n\n\n\n\n UNITED STATES DISTRICT COURT\nPHILIP J. DEVLIN WESTERN DISTRICT OF TEXAS ANNETTE FRENCH\n CLERK OF COURT 200 E. Wall Street, Suite 222 CHIEF DEPUTY\n Midland, TX 79701\n August 18, 2026\n\n Tamar Lusztig\n One Manhattan West, 50th Floor\n New Your, NY 10001\n\n Re: 7:26-MC-00318 Neural AI, LLC v. Tesla Inc.\n\n Dear Counsel:\n\n Our records indicate that you are not admitted to practice in the Western District of Texas. Local\n Court Rule AT-1(f)(1) states:\n\n In General: An attorney who is licensed by the highest court of a state or another\n federal district court, but who is not admitted to practice before this court, may\n represent a party in this court pro hac vice only by permission of the judge\n presiding. Unless excused by the judge presiding, an attorney is ordinarily required\n to apply for admission to the bar of this court.\n\n Pursuant to Local Court Rule AT-2, if you are an attorney residing outside the Western District\n of Texas, the court may require you to designate local counsel in this case.\n\n Please submit a Motion to Appear Pro Hac Vice requesting the Court\u2019s permission to appear in\n the above-captioned cause, along with a $100 filing fee. Applications for permanent admission\n to the bar and Local Court Rules for the Western District of Texas are available on our website at\n www.txwd.uscourts.gov.\n\n Please be advised that the Clerk no longer automatically adds non-admitted attorneys to a\n case. Thus, non-admitted attorneys will not receive orders or notices (which may include\n deadlines, hearing dates, etc.) filed in their case. The Clerk will only add said attorneys to\n the case upon the granting of a Motion to Appear pro hac vice.\n\n Sincerely,\n\n KM\n Deputy Clerk\n\n KM\n\f","ocr_status":2,"date_upload":"2026-08-18T12:43:01.459973-07:00","document_number":"4","attachment_number":null,"pacer_doc_id":"181037215045","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Pro Hac Vice Letter","acms_document_guid":""}],"date_created":"2026-08-18T11:06:17.253399-07:00","date_modified":"2026-08-19T09:12:24.257151-07:00","date_filed":"2026-08-18","time_filed":"12:21:52","entry_number":4,"recap_sequence_number":"2026-08-18.003","pacer_sequence_number":18,"description":"Pro Hac Vice Letter for Attorney Tamar Lusztig. (ktm) (Entered: 08/18/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474887113/","id":474887113,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490401415/","id":490401415,"tags":[],"absolute_url":"/docket/74659430/5/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-18T11:06:16.436163-07:00","date_modified":"2026-08-22T22:50:18.478520-07:00","sha1":"fe98ee49e57fcad39adfbccb7f90e0b3226939d8","page_count":1,"file_size":213114,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.5.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.5.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-mc-00318-DC Document 5 Filed 08/18/26 Page 1 of 1\n\n\n\n\n UNITED STATES DISTRICT COURT\nPHILIP J. DEVLIN WESTERN DISTRICT OF TEXAS ANNETTE FRENCH\n CLERK OF COURT 200 E. Wall Street, Suite 222 CHIEF DEPUTY\n Midland, TX 79701\n August 18, 2026\n\n Emily Portuguese\n One Manhattan West, 50th Floor\n New Your, NY 10001\n\n Re: 7:26-MC-00318 Neural AI, LLC v. Tesla Inc.\n\n Dear Counsel:\n\n Our records indicate that you are not admitted to practice in the Western District of Texas. Local\n Court Rule AT-1(f)(1) states:\n\n In General: An attorney who is licensed by the highest court of a state or another\n federal district court, but who is not admitted to practice before this court, may\n represent a party in this court pro hac vice only by permission of the judge\n presiding. Unless excused by the judge presiding, an attorney is ordinarily required\n to apply for admission to the bar of this court.\n\n Pursuant to Local Court Rule AT-2, if you are an attorney residing outside the Western District\n of Texas, the court may require you to designate local counsel in this case.\n\n Please submit a Motion to Appear Pro Hac Vice requesting the Court\u2019s permission to appear in\n the above-captioned cause, along with a $100 filing fee. Applications for permanent admission\n to the bar and Local Court Rules for the Western District of Texas are available on our website at\n www.txwd.uscourts.gov.\n\n Please be advised that the Clerk no longer automatically adds non-admitted attorneys to a\n case. Thus, non-admitted attorneys will not receive orders or notices (which may include\n deadlines, hearing dates, etc.) filed in their case. The Clerk will only add said attorneys to\n the case upon the granting of a Motion to Appear pro hac vice.\n\n Sincerely,\n\n KM\n Deputy Clerk\n\n KM\n\f","ocr_status":2,"date_upload":"2026-08-18T12:43:00.349507-07:00","document_number":"5","attachment_number":null,"pacer_doc_id":"181037215064","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Pro Hac Vice Letter","acms_document_guid":""}],"date_created":"2026-08-18T11:06:16.415747-07:00","date_modified":"2026-08-19T09:12:24.264708-07:00","date_filed":"2026-08-18","time_filed":"12:25:06","entry_number":5,"recap_sequence_number":"2026-08-18.004","pacer_sequence_number":20,"description":"Pro Hac Vice Letter for Attorney Emily Portuguese. (ktm) (Entered: 08/18/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474877224/","id":474877224,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490391391/","id":490391391,"tags":[],"absolute_url":"","date_created":"2026-08-18T10:18:50.856936-07:00","date_modified":"2026-08-18T10:18:50.856945-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"","attachment_number":null,"pacer_doc_id":"","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Case Assigned/Reassigned","acms_document_guid":""}],"date_created":"2026-08-18T10:18:50.847850-07:00","date_modified":"2026-08-18T10:18:50.847858-07:00","date_filed":"2026-08-18","time_filed":"12:02:05","entry_number":null,"recap_sequence_number":"2026-08-18.001","pacer_sequence_number":null,"description":"","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474863506/","id":474863506,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490377456/","id":490377456,"tags":[],"absolute_url":"/docket/74659430/2/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-18T09:13:39.028415-07:00","date_modified":"2026-08-18T09:13:39.028448-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"2","attachment_number":null,"pacer_doc_id":"","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"","acms_document_guid":""}],"date_created":"2026-08-18T09:13:38.984171-07:00","date_modified":"2026-08-18T09:13:38.993090-07:00","date_filed":"2026-08-18","time_filed":null,"entry_number":2,"recap_sequence_number":"2026-08-18.001","pacer_sequence_number":null,"description":"Miscellaneous Filing fee received in the amount of $52, receipt number ATXWDC-22491522. (Magni, Rocco) (Entered: 08/18/2026)","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474852757/","id":474852757,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490366146/","id":490366146,"tags":[],"absolute_url":"","date_created":"2026-08-18T08:19:55.079524-07:00","date_modified":"2026-08-18T08:19:55.079544-07:00","sha1":"","page_count":null,"file_size":null,"filepath_local":null,"filepath_ia":"","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"","ocr_status":null,"date_upload":null,"document_number":"","attachment_number":null,"pacer_doc_id":"","is_available":false,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Payment of Miscellaneous Filing Fee","acms_document_guid":""}],"date_created":"2026-08-18T08:19:55.061503-07:00","date_modified":"2026-08-18T08:19:55.061517-07:00","date_filed":"2026-08-18","time_filed":"10:10:07","entry_number":null,"recap_sequence_number":"2026-08-18.001","pacer_sequence_number":null,"description":"","tags":[]},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/docket-entries/474781011/","id":474781011,"docket":"https://www.courtlistener.com/api/rest/v4/dockets/74659430/","recap_documents":[{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490292003/","id":490292003,"tags":[],"absolute_url":"/docket/74659430/1/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:05:48.260500-07:00","date_modified":"2026-08-21T19:46:10.028613-07:00","sha1":"7797cd91bd53919dca67bc54cbc040da9e5a477c","page_count":15,"file_size":673086,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.0.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.0.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 1 of 15\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n\nNEURAL AI, LLC,\n Misc. Case No. 7:26-318\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\nTESLA, INC.,\n\n Respondent.\n\n\n NEURAL AI\u2019S MOTION TO COMPEL COMPLIANCE WITH\n SUBPOENA SERVED ON THIRD-PARTY TESLA, INC.\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 2 of 15\n\n\n\n\n TABLE OF CONTENTS\n\nI. FACTUAL BACKGROUND ............................................................................................. 2\n\n A. The Underlying Litigation ...................................................................................... 2\n\n B. The Rule 45 Subpoena to Tesla and Tesla\u2019s Initial Objections .............................. 3\n\n C. NAI\u2019s Meet-and-Confer Efforts and Tesla\u2019s Continued Non-Compliance ............ 4\n\nII. THE COURT HAS JURISDICTION BECAUSE THE PLACE OF COMPLIANCE IN\n AUSTIN, TEXAS IS PROPER. ......................................................................................... 5\n\nIII. TESLA MUST PRODUCE DOCUMENTS RESPONSIVE TO THE SUBPOENA. ....... 6\n\n A. The Requested Discovery Goes to the Heart of NAI\u2019s Infringement Claims......... 7\n\n B. Tesla Has Not Substantiated Burden and NAI Offered Narrower Alternatives. .... 9\n\nIV. CONCLUSION ................................................................................................................. 10\n\n\n\n\n i\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 3 of 15\n\n\n\n\n TABLE OF AUTHORITIES\n\n Page(s)\n\nCases\n\n611 Carpenter LLC v. Atlantic Casualty Ins. Co.,\n 2024 WL 1977160 (W.D. Tex. April 30, 2024) ....................................................................6, 9\n\nConservation L. Found., Inc. v. Equilon Enters. LLC,\n 2025 WL 2821238 (D.R.I. Oct. 3, 2025) ...................................................................................5\n\nFG SRC LLC v. Xilinx, Inc.,\n 2022 WL 22997130 (D. Del. Apr. 11, 2022) .............................................................................8\n\nFlores v. Lowes Home Centers, L.L.C.,\n 2023 WL 3959399 (W.D. Tex. June 12, 2023) .........................................................................6\n\nKim v. NuVasive, Inc.,\n 2011 WL 3844106 (S.D. Cal. Aug. 29, 2011) ...........................................................................8\n\nLucent Technologies, Inc. v. Gateway, Inc.,\n 580 F.3d 1301 (Fed. Cir. 2009)..................................................................................................8\n\nMeritage Homes, LLC v. AIG Specialty Ins. Co.,\n 2024 WL 221448 (W.D. Tex. Jan. 18, 2024) ............................................................................5\n\nNeural AI, LLC v. NVIDIA Corp.,\n Case No. 7:24-cv-00221-LS-DTG (W.D. Tex.) ................................................................1, 2, 3\n\nVelocity Pat. LLC v. FCA US LLC,\n 2017 WL 11893112 (N.D. Ill. Nov. 2, 2017) ............................................................................5\n\nWaller v. Jet Specialty, Inc.,\n 2024 WL 7050192 (W.D. Tex. Nov. 19, 2024) .........................................................................6\n\nRules\n\nFed. R. Civ. P. 26 .............................................................................................................................6\n\nFed. R. Civ. P. 26(b)(1)....................................................................................................................6\n\nFed. R. Civ. P. 45 .................................................................................................................1, 3, 5, 6\n\nFed. R. Civ. P. 45(c)(2)(A) ..........................................................................................................5, 6\n\nFed. R. Civ. P. 45(d)(2)(B)(i) ......................................................................................................5, 6\n\n\n\n\n ii\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 4 of 15\n\n\n\n\n This motion arises from Neural AI, LLC v. NVIDIA Corp., Case No. 7:24-cv-00221-LS-\n\nDTG (W.D. Tex.) (the \u201cUnderlying Action\u201d), a patent-infringement action pending in this District.\n\nPlaintiff Neural AI, LLC (\u201cNAI\u201d) alleges that Defendant NVIDIA Corporation (\u201cNVIDIA\u201d)\n\ninfringes through its GPU-accelerated hardware and software: U.S. Patent No. 8,648,867 (the\n\n\u201c\u2019867 Patent\u201d), Reissue Patent No. RE48,438 (the \u201c\u2019438 Patent\u201d), and Reissue Patent No.\n\nRE49,461 (the \u201c\u2019461 Patent\u201d) (collectively, the \u201cAsserted Patents\u201d). See Exhibits 1\u20134.\n\n NVIDIA has made customer deployment central to the Underlying Action. It disputes\n\nwhether customers configure and use the accused products as NAI alleges and contends that NAI\n\nmust obtain customer evidence to prove indirect infringement and its extent. Tesla, Inc. (\u201cTesla\u201d),\n\na major NVIDIA customer and end user, therefore possesses evidence NVIDIA says NAI must\n\nobtain. See Exhibits 8, 13\u201316; Laiche Decl. \u00b6\u00b6 24\u201325.\n\n NAI served Tesla on June 25, 2026, with Rule 45 subpoenas for documents and corporate\n\ntestimony. The subpoenas seek targeted information about how Tesla configures, integrates, and\n\nuses NVIDIA GPUs and software. Tesla objected to every request and topic, produced no\n\ndocuments, and designated no witness. See Exhibits 5\u20136, 9.\n\n NAI met and conferred repeatedly, identified implicated Tesla systems, supplied focused\n\ntechnical questions, and offered a revised declaration in lieu of broader discovery. See Exhibits\n\n10\u201312. Tesla instead sent a five-paragraph declaration on August 10 identifying six GPU models\n\nand partial software, while omitting the core deployment facts. See Exhibit 20. Tesla then\n\nunilaterally declared the matter concluded. NAI identified the material gaps on August 11, but\n\nTesla did not respond. See Exhibit 21.\n\n NAI therefore respectfully requests an order compelling Tesla to produce documents\n\nresponsive to Requests 1\u201312 on a rolling basis by dates certain and to designate and produce a\n\n\n\n 1\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 5 of 15\n\n\n\n\nknowledgeable witness on Deposition Topics 1\u20135. Tesla\u2019s unilateral declaration does not moot\n\nthat relief, and only an order can secure the discovery Tesla has refused to provide.\n\nI. FACTUAL BACKGROUND\n\n A. The Underlying Litigation\n\n The Underlying Action involves direct, indirect, and induced infringement claims\n\nconcerning claims 16\u201319 of the \u2019867 Patent; claims 1, 3\u20136, 8\u20139, 12, 14, 17\u201318, 21\u201323, 29\u201330, 32,\n\n40, 43\u201344, 46, 48, 51\u201352, and 55\u201357 of the \u2019438 Patent; and claims 21\u201325 and 27\u201330 of the \u2019461\n\nPatent (collectively, the \u201cAsserted Claims\u201d). See Exhibit 1. The Asserted Patents teach GPU-\n\naccelerated computing systems and methods. See Exhibits 2\u20134.\n\n NAI\u2019s January 20, 2026 Final Infringement Contentions and the Subpoenas identify\n\nintegrated combinations of NVIDIA hardware\u2014including GPUs, super computers and servers\u2014\n\nand GPU-accelerated software, including CUDA, cuDNN, TensorRT, and higher-level\n\nframeworks, as the Accused Products. See Exhibit 7. NAI alleges NVIDIA encourages customers\n\nto combine, configure, and use those products in an infringing manner. See id. at 12\u201313.\n\n NVIDIA disputes that selling the Accused Products proves customers configure and deploy\n\nthem as NAI alleges. Although NVIDIA admits it sells the products, partners with resellers,\n\nmanaged-service providers, and data-center providers, and supports customers and partners, it\n\ndenies infringement and disclaims knowledge of customer use. See Exhibit 8 \u00b6\u00b6 18, 20\u201323, 87,\n\n131, 171. NVIDIA has stated throughout discovery that NAI must seek evidence of actual\n\ndeployment from customers and end users. Laiche Decl. \u00b6 25. Customer evidence is therefore\n\ndirectly relevant to NAI\u2019s infringement claims.\n\n Tesla is a major NVIDIA customer and end user. It unveiled a supercomputer with 5,760\n\nNVIDIA A100 GPUs in 2021, expanded it to 7,360 NVIDIA A100 GPUs in 2022, and later\n\n\n\n\n 2\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 6 of 15\n\n\n\n\ndeployed Cortex, a training cluster of approximately 50,000 H100 GPUs at Gigafactory Texas.\n\nSee Exhibits 13\u201315. Tesla uses these systems to train neural networks for Full Self-Driving,\n\nAutopilot, and other AI applications. See Exhibits 13\u201316. Tesla\u2019s records concerning those\n\nsystems\u2019 configuration, deployment, and use bear directly on whether the accused products are\n\nused in an infringing manner and whether NVIDIA induced or contributed to that use.\n\n NAI has subpoenaed multiple NVIDIA customers for evidence uniquely within their\n\npossession. Document discovery in the Underlying Action closed August 11, 2026, and deposition\n\ndiscovery closes September 16, 2026. See Neural AI, LLC v. NVIDIA Corp., Case No. 7:24-cv-\n\n00221-LS-DTG, Dkt. 181 (W.D. Tex.). NVIDIA stipulated that motions to compel third parties\n\nrelated to document discovery could be filed by August 18, 2026. Laiche Decl. \u00b6 28.\n\n B. The Rule 45 Subpoena to Tesla and Tesla\u2019s Initial Objections\n\n NAI served Tesla on June 25, 2026, with Rule 45 subpoenas for documents and corporate\n\ntestimony, designating Austin as the place of compliance. See Exhibits 5\u20136. Tesla\u2019s global\n\nheadquarters and principal place of business are in Austin; its Gigafactory Texas occupies over 10\n\nmillion square feet and employs approximately 20,000 people. See Exhibits 17\u201319.\n\n The document subpoena contains twelve targeted requests, and the accompanying\n\ndeposition subpoena contains five topics, concerning Tesla\u2019s configuration, integration, and use of\n\nNVIDIA GPUs and related software. See Exhibit 5. Both are limited to the relevant damages\n\nperiod\u2014September 13, 2018 to the present\u2014and to U.S.-based activity or activity supporting or\n\ndirected toward U.S. operations. Id.\n\n Tesla\u2019s responses were due July 14, 2026. At Tesla\u2019s request, NAI extended the deadline\n\nto July 21. See Exhibit 10. Tesla then objected to every document request and deposition topic and\n\ndeclined to produce documents or designate a witness. For each request and topic, Tesla said it\n\n\n\n\n 3\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 7 of 15\n\n\n\n\nwas merely \u201cwilling to meet and confer regarding the scope of this Request and the burden it\n\nimposes on Tesla\u201d. See Exhibit 9.\n\n C. NAI\u2019s Meet-and-Confer Efforts and Tesla\u2019s Continued Non-Compliance\n\n The parties first met and conferred on July 28. NAI explained the targeted discovery,\n\nidentified implicated Tesla systems, and discussed categories likely to satisfy the requests. Tesla\n\nsaid it was still investigating, needed technical personnel, and objected to the Subpoenas\u2019 breadth\n\nand requests for confidential technical information.\n\n On August 4, NAI supplied focused technical questions designed to identify Tesla\u2019s actual\n\nNVIDIA products, configurations, data paths, memory use, runtime compilation and TensorRT\n\nuse, frequency, scale, and U.S. nexus, and to narrow collection. See Exhibit 11.\n\n NAI also supplied a draft declaration as an alternative to broader document production and\n\ndeposition testimony. It invited a knowledgeable Tesla declarant to revise the draft after a\n\nreasonable investigation and address the NVIDIA products and software Tesla uses, operation as\n\ndesigned, CPU/GPU memory, input and output paths, pretrained models, custom code or\n\nconfigurations, and frequency. NAI said it would consider an executed declaration in lieu of further\n\ndiscovery, subject to resolving material gaps. See Exhibits 10, 12.\n\n The parties met and conferred again on August 7, 2026. Tesla responded that its\n\ninvestigation remained ongoing, that it was not prepared to provide substantive answers, and that\n\nit could not commit to providing a declaration with sufficient specificity to address the relevant\n\nissues. See Exhibit 10; Laiche Decl. \u00b6\u00b6 26\u201327.\n\n On August 10, Tesla sent NAI a five-paragraph declaration signed by Alon Daks, Tesla\u2019s\n\nSenior Staff Software Engineer, without giving NAI an opportunity to review it before submission.\n\nTesla\u2019s counsel stated that \u201cTesla considers this matter concluded.\u201d See Exhibits 20\u201321. The\n\n\n\n\n 4\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 8 of 15\n\n\n\n\ndeclaration, however, was materially deficient. It merely identifies six NVIDIA GPU models and\n\na partial software list, but omits unmodified use of NVIDIA software, whether the hardware and\n\nsoftware function as designed, NVIDIA-distributed pretrained models, CPU/GPU memory\n\narchitecture, and the standard data path including GPUDirect. Exhibit 20. It also collapses the\n\nsoftware categories and misidentifies Triton, without confirming whether PyTorch or Triton are\n\nused on NVIDIA GPUs. Id.\n\n On August 11, NAI advised Tesla that the declaration was materially insufficient,\n\nidentified these gaps and other deficiencies, and requested another meet and confer. See Exhibit\n\n21. Tesla did not respond, supplement the declaration, produce documents, designate a witness, or\n\npropose a conference.\n\n After explaining the requests, narrowing the issues, and offering alternatives, NAI now\n\nseeks court intervention.\n\nII. THE COURT HAS JURISDICTION BECAUSE THE PLACE OF COMPLIANCE\n IN AUSTIN, TEXAS IS PROPER.\n\n Rule 45 directs a party seeking to compel compliance to apply to \u201cthe court for the district\n\nwhere compliance is required.\u201d Fed. R. Civ. P. 45(d)(2)(B)(i); see also Meritage Homes, LLC v.\n\nAIG Specialty Ins. Co., 2024 WL 221448, at *4 (W.D. Tex. Jan. 18, 2024). The Subpoenas\n\ndesignate Planet Depos, Downtown Austin, 100 Congress Ave., Ste. 2000, Austin, TX 78701, as\n\nthe place of compliance.\n\n Rule 45(c)(2)(A) permits document production at a place within 100 miles of where the\n\nperson resides, is employed, or regularly transacts business in person. The rule does not limit the\n\nplace of compliance to an entity\u2019s headquarters or the location of particular documents or custodians.\n\nSee, e.g., Conservation L. Found., Inc. v. Equilon Enters. LLC, 2025 WL 2821238, at *1 (D.R.I. Oct.\n\n3, 2025); Velocity Pat. LLC v. FCA US LLC, 2017 WL 11893112, at *4 (N.D. Ill. Nov. 2, 2017).\n\n\n\n 5\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 9 of 15\n\n\n\n\n Austin satisfies Rule 45(c)(2)(A). Tesla\u2019s global headquarters is at 1 Tesla Road, Austin,\n\nTexas 78725; its Gigafactory Texas covers 2,500 acres and more than 10 million square feet; and\n\nTesla employs approximately 20,000 people there. See Exhibits 18\u201319. The designated place of\n\ncompliance is proper, so this Court has jurisdiction to resolve the motion.\n\nIII. TESLA MUST PRODUCE DOCUMENTS RESPONSIVE TO THE SUBPOENA.\n\n Federal Rule of Civil Procedure 26 provides that a party may obtain discovery regarding\n\nany nonprivileged matter that is relevant to the parties\u2019 claims or defenses and proportional to the\n\nneeds of the case. Fed. R. Civ. P. 26(b)(1). \u201cAt the discovery stage, relevancy is broadly\n\nconstrued.\u201d Flores v. Lowes Home Centers, L.L.C., 2023 WL 3959399, at *2 (W.D. Tex. June 12,\n\n2023). \u201c\u2018[A] request for discovery should be considered relevant if there is any possibility that the\n\ninformation sought may be relevant to the claim or defense of any party.\u2019\u201d Id.\n\n Where, as here, a non-party refuses discovery in response to a validly issued subpoena,\n\nFederal Rule of Civil Procedure 45 provides the Court for the district where compliance is required\n\nwith broad discretion to compel the production of documents and information from third parties.\n\nFed. R. Civ. P. 45(d)(2)(B)(i). Once a party moving to compel discovery establishes that the\n\nmaterials are relevant, \u201c[t]he party opposing discovery bears the burden of stating \u2018with specificity\n\nthe grounds for objecting to the request.\u2019\u201d Waller v. Jet Specialty, Inc., 2024 WL 7050192, at *1\n\n(W.D. Tex. Nov. 19, 2024); see also 611 Carpenter LLC v. Atlantic Casualty Ins. Co., 2024 WL\n\n1977160, at *1 (W.D. Tex. April 30, 2024) (granting party\u2019s motion to compel non-party\n\nsubpoena). The non-party \u201cmust state with specificity the objection and how it relates to the\n\nparticular request being opposed, and not merely that it is overly broad and burdensome.\u201d 611\n\nCarpenter, 2024 WL 1977160, at *1.\n\n\n\n\n 6\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 10 of 15\n\n\n\n\n A. The Requested Discovery Goes to the Heart of NAI\u2019s Infringement Claims.\n\n NVIDIA has made customer deployment a disputed issue. Tesla is a significant customer\n\nand end user. It built an NVIDIA-based supercomputer with 5,760 A100 GPUs, expanded it to 7,360\n\nA100s, and later deployed Cortex with approximately 50,000 H100s and 16,000 H200s. See Exhibits\n\n13\u201315. Tesla uses these systems to train neural networks for Full Self-Driving, Autopilot, and other\n\nAI applications. See Exhibits 13\u201316. Tesla\u2019s records and testimony can show whether and how often\n\nthe accused methods are actually performed.\n\n The requested information is also uniquely within Tesla\u2019s possession. NVIDIA may know\n\nwhat it designed and distributed, but according to NVIDIA, only Tesla knows what it selected,\n\nconfigured, deployed, and actually ran. NVIDIA claims it does not possess Tesla\u2019s internal\n\narchitecture diagrams, environment configurations, profiler traces, or memory-allocation records,\n\nand NVIDIA has repeatedly disclaimed meaningful insight into customer deployments. See Laiche\n\nDecl. \u00b6 25. At a discovery hearing, NVIDIA told the Court that NAI would need to seek real-world\n\ndeployment information from its customers and end users. See Laiche Decl. \u00b6 24. NAI has done so.\n\n The twelve document requests track the accused computation from software selection\n\nthrough input, execution, memory, and output to establish infringement. See Exhibit 5.\n\n \uf0b7 Requests 1\u20134 seek documents sufficient to identify the software, frameworks, libraries,\n APIs, scripts, configurations, and custom code Tesla uses on NVIDIA GPUs; the NVIDIA\n software and sample code it uses; and how Tesla calls, interfaces with, wraps, depends on,\n modifies, or extends that functionality. They show which accused products Tesla deployed\n and whether Tesla used them as provided and intended or materially modified them.\n\n \uf0b7 Requests 5\u20136 seek architecture, design, data-flow, control-flow, and execution-flow\n materials and documents showing whether the relevant computations involve neural\n networks, layers, and outputs used as inputs to later neurons or layers. They show whether\n Tesla\u2019s systems perform the claimed GPU-accelerated operations and how the accused\n hardware and software work together as relevant to the Asserted Claims.\n\n \uf0b7 Requests 7\u201310 address the claimed memory and data paths: pointer, buffer, and\n intermediate-result reuse; GPU-memory allocation or partitioning; movement of inputs\n\n\n 7\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 11 of 15\n\n\n\n\n from CPU, host, storage, sensor, camera, or network sources into GPU memory; and\n storage, transfer, accumulation, reuse, or return of outputs and intermediate results. They\n target the memory-management and data-transfer limitations at the heart of the Asserted\n Claims.\n\n \uf0b7 Requests 11\u201312 seek documents sufficient to show how GPU computations are scheduled,\n ordered, queued, synchronized, parallelized, launched, interrupted, resumed, and executed,\n and how inputs, commands, model or parameter changes, interruptions, or output changes\n affect GPU-queue placement. They target the claimed coordination and control of GPU\n computations.\n\nThese requests thus target the facts relevant to whether Tesla\u2019s systems perform the asserted\n\nmethod steps and whether NVIDIA\u2019s software, instructions, and support cause or encourage that\n\nuse. See Lucent Technologies, Inc. v. Gateway, Inc., 580 F.3d 1301, 1321\u201322, 1333-34 (Fed. Cir.\n\n2009) (stating patentee must show all steps of claimed method were performed to prove indirect\n\ninfringement and that damages \u201cought to be correlated, in some respect, to the extent the infringing\n\nmethod is used by\u201d directly infringing third parties); FG SRC LLC v. Xilinx, Inc., 2022 WL\n\n22997130, at *5 (D. Del. Apr. 11, 2022) (\u201c[I]t is clear that SRC has a need for information from\n\nthird parties demonstrating that the parties practice the asserted claims.\u201d); Kim v. NuVasive, Inc.,\n\n2011 WL 3844106, at *3 (S.D. Cal. Aug. 29, 2011) (\u201cThe Court finds the information sought in\n\nthe subpoenas is relevant to NuVasive's claim of induced infringement and damages therefrom\u201d).\n\n The five deposition topics seek testimony on the same subjects as the document requests\n\nand directs Tesla to designate one or more knowledgeable persons on five topics. See Exhibit 5.\n\n \uf0b7 Topic 1: the NVIDIA software and libraries Tesla uses to perform computations on\n NVIDIA GPUs;\n\n \uf0b7 Topic 2: NVIDIA sample source code Tesla uses, in whole or in part;\n\n \uf0b7 Topic 3: Tesla\u2019s customizations and data inputs that alter how NVIDIA software\n performs computations;\n\n \uf0b7 Topic 4: identification of Tesla software that uses NVIDIA GPUs to perform\n computations; and\n\n\n\n\n 8\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 12 of 15\n\n\n\n\n \uf0b7 Topic 5: how output and intermediate GPU results are stored, referenced by pointers,\n transferred, copied, streamed, written back, returned, accumulated, reused, or otherwise\n made available between GPU memory and CPU, host, system, storage, display,\n network, or other memory locations.\n\n NAI also offered the draft declaration as a less burdensome alternative and said it would\n\nconsider it in lieu of document production and testimony, subject to resolving material gaps. See\n\nExhibits 10, 12. Tesla\u2019s five-paragraph response was materially deficient and omitted core\n\ndeployment facts and therefore did not eliminate the need for discovery. See Exhibits 20\u201321.\n\n The information is central to the issues in dispute, uniquely within Tesla\u2019s possession, and\n\nsought through targeted, time-limited requests and topics. Tesla should be compelled to produce it.\n\n B. Tesla Has Not Substantiated Burden and NAI Offered Narrower Alternatives.\n\n Tesla\u2019s burden objections do not state with specificity the burden associated with any\n\nrequest or topic. See 611 Carpenter, 2024 WL 1977160, at *1. Tesla repeated the same objections\n\nacross all twelve requests and five topics and said only that it was \u201cwilling to meet and confer\n\nregarding the scope of this Request and the burden it imposes on Tesla.\u201d See Exhibit 9. Boilerplate\n\nobjections do not substantiate undue burden.\n\n Tesla also objected that the definition of \u201cNVIDIA GPUs\u201d encompasses multiple\n\narchitectures and \u201chundreds of individual product SKUs.\u201d See Exhibit 9 at 6. That reflects the\n\nscope of NVIDIA\u2019s accused product line, which NAI\u2019s infringement contentions identify as\n\nAccused Products. See Exhibit 7. The Subpoenas seek documents \u201csufficient to show\u201d identified\n\nfacts about products Tesla actually deploys; they do not require Tesla to identify every GPU it has\n\never used. See Exhibit 5.\n\n NAI minimized burden by agreeing to a one-week extension, identifying implicated\n\nproducts and systems, supplying focused technical questions, and offering a draft declaration as\n\nan alternative to broader discovery. See Exhibits 10\u201312. NAI also invited Tesla to identify specific\n\n\n\n 9\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 13 of 15\n\n\n\n\nburdens. Tesla has offered no concrete search, production date, revised declaration, or witness.\n\n Tesla\u2019s August 10 declaration does not satisfy the subpoenas or the compromise NAI\n\noffered. It is only five paragraphs and covers the bare minimum: six NVIDIA GPU models and a\n\npartial list of software and libraries. Exhibit 20. It does not address the core technical issues within\n\nthe Subpoenas, which are unmodified use of NVIDIA software; whether the hardware and\n\nsoftware function as NVIDIA designed; NVIDIA-distributed pretrained models; CPU/GPU\n\nmemory architecture; and the standard CPU-memory-to-GPU-memory data path, including\n\nGPUDirect. See Exhibits 12, 20. Tesla compounded its noncompliance by submitting the\n\ndeclaration unilaterally, without giving NAI an opportunity to review it for completeness. NAI\u2019s\n\noffer was expressly conditional, and stated NAI would consider a declaration in lieu of broader\n\ndiscovery only \u201csubject to resolving any material gaps.\u201d Exhibit 10 at 1. Tesla had no basis to\n\nconvert that conditional compromise into a unilateral declaration that its obligations were satisfied.\n\nSee Exhibits 10, 20\u201321.\n\nIV. CONCLUSION\n\n For the foregoing reasons, NAI requests that the Court enter an order (1) compelling Tesla\n\nto produce nonprivileged documents responsive to Requests for Production Nos. 1\u201312 on a rolling\n\nbasis to begin within 7 days of the Court\u2019s order and to complete production within twenty-one\n\n(21) days of the Court\u2019s order, and (2) compel Tesla to designate and produce a knowledgeable\n\nwitness on Deposition Topics 1\u20135 within thirty (30) days of the Court\u2019s order.\n\n\n\n\n 10\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 14 of 15\n\n\n\n\nDated: August 17, 2026\n\n\n Respectfully submitted,\n\n /s/ Rocco Magni\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Attorneys for Petitioner Neural AI, LLC\n\n\n\n\n 11\n\f Case 7:26-cv-00318 Document 1 Filed 08/17/26 Page 15 of 15\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that on August 17, 2026, a true and correct copy of\n\nthe foregoing document was served on counsel for Tesla, Inc. and all counsel of record in the\n\nunderlying action.\n\n\n /s/ Rocco Magni\n Rocco Magni\n\n\n\n CERTIFICATE OF CONFERENCE\n\n The undersigned certifies that counsel for Neural AI, LLC conferred in good faith with\n\ncounsel for non-party Tesla, Inc. regarding the issues raised in this Motion, including through\n\nZoom conferences on or about July 28 and August 7, 2026, and related email correspondence.\n\nDespite those efforts, the parties were unable to resolve the dispute.\n\n /s/ Rocco Magni\n Rocco Magni\n\n\n\n\n 12\n\f","ocr_status":2,"date_upload":"2026-08-17T14:19:17.023378-07:00","document_number":"1","attachment_number":null,"pacer_doc_id":"181037209210","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":1,"description":"Compel","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294671/","id":490294671,"tags":[],"absolute_url":"/docket/74659430/1/1/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.003854-07:00","date_modified":"2026-08-21T18:43:34.456720-07:00","sha1":"67c88231dc52b33f27ede6c03173b26b0f5fe18e","page_count":5,"file_size":383103,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.1.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.1.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-cv-00318 Document 1-1 Filed 08/17/26 Page 1 of 5\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n\n NEURAL AI, LLC,\n Misc. Case No.\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\n TESLA, INC.,\n\n Respondent.\n\n\n DECLARATION OF TANNER LAICHE IN SUPPORT OF NEURAL AI\u2019S\n MOTION TO COMPEL COMPLIANCE WITH SUBPOENA\n SERVED ON THIRD-PARTY TESLA, INC.\n\n I, Tanner Laiche, declare as follows:\n\n 1. I am an attorney duly licensed to practice in the States of California and Washington\n\nand am admitted to the Western District of Texas. I am an associate at the law firm of Susman\n\nGodfrey LLP, and am a counsel of record for Petitioner Neural AI, LLC (\u201cNAI\u201d) in the above-\n\ncaptioned matter. I make this declaration in support of NAI\u2019s Motion to Compel Compliance with\n\nSubpoena Served on Third-Party Tesla, Inc. (\u201cTesla\u201d). Unless otherwise stated, I have personal\n\nknowledge of the facts set forth herein and could competently testify thereto.\n\n 2. Attached hereto as Exhibit 1 is a true and correct copy of NAI\u2019s First Amended\n\nComplaint for Patent Infringement filed in the Underlying Action on December 12, 2024.\n\n 3. Attached hereto as Exhibit 2 is a true and correct copy of U.S. Patent No.\n\n8,648,867, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n 4. Attached hereto as Exhibit 3 is a true and correct copy of Reissue Patent No.\n\nRE48,438, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n 5. Attached hereto as Exhibit 4 is a true and correct copy of Reissue Patent No.\n\n\n\n\n 1\n\f Case 7:26-cv-00318 Document 1-1 Filed 08/17/26 Page 2 of 5\n\n\n\n\nRE49,461, entitled \u201cGraphic Processor Based Accelerator System and Method.\u201d\n\n 6. Attached hereto as Exhibit 5 is a true and correct copy of NAI\u2019s Subpoenas to\n\nProduce Documents and to Testify at a Deposition served on Tesla, Inc., served on June 25, 2026,\n\nincluding the accompanying definitions, instructions, requests for production, and deposition\n\ntopics.\n\n 7. Attached hereto as Exhibit 6 is a true and correct copy of the Affidavit of Service\n\nconfirming service of the subpoenas on Tesla\u2019s registered agent, CT Corporation System, in\n\nDallas, Texas, on June 25, 2026.\n\n 8. Attached hereto as Exhibit 7 is a true and correct copy of NAI\u2019s January 20, 2026\n\nFinal Infringement Contentions Cover Page filed in the Underlying Action.\n\n 9. Attached hereto as Exhibit 8 is a true and correct copy of NVIDIA\u2019s Amended\n\nAnswer to NAI\u2019s Amended Complaint, filed on November 25, 2025, as docketed at Dkt. 130 in\n\nthe Underlying Action (public, redacted version).\n\n 10. Attached hereto as Exhibit 9 is a true and correct copy of Non-Party Tesla, Inc.\u2019s\n\nObjections and Responses to Plaintiff Neural AI, LLC\u2019s Subpoena, dated July 21, 2026.\n\n 11. Attached hereto as Exhibit 10 is a true and correct copy of the meet-and-confer and\n\nextension correspondence between counsel for NAI and counsel for Tesla regarding NAI\u2019s\n\nsubpoenas.\n\n 12. Attached hereto as Exhibit 11 is a true and correct copy of NAI\u2019s Third-Party\n\nQuestions provided to Tesla on August 4, 2026.\n\n 13. Attached hereto as Exhibit 12 is a true and correct copy of NAI\u2019s Draft Third-Party\n\nDeclaration provided to Tesla on August 4, 2026.\n\n 14. Attached hereto as Exhibit 13 is a true and correct copy of an NVIDIA Blog post\n\n\n\n\n 2\n\f Case 7:26-cv-00318 Document 1-1 Filed 08/17/26 Page 3 of 5\n\n\n\n\ntitled \u201cTesla Unveils Supercomputer Powered by NVIDIA GPUs,\u201d printed from\n\nhttps://blogs.nvidia.com. I obtained this document from NVIDIA\u2019s publicly accessible website on\n\nAugust 17, 2026.\n\n 15. Attached hereto as Exhibit 14 is a true and correct copy of a Tom\u2019s Hardware\n\narticle titled \u201cTesla Brags About In-House Supercomputer, Now With 7,360 A100 GPUs,\u201d printed\n\nfrom https://www.tomshardware.com. I obtained this document from Tom\u2019s Hardware\u2019s publicly\n\naccessible website on August 17, 2026.\n\n 16. Attached hereto as Exhibit 15 is a true and correct copy of a TechCrunch article\n\ntitled \u201cTesla Dojo: The rise and fall of Elon Musk\u2019s AI supercomputer,\u201d printed from\n\nhttps://techcrunch.com. I obtained this document from TechCrunch\u2019s publicly accessible website\n\non August 17, 2026.\n\n 17. Attached hereto as Exhibit 16 is a true and correct copy of a Popular Science article\n\ntitled \u201cWhat we know about Tesla\u2019s supercomputer,\u201d printed from https://www.popsci.com. I\n\nobtained this document from Popular Science\u2019s publicly accessible website on August 17, 2026.\n\n 18. Attached hereto as Exhibit 17 is a true and correct copy of the Tesla Careers\n\nwebpage, printed from https://www.tesla.com/careers. I obtained this document from Tesla\u2019s\n\npublicly accessible website on August 17, 2026.\n\n 19. Attached hereto as Exhibit 18 is a true and correct copy of an article titled \u201cTesla\n\nemploys 20,000 in Austin, could triple amid Cybertruck ramp-up.\u201d I obtained this document from\n\na publicly accessible website on August 17, 2026.\n\n 20. Attached hereto as Exhibit 19 is a true and correct copy of the Giga Texas webpage\n\nfrom Tesla\u2019s website, printed from https://www.tesla.com. I obtained this document from Tesla\u2019s\n\npublicly accessible website on August 17, 2026.\n\n\n\n\n 3\n\f Case 7:26-cv-00318 Document 1-1 Filed 08/17/26 Page 4 of 5\n\n\n\n\n 21. Attached hereto as Exhibit 20 is a true and correct copy of the Declaration of Alon\n\nDaks, a Senior Staff Software Engineer at Tesla, Inc., executed on August 10, 2026, regarding\n\nTesla\u2019s use of NVIDIA GPUs and software identified in NAI\u2019s subpoena.\n\n 22. Attached hereto as Exhibit 21 is a true and correct copy of additional meet-and-\n\nconfer correspondence between counsel for NAI and counsel for Tesla, dated August 10\u201311, 2026,\n\nregarding Tesla\u2019s declaration and the outstanding subpoena issues.\n\n 23. The Underlying Action\u2014Neural AI, LLC v. NVIDIA Corporation, Case No. 7:24-\n\ncv-00221-LS-DTG (W.D. Tex.)\u2014is a patent infringement action in which NAI alleges that\n\nNVIDIA, Corp.\u2019s (\u201cNVIDIA\u201d) GPU-accelerated computing hardware and software infringe U.S.\n\nPatent No. 8,648,867 (the \u201c\u2019867 Patent\u201d), Reissue Patent No. RE48,438 (the \u201c\u2019438 Patent\u201d), and\n\nReissue Patent No. RE49,461 (the \u201c\u2019461 Patent\u201d) (collectively, the \u201cAsserted Patents\u201d).\n\n 24. Throughout discovery in the above-captioned action, NVIDIA has taken the\n\nposition that NAI must obtain evidence from NVIDIA\u2019s customers and end users to prove how the\n\naccused products are actually deployed and configured. During a sealed discovery conference on\n\nSeptember 8, 2025, NVIDIA\u2019s counsel represented to the Court that it only provides tools for end-\n\nuser companies to build their own AI applications on NVIDIA hardware and that if NAI wanted\n\nto learn how the accused software was deployed in real-world systems, NAI would need to seek\n\nthat information directly from NVIDIA\u2019s customers and end users.\n\n 25. Throughout discovery NVIDIA has also disclaimed knowledge concerning how its\n\ncustomers, partners, and end users ultimately deploy, configure, and operate NVIDIA hardware\n\nand software. NAI accordingly turned to third-party discovery, including subpoenas to Tesla, to\n\nobtain the evidence NVIDIA contends NAI must have.\n\n 26. On August 10, 2026, Tesla provided a declaration from Alon Daks, a Senior Staff\n\n\n\n\n 4\n\f Case 7:26-cv-00318 Document 1-1 Filed 08/17/26 Page 5 of 5\n\n\n\n\nSoftware Engineer (Exhibit 20). However, as set forth in Exhibit 21, the declaration is materially\n\ninsufficient because it addresses only GPU identification and partial software identification while\n\nomitting entirely the substantive technical topics set forth in NAI\u2019s draft declaration (draft\n\nparagraphs 7\u201315), including unmodified use of NVIDIA software, hardware/software functioning\n\nas designed by NVIDIA, use of NVIDIA-distributed pretrained models, CPU/GPU memory\n\narchitecture, and the standard data flow.\n\n 27. As of the date of this declaration, Tesla has not: (a) identified any specific document\n\nrequests to which it will respond; (b) agreed to produce any responsive documents; (c) fully\n\nanswered NAI\u2019s technical questions; (d) designated a witness for deposition; or (e) committed to\n\nany date by which it will do any of the foregoing.\n\n 28. Document discovery in the Underlying Action closed on August 11, 2026.\n\nDefendant NVIDIA, Corp. in the underlying action stipulated to extend the deadline for third-party\n\nmotions to compel to August 18, 2026. Deposition discovery closes on September 16, 2026.\n\n I declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct.\n\n Executed on August 17, 2026, in Seattle, Washington.\n\n\n\n\n Tanner Laiche\n\n\n\n\n 5\n\f","ocr_status":2,"date_upload":"2026-08-17T14:36:18.896286-07:00","document_number":"1","attachment_number":1,"pacer_doc_id":"181037209211","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Affidavit of Tanner Laiche","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294672/","id":490294672,"tags":[],"absolute_url":"/docket/74659430/1/2/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.036233-07:00","date_modified":"2026-08-21T19:21:56.463963-07:00","sha1":"ce8e750aa5a6b73ce94db5f914cf8c70c0685c3f","page_count":95,"file_size":2297066,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.2.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.2.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-2 Filed 08/17/26 Page 1 of 95\n\n\n\n\n EXHIBIT\n\n 1\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 2 of 95\n 1 of 94\n\n\n\n\n UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND-ODESSA DIVISION\n\nNEURAL AI, LLC )\n )\n )\n Plaintiff, )\nv. ) Civil Action No. 7:24-cv-00221\n )\nNVIDIA CORPORATION )\n ) JURY TRIAL DEMANDED\n )\n Defendant. )\n )\n\n FIRST AMENDED COMPLAINT FOR PATENT INFRINGEMENT\n\n Neural AI, LLC (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) alleges against Defendant Nvidia Corporation\n\n(\u201cNvidia\u201d or \u201cDefendant\u201d) the following:\n\n 1. This case involves patented technologies that revolutionized, and have become\n\nwidely adopted in, the field of graphical processor unit (\u201cGPU\u201d)-accelerated computing for\n\nartificial intelligence, machine learning, and complex numerical simulations. GPU-accelerated\n\ncomputing powers many of the most advanced and powerful forms of artificial intelligence that\n\nhave exploded over the past decade.\n\n 2. Highly complex numerical simulations, such as the prediction of protein chains,\n\ngenetic sequences and cryptographic sequences, and advanced machine learning techniques such\n\nas deep learning neural networks, require hardware capable of a high degree of parallel processing\n\nfor efficient computation. GPUs, which generally have hundreds to thousands more computational\n\nprocessors or \u201ccores\u201d than central processing units (\u201cCPUs\u201d), are the preferred hardware for\n\nexecuting such simulations and machine learning techniques. Indeed, the parallel operation of\n\nthousands of high-performance GPUs have become a basic necessity for the execution and training\n\n\n\n\n 1\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 3 of 95\n 2 of 94\n\n\n\n\nof complex natural language and image-generation models, such as ChatGPT\u2019s GPT-4 and\n\nSora.AI. (See https://www.fierceelectronics.com/sensors/chatgpt-runs-10k-nvidia-training-gpus-\n\npotential-thousands-more.)\n\n 3. Before Plaintiff\u2019s innovations, the conventional wisdom in the field of GPU-\n\naccelerated computing was that the exchange of intermediate outputs between a GPU and a CPU\n\nwas too computationally expensive. This was so because the GPU, adapted for highly parallel\n\nprocessing tasks (e.g., graphically modeling a physics engine or rendering complex moving\n\nimages), was ill-suited for handling operations better left to the CPU, like interacting with a user\u2019s\n\nmouse and keyboard or sending and receiving simple datasets. Plaintiff\u2019s foundational technology\n\nchanged this by inventing techniques that leveraged the unique advantages of both the CPU and\n\nthe GPU to enable their efficient interplay in hardware-accelerated computing.\n\n 4. Plaintiff\u2019s patented technologies are enshrined in U.S. Patent Nos. 8,648,867 (\u201cthe\n\n\u2019867 Patent\u201d), RE49,461 (\u201cthe \u2019461 Patent\u201d), and RE48,438 (\u201cthe \u2019438 Patent\u201d) (collectively, \u201cthe\n\nAsserted Patents\u201d or \u201cThe GPU-Based Acceleration Patents\u201d).\n\n NATURE OF THE CASE\n\n 5. Plaintiff brings claims under the patent laws of the United States, 35 U.S.C. \u00a7 1, et\n\nseq., for infringement of the Asserted Patents. Defendant has infringed and continues to infringe\n\neach of the Asserted Patents under at least 35 U.S.C. \u00a7\u00a7271(a), 271(b) and 271(c).\n\n THE PARTIES\n\n 6. Plaintiff Neural AI, LLC, is the owner by assignment of each of the Asserted\n\nPatents.\n\n 7. The technology of the Asserted Patents underpins multiple artificial intelligence\n\nand accelerated computing products that incorporate the patented technology, such as Neurala,\n\n\n\n\n 2\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 4 of 95\n 3 of 94\n\n\n\n\nInc.\u2019s Vision Inspection Automation (VIA), Vision AI software, and Brain Builder platform.\n\n 8. Neural AI is a Texas limited liability company and is a registered business in Texas.\n\nNeural AI maintains its principal office in this District, at 510 Austin Avenue, Suite 2554, Waco,\n\nTX 76701.\n\n 9. Defendant Nvidia Corporation is a Delaware corporation with its headquarters and\n\nprincipal place of business in Santa Clara, California. (See https://investor.nvidia.com/financial-\n\ninfo/sec-filings/sec-filings-details/default.aspx?FilingId=17293267, U.S. Securities and Exchange\n\nCommission Form 10-K for Fiscal Year Ended January 28, 2024;\n\nhttps://nvidianews.nvidia.com/multimedia/santa-clara-headquarters.) Defendant Nvidia\n\nCorporation is registered with the Secretary of State to conduct business in Texas. Nvidia has an\n\noffice in this District located in Austin, Texas. (See https://www.nvidia.com/en-us/contac.)\n\n JURISDICTION & VENUE\n\n 10. This action arises under the Patent Laws of the United States, 35 U.S.C. \u00a7 1, et seq.\n\nThe Court has subject matter jurisdiction pursuant to 28 U.S.C. \u00a7\u00a7 1331 and 1338(a).\n\n 11. This Court has personal jurisdiction over Defendant because it regularly conducts\n\nbusiness in the State of Texas and in this District. This business includes operating systems, using\n\nand/or providing computer hardware, software, firmware, and platforms, and/or providing services\n\nand/or engaging in activities in Texas and in this District that infringe one or more claims of the\n\nAsserted Patents, as well as inducing and contributing to the direct infringement of others through\n\nacts in this District.\n\n 12. Nvidia has also, directly and through its extensive network of partnerships,\n\nincluding with local IT service providers, purposefully and voluntarily placed products and/or\n\nprovided services that practice and/or implement the methods, systems, and apparatuses claimed\n\n\n\n\n 3\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 5 of 95\n 4 of 94\n\n\n\n\nin the Asserted Patents into the stream of commerce with the intention and expectation that they\n\nwill be purchased and used by customers in this District, as detailed below. (See\n\nhttps://www.nvidia.com/en-us/about-nvidia/partners/.)\n\n 13. Defendant has also acknowledged that this Court has personal jurisdiction over it\n\nin cases filed against it in this District. (See, e.g., Vantage Micro LLC v. NVIDIA Corporation,\n\nCase No. 6:19-cv-00582-RP, ECF 22 (W.D. Tex., Jan. 4, 2020) (admitting to personal\n\njurisdiction); Ocean Semiconductor LLC v. NVIDIA Corporation, Case No. 6:20-cv-01211-ADA,\n\nECF 14 (W.D. Tex., Mar. 12, 2021) (same).) Defendant has admitted \u201cit is subject to this Court\u2019s\n\ngeneral personal jurisdiction.\u201d (Id.)\n\n 14. Venue is proper in this District pursuant to 28 U.S.C. \u00a7\u00a7 1391(b) and (c) and 28\n\nU.S.C. \u00a7 1400(b) because Defendant Nvidia Corporation has regular and systematic contacts\n\nwithin this District and has committed acts of infringement within this District.\n\n 15. Defendant Nvidia Corporation is a registered business in Texas and has regular and\n\nestablished places of business in this District. Nvidia has an office in this District located at 11001\n\nLakeline Blvd, Suite 100 Bldg. 2, Austin, Texas 78717. (See https://craft.co/nvidia.) Nvidia\u2019s\n\nAustin office has \u201c54,000 SF of new shell office and DVS labs\u201d and \u201c35,000 SF of offices, testing\n\nand software labs.\u201d (See https://kiddgrp.com/project/nvidia-corporation/.)\n\n 16. Defendant Nvidia Corporation has hundreds of employees in this District\u2014\n\nincluding positions in engineering, sales, marketing, and finance. LinkedIn lists approximately 792\n\npersons associated with Nvidia and identified as being located in the Austin or Austin metropolitan\n\narea. (See\n\nhttps://www.linkedin.com/company/nvidia/people/?facetGeoRegion=104472865%2C90000064.)\n\nLinkedIn also lists approximately 1,158 persons associated with Nvidia and identified as being\n\n\n\n\n 4\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 6 of 95\n 5 of 94\n\n\n\n\nlocated in the State of Texas. (See id.)\n\n 17. In addition, Defendant Nvidia Corporation has over 100 jobs posted for the State\n\nof Texas on its affiliated Workday page with approximately 93 of those jobs\u2014the vast majority of\n\nwhich are engineering jobs\u2014listed for Austin, Texas. (See\n\nhttps://nvidia.wd5.myworkdayjobs.com/NVIDIAExternalCareerSite?locations=91336993fab910\n\naf6d702939a7fcc2d9&locations=91336993fab910af6d702b631b94c2de (approximately 111\n\nNvidia job postings for Texas).) These jobs are particularly relevant to the Asserted Patents and\n\nAccused Products, as defined below, because they pertain to artificial intelligence, machine\n\nlearning, deep learning, data centers, accelerated computing, high performance computing\n\n(\u201cHPC\u201d), and related hardware, software, and/or firmware\u2014including Nvidia\u2019s GPUs, CPUs,\n\nsystems-on-a-chip (\u201cSoCs\u201d), platforms, and application programming interfaces.\n\n 18. Nvidia\u2019s operations in this District include client outreach and sales for each of the\n\nAccused Products and related or supporting services. As detailed above, Nvidia has customer-\n\nfacing personnel and operations in this District. Nvidia also provides technical support to partners\n\nand customers for its products in the District.\n\n 19. Nvidia has committed acts of infringement within this District. Nvidia uses the\n\nAccused Products in this District in manners that practice the Asserted Patents, including by testing\n\nthe Accused Products and by using the Accused Products at its offices and premises in this District.\n\n 20. Defendant makes, uses, advertises, offers for sale, and/or sells hardware for\n\naccelerated computing, including GPUs, CPUs, and SoCs; computers for accelerated computing\n\n(e.g., supercomputers, servers, and data centers for high performance computing); and computer\n\nplatform software-as-a-service (\u201cSaaS\u201d) that implements accelerated computing (including the\n\nAccused Products) in the State of Texas and in this District directly and/or through its partnerships\n\n\n\n\n 5\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 7 of 95\n 6 of 94\n\n\n\n\nwith businesses in the State of Texas and in this District. Defendant also provides data center and\n\nHPC services that practice the Asserted Patents in the State of Texas and in this District directly\n\nand/or through its partnerships with businesses in the State of Texas and in this District.\n\n 21. Nvidia sells, offers for sale, advertises, makes, installs, and/or otherwise provides\n\nhardware, software, firmware, and/or computer platforms for accelerated computing and data\n\ncenter and HPC services, including the Accused Products, the use of which infringes the Asserted\n\nPatents in this District and the State of Texas. (See https://www.nvidia.com/en-us/data-\n\ncenter/solutions/accelerated-computing/.) Nvidia performs these acts directly and/or through its\n\npartnerships with other entities. (See id. (\u201cNVIDIA has defined a range of accelerated platforms\n\nthat each consist of hardware systems designed according to the needs of the use case as well as\n\nthe software stack that enables the operation and management of the business applications. These\n\nhardware systems and software are available from NVIDIA and our partners.\u201d).)\n\n 22. Nvidia also uses a network of partners, which comprise re-sellers, managed service\n\nproviders, and product and solution experts, to provide the Accused Products and implementation\n\nservices for the Accused Products to customers in this District. Each of these partners sells, offers\n\nfor sale, installs, and/or implements Nvidia\u2019s accelerated computing hardware, software, and/or\n\ncomputer platform services. (See https://www.nvidia.com/en-us/about-nvidia/partners/.)\n\n 23. Nvidia\u2019s partners include \u201cData Center Provider[s].\u201d (See\n\nhttps://www.nvidia.com/en-us/about-nvidia/partners/.) Nvidia\u2019s Data Center Provider partners\n\n\u201coffer colocation services such as high-density data center facilities, interconnected infrastructure,\n\nand state-of-art cooling technologies for hosting NVIDIA DGX\u2122 servers globally.\u201d (See id.)\n\nNvidia\u2019s Data Center Provider partners in the \u201cNVIDIA DGX-Ready Data Center program, built\n\non the NVIDIA DGX\u2122 platform and delivered by NVIDIA partners,\u201d help \u201caccelerate the scaling\n\n\n\n\n 6\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 8 of 95\n 7 of 94\n\n\n\n\nof AI across [a customer\u2019s] organization.\u201d (See https://www.nvidia.com/en-us/data-\n\ncenter/colocation-partners/#aligned-energy.)\n\n 24. As further detailed below, Nvidia engages in activities that directly infringe the\n\nAsserted Patents within this District. For example, Nvidia\u2019s operation and use of its accelerated\n\ncomputing hardware, software, and/or computer platform services, including its data center-scale\n\naccelerated computing platforms, within this District infringe the Asserted Patents.\n\n 25. Nvidia also infringes (directly or indirectly) the Asserted Patents by providing\n\nservices in connection with the Accused Products including installing, maintaining, supporting,\n\noperating, providing instructions, and/or advertising Nvidia\u2019s computer platform, data center, and\n\nHPC services within this District. For example, under Nvidia\u2019s cloud and data center line of\n\nproducts and services, the Nvidia DGX platform is a \u201ca fully integrated hardware and software AI\n\nplatform\u201d and \u201ccombines the best of NVIDIA software, infrastructure, and expertise in a modern,\n\nunified AI development solution.\u201d (See https://www.nvidia.com/en-us/data-center/dgx-platform/.)\n\nIndeed, \u201cDGX infrastructure is a complete AI solution, and includes NVIDIA AI Enterprise\n\nsoftware to accelerate data science pipelines and streamline development and deployment of\n\nproduction-grade AI applications.\u201d (See id.) Nvidia platform user and partner customers infringe\n\nthe Asserted Patents by installing and operating Nvidia\u2019s computer platform software, which\n\nperforms the claimed methods in the Asserted Patents within this District. (See also, e.g.,\n\nhttps://www.nvidia.com/en-us/data-center/products/ai-enterprise/ (Nvidia AI Enterprise);\n\nhttps://developer.nvidia.com/cuda-zone (Nvidia CUDA Toolkit); https://www.nvidia.com/en-\n\nus/data-center/gpu-cloud-computing/ (GPU Cloud Computing).)\n\n 26. Defendant encourages and induces its customers of the Accused Products to\n\nperform the methods claimed in the Asserted Patents. For example, Nvidia makes its accelerated\n\n\n\n\n 7\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 9 of 95\n 8 of 94\n\n\n\n\ncomputing platforms and services available on its website, widely advertises those platforms and\n\nservices, provides applications that allow partners and users to access those platforms and services,\n\nprovides instructions for installing, and maintaining those platforms and services and supporting\n\nsoftware and/or firmware, and provides technical support to users. (See\n\nhttps://www.nvidia.com/en-us/data-center/dgx-support/.)\n\n 27. Nvidia further encourages and induces its customers to operate Nvidia\u2019s hardware\n\nand software in an infringing manner, and to use Nvidia\u2019s infringing computer platforms, by\n\nproviding directions for and encouraging customers to install software, such as software for\n\nNVIDIA AI Enterprise and CUDA, (see https://docs.nvidia.com/ai-enterprise/deployment-guide-\n\nvmware/0.1.0/software.html; https://developer.nvidia.com/cuda-downloads), which offers\n\nevaluation, installation, configuration, customization, and development of Nvidia\u2019s infringing\n\nsoftware products and services.\n\n 28. Defendant also contributes to the infringement of its customers and end users of the\n\nAccused Products by offering within the United States or importing into the United States the\n\nAccused Products, which are for use in practicing, and under normal operation practice, one or\n\nmore of the methods claimed in the Asserted Patents, constituting a material part of the inventions\n\nclaimed, and not a staple article or commodity of commerce suitable for substantial non-infringing\n\nuses. Indeed, as shown herein, the Accused Products and the example functionality described\n\nbelow have no substantial non-infringing uses and are specifically designed to practice the methods\n\nclaimed in the Asserted Patents.\n\n 29. On information and belief, Defendant has not disputed that venue is proper in this\n\nDistrict in cases filed against it in this District. (See, e.g., Vantage Micro LLC v. NVIDIA Corp.,\n\nNo. 6:19-cv-00582, ECF 22; Polaris Innovations Ltd. v. Dell Inc. et al., No. 5:16-cv-00451, ECF\n\n\n\n\n 8\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 Filed\n 30 08/17/26\n Filed 12/12/24\n Page Page\n 10 of 95of 94\n\n\n\n\n19; Cirrus Logic, Inc. v. ATI Techs., et al., No. 1:03-cv-00302, ECF 6.)\n\n 30. Defendant\u2019s infringement adversely impacts Plaintiff in this District.\n\n PLAINTIFF\u2019S PATENTED INNOVATIONS\n\n 31. The Asserted Patents pioneered the adaptation of GPU-acceleration technology to\n\nthe supervised execution of complex artificial intelligence algorithms and numerical simulations,\n\nsuch that it became possible for the first time to dynamically supervise, review, and correct\n\nintermediate \u201csolutions\u201d that were produced by these accelerated algorithms and simulations\n\nwithout performance loss.\n\n The GPU-Based Acceleration Patents\n U.S. Patent Nos. 8,648,867, RE49,461, and RE48,438\n\n 32. The \u2019867, \u2019461, and \u2019438 Patents are part of the same patent family and generally\n\ndisclose and claim systems and methods related to the accelerated execution of numerical\n\nsimulations and neural networks such that the intermediate outputs of a given execution \u201cstep\u201d can\n\nbe dynamically transferred from the GPU to the CPU, reviewed, and corrected within the same\n\ncomputational cycle before being fed as inputs to the next execution step.\n\n 33. The \u2019867 Patent is entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on September 24, 2007, and was duly and legally issued by the United States\n\nPatent and Trademark Office (\u201cUSPTO\u201d) on February 11, 2014. The \u2019867 Patent claims priority\n\nto Provisional Application No. 60/826,892, filed on September 25, 2006. A true and correct copy\n\nof the \u2019867 Patent is attached as Exhibit 1.\n\n 34. The \u2019438 Patent is entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on November 9, 2017, and was duly and legally issued by the USPTO on\n\nFebruary 16, 2021. The \u2019438 Patent is a re-issue of the \u2019867 Patent and claims priority to\n\nProvisional Application No. 60/826,892, filed on September 25, 2006. A true and correct copy of\n\n\n\n 9\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 11 of10\n 95of 94\n\n\n\n\nthe \u2019438 Patent is attached as Exhibit 2.\n\n 35. The \u2019461 Patent is also entitled \u201cGraphic Processor Based Accelerator System and\n\nMethod,\u201d was filed on December 29, 2020, and was duly and legally issued by the USPTO on\n\nMarch 14, 2023. The \u2019461 Patent is a re-issue of the \u2019867 Patent and claims priority to Provisional\n\nApplication No. 60/826,892, filed on September 25, 2006. A true and correct copy of the \u2019461\n\nPatent is attached as Exhibit 3.\n\n 36. The \u2019867 Patent improves upon prior GPU acceleration technology by disclosing\n\nand claiming a novel hardware and firmware system for performing a numerical simulation that\n\npermits dynamic editing of the outputs that flow from intermediate \u201csteps\u201d of that simulation,\n\nbefore they become inputs to the next \u201cstep.\u201d In particular, the \u2019867 patent discloses a CPU tethered\n\nto a GPU-based accelerator, each with their own corresponding memories, and an accelerator\n\n\u201ccontroller\u201d that coordinates transfers of data between the CPU and the GPU-based accelerator,\n\nsuch that the intermediate results from one step can be transferred from the GPU-based accelerator\n\nto the CPU, reviewed and corrected by the CPU, and transferred back to the GPU-based accelerator\n\nbefore the next computational cycle begins.\n\n 37. The \u2019867 Patent explains that performing the numerical computation in this\n\nstepwise fashion enables the system to eliminate \u201crace conditions,\u201d i.e., conflicts that occur when\n\ntwo programmatic \u201cthreads\u201d attempt to change the same shared data at the same time, which would\n\notherwise occur when other system elements attempt to access intermediate outputs of the\n\nnumerical computation. (See \u2019867 Patent, 5:60-6:31.) This avoids the computational overhead\n\nprevalent in conventional GPU-based accelerator architectures when transferring data from the\n\naccelerator to the CPU.\n\n\n\n\n 10\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 12 of11\n 95of 94\n\n\n\n\n 38. By enabling such \u201ccontroller-driven data exchange\u201d between the GPU-based\n\naccelerator and the CPU, the system described in the \u2019867 Patent allows for an \u201cinput parser\u201d\n\nexecuting on a CPU core to \u201cchange input\u2026on the fly during the simulation,\u201d thus enabling\n\nautomatic review and dynamic error correction of the numerical simulations or neural networks\n\nthat are being executed on the claimed system. (See id., 9:8-17.) Such dynamic, in-execution\n\nreview and error correction of whether each intermediate \u201cstep\u201d of a simulation or neural network\n\nis generating correct results is essential to the performance and reliability of large language models,\n\nimage classification, and image generation models that have become prevalent today. Because of\n\nthe scale to which such simulations and models have grown, it is no longer feasible to \u201crestart\u201d\n\nthem from scratch, only to correct them as they execute.\n\n 39. The \u2019461 and \u2019438 Patents disclose hardware and firmware configurations similar\n\nto those of the \u2019867 Patent, but are directed to using those configurations to process the layers of\n\nan artificial neural network (\u201cANN\u201d). The \u2019461 Patent is directed to further interplay between the\n\nCPU and the GPU-based accelerator: separating the CPU and GPU-based accelerator into separate\n\n\u201cstreams,\u201d whereby the CPU executes a \u201cuser interaction stream\u201d (e.g., enabling the parsing and\n\ndynamic editing of intermediate outputs, or for the ANN to be paused and resumed), while the\n\naccelerator executes a \u201ccomputational stream\u201d that executes the layers of the artificial neural\n\nnetwork. When the ANN is initialized, control over the generation of outputs shifts to the\n\ncomputational stream. However, once a pre-defined layer of the ANN has completed execution,\n\nor is interrupted, control over the generation of outputs and feeding of inputs is shifted back to the\n\nCPU\u2019s user interaction stream.\n\n 40. The Asserted Patents describe this \u201cshift of priorities\u201d as \u201c[t]he crucial feature of\n\nthe interaction between the User Interaction Stream and the Computational Stream.\u201d (\u2019867 Patent,\n\n\n\n\n 11\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 13 of12\n 95of 94\n\n\n\n\n7:45-47.) Even though the computational stream is in control during the ANN computation,\n\npriority shifting enables \u201c[t]he user [to] retain[] the ability to interrupt the simulation, change the\n\ninput, or to change the display properties of the framework\u201d because the user\u2019s \u201cinteractions are\n\nqueued to be performed at times determined by the controller-driven data exchange to avoid\n\ncorruption of the data.\u201d (Id., 8:47-57.)\n\n ACCUSED PRODUCTS\n\n 41. Nvidia offers, sells, and uses several products that provide and implement GPU-\n\nacceleration hardware, software, platforms, and services for individuals and enterprises and\n\nincorporate Plaintiff\u2019s patented technologies. (See https://www.nvidia.com/en-\n\nus/solutions/ai/inference/; https://marketplace.nvidia.com/en-us/data-center/?page=4;\n\nhttps://marketplace.nvidia.com/en-us/laptops-workstations/?page=9;\n\nhttps://marketplace.nvidia.com/en-us/software/?page=3.)\n\n 42. The Accused Products include Nvidia\u2019s GPU accelerators and superchips. (See\n\nhttps://resources.nvidia.com/l/en-us-gpu.) Nvidia\u2019s GPU accelerators include Nvidia\u2019s GPUs with\n\nNvidia\u2019s \u201cHopper,\u201d \u201cAda Lovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d \u201cPascal,\u201d and \u201cMaxwell\u201d\n\nGPU architectures. (See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-\n\nmatrix/index.html.) These GPUs are specifically designed to run and implement GPU-based\n\nhardware acceleration using Nvidia\u2019s proprietary CUDA (Compute Unified Device Architecture)\n\nplatform and CUDA libraries for GPU acceleration. (See id. (Nvidia GPU architectures\n\nimplementing Nvidia\u2019s cuDNN (CUDA Deep Neural Network) library for GPU acceleration.);\n\nhttps://developer.nvidia.com/cuda-gpus.)\n\n 43. Nvidia\u2019s Hopper GPUs include the H100 and H200 GPUs. (See\n\nhttps://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/ (Hopper\n\n\n\n\n 12\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 14 of13\n 95of 94\n\n\n\n\narchitecture); https://www.nvidia.com/en-us/data-center/h100/ (H100);\n\nhttps://www.nvidia.com/en-us/data-center/h200/ (H200).) In addition, Nvidia\u2019s superchips that\n\nimplement GPU accelerators include the GH200, or Grace Hopper Superchip, which implements\n\nthe Hopper-GPU architecture. (See https://www.nvidia.com/en-us/data-center/grace-hopper-\n\nsuperchip/ (GH200).)\n\n 44. Nvidia\u2019s Ada Lovelace (or Lovelace) GPUs include Nvidia Data Center GPUs,\n\nincluding L40, L40S, and L4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including\n\nRTX Ada Generations series GPUs and Laptop GPUs; and GeForce RTX 40 series GPUs and\n\nLaptop GPUs. (See https://www.nvidia.com/en-us/technologies/ada-architecture/ (Ada Lovelace\n\narchitecture). See https://www.nvidia.com/en-us/data-center/l40/ (L40);\n\nhttps://www.nvidia.com/en-us/data-center/l40s/ (L40S); https://www.nvidia.com/en-us/data-\n\ncenter/l4/ (L4). See https://resources.nvidia.com/en-us-design-viz-stories-ep/l40-linecard (Nvidia\n\nProfessional GPUs); https://www.nvidia.com/en-us/ai-on-rtx/ (RTX GPUs featuring \u201cAccelerated\n\nDevelopment\u201d); https://www.nvidia.com/en-us/design-visualization/desktop-graphics/ (RTX Ada\n\nGeneration GPUs); https://www.nvidia.com/en-us/design-visualization/rtx-professional-\n\nlaptops/compare-table/ (RTX Ada Generation Laptop GPUs). See https://www.nvidia.com/en-\n\nus/geforce/graphics-cards/40-series/ (GeForce RTX 40 GPUs); https://www.nvidia.com/en-\n\nus/geforce/graphics-cards/compare/ (GeForce RTX 40 GPUs); https://www.nvidia.com/en-\n\nus/geforce/laptops/compare/ (GeForce RTX 40 Laptop GPUs).)\n\n 45. Nvidia\u2019s Ampere GPUs include Nvidia Data Center GPUs, including A100, A40,\n\nA30, A16, A10, and A2 GPUs; Nvidia Workstation and Professional Laptop GPUs, including\n\nRTX A series GPUs and Laptop GPUs; GeForce RTX 30 series GPUs and Laptop GPUs; and\n\nGeForce MX570 Laptop GPU. (See https://www.nvidia.com/en-us/data-center/ampere-\n\n\n\n\n 13\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 15 of14\n 95of 94\n\n\n\n\narchitecture/ (Ampere architecture). See https://www.nvidia.com/en-us/data-center/a100/ (A100);\n\nhttps://www.nvidia.com/en-us/data-center/a40/ (A40); https://www.nvidia.com/en-us/data-\n\ncenter/a30/ (A30); https://www.nvidia.com/en-us/data-center/a16/ (A16);\n\nhttps://www.nvidia.com/en-us/data-center/a10/ (A10); https://www.nvidia.com/en-us/data-\n\ncenter/a2/ (A2). See https://www.nvidia.com/en-us/design-visualization/desktop-graphics/ (RTX\n\nA GPUs); https://www.nvidia.com/en-us/design-visualization/rtx-professional-laptops/compare-\n\ntable/ (RTX A Laptop GPUs). See https://www.nvidia.com/en-us/geforce/graphics-cards/30-\n\nseries/ (GeForce RTX 30 GPUs); https://www.nvidia.com/en-us/geforce/graphics-cards/compare/\n\n(GeForce RTX 30 GPUs); https://www.nvidia.com/en-us/geforce/laptops/compare/30-series/\n\n(GeForce RTX 30 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n570/ (GeForce MX570 Laptop GPU).)\n\n 46. Nvidia\u2019s Turing GPUs include Nvidia Data Center GPUs, including Tesla T4 GPUs\n\nand Quadro RTX 8000 (passive) and Quadro RTX 6000 (passive) GPUs; Nvidia Workstation and\n\nProfessional Laptop GPUs, including T series GPUs and Laptop GPUs, Quadro T series Laptop\n\nGPUs, and Quadro RTX series GPUs and Laptop GPUs; Titan series Titan RTX GPU;\n\nGeForce RTX 20 series and GeForce GTX 16 series GPUs and Laptop GPUs; and GeForce\n\nMX550, MX450, and MX430 Laptop GPUs. (See https://www.nvidia.com/en-us/geforce/turing/\n\n(Turing architecture). See https://www.nvidia.com/en-us/data-center/tesla-t4/ (Tesla T4);\n\nhttps://www.nvidia.com/en-gb/design-visualization/quadro-data-center/ (Quadro RTX 8000\n\n(passive) and Quadro RTX 6000 (passive). See https://www.nvidia.com/en-us/design-\n\nvisualization/quadro/ (T series GPUs/Laptop GPUs, Quadro T series Laptop GPUs, and Quadro\n\nRTX GPUs/Laptop GPUs); https://www.nvidia.com/en-us/design-visualization/desktop-graphics\n\n(T series GPUs/Laptop GPUs); https://www.nvidia.com/content/dam/en-\n\n\n\n\n 14\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 16 of15\n 95of 94\n\n\n\n\nzz/Solutions/titan/documents/titan-rtx-for-creators-us-nvidia-1011126-r6-web.pdf (Titan RTX);\n\nhttps://www.nvidia.com/en-us/geforce/20-series/ (GeForce RTX 20 GPUs);\n\nhttps://www.nvidia.com/en-us/geforce/graphics-cards/compare/ (GeForce RTX 20 GPUs and\n\nGeForce GTX 16 GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/compare-20-\n\nseries/ (GeForce RTX 20 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-\n\nlaptops/compare-16-series/ (GeForce GTX 16 Laptop GPUs); https://www.nvidia.com/en-\n\nus/geforce/gaming-laptops/mx-550/ (GeForce MX550 Laptop GPU); https://www.nvidia.com/en-\n\nus/geforce/gaming-laptops/mx-450/ (GeForce MX450 Laptop GPU);\n\nhttps://wccftech.com/nvidia-geforce-mx450-turing-discrete-notebook-gpu-gddr6-pcie-4/\n\n(GeForce M Laptop GPUs).)\n\n 47. Nvidia\u2019s Volta GPUs include Nvidia Data Center GPUs, including the Tesla V100\n\nGPU; Nvidia Workstation GPUs, including Quadro GV100; and Titan series Titan V GPU. (See\n\nhttps://www.nvidia.com/en-us/data-center/volta-gpu-architecture/ (Volta architecture);\n\nhttps://www.nvidia.com/en-us/data-center/v100/ (Tesla V100);\n\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\nvisualization/productspage/quadro/quadro-desktop/quadro-volta-gv100-data-sheet-us-nvidia-\n\n704619-r3-web.pdf (Quadro GV100); https://nvidianews.nvidia.com/news/nvidia-titan-v-\n\ntransforms-the-pc-into-ai-supercomputer (Titan V).)\n\n 48. Nvidia\u2019s Pascal GPUs include Nvidia Data Center GPUs, including Tesla P100,\n\nP40, and P4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including the\n\nQuadro GP100 GPU and Quadro P series GPUs and Laptop GPUs; Titan series Titan Xp and Titan\n\nX GPUs; GeForce GTX 10 series GPUs and Laptop GPUs; and GeForce MX300 series, MX200\n\nseries, and MX150 Laptop GPUs. (See https://developer.nvidia.com/pascal;\n\n\n\n\n 15\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 17 of16\n 95of 94\n\n\n\n\nhttps://www.nvidia.com/en-us/data-center/pascal-gpu-architecture/ (Pascal architecture). See\n\nhttps://www.nvidia.com/en-us/data-center/tesla-p100 (Tesla P100);\n\nhttps://developer.nvidia.com/cuda-gpus (Tesla P40 and P4);\n\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\nvisualization/productspage/quadro/quadro-desktop/quadro-pascal-gp100-data-sheet-us-nv-\n\n704562-r1.pdf (Quadro GP100); https://www.nvidia.com/en-us/design-visualization/quadro/\n\n(Quadro P series GPUs/Laptop GPUs). See https://www.nvidia.com/content/geforce-\n\ngtx/NVIDIA_TITAN_X_USER_GUIDE_v02.pdf (Titan X);\n\nhttps://www.nvidia.com/content/geforce-gtx/NVIDIA_TITAN_Xp_USER_GUIDE_v02.pdf\n\n(Titan Xp); https://www.nvidia.com/en-us/geforce/10-series/ (GeForce GTX 10);\n\nhttps://www.nvidia.com/en-us/geforce/graphics-cards/compare/ (GeForce GTX 10 GPUs);\n\nhttps://www.nvidia.com/en-us/geforce/news/gfecnt/nvidia-geforce-gtx-10-series-laptops/\n\n(GeForce GTX 10 Laptop GPUs); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n350/ (GeForce MX350 Laptop GPU); https://www.nvidia.com/en-us/geforce/gaming-laptops/mx-\n\n330/ (GeForce MX330 Laptop GPU); https://wccftech.com/nvidia-geforce-mx450-turing-\n\ndiscrete-notebook-gpu-gddr6-pcie-4/ (GeForce M Laptop GPUs).)\n\n 49. Nvidia\u2019s Maxwell GPUs include Nvidia Data Center GPUs, including Tesla M60,\n\nM40, and M10 GPUs; Nvidia Workstation and Professional Laptop GPUs, including Quadro M\n\nseries GPUs and Laptop GPUs, the NVS 810 GPU, and Tesla M6 series Laptop GPUs; Titan series\n\nGTX Titan X GPU; GeForce GTX 900 series and GeForce GTX 700 series GPUs and Laptop\n\nGPUs; and GeForce MX130 series and MX110 Laptop GPUs. (See\n\nhttps://developer.nvidia.com/blog/maxwell-most-advanced-cuda-gpu-ever-made/ (Maxwell\n\narchitecture); https://www.nvidia.com/content/dam/en-zz/Solutions/design-\n\n\n\n\n 16\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 18 of17\n 95of 94\n\n\n\n\nvisualization/solutions/resources/documents1/nvidia-m60-datasheet.pdf (M60);\n\nhttps://images.nvidia.com/content/tesla/pdf/78071_Tesla_M40_24GB_Print_Datasheet_LR.PDF\n\n(M40); https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-\n\nm10/pdf/188359-Tesla-M10-DS-NV-Aug19-A4-fnl-Web.pdf (M10);\n\nhttps://www.nvidia.com/en-us/design-visualization/quadro/ (Quadro M GPUs/Laptop GPUs);\n\nhttps://www.nvidia.com/docs/IO/146527/nvs-810-datasheet.pdf (NVS 810);\n\nhttps://images.nvidia.com/content/tesla/pdf/188300-Tesla-M6-DS-Aug19-A4-fnl-Web.pdf (Tesla\n\nM6); https://www.nvidia.com/content/geforce-gtx/GTX_TITAN_X_User_Guide.pdf (GTX Titan\n\nX); https://developer.nvidia.com/maxwell-compute-architecture (GeForce GTX 900 and 700\n\nGPUs/Laptop GPUs); https://wccftech.com/nvidia-geforce-mx450-turing-discrete-notebook-gpu-\n\ngddr6-pcie-4/ (GeForce M Laptop GPUs).)\n\n 50. These GPUs and superchips implement, and are specifically designed for, GPU-\n\nacceleration for artificial intelligence and neural networks. Nvidia\u2019s proprietary CUDA platform\n\nfor parallel computing, which includes GPU-acceleration libraries such as cuDNN (CUDA Deep\n\nNeural Network), is implemented in the Nvidia Hopper, Ada Lovelace, Ampere, Turing, Volta,\n\nPascal, and Maxwell GPU architectures.\n\n\n\n\n 17\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 19 of18\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 51. The Accused Products further include Nvidia\u2019s supercomputers and servers that\n\nimplement its GPU accelerators and superchips. These supercomputers and servers include: the\n\nEGX line of servers for data centers and edge devices, the HGX line of supercomputers, the DGX\n\nline of supercomputers, and the OVX line of supercomputers. (See https://www.nvidia.com/en-\n\nus/data-center/solutions/accelerated-computing/.)\n\n 52. Nvidia\u2019s \u201cEGX hardware portfolio\u201d includes \u201caccelerators [that] combine the\n\nperformance of NVIDIA Ampere GPUs.\u201d (See https://www.nvidia.com/en-us/data-\n\ncenter/products/egx/; see https://www.nvidia.com/en-us/design-visualization/egx-graphics/.)\n\nNvidia\u2019s HGX \u201cAI supercomputing platform brings together the full power of NVIDIA GPUs,\n\nNVIDIA NVLink\u2122, NVIDIA networking, and fully optimized AI and high-performance\n\ncomputing (HPC) software stacks.\u201d (See https://www.nvidia.com/en-us/data-center/hgx/;\n\nhttps://nvdam.widen.net/s/5kgbjq2v2t/hpc-hgx-h100-datasheet-nvidia-web.) One example\n\n\n 18\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 20 of19\n 95of 94\n\n\n\n\nconfiguration includes \u201cfour or eight H200 or H100 GPUs.\u201d (See id.; see\n\nhttps://nvdam.widen.net/s/5kgbjq2v2t/hpc-hgx-h100-datasheet-nvidia-web.) Nvidia\u2019s DGX\n\nsupercomputers include the DGX H200, DGX BasePOD, and DGX SuperPOD with DGX GB200.\n\n(See https://www.nvidia.com/en-us/data-center/dgx-platform/; see also\n\nhttps://www.nvidia.com/en-us/data-center/base-command/; https://resources.nvidia.com/en-us-\n\ndgx-software/nvidia-base-command (DGX Base Command operating system for DGX data\n\ncenters.) And Nvidia\u2019s OVX supercomputers implement \u201cL40S GPUs . . . for both complex AI\n\nand graphics-intensive workloads.\u201d (See https://www.nvidia.com/en-us/data-center/products/ovx/;\n\nsee https://resources.nvidia.com/en-us-ovx/ovx-datasheet.)\n\n 53. The Accused Products further include Nvidia\u2019s software, platforms, and services\n\nfor accelerated computing. These include CUDA, Nvidia AI Enterprise, the DGX Platform, Nvidia\n\nOmniverse, Nvidia Drive, Nvidia Isaac Sim, and Nvidia NGC.\n\n 54. CUDA is Nvidia\u2019s proprietary \u201cparallel computing platform and programming\n\nmodel.\u201d (See https://developer.nvidia.com/cuda-zone.) CUDA is designed to support Nvidia\u2019s\n\nGPU accelerators and superchips and includes software specifically for GPU-acceleration such as\n\nthe cuDNN \u201cGPU-accelerated library.\u201d (See id.; https://developer.nvidia.com/cudnn.) In addition,\n\nNvidia\u2019s CUDA-X, built on top of CUDA, is a collection of \u201cGPU-accelerated microservices and\n\nlibraries for AI.\u201d (See https://www.nvidia.com/en-us/technologies/cuda-x/.) Nvidia also offers the\n\nCUDA Toolkit and SDK Manager for developing GPU-accelerated applications. (See\n\nhttps://developer.nvidia.com/cuda-toolkit; https://developer.nvidia.com/sdk-manager.)\n\n 55. In addition, Nvidia AI Enterprise is Nvidia\u2019s \u201cend-to-end, cloud-native software\n\nplatform\u201d for \u201caccelerat[ing] data science pipelines . . . and other generative AI applications.\u201d (See\n\nhttps://www.nvidia.com/en-us/data-center/products/ai-enterprise/.) It is Nvidia\u2019s \u201c\u2018operating\n\n\n\n\n 19\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 21 of20\n 95of 94\n\n\n\n\nsystem\u2019 for enterprise AI.\u201d (See id.)\n\n 56. In addition, Nvidia\u2019s DGX platform is \u201cis a complete AI solution, and includes\n\nNVIDIA AI Enterprise software.\u201d (See https://www.nvidia.com/en-us/data-center/dgx-platform/.)\n\nNvidia DGX Cloud is \u201can AI-training-as-a-service platform which includes cloud-based\n\ninfrastructure and software for AI, customizable pretrained AI models, and access to NVIDIA\n\nexperts.\u201d (See https://d18rn0p25nwr6d.cloudfront.net/CIK-0001045810/1cbe8fe7-e08a-46e3-\n\n8dcc-b429fc06c1a4.pdf, Nvidia U.S. Securities and Exchange Commission Form 10-K for Fiscal\n\nYear Ended January 28, 2024 at 6.)\n\n 57. In addition, Nvidia Omniverse is \u201ca development platform and operating system\n\nfor building virtual world simulation applications, available as a software subscription.\u201d (See\n\nhttps://d18rn0p25nwr6d.cloudfront.net/CIK-0001045810/1cbe8fe7-e08a-46e3-8dcc-\n\nb429fc06c1a4.pdf, Nvidia U.S. Securities and Exchange Commission Form 10-K for Fiscal Year\n\nEnded January 28, 2024 at 6.) Nvidia Omniverse implements software and services \u201cinto existing\n\nsoftware tools and simulation workflows for building AI systems.\u201d (See\n\nhttps://www.nvidia.com/en-us/omniverse/.)\n\n 58. In addition, Nvidia Drive is a platform that \u201cconsists of both the AI infrastructure\n\nand in-vehicle hardware and software\u201d for autonomous vehicles. (See https://www.nvidia.com/en-\n\nus/self-driving-cars/.) \u201cNVIDIA DRIVE Infrastructure encompasses data center hardware,\n\nsoftware, and workflows\u2014both on premises and in NVIDIA DGX Cloud & Omniverse.\u201d (See id.)\n\n 59. In addition, Nvidia Isaac Sim is a platform that enables \u201cdevelopers to design,\n\nsimulate, test, and train AI-based robots and autonomous machines in a physically-based virtual\n\nenvironment.\u201d (See https://developer.nvidia.com/isaac/sim.) It is built on Nvidia Omniverse. (See\n\nid.)\n\n\n\n\n 20\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 22 of21\n 95of 94\n\n\n\n\n 60. In addition, Nvidia NGC is a collection of software services and tools that support\n\n\u201cend-to-end AI and digital twin workflows\u201d that runs on \u201cNVIDIA GPU-accelerated platforms.\u201d\n\n(See https://www.nvidia.com/en-us/gpu-cloud/.) NGC \u201coffers a collection of cloud services . . .\n\nfor generative AI, drug discovery, and speech AI solutions, and the NGC Private Registry for\n\nsecurely sharing proprietary AI software.\u201d (See id.)\n\n FIRST CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019867 PATENT)\n\n 61. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n 62. Defendant has infringed and continues to infringe one or more claims of the \u2019867 Patent\n\nin violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will continue to\n\ndo so. The Accused Products, including features of, e.g., the Grace Hopper Superchip (GH200), at least\n\nwhen used for their ordinary and customary purposes, practice each element of at least claim 16 of the\n\n\u2019867 Patent as demonstrated below.\n\n 63. For example, claim 16 of the \u2019867 Patent recites:\n\n 16. A method for performing a numerical simulation on input data\n in a computer system including a central processing unit and an\n accelerator, the method comprising:\n\n receiving, by an accelerator, first input data from the central\n processing unit;\n\n transferring, by an accelerator controller, the first input data into a\n first partition, referenced by first pointer, of an accelerator memory\n before a first computational cycle of the numerical simulation;\n\n performing, by at least one graphics processing unit during the first\n computational cycle, at least one calculation on the first portion of\n the input data as to generate first output data;\n\n\n\n\n 21\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 23 of22\n 95of 94\n\n\n\n\n storing, by the accelerator controller, the first output data into a\n second partition, referenced by a second pointer, of the accelerator\n memory; and\n\n swapping the first pointer with the second pointer at the end of the\n first computational cycle, such that the first output data becomes an\n input for a second computational cycle of the numerical simulation.\n\n 64. The Accused Products perform each step of the method of claim 16 of the \u2019867\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod for performing a numerical simulation on input data in a computer system including a\n\ncentral processing unit and an accelerator, as further explained below. For instance, the Grace\n\nHopper Superchip (GH200) \u201cbrings together the groundbreaking performance of the NVIDIA\n\nHopper GPU with the versatility of the NVIDIA Grace\u2122 CPU . . . in a single Superchip.\u201d\n\n\n\n\n 22\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 24 of23\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 65. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 66. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA enables NVIDIA GPUs to be used for general purpose computing tasks. CUDA further\n\nincludes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a compiler, development\n\ntools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural Network library\n\n(cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d It \u201cprovides\n\nhighly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n 23\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 25 of24\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 67. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 24\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 26 of25\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 68. The Accused Products perform a method that includes receiving, by an accelerator,\n\nfirst input data from the central processing unit. For instance, the \u201cCUDA programming model\u201d\n\nimplements programming functions and instructions for CPUs and GPUs. \u201cThe host is the CPU\n\navailable in the system\u201d and \u201csystem memory associated with the CPU is called host memory.\u201d\n\n\u201cThe GPU is called a device and GPU memory likewise called device memory.\u201d As an example,\n\nthe first main CUDA program execution step is \u201c[c]opy[ing] the input data from host [CPU]\n\nmemory to device [GPU] memory, also known as host-to-device transfer.\u201d\n\n\n\n\n 25\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 27 of26\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 69. The Accused Products practice a method that includes transferring, by an\n\naccelerator controller, the first input data into a first partition, referenced by first pointer, of an\n\naccelerator memory before a first computational cycle of the numerical simulation. For instance,\n\nthe GPU architecture of the Accused Products implements a controller. As an example, the\n\nHopper-GPU architecture implements \u201cHBM3 memory controllers\u201d including \u201c12 512-bit\n\nmemory controllers\u201d coupled GPU memory including \u201c6 HBM3 or HBM2e stacks,\u201d \u201c80 GB\n\nHBM3, 5 HBM3 stacks,\u201d and \u201c80 GB HBM2e, 5 HBM2e stacks.\u201d\n\n\n\n\n 26\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 28 of27\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ (emphasis added).)\n\n 70. The Accused Products implement CUDA, Nvidia\u2019s parallel computing platform.\n\nCUDA enables NVIDIA GPUs to be used for general purpose computing tasks and includes\n\nspecialized GPU-acceleration libraries such as cuDNN. Examples of parameters used in CUDA\n\ninclude pointers \u201cdst\u201d (\u201cDestination memory address\u201d) and \u201csrc\u201d (\u201cSource memory address\u201d). For\n\ninstance, exemplary CUDA function \u201ccudaMemcpy\u201d copies \u201cbytes [data] from the memory area\n\npointed to by src [source memory address pointer] to the memory area pointed to by dst\n\n\n\n 27\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 29 of28\n 95of 94\n\n\n\n\n[destination memory address pointer], where kind [type of transfer] specifies the direction of the\n\ncopy.\u201d One of the destinations is \u201ccudaMemcpyHostToDevice,\u201d or host (CPU) to device (GPU).\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 71. The Accused Products practice a method that includes performing, by at least one\n\ngraphics processing unit during the first computational cycle, at least one calculation on the first\n\nportion of the input data as to generate first output data. For instance, CUDA uses \u201cstreams\u201d to\n\nexecute a sequence of commands in order. As shown below, an exemplary CUDA function\n\n\u201ccudaMemcpyAsync\u201d is used to copy data between a host (CPU) and a device (GPU). \u201cEach\n\nstream copies its portion of input array hostPtr [pointer for CPU] to array inputDevPtr in device\n\n[GPU] memory.\u201d The stream then \u201cprocesses inputDevPtr on the device [GPU] by calling\n\nMyKernel(), and copies the result outputDevPtr back to the same portion of hostPtr.\u201d\n\n\n\n\n 28\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 30 of29\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#creation-and-\n\ndestruction-of-streams (emphasis added).)\n\n 72. The Accused Products practice a method that includes storing, by the accelerator\n\ncontroller, the first output data into a second partition, referenced by a second pointer, of the\n\naccelerator memory. For instance, exemplary excerpts of CUDA code shown below demonstrate\n\nCUDA being used to calculate a square sub-matrix Csub of matrix C using the function MatMul.\n\nAn exemplary CUDA stream \u201callocate[s] [matrix] C in device memory.\u201d After Matrices A and B\n\nare synchronized and multiplied, the exemplary CUDA stream \u201c[w]rite[s] Csub to device [GPU]\n\nmemory.\u201d\n\n\n\n\n 29\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 31 of30\n 95of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory\n\n(emphasis added).)\n\n 73. In addition, exemplary CUDA function \u201ccudaMalloc\u201d is used to \u201callocate weight,\n\nwork, and reserve space buffer sizes in the GPU memory.\u201d \u201cThe work-space buffer is used for\n\ntemporary storage\u201d and the \u201c content can be discarded or modified after all GPU kernels launched\n\nby the corresponding API complete.\u201d The \u201creserve-space buffer\u201d used to transfer intermediate\n\nresults is used for transferring \u201cintermediate results\u201d as used in the cuDNN GPU-acceleration\n\nlibrary for CUDA.\n\n\n 30\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 32 of31\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-\n\n893/api/index.html#cudnnGetMultiHeadAttnBuffers (emphasis added).)\n\n 74. The Accused Products practice a method that includes swapping the first pointer\n\nwith the second pointer at the end of the first computational cycle, such that the first output data\n\nbecomes an input for a second computational cycle of the numerical simulation. For example, the\n\ncuDNN GPU-acceleration library of the Accused Products implement operations that \u201ctake tensors\n\nas input and produce tensors as output.\u201d\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n\n\n\n 31\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 33 of32\n 95of 94\n\n\n\n\n 75. Nvidia confirms that CUDA implements pointer swapping for device (GPU) pointers.\n\n\n\n\n *****\n\n\n\n\n(See https://forums.developer.nvidia.com/t/swap-device-pointers/38964 (emphasis added).)\n\n\n\n\n 32\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 34 of33\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-c-programming-guide/#creation-and-destruction-of-\n\nstreams (emphasis added).)\n\n 76. Each claim in the \u2019867 Patent recites an independent invention. Neither claim 16,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019867 Patent.\n\n 77. Defendant has been aware of the technology patented by the \u2019867 Patent since at\n\nleast 2007, when the inventors of the Asserted Patents first discussed their patented technologies\n\n\n\n 33\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 35 of34\n 95of 94\n\n\n\n\nwith Mr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited the application for the \u2019867 Patent in its own patent portfolio since at least June 28, 2010.\n\n\n\n\n *****\n\n\n\n\n(See https://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy (emphasis\n\nadded).)\n\n\n\n\n *****\n\n\n\n\n 34\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 36 of35\n 95of 94\n\n\n\n\n(See https://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf\n\n(Nvidia U.S. Patent No. 8,922,566) (emphasis added).)\n\n\n\n\n(See\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner)\n\n(emphasis added).)\n\n 78. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019867 Patent and its related patents and applications. These discussions included at\n\nleast Mr. Alvin Lin, an Nvidia Senior Director of Business Development, and Mr. Jeff Herbst, then\n\nan Nvidia Vice President of Business Development and head of Nvidia\u2019s Inception GPU Ventures,\n\nin or around September 6, 2016. In or around October 2016, Nvidia, through its representatives,\n\n\n 35\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 37 of36\n 95of 94\n\n\n\n\ninitiated discussions with the inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 79. The inventors also discussed their patented technology, in addition to the \u2019867\n\nPatent and its family, with Defendant\u2019s representatives at Nvidia\u2019s artificial intelligence\n\nconference in or around June 2017. On or about June 26, 2017, Defendant received materials from\n\nthe inventors, in lieu of a meeting on or about June 29, that identified the \u2019867 Patent and its family\n\nand described the technology in detail. Defendant had previously stated it was interested in the\n\ninventors\u2019 solutions. Defendant also featured the inventors on its website as members of\n\nDefendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n 80. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\n 36\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 38 of37\n 95of 94\n\n\n\n\n\u2019867 Patent.\n\n 81. Defendant directly infringes at least claim 16 of the \u2019867 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 82. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 16 of the \u2019867 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 83. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 16 of the \u2019867 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 16 of the \u2019867 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 16 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\ninstallation, support, and distribution of the Accused Products.\n\n 84. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\n\n\n\n 37\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 39 of38\n 95of 94\n\n\n\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 85. Defendant further encourages and induces its customers to infringe claim 16 of the\n\n\u2019867 Patent: 1) by making its accelerated computing and data center services available on its\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 86. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 87. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\n\n\n\n 38\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 40 of39\n 95of 94\n\n\n\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019867 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 88. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019867 Patent.\n\n 89. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019867 Patent.\n\n 90. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\n\n\n\n 39\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 41 of40\n 95of 94\n\n\n\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 16 of the \u2019867 Patent.\n\n 91. Plaintiff has suffered and continue to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019867 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 92. Defendant\u2019s infringement of the \u2019867 Patent is knowing and willful. Defendant had\n\nactual knowledge of the \u2019867 Patent application since at least 2010 and actual knowledge of the\n\n\u2019867 Patent, and its family, since at least 2017.\n\n 93. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents\n\nand Plaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and\n\nservices that it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019867 Patent\n\nwith knowledge of the \u2019867 Patent constitutes willful infringement.\n\n SECOND CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019438 PATENT)\n\n 94. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n 95. Defendant has infringed and continues to infringe one or more claims of the \u2019438 Patent\n\nin violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will continue to\n\ndo so. The Accused Products, including features of, e.g., the Grace Hopper Superchip (GH200), at least\n\n\n\n 40\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 42 of41\n 95of 94\n\n\n\n\nwhen used for their ordinary and customary purposes, practice each element of at least claim 21 of the\n\n\u2019438 Patent as demonstrated below.\n\n 96. For example, claim 21 of the \u2019438 Patent recites:\n\n 21. A method of performing a sequence of computations\n representing an artificial neural network, the method comprising:\n\n receiving, at a central processing unit (CPU), first input data\n acquired from an external system in real time;\n\n initializing, by a controller operably coupled to a graphics\n processing unit (GPU), textures and shaders in a memory operably\n coupled to the GPU;\n\n transferring the first input data received by the CPU to the memory\n operably coupled to the GPU;\n\n performing, by the graphics processing unit (GPU), a first\n computation in the sequence of computations on the first input data\n based on the textures and shaders to generate first output data,\n computations in the sequence of computations representing\n respective layers of neurons in the artificial neural network, an\n output of the first computation in the sequence of computations\n representing an output of a first neuron in a first layer in the artificial\n neural network;\n\n storing, in the memory operably coupled to the GPU, the first input\n data and the first output data; and\n\n transferring second input data acquired from the external system in\n real time into the memory operably coupled to the GPU after the\n GPU starts the first computation and before the GPU starts a second\n computation of the sequence of computations, an output of the\n second computation in the sequence of computations representing\n an output of a second neuron in a second layer in the artificial neural\n network.\n\n 97. The Accused Products perform each step of the method of claim 21 of the \u2019438\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod of performing a sequence of computations representing an artificial neural network, as\n\nfurther explained below. For instance, the Grace Hopper Superchip (GH200) \u201cbrings together the\n\n\n\n 41\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 43 of42\n 95of 94\n\n\n\n\ngroundbreaking performance of the NVIDIA Hopper GPU with the versatility of the NVIDIA\n\nGrace\u2122 CPU . . . in a single Superchip.\u201d It includes the cuDNN (CUDA Deep Neural Network)\n\nlibrary for \u201c[d]eep neural networks.\u201d\n\n\n\n\n *****\n\n\n\n\n 42\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 44 of43\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 98. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 99. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA further includes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a\n\ncompiler, development tools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural\n\nNetwork library (cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d\n\n\n\n 43\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 45 of44\n 95of 94\n\n\n\n\nIt \u201cprovides highly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 100. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 44\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 46 of45\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 101. The Accused Products perform a method that includes receiving, at a central\n\nprocessing unit (CPU), first input data acquired from an external system in real time. For instance,\n\nas illustrated below, a diagram describing the architecture of the Grace Hopper Superchip depicts\n\na CPU (\u201cGrace CPU\u201d) with \u201c[u]p to 72 cores\u201d that receives and sends input and output data via\n\n\u201cHigh-Speed IO\u201d (\u201cPCIe-5\u201d). The Grace CPU is further depicted as being coupled to memory\n\n\u201cCPU LPDDR5X\u201d up to 480GB via a link up to \u201c500 GB/s.\u201d\n\n\n\n\n 45\n\fCaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 47 of46\n 95of 94\n\n\n\n\n *****\n\n\n\n\n 46\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 48 of47\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 102. In a related diagram example, \u201cCPU PHYSICAL MEMORY\u201d (LPDDR5X) is\n\nillustrated as being accessed by a CPU (\u201cCPU-resident access\u201d).\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 103. As previously stated, the Accused Products implement CUDA and specialized\n\nGPU-acceleration libraries such as cuDNN. \u201cCUDA\u00ae is a parallel computing platform and\n\n\n\n 47\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 49 of48\n 95of 94\n\n\n\n\nprogramming model developed by NVIDIA for general computing on graphical processing units\n\n(GPUs)\u201d including \u201cGPU-accelerated applications.\u201d In GPU-accelerated applications, \u201cthe\n\nsequential part of the workload runs on the CPU \u2013 which is optimized for single-threaded\n\nperformance \u2013 while the compute intensive portion of the application runs on thousands of GPU\n\ncores in parallel.\u201d\n\n\n\n\n(See https://developer.nvidia. com/cuda-zone (emphasis added).)\n\n 104. The \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs and GPUs. \u201cThe host is the CPU available in the system\u201d and \u201csystem\n\nmemory associated with the CPU is called host memory.\u201d As an example, the first main CUDA\n\nprogram execution step is \u201c[c]opy[ing] the input data from host [CPU] memory.\u201d\n\n\n\n\n 48\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 50 of49\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 105. The Accused Products perform a method that includes initializing, by a controller\n\noperably coupled to a graphics processing unit (GPU), textures and shaders in a memory operably\n\ncoupled to the GPU. For instance, as illustrated below, a GPU (\u201cHopper GPU\u201d) for the Grace\n\nHopper Superchip is depicted as accessing both CPU and GPU memory using \u201cNVLINK\u201d for both\n\naccessing and storing data. Indeed, \u201cNVIDIA GH200 is designed to accelerate applications with\n\nexceptionally large memory footprints.\u201d As illustrated, the Hopper GPUs are illustrated coupled\n\nto \u201cGPU HBM3\u201d high bandwidth memory or \u201cGPU HBM3e\u201d high bandwidth memory.\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 106. The GPU architecture of the Accused Products implements a controller. As an\n\nexample, the Hopper-GPU architecture includes \u201cGPU processing clusters\u201d and \u201ctexture\n\nprocessing clusters\u201d and implements \u201cHBM3 memory controllers\u201d including \u201c12 512-bit memory\n\ncontrollers\u201d coupled GPU memory including \u201c6 HBM3 or HBM2e stacks,\u201d \u201c80 GB HBM3, 5\n\nHBM3 stacks,\u201d and \u201c80 GB HBM2e, 5 HBM2e stacks.\u201d\n\n\n\n\n 49\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 51 of50\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ (emphasis added).)\n\n 107. In addition, CUDA includes the exemplary NPP (Nvidia Performance Primitives)\n\nlibrary \u201cfor performing CUDA accelerated processing\u201d and \u201cperforming CUDA accelerated\n\nprocessing for 2D image and signal processing.\u201d\n\n\n\n\n 50\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 52 of51\n 95of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/index.html (emphasis added).)\n\n\n\n\n(See https://docs.nvidia.com/cuda/npp/introduction.html (emphasis added).)\n\n 108. As an example, the exemplary CUDA NPP library passes image data using \u201c[a]\n\npointer to the image\u2019s underlying data type\u201d and \u201c[a] line step in bytes.\u201d In this example, the\n\npointer is passed \u201cto the underlying pixel data type\u201d and the pointer and line step are passed\n\nindividually for processing involving \u201cimage data.\u201d\n\n\n\n\n 51\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 53 of52\n 95of 94\n\n\n\n\n(See\n\nhttps://docs.nvidia.com/cuda/npp/introduction.html#nppi_conventions_lb_1passing_image_data\n\n(emphasis added).)\n\n 109. As another example, the exemplary CUDA NPP library implements function for\n\nimage color conversion. These functions \u201cmanipulat[e] an image\u2019s color model and sampling\n\nformat\u201d and \u201ccan be found in the nppicc [NVIDIA Performance Primitives Image Color\n\nConversion] library.\u201d As shown, these functions save \u201capplication load time\u201d and \u201cCUDA\n\nruntime.\u201d\n\n\n\n\n(See https://docs.nvidia.com/cuda/npp/image_color_conversion.html#image-color-model-\n\nconversion-functions (emphasis added).)\n\n 110. The Accused Products perform a method that includes transferring the first input\n\n\n 52\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 54 of53\n 95of 94\n\n\n\n\ndata received by the CPU to the memory operably coupled to the GPU. For instance, the \u201cCUDA\n\nprogramming model\u201d implements programming functions and instructions for CPUs and GPUs.\n\n\u201cThe host is the CPU available in the system\u201d and \u201csystem memory associated with the CPU is\n\ncalled host memory.\u201d \u201cThe GPU is called a device and GPU memory likewise called device\n\nmemory.\u201d As an example, the first main CUDA program execution step is \u201c[c]opy[ing] the input\n\ndata from host [CPU] memory to device [GPU] memory, also known as host-to-device transfer.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 111. The Accused Products perform a method that includes performing, by the graphics\n\nprocessing unit (GPU), a first computation in the sequence of computations on the first input data\n\nbased on the textures and shaders to generate first output data, computations in the sequence of\n\ncomputations representing respective layers of neurons in the artificial neural network, an output\n\nof the first computation in the sequence of computations representing an output of a first neuron\n\nin a first layer in the artificial neural network. For instance, the CUDA platform programming\n\nimplemented in the Accused Products utilizes the GPU and GPU memory. As an example, after\n\nthe \u201chost-to-device transfer\u201d (host (CPU) memory to device (GPU) memory), the second main\n\n\n 53\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 55 of54\n 95of 94\n\n\n\n\nstep is \u201c[l]oad the GPU program and execute, caching data on-chip for performance.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 112. In addition, as previously stated, cuDNN is a CUDA \u201cGPU-acceleration library of\n\nprimitives for deep neural networks.\u201d (See https://developer.nvidia.com/cudnn.) The exemplary\n\ncuDNN release notes below demonstrate computations implemented for RNNs and related data\n\nbeing transferred to GPU memory. As shown, users do \u201cnot need to transfer [an] array [from RNN\n\ndata descriptors] to device memory; the operation will be performed automatically by RNN APIs.\u201d\n\n\n\n\n *****\n\n\n\n\n 54\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 56 of55\n 95of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-893/release-\n\nnotes/index.html#abstract (emphasis added).)\n\n 113. Relatedly, cuDNN operations below exemplify tensors being used as inputs and\n\noutputs (e.g., Tmp0). Exemplary \u201ccuDNN operations take tensors as input and produce tensors as\n\noutput.\u201d As part of CUDA, these cuDNN operations implement computer tasks performed by a\n\nGPU.\n\n\n\n\n 55\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 57 of56\n 95of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n 114. The Accused Products perform a method that includes storing, in the memory\n\noperably coupled to the GPU, the first input data and the first output data. For instance, as\n\nillustrated below, a diagram describing the architecture of the Grace Hopper Superchip depicts a\n\nGPU (\u201cHopper GPU\u201d) in communication with GPU memory (\u201cGPUHBM3 or HBm3e\u201d high\n\nbandwidth memory).\n\n\n\n\n 56\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 58 of57\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 115. As previously stated, the \u201cCUDA programming model\u201d implements programming\n\nfunctions and instructions for CPUs (host) and GPUs (device). For example, after the \u201chost-to-\n\ndevice transfer\u201d (CPU to GPU) first main step and \u201c[l]oad[ing] the GPU program and execut[ing]\u201d\n\n\n 57\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 59 of58\n 95of 94\n\n\n\n\nand \u201ccaching data on-chip for performance\u201d for the second main step, the \u201cresults\u201d are stored on\n\nGPU \u201cdevice memory.\u201d\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded.)\n\n 116. The Accused Products perform a method that includes transferring second input\n\ndata acquired from the external system in real time into the memory operably coupled to the GPU\n\nafter the GPU starts the first computation and before the GPU starts a second computation of the\n\nsequence of computations, an output of the second computation in the sequence of computations\n\nrepresenting an output of a second neuron in a second layer in the artificial neural network. For\n\ninstance, as illustrated below, a diagram describing the architecture of the Grace Hopper Superchip\n\ndepicts a CPU (\u201cGrace CPU\u201d) in communication with a GPU (\u201cHopper GPU\u201d) via \u201cNVLink-C2C\u201d\n\n(chip-to-chip). \u201cHigh-Speed IO\u201d input and output data is received by the CPU via \u201cPCIe-5.\u201d\n\n\n\n\n 58\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 60 of59\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 117. Furthermore, the Accused Products, including the Grace Hopper Superchip,\n\nimplement libraries and SDKs designed for neural networks that \u201care created from large numbers\n\nof identical neurons [that] are highly parallel by nature.\u201d The Accused Products implement\n\n\n 59\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 61 of60\n 95of 94\n\n\n\n\ncuDNN, a library \u201cmakes it easy to obtain state-of-the-art performance with Deep Neural\n\nNetworks,\u201d and TensorRT, a platform accelerator and runtime for optimizing, validating, and\n\ndeploying neural networks for inference (e.g., applying knowledge from a trained neural network\n\nmodel and inferring a result).\n\n\n\n\n(See https://developer.nvidia.com/discover/artificial-neural-network (emphasis added).)\n\n 118. An example below illustrates an exemplary neural network the Accused Products\n\nare designed to accelerate using parallel computations. \u201cInput\u201d (four) and \u201cOutput\u201d (eight) neurons\n\nare depicted below in a full-connected or linear layer structure in which all of the input neurons\n\ndepicted in a first layer are connected to all of the output neurons depicted in a second layer.\n\nComputations for the neural network are performed using, for example, \u201cNVIDIA Matrix\n\nMultiplication.\u201d Examples of inputs and outputs for forward propagation, activation gradient\n\ncomputation, and weight gradient computation (as matrix by matrix multiplications) are shown\n\nbelow.\n\n\n\n\n 60\n\fCaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 62 of61\n 95of 94\n\n\n\n\n 61\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 63 of62\n 95of 94\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/performance/dl-performance-fully-\n\nconnected/index.html#performance (annotations added).)\n\n\n 62\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 64 of63\n 95of 94\n\n\n\n\n 119. Indeed, the cuDNN GPU-acceleration library of the Accused Products implement\n\noperations that \u201ctake tensors as input and produce tensors as output.\u201d\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-881/developer-\n\nguide/index.html#tensors-layouts (emphasis added).)\n\n 120. Each claim in the \u2019438 Patent recites an independent invention. Neither claim 21,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019438 Patent.\n\n 121. Defendant has been aware of the \u2019438 Patent since at least the filing of the original\n\nComplaint. Defendant has been aware of the technology patented by the \u2019438 Patent since at least\n\n2007, when the inventors of the Asserted Patents first discussed their patented technologies with\n\nMr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited an ancestor of the \u2019438 Patent in its own patent portfolio since at least June 28, 2010 (See\n\nhttps://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy;\n\nhttps://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf;\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner).)\n\n 122. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019438 Patent family. These discussions included at least Mr. Alvin Lin, an Nvidia\n\nSenior Director of Business Development, and Mr. Jeff Herbst, then an Nvidia Vice President of\n\nBusiness Development and head of Nvidia\u2019s Inception GPU Ventures, in or around September 6,\n\n\n 63\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 65 of64\n 95of 94\n\n\n\n\n2016. In or around October 2016, Nvidia, through its representatives, initiated discussions with\n\nthe inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 123. The inventors also discussed their patented technology, including the underlying\n\ntechnology and family to the \u2019438 Patent (including U.S. Patent No. 9,189,828, the patent the \u2019438\n\nPatent reissued from), with Defendant\u2019s representatives at Nvidia\u2019s artificial intelligence\n\nconference in or around June 2017. On or about June 26, 2017, Defendant received materials from\n\nthe inventors, in lieu of a meeting on or about June 29, that identified patents related to the \u2019438\n\nPatent and described the technology in detail. Defendant had previously stated it was interested in\n\nthe inventors\u2019 solutions. Defendant also featured the inventors on its website as members of\n\nDefendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n\n 64\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 66 of65\n 95of 94\n\n\n\n\n 124. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\u2019438 Patent.\n\n 125. Defendant directly infringes at least claim 21 of the \u2019438 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 126. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 21 of the \u2019438 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 127. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 21 of the \u2019438 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 21 of the \u2019438 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 21 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\n\n\n\n 65\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 67 of66\n 95of 94\n\n\n\n\ninstallation, support, and distribution of the Accused Products.\n\n 128. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 129. Defendant further encourages and induces its customers to infringe claim 21 of the\n\n\u2019438 Patent: 1) by making its accelerated computing and data center services available on its\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 130. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 131. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\n\n\n\n 66\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 68 of67\n 95of 94\n\n\n\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019438 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 132. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019438 Patent.\n\n 133. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019438 Patent.\n\n\n\n\n 67\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 69 of68\n 95of 94\n\n\n\n\n 134. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 21 of the \u2019438 Patent.\n\n 135. Plaintiff has suffered and continues to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019438 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 136. Defendant\u2019s infringement of the \u2019438 Patent is knowing and willful. Defendant\n\nacquired actual knowledge of the patent that the \u2019438 Patent reissued from, and its family, since at\n\nleast 2017 and has acquired additional knowledge of the \u2019438 Patent since at least the filing of this\n\nlawsuit.\n\n 137. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents and\n\nPlaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and services\n\nthat it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019438 Patent with\n\nknowledge of the \u2019438 Patent constitutes willful infringement.\n\n THIRD CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019461 PATENT)\n\n 138. Plaintiff realleges and incorporates by reference the allegations of the preceding\n\nparagraphs of this Complaint.\n\n\n\n 68\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 70 of69\n 95of 94\n\n\n\n\n 139. Defendant has infringed and continues to infringe one or more claims of the \u2019461\n\nPatent in violation of 35 U.S.C. \u00a7 271 in this District and elsewhere in the United States and will\n\ncontinue to do so. The Accused Products, including features of, e.g., the Grace Hopper Superchip\n\n(GH200), at least when used for their ordinary and customary purposes, practice each element of\n\nat least claim 21 of the \u2019461 Patent as demonstrated below.\n\n 140. For example, claim 21 of the \u2019461 Patent recites:\n\n 21. A method of executing computations representing an artificial\n neural network on a computer system comprising at least one central\n processing unit (CPU), a processing unit, a first memory partition,\n and a second memory partition, the method comprising:\n\n executing, by the at least one CPU, a user interaction stream, the\n user interaction stream controlling transfer of inputs to the artificial\n neural network to the first memory partition and the second memory\n partition;\n\n executing, by the processing unit, a computational stream, the\n computational stream controlling data exchange between the user\n interaction stream and the computational stream during execution of\n the computations representing the artificial neural network;\n\n shifting control of a data exchange between the user interaction\n stream and the computational stream to the computational stream in\n response to starting execution of the computations representing the\n artificial neural network;\n\n shifting control of the data exchange between the user interaction\n stream and the computational stream to the user interaction stream\n in response to completion or interruption of the computations\n representing the artificial neural network;\n\n queueing a user command received by the user interaction stream\n during execution of the computations representing the artificial\n neural network; and\n\n executing the user command during execution of the computations\n representing the artificial neural network at times determined by the\n computational stream.\n\n 141. The Accused Products perform each step of the method of claim 21 of the \u2019461\n\n\n\n 69\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 71 of70\n 95of 94\n\n\n\n\nPatent. To the extent the preamble is construed to be limiting, the Accused Products perform a\n\nmethod of executing computations representing an artificial neural network on a computer system\n\ncomprising at least one central processing unit (CPU), a processing unit, a first memory partition,\n\nand a second memory partition, as further explained below. For instance, the Grace Hopper\n\nSuperchip (GH200) \u201cbrings together the groundbreaking performance of the NVIDIA Hopper\n\nGPU with the versatility of the NVIDIA Grace\u2122 CPU . . . in a single Superchip.\u201d It includes the\n\ncuDNN (CUDA Deep Neural Network) library for \u201c[d]eep neural networks.\u201d\n\n\n\n\n *****\n\n\n\n\n 70\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 72 of71\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 142. As illustrated below, a diagram describing the architecture of the Grace Hopper\n\nSuperchip depicts a CPU (\u201cGrace CPU\u201d) with \u201c[u]p to 72 cores\u201d and CPU memory (\u201cCPU\n\nLPDDR5X\u201d) and a GPU (\u201cHopper GPU\u201d) and GPU memory (\u201cGPUHBM3 or HBm3e\u201d).\n\n\n\n\n 71\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 73 of72\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 143. The \u201cGrace Hopper Superchip is the first true heterogeneous accelerated platform\n\nfor high-performance computing (HPC) and AI workloads. It accelerates applications with the\n\nstrengths of both GPUs and CPUs while providing the simplest and most productive heterogeneous\n\nprogramming model to date.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 144. In addition, the Accused Products, including the Grace Hopper Superchip,\n\nimplement CUDA, Nvidia\u2019s proprietary \u201cparallel computing platform and programming model.\u201d\n\nCUDA further includes the CUDA Toolkit, which \u201cincludes GPU-accelerated libraries, a\n\ncompiler, development tools and the CUDA runtime.\u201d As an example, the \u201cCUDA\u00ae Deep Neural\n\n\n\n 72\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 74 of73\n 95of 94\n\n\n\n\nNetwork library (cuDNN) is a GPU-acceleration library of primitives for deep neural networks.\u201d\n\nIt \u201cprovides highly tuned implementations for standard routines\u201d for GPU-based acceleration.\n\n\n\n\n(See https://developer.nvidia.com/cuda-zone (emphasis added).)\n\n\n\n\n(See https://developer.nvidia.com/cudnn (emphasis added).)\n\n 145. Nvidia GPU architectures that implement CUDA and cuDNN include the Hopper\n\n(e.g., Grace Hopper Superchip (GH200), H100), Ada Lovelace, Ampere, Turing, Volta, Pascal,\n\nand Maxwell GPU architectures of the Accused Products.\n\n\n\n\n 73\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 75 of74\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-896/support-matrix/index.html\n\n(emphasis added).)\n\n 146. The Accused Products perform a method that includes executing, by the at least\n\none CPU, a user interaction stream, the user interaction stream controlling transfer of inputs to\n\nthe artificial neural network to the first memory partition and the second memory partition. For\n\ninstance, as shown in the Grace Hopper Superchip architecture diagram below, the Grace Hopper\n\nSuperchip is illustrated below with a CPU (\u201cGRACE CPU\u201d). The CPU \u201cshare[s] a single per-\n\nprocess page table\u201d with a GPU (\u201cHopper GPU\u201d), \u201cenabling all CPU and GPU threads to access\n\nall system-allocated memory.\u201d The CPU is depicted as coupled to the GPU via \u201cNVLINK C2C\n\n[chip-to-chip],\u201d and can access the \u201cSystem Page Table\u201d and \u201cCPU PHYSICAL MEMORY\u201d via\n\n\u201cCPU-resident access\u201d and \u201cGPU PHYSICAL MEMORY\u201d via \u201c[r]emote access\u201d and \u201cPTE [page\n\ntable entry] B.\u201d The GPU can also access the System Page Table, and it can access \u201cGPU\n\nPHYSICAL MEMORY\u201d via \u201cGPU-resident access\u201d and \u201cCPU PHYSICAL MEMORY\u201d via\n\n\u201c[r]emote access\u201d and \u201cPTE A.\u201d Moreover, the \u201cSystem Page Table\u201d \u201c[t]ranslates CPU malloc()\n\n\n 74\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 76 of75\n 95of 94\n\n\n\n\n[memory allocation] to CPU or GPU.\u201d \u201cThe CPU heap, CPU thread stack, global variables\n\nmemory-mapped files, and inter-process memory are accessible to all CPU and GPU threads.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 147. The \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs and GPUs. \u201cThe host is the CPU available in the system\u201d and \u201csystem\n\nmemory associated with the CPU is called host memory.\u201d \u201cThe GPU is called a device and GPU\n\nmemory likewise called device memory.\u201d As an example, the first main CUDA program execution\n\nstep is \u201c[c]opy[ing] the input data from host [CPU] memory to device [GPU] memory, also known\n\nas host-to-device transfer.\u201d\n\n\n\n\n 75\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 77 of76\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 148. The Accused Products perform a method that includes executing, by the processing\n\nunit, a computational stream, the computational stream controlling data exchange between the\n\nuser interaction stream and the computational stream during execution of the computations\n\nrepresenting the artificial neural network. For instance, the \u201cCUDA programming model\u201d\n\nimplements programming functions and instructions for CPUs and GPUs. As previously stated,\n\nthe host is the CPU and the device is the GPU. After \u201c[c]opy[ing] the input data from host [CPU]\n\nmemory to device [GPU] memory,\u201d the second main CUDA program execution step is\n\n\u201c[l]oad[ing] the GPU program and execut[ing].\u201d\n\n\n\n\n 76\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 78 of77\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded).)\n\n 149. Indeed, the Grace Hopper Superchip \u201cis designed to accelerate applications\u201d using\n\n\u201cExtended GPU Memory.\u201d As depicted in the gram of the Grace Hopper architecture below, a\n\nGPU (\u201cHOPPER GPU\u201d) can access \u201cLocal CPU,\u201d \u201cPeer CPU,\u201d and \u201cPeer GPU\u201d memory via\n\n\u201cNVLink.\u201d\n\n\n\n\n 77\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 79 of78\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 150. For instance, exemplary CUDA library cuDNN function\n\n\u201ccudnnSetRNNDescriptor_v8\u201d \u201cinitializes a previously created RNN [recurrent neural network]\n\ndescriptor object.\u201d This function \u201cstore[s] all information needed to compute the total number of\n\nadjustable weights/biases in the RNN model.\u201d In addition, the parameters \u201cdirMode,\u201d\n\n\u201cinputMode,\u201d and \u201cdatatype\u201d confirm the exchange of calculations and values between the hidden\n\nlayers of an RNN.\n\n\n\n\n *****\n\n\n\n\n 78\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 80 of79\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-891/pdf/cuDNN-API.pdf\n\n(emphasis added).)\n\n 151. The Accused Products perform a method that includes shifting control of a data\n\nexchange between the user interaction stream and the computational stream to the computational\n\nstream in response to starting execution of the computations representing the artificial neural\n\nnetwork. For instance, the \u201cCUDA programming model\u201d implements programming functions and\n\ninstructions for CPUs (host) and GPUs (device). As an example, the \u201chost-to-device transfer\u201d\n\n(CPU to GPU) first main step, the second main step is \u201c[l]oad the GPU program and execute\u201d and\n\nthe third main step is \u201c[c]opy the results from device [GPU] memory to host [CPU] memory, also\n\nknown as device-to-host transfer\u201d (GPU to CPU).\n\n\n\n\n 79\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 81 of80\n 95of 94\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\nadded.)\n\n 152. As previously stated, the Grace Hopper Superchip \u201cis designed to accelerate\n\napplications\u201d using \u201cExtended GPU Memory\u201d and the GPU can access local/peer CPU and peer\n\nGPU memory via \u201cNVLink.\u201d The Grace Hopper Superchip\u2019s Extended GPU Memory feature\n\n\u201cenables GPUs to access all the system memory efficiently\u201d and \u201cphysical memory in the system\n\ncan be allocated to be accessible from any GPU thread.\u201d\n\n\n\n\n 80\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 82 of81\n 95of 94\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 153. The Accused Products perform a method that includes shifting control of the data\n\nexchange between the user interaction stream and the computational stream to the user interaction\n\nstream in response to completion or interruption of the computations representing the artificial\n\nneural network. For instance, after the \u201cCUDA programming model\u201d \u201chost-to-device transfer\u201d\n\n(CPU to GPU) and GPU program load and execution steps, the third main step is \u201c[c]opy the\n\nresults from device [GPU] memory to host [CPU] memory, also known as device-to-host transfer\u201d\n\n(GPU to CPU). The \u201chost-to-device transfer\u201d (CPU to GPU) first main step can be reintroduced\n\nfor additional computations.\n\n\n\n\n(See https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/ (emphasis\n\n\n 81\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 83 of82\n 95of 94\n\n\n\n\nadded.)\n\n 154. In addition, as shown in the Grace Hopper Superchip architecture diagram below,\n\nthe CPU (\u201cGRACE CPU\u201d) \u201cshare[s] a single per-process page table\u201d with a GPU (\u201cHopper\n\nGPU\u201d), \u201cenabling all CPU and GPU threads to access all system-allocated memory.\u201d \u201cThe CPU\n\nheap, CPU thread stack, global variables memory-mapped files, and inter-process memory are\n\naccessible to all CPU and GPU threads.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n 155. The Accused Products perform a method that includes queueing a user command\n\nreceived by the user interaction stream during execution of the computations representing the\n\nartificial neural network. For instance, as shown by publicly available CUDA toolkit\n\n\n\n 82\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 84 of83\n 95of 94\n\n\n\n\ndocumentation, CUDA implements exemplary \u201cmemory management functions\u201d that \u201c[c]op[y]\n\ndata between host [CPU] and device [GPU].\u201d This includes CUDA functions \u201ccudaMemcpy\u201d and\n\n\u201ccudaMemcpyAsync.\u201d\n\n\n\n\n *****\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 156. As an example, exemplary CUDA memory management function\n\n\u201ccudaMemcpyAsync\u201d \u201c[c]opies count bytes [data] from the memory area pointed to by src [source\n\nmemory address pointer] to the memory area pointed to by dst [destination memory address\n\npointer], where kind [type of transfer] specifies the direction of the copy.\u201d Destinations includes\n\n\u201ccudaMemcpyHostToDevice [CPU to device GPU], cudaMemcpyDeviceToHost [GPU to CPU],\n\ncudaMemcpyDeviceToDevice [GPU to GPU]. Because the function \u201ccudaMemcpyAsync() is\n\nasynchronous with respect to the host, [] the call may return before the copy is complete. The copy\n\ncan optionally be associated to a stream [identified stream] by passing a non-zero stream\n\nargument.\u201d\n\n\n\n\n 83\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 85 of84\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 157. The Accused Products perform a method that includes executing the user command\n\nduring execution of the computations representing the artificial neural network at times\n\ndetermined by the computational stream. For instance, as shown by exemplary and publicly\n\navailable CUDA toolkit documentation, CUDA implements \u201cmemory management functions\u201d that\n\n\u201c[c]op[y] data between host [CPU] and device [GPU].\u201d\n\n\n\n\n 84\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 86 of85\n 95of 94\n\n\n\n\n *****\n\n\n\n\n *****\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 158. As an example, exemplary CUDA memory management function\n\n\u201ccudaMemcpyAsync\u201d \u201c[c]opies count bytes [data] from the memory area pointed to by src [source\n\nmemory address pointer] to the memory area pointed to by dst [destination memory address\n\npointer], where kind [type of transfer] specifies the direction of the copy.\u201d Because the function\n\n\u201ccudaMemcpyAsync() is asynchronous with respect to the host, [] the call may return before the\n\ncopy is complete. The copy can optionally be associated to a stream [identified stream].\u201d\n\n\n\n\n 85\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 87 of86\n 95of 94\n\n\n\n\n(See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html\n\n(emphasis added).)\n\n 159. In another example, CUDA implements \u201cCUDA-specific memory APIs [that]\n\nprovide users with guarantees about where the memory resides, which threads can access it,\n\nwhether it is migratable, and many other features that enable users to extract all the performance\n\nthe hardware has to offer.\u201d\n\n\n\n\n(See https://resources.nvidia.com/en-us-grace-cpu/nvidia-grace-hopper (emphasis added).)\n\n\n 86\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 88 of87\n 95of 94\n\n\n\n\n 160. Each claim in the \u2019461 Patent recites an independent invention. Neither claim 21,\n\ndescribed above, nor any other individual claim is representative of all claims in the \u2019461 Patent.\n\n 161. Defendant has been aware of the \u2019461 Patent since at least the filing of the original\n\nComplaint. Defendant has been aware of the technology patented by the \u2019461 Patent since at least\n\n2007, when the inventors of the Asserted Patents first discussed their patented technologies with\n\nMr. Sanford Russell, then the CTO of Nvidia. At the time, the inventors asked Defendant to\n\ncollaborate with them on training neural networks using Nvidia\u2019s GPUs. Defendant informed the\n\ninventors, through Mr. Russell, that it was not interested in the collaboration. Defendant has also\n\ncited an ancestor of the \u2019461 Patent in its own patent portfolio since at least June 28, 2010 (See\n\nhttps://patents.google.com/patent/US8648867B2/en?oq=8648867#citedBy;\n\nhttps://patentimages.storage.googleapis.com/ee/13/e9/61df149c3fddc7/US8922566.pdf;\n\nhttps://patentcenter.uspto.gov/applications/13335850/displayReferences/referenceForms?applicat\n\nion= (Nvidia U.S. Appl. No. 13/335,850 August 12, 2014, List of References Cited by Examiner).)\n\n 162. Starting in or around 2016, the inventors of the Asserted Patents held multiple\n\ndiscussions with Nvidia to invest in or purchase their AI company, Neurala, Inc., and all its assets,\n\nincluding the \u2019461 Patent family. These discussions included at least Mr. Alvin Lin, an Nvidia\n\nSenior Director of Business Development, and Mr. Jeff Herbst, then an Nvidia Vice President of\n\nBusiness Development and head of Nvidia\u2019s Inception GPU Ventures, in or around September 6,\n\n2016. In or around October 2016, Nvidia, through its representatives, initiated discussions with\n\nthe inventors to invest in Neurala, Inc. for approximately $10 million.\n\n 163. The inventors also discussed their patented technology, including the underlying\n\ntechnology and family to the \u2019461 Patent, with Defendant\u2019s representatives at Nvidia\u2019s artificial\n\nintelligence conference in or around June 2017. On or about June 26, 2017, Defendant received\n\n\n\n\n 87\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 89 of88\n 95of 94\n\n\n\n\nmaterials from the inventors, in lieu of a meeting on or about June 29, that identified patents related\n\nto the \u2019461 Patent and described the technology in detail. Defendant had previously stated it was\n\ninterested in the inventors\u2019 solutions. Defendant also featured the inventors on its website as\n\nmembers of Defendant\u2019s start-up incubator on or about September 25, 2019.\n\n\n\n\n *****\n\n\n\n\n(See https://developer.nvidia.com/blog/inception-spotlight-ai-startup-neurala-sees-7x-speedup-\n\nwith-ngc/ (September 25, 2019); see also https://www.youtube.com/watch?v=-WBtxGLoQNs\n\n(\u201cNeurala Accelerating AI Video Annotation with NGC Containers\u201d posted by Defendant\u2019s\n\nYouTube account).)\n\n 164. Neural AI and/or its predecessors-in-interest have satisfied all statutory obligations\n\nrequired to collect pre-filing damages for the full period allowed by law for infringement of the\n\n\u2019461 Patent.\n\n 165. Defendant directly infringes at least claim 21 of the \u2019461 Patent, either literally or\n\nunder the doctrine of equivalents, by performing the steps described above. For example,\n\nDefendant performs the claimed method in an infringing manner as described above by\n\n\n 88\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 90 of89\n 95of 94\n\n\n\n\nimplementing the Accused Products as part of its accelerated computing operations and running\n\ncorresponding software that implements the infringing performance. Defendant also performs the\n\nclaimed method in an infringing manner when testing the operation of the Accused Products and\n\ncorresponding systems. As another example, Defendant performs the claimed method when\n\nproviding or administering services to third parties, customers, and partners using the Accused\n\nProducts.\n\n 166. Defendant\u2019s partners, customers, and users of its Accused Products and\n\ncorresponding systems and services directly infringe at least claim 21 of the \u2019461 Patent, literally\n\nor under the doctrine of equivalents, at least by using the Accused Products and corresponding\n\nsystems and services, as described above.\n\n 167. Defendant has actively induced and is actively inducing infringement of at least\n\nclaim 21 of the \u2019461 Patent with specific intent to induce infringement, and/or willful blindness to\n\nthe possibility that its acts induce infringement, in violation of 35 U.S.C. \u00a7 271(b). For example,\n\nDefendant encourages and induces customers to use Nvidia\u2019s CUDA platform in a manner that\n\ninfringes claim 21 of the \u2019461 Patent at least by offering and providing software that performs a\n\nmethod that infringes claim 21 when installed and operated by the customer using the Accused\n\nProducts, and by engaging in activities relating to selling, marketing, advertising, promotion,\n\ninstallation, support, and distribution of the Accused Products.\n\n 168. Defendant encourages, instructs, directs, and/or requires third parties\u2014including\n\nits certified partners and/or customers\u2014to perform the claimed method using the software,\n\nplatform, services, and systems in infringing ways, as described above.\n\n 169. Defendant further encourages and induces its customers to infringe claim 21 of the\n\n\u2019461 Patent: 1) by making its accelerated computing and data center services available on its\n\n\n\n\n 89\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 91 of90\n 95of 94\n\n\n\n\nwebsite, providing applications that allow users to access those services, widely advertising those\n\nservices, and providing technical support and instructions to users (see\n\nhttps://www.nvidia.com/en-us/data-center/data-center-gpus/gpu-test-drive/); and 2) through\n\nactivities relating to marketing, advertising, promotion, installation, support, and distribution of\n\nthe Accused Products, including its CUDA platform, and services in the United States. (See\n\nhttps://www.nvidia.com/en-us/; see https://www.nvidia.com/en-us/about-nvidia/partners/;\n\nhttps://www.nvidia.com/en-us/data-center/where-to-buy/; https://www.nvidia.com/en-us/data-\n\ncenter/where-to-buy-tesla/.)\n\n 170. For example, Defendant shares instructions, guides, and manuals, which advertise\n\nand instruct third parties on how to use its hardware and platform as described above, including at\n\nleast customers and partners. (See https://docs.nvidia.com/cuda/cuda-c-programming-guide/.)\n\nDefendant also provides customer service and technical support to purchasers of the Accused\n\nProducts and corresponding systems and services, which directs and encourages customers to\n\nperform certain actions that use the Accused Products in an infringing manner. (See\n\nhttps://www.nvidia.com/en-us/support/; https://www.nvidia.com/en-\n\nus/support/enterprise/services/.)\n\n 171. Defendant and/or Defendant\u2019s partners recommend and sell the Accused Products\n\nand provide technical support for the installation, implementation, integration, and ongoing\n\noperation of the Accused Products for each individual customer. On information and belief, each\n\ncustomer enters into a contractual relationship with Defendant and/or one of Defendant\u2019s partners,\n\nwhich obligates each customer to perform certain actions in order to use the Accused Products.\n\n(See https://www.nvidia.com/en-us/agreements/; https://www.nvidia.com/en-\n\nus/agreements/cloud-services/nvidia-cloud-agreement/; https://www.nvidia.com/en-\n\n\n\n\n 90\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 92 of91\n 95of 94\n\n\n\n\nus/agreements/cloud-services/service-specific-terms-for-nvidia-dgx-cloud/.) Further, in order to\n\nreceive the benefit of Defendant\u2019s and/or its partner\u2019s continued technical support and their\n\nspecialized knowledge and guidance of the operability of the Accused Products, each customer\n\nmust continue to use the Accused Products in a way that infringes the \u2019461 Patent. (See\n\nhttps://www.nvidia.com/en-us/support/.)\n\n 172. Further, as the entity that provides installation, implementation, and integration of\n\nthe Accused Products in addition to ensuring the Accused Product remains operational for each\n\ncustomer through ongoing technical support, on information and belief, Defendant and/or\n\nDefendant\u2019s partners affirmatively aid and abet each customer\u2019s use of the Accused Products in a\n\nmanner that performs the claimed method of, and infringes, the \u2019461 Patent.\n\n 173. Defendant also contributes to the infringement of its partners, customers, and users\n\nof the Accused Products by providing within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice, the\n\nmethods, systems, and devices claimed in the Asserted Patents, constituting a material part of the\n\ninventions claimed, and not a staple article or commodity of commerce suitable for substantial\n\nnon-infringing uses. Indeed, as shown above, the Accused Products and the example functionality\n\nhave no substantial non-infringing uses but are specifically designed to practice the \u2019461 Patent.\n\n 174. On information and belief, the infringing actions of each partner, customer, and/or\n\nuser of the Accused Products are attributable to Defendant. For example, on information and belief,\n\nDefendant directs and controls the activities or actions of its partners or others in connection with\n\nthe Accused Products by contractual agreement or otherwise requiring partners or others to provide\n\ninformation and instructions to customers who acquire the Accused Products which, when\n\nfollowed, results in infringement. Defendant further directs and controls the operation of devices\n\n\n\n\n 91\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 93 of92\n 95of 94\n\n\n\n\nexecuting the Accused Products by programming the software which, when executed by a\n\ncustomer or user, performs the claimed method of at least claim 21 of the \u2019461 Patent.\n\n 175. Plaintiff has suffered and continues to suffer damages as a result of Defendant\u2019s\n\ninfringement of the \u2019461 Patent. Defendant is therefore liable to Plaintiff under 35 U.S.C. \u00a7 284\n\nfor damages in an amount that adequately compensates Plaintiff for Defendant\u2019s infringement, but\n\nno less than a reasonable royalty.\n\n 176. Defendant\u2019s infringement of the \u2019461 Patent is knowing and willful. Defendant\n\nacquired actual knowledge of the family of the \u2019461 Patent since at least 2017 and has acquired\n\nadditional knowledge of the \u2019461 Patent since at least the filing of this lawsuit.\n\n 177. On information and belief, despite Defendant\u2019s knowledge of the Asserted Patents and\n\nPlaintiff\u2019s patented technology, Defendant made the deliberate decision to sell products and services\n\nthat it knew infringe these patents. Defendant\u2019s continued infringement of the \u2019461 Patent with\n\nknowledge of the \u2019461 Patent constitutes willful infringement.\n\n\n\n\n 92\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 94 of93\n 95of 94\n\n\n\n\n PRAYER FOR RELIEF\n\nWHEREFORE, Plaintiff respectfully requests the following relief:\n\n a) That this Court adjudge and decree that Defendant has been, and is currently,\n\n infringing each of the Asserted Patents;\n\n b) That this Court award Plaintiff damages to compensate for Defendant\u2019s past and\n\n future infringement of the Asserted Patents, through the life of the Asserted Patents;\n\n c) That this Court award Plaintiff pre- and post-judgment interest on such;\n\n d) That this Court order an accounting of damages incurred by Plaintiff from six years\n\n prior to the date this lawsuit was filed through entry of a final, non-appealable\n\n judgment;\n\n e) That this Court determine that this patent infringement case is exceptional and\n\n award Plaintiff its costs and attorneys\u2019 fees incurred in this action;\n\n f) That this Court award increased damages under 35 U.S.C. \u00a7 284; and\n\n g) That this Court award such other relief as the Court deems just and proper.\n\n DEMAND FOR JURY TRIAL\n\n Plaintiff respectfully requests a trial by jury on all issues triable thereby.\n\n\n\n\n 93\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-2 30\n Filed Filed\n 08/17/26\n 12/12/24\n PagePage\n 95 of94\n 95of 94\n\n\n\nDATED: December 12, 2024\n By:/s/ Mark D. Siegmund\n Mark D. Siegmund\n Texas Bar No. 24117055\n CHERRY JOHNSON SIEGMUND JAMES\n PLLC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n Telephone: (254) 732-2242\n Facsimile: (866) 627-3509\n msiegmund@cjsjlaw.com\n\n Christopher C. Campbell\n KING & SPALDING LLP\n 1700 Pennsylvania Avenue, NW\n Suite 900\n Washington, DC 20006\n Telephone: (202) 626-5578\n Facsimile: (202) 626-3737\n ccampbell@kslaw.com\n\n Britton F. Davis\n Brian Eutermoser (pro hac vice to be filed)\n KING & SPALDING LLP\n 1401 Lawrence Street\n Suite 1900\n Denver, CO 80202\n Telephone: (720) 535-2300\n Facsimile: (720) 535-2400\n bfdavis@kslaw.com\n beutermoser@kslaw.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned does hereby certify that a true and correct copy of the foregoing document\n\nwas served on all counsel of record via the Court\u2019s electronic filing system on this 12th day of\n\nDecember 2024.\n\n By:/s/ Mark D. Siegmund\n Mark D. Siegmund\n\n\n\n 94\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:26.137167-07:00","document_number":"1","attachment_number":2,"pacer_doc_id":"181037209212","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 1","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294673/","id":490294673,"tags":[],"absolute_url":"/docket/74659430/1/3/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.059147-07:00","date_modified":"2026-08-21T18:46:57.396075-07:00","sha1":"d4be6a4b94edbceb3867a17b7d85272438aceb78","page_count":15,"file_size":1311450,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.3.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.3.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 1 of 15\n\n\n\n\n EXHIBIT\n\n 2\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 2 of 15\n\n USOO8648867B2\n\n\n(12) United States Patent (10) Patent No.: US 8,648,867 B2\n Gorchetchnikov et al. (45) Date of Patent: Feb. 11, 2014\n(54) GRAPHIC PROCESSORBASED (56) References Cited\n ACCELERATOR SYSTEMAND METHOD\n U.S. PATENT DOCUMENTS\n(75) Inventors: Anatoli Gorchetchnikov, Belmont, MA 5,388,206 A * 2/1995 Poulton et al. ................ 345,505\n (US); Heather Marie Ames, South 2005, 0166042 A1* 7, 2005 Evans ............ T13,150\n Boston, MA (US); Massimiliano 2007/0052713 A1* 3/2007 Chung et al. ... 34.5/5O1\n Versace, South Boston, MA (US); 2007/0279429 A1* 12/2007 Ganzer ......................... 345,582\n Fabrizio Santini, Jamaica Plain, MA\n (US) * cited by examiner\n(73) Assignee: Neurala LLC, Boston, MA (US) Primary Examiner \u2014 Maurice L. McDowell, Jr.\n(*) Notice: Subject to any disclaimer, the term of this (57) ABSTRACT\n patent is extended or adjusted under 35 An accelerator system is implemented on an expansion card\n U.S.C. 154(b) by 1030 days. comprising a printed circuit board having (a) one or more\n graphics processing units (GPU), (b) two or more associated\n(21) Appl. No.: 11/860,254 memory banks (logically or physically partitioned), (c) a\n(22) Filed: Sep. 24, 2007 specialized controller, and (d) a local bus providing signal\n coupling compatible with the PCI industry standards (this\n(65) Prior Publication Data includes but is not limited to PCI-Express, PCI-X, USB 2.0,\n or functionally similar technologies). The controller handles\n US 2008/O 117220 A1 May 22, 2008 most of the primitive operations needed to set up and control\n Related U.S. Application Data GPU computation. As a result, the computer's central pro\n cessing unit (CPU) is freed from this function and is dedicated\n(60) Provisional application No. 60/826,892, filed on Sep. to other tasks. In this case a few controls (simulation start and\n 25, 2006. stop signals from the CPU and the simulation completion\n signal back to CPU), GPU programs and input/output data are\n(51) Int. C. the information exchanged between CPU and the expansion\n G06F 5/00 (2006.01) card. Moreover, since on every time step of the simulation the\n(52) U.S. C. results from the previous time step are used but not changed,\n USPC ........................................... 345/501; 34.5/503 the results are preferably transferred back to CPU in parallel\n(58) Field of Classification Search with the computation.\n USPC .................................................. 345/501,503\n See application file for complete search history. 19 Claims, 5 Drawing Sheets\n\n 140 200\n 8O 250\n -4\n\n\n SHADERE.K.\n MEMORY\n\n 50\n\n O EXRE\n - icon TROLLER w NEMORY\n\n C\n -\n\n\n\n\n 2\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 3 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 1 of 5 US 8,648,867 B2\n\n\n\n\n 7.\n\n\n\n\n $3.\n\n\n\n\n s\n\n\n\n\n FG,\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 4 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 2 of 5 US 8,648,867 B2\n\n\n\n\n g\n i.\n :\n i\n : r\n s\n al e3esiasti Od\n\n\n s\n\n 3\n\n\n s\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 5 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 3 of 5 US 8,648,867 B2\n\n\n\n\n 300\n CPU 120 Start\n 304\n Expansion Cardiso :\n iaia fict iss: Risitio: &::::::::8:::::::::\n Sires 301 Se::303\n 320\n Disk:O Graphic User Controller\n Interface initialization\n initialization Initialization\n\n\n\n\n 325\n User input Parser External input\n Texture textures from RAM to\n interaction Generator texture memory bank\n 326 :\n Population Parser Population shader\n Shader Generato binaries from RAM to\n and Compiler shader memory bank\n\n\n\n\n Simulation\n Initialization\n\n\n 330\n\n\n Progress Input Parser\n monitor Texture\n Generator Computation\n (see Fig. 4)\n Data output\n to Disk\n Output Data Data\n ACCumulation issils\n in RAM\n\n Last\n iteration\n 318 :\n\n Yes\n 350\n\n\n\n\n 3\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 6 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 4 of 5 US 8,648,867 B2\n\n\n Expansion Card 180\n\n Simulation\n start\n\n iDesi: {34}{ut : {iciplisiii) 2. iaia iiii:\n Si:Sires: Siii:Sire:: Silisirai?\n 402 403 404\n Input textures from\n texture memory\n bank to GPU\n\n\n\n Shaders from\n shader memory\n bank to GPU\n\n\n ... ...........\n\n\n\n\n Shader\n execution\n\n\n\n\n Output texture\n upload to texture\n memory bank\n\n New external input Data\n textures from RAM told -----------\n\n\n\n\n Wait for swap Swap input Wait for swap\n of input/output and output of input/output\n texture pointers texture pointers texture pointers\n\n 475\n Input textures from NO Last\n texture memory iteration?\n bank to RAM\n Yes\n\n\n\n iteration?\n Yes 490\n Wait for all three\n streams of\n execution to finish\n\n 499\n\n\n\n EIG 4\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 7 of 15\n\n\nU.S. Patent Feb. 11, 2014 Sheet 5 of 5 US 8,648,867 B2\n\n\n\n\n (p)\n\n\n\n\n *\u00b7,,?)&xaddO2OInputDependency(m gate1, O., O., 0.004, O., 0, 0);\nInteraction Stream, Data Output Stream, and Computational u->addFullDependency(m gate2, population());\nStream. Through TSimulator:timestep(), TSimulator:out 40\nfileInterval ( ), and TSimulator:outmode(), the application public: TCablePopRCF(): TPopulation(\u201ccompCPU RCF, w, h, true)\ncan set the time step of the simulation, the time step of disk { };\noutput, and the mode of the disk output. The external input ~TCablePopRCF() {if(m equation) delete m equation;\n if(m gate1) delete m gate1;\npattern should be packed into a TPattern object and bound to\nthe simulation object through TSimulator:resetInputs( ) 45 bool if(m gate2) delete m gate2:};\n fillElements(TSimulator sim);\nmethod. TSimulator::simLength( ) sets the length of the }:\nsimulation. bool TCablePopRCF::fillElements(TSimulatior sim)\n The second step is to create at least one population of { equation = new TEq RCF (this, m compet, m persist);\nequations (TPopulation object). Population holds one equa mcreateCatingStructure();\ntion object TEquation. This object contains only a formula 50 for(size ti = 0; i < x.Size(); ++i)\nand does not hold element-specific data, so all elements of the for(size tj = 0; j < ySize(); ++)\npopulation can share single TEquation. {\n The TEquation object is converted to a GPU program TElement u = new TCPUElement(this, m equation, i,j):\n sim->addUnit(u);\nbefore execution. GPU programs have to be executed within createUnitStructure(u);\na graphical context, which is stream specific. TSimulator 55 return true:\ncreates this context within a Computational Stream, therefore\nall programs and data arrays that are necessary for computa int\ntion have to be initialized within Computational Stream. Con main()\nstructor of TPopulation is called from User Interaction //{ Input pattern generation (309 in FIG. 3)\nStream, so no GPU-related objects can be initialized in this 60 uint32 t pat = new uint32 twh;\nCOnStructOr. TRandom-float randGen (O);\n TPopulation::fillElements( ) is a virtual method designed for(uint32 ti = 0; i < wh; ++i)\nto overcome this difficulty. It is called from within the Com pati = randGen.random ();\nputational Stream after TSimulator::networkCreate( ) is TPattern p = new TPattern (pat, w, h);\n if Setting up the simulation\ncalled in the User Interaction Stream. A user has to override 65 TSimulator cableSim = new TSimulator(\u201cdata'); fi(308 and 320 in\nTPopulation::fillElements( ) to create TEquation and other FIG. 3)\ncomputation related objects both element independent and\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 13 of 15\n\n\n US 8,648,867 B2\n 11 12\n -continued need to perform the computations and initiates the upload 435\n of them onto the GPU 240. The GPU 240 can communicate\ncableSim->timestep(0.05); fi(320 in FIG. 3)\ncableSim->resetInputs(p); (325 in FIG. 3) directly with the texture memory bank 250 to upload the\ncableSim->OutfileInterval(0.1); (308 in FIG. 3) appropriate texture to perform the computations. The control\ncableSim->Outmode(SANNDRA::timefunc); (308 in FIG. 3)\ncableSim->simLength(60.0); (320 in FIG. 3)\n ler 220 also pulls the first shader (known by the stored order)\nif Preparing the population from the shader memory bank 210 and uploads 450 it onto the\nTPopulation* cablePop = new TCablePopRCF(); //(310 in FIG. 3) GPU 240.\ncableSim->networkCreate(); //(326 in FIG. 3) The GPU 240 executes the following operations in this\nuint16 t user = 1; order: performs the computation (execution of the shader)\nwhile(user) 10\n 470; tells the controller 220 that it is done with the computa\nif(cableSim->simulationStart(true, 1)) (311 in FIG. 3) tions for the current shader; and after all shaders for this\nexit(1): particular equation are executed sends 480 the output textures\nstd::cout-\u201cRepeat?\\n: //(305 in FIG. 3)\nstd::cin>user; //(305 in FIG. 3) to the output portion of the texture memory bank 250. This\nif(user == 1) 15 cycle continues through all of the equations based on the\ncableSim->networkReset(); //(305 in FIG. 3) branching step 482.\nif cableSim) An example shader that performs fourth order Runge\ndelete cableSim; Also deletes cablePop and its internals Kutta numerical integration is shown in Listing 2 using GLSL\nexit(0); notation;\nListing 1.\n\n FIG. 4 is a detailed flow diagram illustrating a part of an uniform sampler2DRect Variable;\n uniform float integration step;\nexemplary implementation of the bottom level system and float halfstep = integration step*0.5;\nmethod performed during the computation on the GPU accel 25 float fl. 6 step = integration stepf 6.0:\nerator of the expansion card 180 and is a more detailed view vec4 output = texture2DRect(Variable, gl TexCoord O.st);\n if define equation() here\nof the computational box 330 in FIG. 3. FIG. 4 is a represen vec4 rungekutta4(vec4 x)\ntation of one of several ways in which a system and method\nfor processing numerical techniques can be implemented. const vecA k1 = equation(x);\n const vec4 k2 = equation(x + halfstep*k1);\n With systems of equations that have complex interdepen 30 const vec4 k3 = equation(x + halfstep*k2);\ndencies it is likely that the variable in Some equation from a const vecA k4 = equation(x + integration Step*k3);\nprevious time step has to be used by some other equation after return fl. 6step*(k1 + 2.0* (k2 + k3) + k4);\nthe new values of this variable are already computed for new\ntime step. To avoid data confusion, the new values of variables void main (void)\n {\nshould be rendered in a separate texture. After the time step is 35 output += rungekutta4(output);\ncompleted for all equations, these new values should be cop gl FragColor = output;\nied over old values so that they are used as input during the Listing 2.\nnext time step. Copying textures is an expensive operation,\ncomputationally, but since the textures are referred to by\ntexture IDs (pointers), Swapping these pointers for input and 40 The shader in Listing 2 can be executed on conventional\noutput textures after each time step achieves the same resultat video card. Using the controller 220 this code can be further\na much lesser cost. optimized, however. Since the integration step does not\n In the hardware solution suggested herein, ID Swapping is change during the simulation, the step itself as well as the\nequivalent to Swapping the base memory address for two halfstep and /6 of the step can be computed once per simula\npartitions of the texture memory bank 250. They are swapped 45 tion, and updated in all shaders by a shader update procedures\n485 during synchronization (485, 430, and 455) so that data\ntransfer 445 and the computation 435-487 proceeds immedi 310,326 discussed above.\nately and in parallel with data transfer as shown in FIG. 4. A computed theofmain\n After all the equations in the computational cycle are\nhardware solution allows this parallelism through access of 220 can switch 485execution the\n substream 403 on the controller\n reference pointers of the input and\nthe controller 220 to the onboard texture memory bank 250. 50 output portions of the texture memory bank 250.\n The main computation and data exchange are executed by The two other substreams of execution on the controller\nthe controller 220. It runs three parallel substreams of execu\ntion: Computational Substream 403, Data Output Substream 220 are waiting (blocks 430 and 455, respectively) for this\n402, and Data Input Substream 404. These streams are syn switch to begin their execution. The Data Input Substream\nchronized with each other during the swap of pointers 485 to 55 404 is controlling 440 the input of additional data from the\nthe input and output texture memory partitions of the texture CPU 120. This is necessary in cases where the simulation is\nmemory bank 250 and the check for the last iteration 487. monitoring the changing input, for example input from a\nAlgorithmically, these two operations are a single atomic video camera or other recording device in the real time. This\noperation, but the block diagram shows them as two separate substream uploads new external input from the CPU 120 to\nblocks for clarity. 60 the texture memory bank 250 so it can be used by the main\n The Computational Substream 403 performs a computa computational Substream 403 on the next computational step\ntional cycle including a sequential execution of all shaders and waits for the next iteration 475. The Data Output Sub\nthat were stored in the shader memory bank 210 using the stream 445 controls the output of simulation results to the\nappropriate input and output textures. To begin the simulation CPU 120 if requested by the user. This substream uploads the\nthe controller 220 initializes three execution substreams 403, 65 results of the previous step to the main RAM 130 so that the\n402, and 404. On every simulation step, the Computational CPU 120 can save them on disk 140 or show them on the\nSubstream 403 determines which textures the GPU 240 will results display 313 and waits for the next iteration 460.\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 14 of 15\n\n\n US 8,648,867 B2\n 13 14\n Since the Computational Substream 403 determines the While this invention has been particularly shown and\ntiming of input 440 and output 445 data transfers, these data described with references to preferred embodiments thereof,\ntransfers are driven by the controller 220. To further reduce it will be understood by those skilled in the art that various\nthe data transfer overhead (and disk 140 overhead also) the changes in form and details may be made therein without\ncontroller 220 initiates transfer only after selected computa departing from the scope of the invention encompassed by the\ntional steps. For example, if the experimental data that is appended claims.\nsimulated was recorded every 10 milliseconds (msec) and the\nsimulation for better precision was computed every 1 mSec, What is claimed is:\nthen only every tenth result has to be transferred to match the 1. A computer system for performing a numerical simula\nexperimental frequency. 10 tion over a plurality of computational cycles including at least\n This solution stores two copies of output data, one in the a first computational cycle and a second computational cycle,\nexpansion card texture memory bank 250 and another in the the computer system comprising:\nsystem RAM 130. The copy in the system RAM 130 is a central processing unit;\naccessed twice: for disk I/O and screen visualization 313. An a main memory, operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a 15 unit, to store input data to be accessed by the central\ndirect read access to the onboard texture memory bank 250 by processing unit in performing the numerical simulation;\nmapping the memory of the hardware onto a global memory a video system, operably coupled to the central processing\nspace. The alternative solution will double the communica unit, to drive a video monitor to display an indication of\ntion through the local bus 190. Since the goal discussed herein the numerical simulation in response to the computer\nis reducing the information transfer through the local bus 190, system performing the numerical simulation;\nthe former solution is favored. an accelerator, operably coupled to the central processing\n The main stream substream 403 determines if this is the last unit, to receive at least a portion of the input data from\niteration 487. If it is the last iteration, the controller 220 waits the central processing unit and to provide first output\nfor the all of the execution substreams to finish 490 and then data generated during the first computational cycle to the\nreturns the control to the CPU 120, otherwise it begins the 25 central processing unit after a conclusion of the first\nnext computational cycle. computational cycle, the accelerator comprising:\n This repeats through all of the computational cycles of the at least one graphics processing unit to generate second\nsimulation. output data, during the second computational cycle,\n Conclusion by performing at least one calculation on the first\n This GPU accelerator system offers the following potential 30 output data; and\nadvantages: an accelerator memory, operably coupled to the at least\n 1. Limited computations on the CPU 120. The CPU 120 is one graphic processing unit, the accelerator memory\nonly used for user input, sending information to the controller comprising:\n220, receiving output after each computational cycle (or less a first partition, referenced by a first pointer, to store\nfrequently as defined by the user), writing this output to disk 35 the first output data during the second computa\n140, and displaying this output on the monitor 170. This frees tional cycle; and\nthe CPU 120 to execute other applications and allows the a second partition, referenced by a second pointer, to\nexpansion card to run at its full capacity without being slowed store the second output data generated during the\ndown by extensive interactions with the CPU 120. second computational cycle; and\n 2. Minimizing data transfer between the expansion card 40 an accelerator controller, operably coupled to the accelera\n180 and the system bus 200. All of the information needed to tor memory and the central processing unit, to transfer\nperform the simulations will be stored on the expansion card the at least the portion of the input data into the accel\n180 and all simulations will take place on it. Furthermore, erator memory before the first computational cycle, to\nwhatever data transfer remains necessary will take place in transfer the first output data from the accelerator\nparallel with the computation, thus reducing the impact of this 45 memory to the main memory during the second compu\ntransfer on the performance. tational cycle, to direct the second output data into the\n 3. New way to execute GPU programs (shaders). Previ second partition during the second computational cycle,\nously, the CPU 120 had full control over the order of shaders and to Swap the first pointer and the second pointer at the\nexecution and was required to produce specific commands on conclusion of the second computational cycle Such that\nevery cycle to tell the GPU 240 which shader to use. With the 50 the second output data becomes an input for a third\ninvention disclosed herein, shaders will initially be stored on computational cycle of the plurality of computational\nthe shader memory bank 210 on the expansion card 180 and cycles.\nwill be sent to the GPU 240 for execution by the general 2. The computer system as claimed in claim 1, wherein the\npurpose controller 220 located on the expansion card. accelerator controller is configured to dictate an order of\n 4. Multiple parallelisms. The GPU 240 is inherently paral 55 execution of instructions to the at least one graphics process\nlel and is well suited to perform parallel computations. In ing unit.\nparallel with the GPU 240 performing the next calculation, 3. The computer system as claimed in claim 1, wherein the\nthe controller 220 is uploading the data from the previous accelerator controller comprises an interface controller to\ncalculation into main memory 130. Furthermore, the CPU communicate with the central processing unit over a bus of\n120 at the same time uses uploaded previous results to save 60 the computer system.\nthem onto disk 140 and to display them on the screen through 4. The computer system as claimed in claim 1, wherein the\nthe system bus 200. accelerator memory comprises a texture memory bank to\n 5. Reuse of existing and affordable technology. All hard store the at least the portion of the input data and the first\nware used in the invention and mentioned here-in are based on output data and a shader memory bank to store instructions\ncurrently available and reliable components. Further advance 65 for performing a set of operations to be performed on the at\nof these components will provide straightforward improve least the portion of the input data by the at least one graphic\nments of the invention. processing unit.\n\f Case 7:26-cv-00318 Document 1-3 Filed 08/17/26 Page 15 of 15\n\n\n US 8,648,867 B2\n 15 16\n 5. The computer system as claimed in claim 4, wherein the bank to store instructions for processing operations to be\ntexture memory is partitioned into the first partition, the sec performed on the input data by the at least one graphic pro\nond partition, a third partition to store internal variables, a cessing unit.\nfourth partition to store data textures used as input at a par 13. The accelerator system as claimed in claim 12, wherein\nticular computation cycle of the plurality of computational 5 the texture memory is partitioned into a first partition to store\ncycles. the input data, a second partition to store internal variables, a\n 6. The computer system as claimed in claim 1, wherein the third partition to store data textures used as input at a particu\naccelerator controller inputs the at least the portion of the lar computation cycle of the numerical simulation, and a\ninput data and a series of instructions into the at least one fourth partition to store the output data.\ngraphic processing unit, wherein the at least one graphics 10\nprocessing unit then executes the instructions on the at least the14. The accelerator system as claimed in claim 9, wherein\n accelerator controller is configured to perform successive\nthe portion of the input data. computational cycles of the numerical simulation by feeding\n 7. The computer system as claimed in claim 1, wherein the the output data generated\naccelerator controller comprises a set of instructions stored ing unit from a previous by the at least one graphic process\n computational cycle and the input\non a memory. 15\n 8. The computer system as claimed in claim 1, wherein the data for a next computational cycle into the at least one\nat least the portion of the input data represents an initial graphic processing unit.\ncondition of the numerical simulation. 15. The accelerator system as claimed in claim 9, wherein\n 9. An accelerator system for a computer system performing the accelerator controller comprises a set of instructions\na numerical simulation, the accelerator system comprising: 20 stored in a memory.\n 16. A method for performing a numerical simulation on\n at least one graphics processing unit to generate output data input data in a computer system including a central process\n by performing at least one computation during a first\n computational cycle of the numerical simulation; ing unit and an accelerator, the method comprising:\n an accelerator memory, operably coupled to the at least one receiving, by an accelerator, first input data from the central\n processing unit;\n graphics processing unit, to store data used to perform 25 transferring,\n the at least one computation; and by an accelerator controller, the first input\n an accelerator controller, operably coupled to the accelera data into a first partition, referenced by first pointer. ofan\n tor memory and the at least one graphics processing unit, accelerator memory before a first computational cycle of\n to execute: the numerical simulation;\n (i) a computational stream controlling performance of 30 performing, by at least one graphics processing unit during\n the first computational cycle, at least one calculation on\n the at least one computation by the at least one graph the first portion of the input data as to generate first\n ics processing unit; output data;\n (ii) an output stream controlling transfer of the output storing, by the accelerator controller, the first output data\n data from the at least one graphics processing unit to into a second partition, referenced by a second pointer,\n the accelerator memory during the first computational 35 of the accelerator memory; and\n cycle; and\n (iii) an input stream controlling transfer of input data to Swapping the first pointer with the second pointer at the end\n the accelerator memory for use by the at least one of the first computational cycle, such that the first output\n graphics processing unit during a second computa data becomes an input for a second computational cycle\n of the numerical simulation.\n tional cycle of the numerical simulation. 40\n 10. The accelerator system as claimed in claim 9, wherein 17. The method as claimed inclaim 16, further comprising:\nthe accelerator controller is configured to dictate an order of sending, by the accelerator controller, instructions for per\nexecution of instructions to the at least one graphics process forming the at least one calculation to the at least one\ning unit. graphics processing unit.\n 11. The accelerator system as claimed in claim 9, wherein 45 18. The method as claimed in claim 16, further comprising:\nthe accelerator controller is configured to send instructions to partitioning the accelerator memory into the first partition,\nthe at least one graphics processing unit, wherein the at least the second partition, a third partition to store internal\none graphics processing unit is configured to execute the Variables, and a fourth partition to store data used as\ninstructions on the input data, and wherein the accelerator input at a particular computation cycle of the numerical\n simulation.\ncontroller is configured to transfer the output data from the 50 19. The method of claim 16, further comprising:\naccelerator memory to a main memory during the execution transferring, by the accelerator controller, the first output\nof the instructions by the at least one graphics processing unit. data to the main memory during the second computa\n 12. The accelerator system as claimed in claim 9, wherein tional cycle.\nthe accelerator memory comprises a texture memory bank to\nStore the input data and the output data and a shader memory\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:19.582774-07:00","document_number":"1","attachment_number":3,"pacer_doc_id":"181037209213","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 2","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294674/","id":490294674,"tags":[],"absolute_url":"/docket/74659430/1/4/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.070720-07:00","date_modified":"2026-08-21T18:40:15.843093-07:00","sha1":"447fef62c4e286e8a7c47e2b9c6b70e303296f35","page_count":22,"file_size":2170082,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.4.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.4.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 1 of 22\n\n\n\n\n EXHIBIT\n\n 3\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 2 of 22\n\n USOORE48438E\n(( 1912)) United\n United States\n Reissued Patent ( 10 ) Patent Number : US RE48,438 E\n Gorchetchnikov et al . (45 ) Date of Reissued Patent : * Feb . 16, 2021\n( 54 ) GRAPHIC PROCESSOR BASED ( 56 ) References Cited\n ACCELERATOR SYSTEM AND METHOD\n U.S. PATENT DOCUMENTS\n( 71 ) Applicant: Neurala, Inc. , Boston, MA (US) 5,063,603 A 11/1991 Burt\n( 72 ) Inventors: Anatoli Gorchetchnikov , Belmont, MA 5,136,687 A 8/1992 Edelman et al .\n (US ) ; Heather Marie Ames , Milton , (Continued )\n MA (US ) ; Massimiliano Versace, FOREIGN PATENT DOCUMENTS\n Milton, MA (US ); Fabrizio Santini ,\n Jamaica Plain, MA (US) EP 1 224 622 B1 7/2002\n WO WO 2014/190208 11/2014\n( 73 ) Assignee : Neurala, Inc. , Boston, MA (US) ( Continued )\n( * ) Notice: This patent is subject to a terminal dis\n claimer . OTHER PUBLICATIONS\n( 21 ) Appl. No .: 15 /808,201 Cornwall et al . , \u201c Automatically Translating a General Purpose C ++\n Image Processing Library for GPUs \u201d , IEEE , Jun . 2006 , 8 pages .\n(22 ) Filed : Nov. 9 , 2017 ( Year: 2006 ) . *\n Related U.S. Patent Documents ( Continued )\nReissue of:\n( 64) Patent No .: 9,189,828 Primary Examiner William H. Wood\n Issued: Nov. 17 , 2015 (74 ) Attorney, Agent, or Firm -Smith Baluch LLP\n Appl. No .: 14 /147,015 ( 57 ) ABSTRACT\n Filed : Jan. 3 , 2014\nU.S. Applications: An accelerator system is implemented on an expansion card\n( 63 ) Continuation of application No. 11 / 860,254 , filed on comprising a printed circuit board having ( a) one or more\n Sep. 24 , 2007 , now Pat . No. 8,648,867 . graphics processing units ( GPUs ) , ( b ) two or more associ\n (Continued ) ated memory banks ( logically or physically partitioned ), (c )\n a specialized controller, and ( d) a local bus providing signal\n( 51 ) Int. Ci. coupling compatible with the PCI industry standards. The\n GO6T 1/60 ( 2006.01 ) controller handles most of the primitive operations to set up\n G06F 9/50 ( 2006.01 ) and control GPU computation. Thus, the computer's central\n (Continued ) processing unit ( CPU) can be dedicated to other tasks . In this\n( 52 ) U.S. Ci . case a few controls ( simulation start and stop signals from\n CPC G06T 1/20 (2013.01 ) ; G06F 9/5027\n the CPU and the simulation completion signal back to CPU ),\n GPU programs and input/output data are exchanged between\n ( 2013.01 ) ; G06T 1/60 ( 2013.01 ) ; CPU and the expansion card . Moreover, since on every time\n ( Continued ) step of the simulation the results from the previous time step\n( 58 ) Field of Classification Search are used but not changed, the results are preferably trans\n CPC ... GO6F 9/5027 ; G06F 2209/509 ; G06T 1/20 ; ferred back to CPU in parallel with the computation .\n GO6T 1/60 ; G06N 99/005 ; GO6N 37063\n See application file for complete search history . 56 Claims , 5 Drawing Sheets\n Expansion Cards 40\n\n 420\n ? e.com\n\n 403 435\n xxtes\n 19x VI 9XY\n SKM 2\n 450\n Swarum\n shokolaty\n bars CPU\n\n\n 19:43 Shader GPU .\n\n 180\n Outdoor\n yoles exi\n Teney bank\n 440\n 482 Now external\n texmex Tom FRAM\n UM\n WOX + vary but\n\n\n Wait for SWRP Svima i wait for swap\n otrputloutpan\n xure points\n ? output\n texture packs\n otinescioutput\n extere gointate\n plexulante\n 478\n ******** * herbimity\n Dank AM leration harasana\n\n\n en?\n 490\n Waxa ste\n tra\n axexkoren\n 499\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 3 of 22\n\n\n US RE48,438 E\n Page 2\n\n Related U.S. Application Data 2014/0032461 Al 1/2014 Weng\n 2014/0089232 A1 3/2014 Buibas et al .\n( 60 ) Provisional application No. 60 /826,892 , filed on Sep. 2014/0052679 Al\n 2015/0127149 Al\n 11/2014 Sinyavskiy et al .\n 5/2015 Sinyavskiy et al .\n 25 , 2006 . 2015/0134232 Al 5/2015 Robinson\n 2015/0224648 A1 8/2015 Lee et al .\n( 51 ) Int . Ci . 2016/0075017 A1 3/2016 Laurent et al.\n GOOT 1/20 ( 2006.01 ) 2016/0082597 A1 3/2016 Gorchetchnikov et al .\n GOON 37063 (2006.01 ) 2016/0096270 A1 4/2016 Gabardos et al .\n GOON 20/00 2016/0198000 Al 7/2016 Gorchetchnikov et al .\n ( 2019.01 ) 2017/0024877 A1 1/2017 Versace et al .\n( 52) U.S. Ci. 2017/0076194 A1 3/2017 Versace et al .\n CPC G06F 2209/509 (2013.01 ) ; GO6N 37063 2017/0193298 Al 7/2017 Versace et al .\n (2013.01 ) ; GOON 20/00 (2019.01 ) FOREIGN PATENT DOCUMENTS\n( 56) References Cited WO WO 2014/204615 12/2014\n U.S. PATENT DOCUMENTS WO WO 2015/143173 9/2015\n WO WO 2016/014137 1/2016\n 5,142,665 A * 8/1992 Bigus GOON 3/04\n 706/16 OTHER PUBLICATIONS\n 5,172,253 A 12/1992 Lynne\n 5,388,206 A 2/1995 Poulton et al . Montrym et al . , \u201c The GeForce 6800 \u201d , IEEE , 2005 , 11 pages. ( Year :\n 6,018,696 A 1/2000 Matsuoka et al . 2005) . *\n 6,336,051 B1 1/2002 Pangels et al. Perumalla , Kalyan S. , \u201c Discrete - event Execution Alternatives on\n 6,647,508 B2 11/2003 Zalewski et al .\n 7,119,810 B2 * 10/2006 Sumanaweera G06T 15/005 General Purpose Graphical Processing Units (GPGPUs)\", IEEE ,\n 345/506 \u201c Proceedings of the 20th Workshop on Principles of Advanced and\n 7,219,085 B2 * 5/2007 Buck GOON 3/08 Distributed Simulation ( PADS'06 ) \u201d , Mar. 2006 , 8 pages . ( Year:\n 706/12 2006 ) . *\n 7,477,256 B1 * 1/2009 Johnson G06F 3/14 Setoain et al . , \u201c Parallel Hyperspectral Image Processing on Com\n 345/501 modity Graphics Hardware \u201d , IEEE , \u201c Proceedings of the 2006\n 7,525,547 B1 * 4/2009 Diard G06T 1/20 International Conference on Parallel Processing Workshops\n 345/502\n 7,765,029 B2 7/2010 Fleischer et al . ( ICPPW'06 ) \" , Mar. 2006 , 8 pages . ( Year: 2006 ) . *\n 7,861,060 B1 * 12/2010 Nickolls et al . 712/22 Luo et al ., \u201c Artificial Neural Network Computation on Graphic\n 7,873,650 B1 1/2011 Chapman et Process Unit \u201d , IEEE , \u201c Proceedings of International Joint Confer\n 8,392,346 B2 3/2013 Ueda et al . ence on Neural Networks, Montreal, Canada, Jul. 31 - Aug. 4 , 2005 \u201d ,\n 8,510,244 B2 8/2013 Carson et al . Feb. 2005 , pp . 622-626 ( Year: 2005 ) . *\n 8,583,286 B2 11/2013 Fleischer et al . Non - Final Office Action dated Jan. 4 , 2018 from U.S. Appl. No.\n 8,648,867 B2 * 2/2014 Gorchetchnikov et al. .. 345/501 15/262,637 , 23 pages.\n 9,031,692 B2 5/2015 Zhu Wu, Yan & J. Cai , H. ( 2010 ) . A Simulation Study of Deep Belief\n 9,177,246 B2 11/2015 Buibas et al .\n 9,626,566 B2 4/2017 Versace et al. Network Combined with the Self -Organizing Mechanism of Adap\n 10,083,523 B2 9/2018 Versace et al. tive Resonance Theory. 10.1109 /CISE.2010.5677265 , 4 pages .\n 2001/0010034 A1 * 7/2001 Burton G06F 17/5022 Al-Kaysi, A. M. et al . , A Multichannel Deep Belief Network for the\n 703/17 Classification of EEG Data, from Ontology -based Information\n 2002/0046271 A1 4/2002 Huang Extraction for Residential Land Use Suitability: A Case Study of the\n 2002/0050518 A1 5/2002 Roustaei City of Regina, Canada, DOI 10.1007/978-3-319-26561-2_5 , 8\n 2002/0064314 Al 5/2002 Comaniciu et al . pages (Nov. 2015 ) .\n 2002/0168100 A1 11/2002 Woodall\n 2003/0026588 A1 2/2003 Elder et al . Apolloni, B. et al . , Training a network of mobile neurons, Proceed\n 2003/0078754 Al 4/2003 Hamza ings of International Joint Conference on Neural Networks, San\n 2004/0015334 Al 1/2004 Ditlow et al. Jose , CA , doi: 10.1109 /IJCNN.2011.6033427, pp . 1683-1691 ( Jul.\n 2005/0166042 Al 7/2005 Evans 31 -Aug. 5 , 2011 ) .\n 2006/0129506 A1 * 6/2006 Edelman GO5D 1/0088 Boddapati, V., Classifying Environmental Sounds with Image Net\n 706/12 works, Thesis , Faculty of Computing Blekinge Institute of Tech\n 2006/0184273 A1 8/2006 Sawada et al. nology, 37 pages . ( Feb. 2017 ) .\n 2007/0052713 A1 3/2007 Chung et al .\n 2007/0198222 A1 8/2007 Schuster et al . Khaligh -Razavi, S.-M. et al . , Deep Supervised, but Not Unsuper\n 2007/0279429 A1 12/2007 Ganzer vised, Models May Explain IT Cortical Represen on, PLoS\n 2008/0033897 A1 2/2008 Lloyd Computational Biology , vol . 10 , Issue 11 , 29 pages . (Nov. 2014 ) .\n 2008/0066065 Al 3/2008 Kim et al . Kim , S. , Novel approaches to clustering , biclustering and algo\n 2008/0258880 A1 10/2008 Smith et al . rithms based on adaptive resonance theory and intelligent control,\n 2009/0080695 Al 3/2009 Yang et al . Doctoral Dissertations, Missouri University of Science and Tech\n 2009/0089030 Al 4/2009 Sturrock et al .\n 2009/0116688 A1 5/2009 Monacos et al . nology, 125 pages . ( 2016 ) .\n 2010/0048242 A1 2/2010 Rhoads et al . Notice of Alllowance dated May 22 , 2018 from U.S. Appl. No.\n 2011/0004341 Al 1/2011 Sarvadevabhatla et al . 15/262,637 , 6 pages.\n 2011/0173015 A1 7/2011 Chapman et al . Ren, Y. et al . , Ensemble Classification and Regression Recent\n 2011/0279682 A1 11/2011 Li et al . Developments, Applications and Future Directions , in IEEE Com\n 2012/0072215 A1 3/2012 Yu et al . putational Intelligence Magazine , 10.1109 /MCI.2015.2471235 , 14\n 2012/0089552 A1 4/2012 Chang et al . pages (2016) .\n 2012/0197596 A1 8/2012 Comi\n 2012/0316786 A1 12/2012 Liu et al . Sun , Z. et al . , Recognition of SAR target based on multilayer\n 2013/0126703 A1 5/2013 Caulfield auto -encoder and SNN , International Journal of Innovative Com\n 2013/0131985 Al 5/2013 Weiland et al . puting, Information and Control, vol . 9 , No. 11 , pp . 4331-4341 ,\n 2014/0019392 A1 1/2014 Buibas et al. Nov. 2013 .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 4 of 22\n\n\n US RE48,438 E\n Page 3\n\n( 56 ) References Cited Coifman , R.R. , Lafon , S. , Lee , A.B. , Maggioni , M. , Nadler, B. ,\n Warner, F. , and Zucker, S.W. Geometric diffusions as a tool for\n OTHER PUBLICATIONS harmonic analysis and structure definition of data : Diffusion maps.\n Proceedings of the National Academy of Sciences of the United\nAdelson , E. H. , Anderson, C. H. , Bergen , J. R. , Burt, P. J. , & Ogden, States of America , 102 ( 21 ) : 7426 , 2005 .\nJ. M. ( 1984 ) . Pyramid methods in image processing. RCA engineer, Davis, C. E. 2005. Graphic Processing Unit Computation of Neural\n29(6) , 33-41 . Networks. Master's thesis, University of New Mexico , Albuquer\nAggarwal, Charu C , Hinneburg, Alexander, and Keim , Daniel A. On que, NM , 121 pages .\nthe surprising behavior of distance metrics in high dimensional Dosher, B.A. , and Lu , Z.L. ( 2010 ) . Mechanisms of perceptual\nspace . Springer, 2001 . attention in precuing of location . Vision Res . , 40 ( 10-12 ) . 1269\nAmes, H , Versace , M. , Gorchetchnikov, A. , Chandler, B. , Livitz , G. , 1292 .\nL\u00e9veill\u00e9 , J. , Mingolla, E. , Carter, D. , Abdalla, H. , and Snider, G. Ellias , S. A. , and Grossberg, S. ( 1975 ) . Pattern formation, contrast\n( 2012 ) Persuading computers to act more like brains. In Advances control and oscillations in the short term memory of shunting\nin Neuromorphic Memristor Science and Applications, Kozma , on -center off - surround networks. Biol Cybern 20 , pp . 69-98 .\nR.Pino , R . , and Pazienza , G. ( eds ) , Springer Verlag . Extended European Search Report and Written Opinion dated Jun .\nAmes, H. Mingolla, E. , Sohail , A. , Chandler, B. , Gorchetchnikov, 1 , 2017 from European Application No. 14813864.7 , 10 pages .\nA. , L\u00e9veill\u00e9 , J. , Livitz , G. and Versace, M. ( 2012 ) The Animat . IEEE Extended European Search Report and Written Opinion dated Oct.\nPulse , Feb. 2012 , 3 ( 1 ) , 47-50 . 12 , 2017 from European Application No. 14800348.6 , 12 pages .\nArtificial Intelligence As a Service. Invited talk, Defrag, Broomfield , Extended European Search Report and Written Opinion Oct. 23 ,\nCO , Nov. 4-6 ( 2013 ). 2017 from European Application No. 15765396.5 , 8 pages .\nAryananda, L. ( 2006 ) . Attending to learn and learning to attend for Fazl , A. , Grossberg , S. , and Mingolla , E. ( 2009 ) . View - invariant\na social robot . Humanoids 06 , pp . 618-623 . object category learning, recognition, and search : How spatial and\nBaraldi, A. and Alpaydin , E. ( 1998 ) . Simplified ART: A new class object attention are coordinated using surface -based attentional\nof ART algorithms. International Computer Science Institute, Berke shrouds . Cognitive Psychology 58 , 1-48 .\nley, CA , TR- 98-004 , 1998 . F\u00f6ldi\u00e1k , P. ( 1990 ) . Forming sparse representations by local anti\nBaraldi, A. and Alpaydin , E. ( 2002 ) . Constructive feedforward ART Hebbian learning, Biological Cybernetics, vol . 64 , pp . 165-170 .\nclustering networks \u2014 Part I. IEEE Transactions on Neural Net Friston K. , Adams R. , Perrinet L. , & Breakspear M. ( 2012 ) . Per\nworks 13 ( 3 ) , 645-661 . ceptions as hypotheses: saccades as experiments . Frontiers in Psy\nBaraldi, A. and Parmiggiani, F. ( 1997 ) . Fuzzy combination of chology, 3 ( 151 ), 1-20 .\nKohonen's and ART neural network models to detect statistical Galbraith , B.V, Guenther, F.H. , and Versace , M. ( 2015 ) A neural\nregularities in a random sequence of multi - valued input patterns. In network -based exploratory learning and motor planning system for\nInternational Conference on Neural Networks, IEEE . co - robots.Frontiers in Neuroscience, in press .\nBaraldi, Andrea and Alpaydin , Ethem . Constructive feedforward George , D. and Hawkins, J. ( 2009 ) . Towards a mathematical theory\nART clustering networks \u2014 part II . IEEE Transactions on Neural of cortical micro - circuits . PLoS Computational Biology 5 ( 10 ) , 1-26 .\nNetworks, 13 ( 3 ) : 662-677 , May 2002. ISSN 1045-9227 . doi: 10.1109/ Georgeii , J. , and Westermann , R. ( 2005 ) . Mass - spring systems on\ntnn.2002.1000131\n1000131 .\n . URL http://dx.doi.org/10.1109/tnn.2002 . the GPU . Simulation Modelling Practice and Theory 13 , pp . 693\n 702 .\nBengio , Y., Courville, A. , & Vincent, P. Representation learning: A Gorchetchnikov A. , Hasselmo M.E. ( 2005 ) . A biophysical imple\nreview and new perspectives , IEEE Transactions on Pattern Analy mentation of a bidirectional graph search algorithm to solve mul\nsis and Machine Intelligence , vol . 35 Issue 8 , Aug. 2013 , pp . tiple goal navigation tasks . Connection Science , 17 ( 1-2 ) , pp . 145\n1798-1828 . 166 .\nBerenson , D. et al ., A robot path planning framework that learns Gorchetchnikov A. , Hasselmo M.E. ( 2005 ) . A simple rule for\nfrom experience, 2012 International Conference on Robotics and spike -timing -dependent plasticity : local influence of AHP current.\nAutomation, 2012 , 9 pages [ retrieved from the internet] URL : http : // Neurocomputing, 65-66 , pp . 885-890 .\nusers.wpi.edu/-dberenson/lightning.pdf. Gorchetchnikov A. , Versace, M. , Hasselmo M.E. ( 2005 ) . A Model\nBernhard , F. , and Keriven, R. ( 2005 ) . Spiking Neurons on GPUs. of STDP Based on Spatially and Temporally Local Information :\nTech . Rep. 05-15 , Ecole Nationale des Ponts et Chauss\u2019es, 8 pages. Derivation and Combination with Gated Decay. Neural Networks,\nBesl , P. J. , & Jain , R. C. ( 1985 ) . Three - dimensional object recog 18 , pp . 458-466 .\nnition . ACM Computing Surveys (CSUR ), 17 ( 1 ) , 75-145 . Gorchetchnikov A. , Versace, M. , Hasselmo M.E. ( 2005 ) . Spatially\nBohn , C.-A. Kohonen . ( 1998 ) . Feature Mapping Through Graphics and temporally local spiketiming -dependent plasticity rule. In :\nHardware. In Proceedings of 3rd Int. Conference on Computational Proceedings of the International Joint Conference on Neural Net\nIntelligence and Neurosciences , 4 pages . works, No. 1568 in IEEE CD - ROM Catalog No. 05CH37662C , pp .\nBradski, G. , & Grossberg, S. ( 1995 ) . Fast -learning Viewnet archi 390-396 .\ntectures for recognizing three - dimensional objects from multiple Gorchetchnikov A. An Approach to a Biologically Realistic Simu\ntwo - dimensional views. Neural Networks, 8 ( 7-8 ) , 1053-1080 . lation of Natural Memory . Master's thesis , Middle Tennessee State\nCanny, J.A. ( 1986 ) . Computational Approach to Edge Detection, University, Murfreesboro, TN , 70 pages .\nIEEE Trans. Pattern Analysis and Machine Intelligence , 8 (6 ):679 Grossberg, S. ( 1973 ) . Contour enhancement, short -term memory ,\n698 . and constancies in reverberating neural networks. Studies in Applied\nCarpenter, G.A. and Grossberg, S. ( 1987 ) . A massively parallel Mathematics 52 , 213-257 .\narchitecture for a self -organizing neural pattern recognition machine . Grossberg , S. , and Huang, T.R. ( 2009 ) . Artscene : A neural system\nComputer Vision, Graphics, and Image Processing 37 , 54-115 . for natural scene classification . Journal of Vision, 9 (4 ) , 6.1-19 .\nCarpenter, G.A. , and Grossberg , S. ( 1995 ) . Adaptive resonance doi : 10.1167 /9.4.6 .\ntheory ( ART ). In M. Arbib ( Ed . ) , The handbook of brain theory and Grossberg, S. , and Versace , M. ( 2008 ) Spikes , synchrony, and\nneural networks. (pp . 79-82 ) . Cambridge , M.A .: MIT press . attentive learning by laminar thalamocortical circuits . Brain Research ,\nCarpenter, G.A. , Grossberg, S. and Rosen , D.B. ( 1991 ) . Fuzzy ART: 1218C , 278-312 [ Authors listed alphabetically ].\nFast stable learning and categorization of analog patterns by an Hagen , T.R. , Hjelmervik, J. , Lie , K.-A. , Natvig , J. and Ofstad\nadaptive resonance system . Neural Networks 4 , 759-771 . Henriksen , M. ( 2005 ) . Visual simulation of shallow -water waves .\nCarpenter, Gail A and Grossberg, Stephen . The art of adaptive Simulation Modelling Practice and Theory 13 , pp . 716-726 .\npattern recognition by a self -organizing neural network . Computer, Hasselt , Hado Van . Double q -learning. In Advances in Neural\n21 ( 3 ) : 77-88 , 1988 . Information Processing Systems , pp . 2613-2621 , 2010 .\nCoifman , R.R. and Maggioni , M. Diffusion wavelets . Applied and Hinton , G. E. , Osindero , S. , and Teh , Y. ( 2006 ) . A fast learning\nComputational Harmonic Analysis, 21 ( 1 ) : 53-94 , 2006 . algorithm for deep belief nets. Neural Computation, 18 , 1527-1554 .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 5 of 22\n\n\n US RE48,438 E\n Page 4\n\n( 56 ) References Cited L\u00e9veill\u00e9 , J. , Ames, H. , Chandler, B. , Gorchetchnikov, A. , Mingolla ,\n E. , Patrick , S. , and Versace, M. ( 2010 ) Learning in a distributed\n OTHER PUBLICATIONS software architecture for large -scale neural modeling. BIONET\n ICS10 , Boston , MA , USA .\nHodgkin , A.L. , and Huxley, A.F. ( 1952 ) . Quantitative description of Livitz , G. , Versace, M. , Gorchetchnikov, A. , Vasilkoski, Z. , Ames,\nmembrane current and its application to conduction and excitation H. , Chandler, B. , L\u00e9veill\u00e9 , J. , Mingolla , E. , Snider, G. , Amerson , R. ,\nin nerve. J Physiol 117 , pp . 500-544 . Carter, D. , Abdalla, H. , and Qureshi , S. ( 2011 ) Visually -Guided\nHopfield , J. ( 1982 ) . Neural networks and physical systems with Adaptive Robot ( ViGUAR ). Proceedings of the International Joint\nemergent collective computational abilities . In Proc Natl Acad Sci Conference on Neural Networks ( IJCNN) 2011 , San Jose , CA ,\nUSA , vol . 79 , pp . 2554-2558 . USA .\nIlie , A. ( 2002 ) . Optical character recognition on graphics hardware . Lowe , D.G. ( 2004 ) . Distinctive Image Features from Scale - Invariant\nTech . Rep . integrative paper, UNCCH , Department of Computer Keypoints. Journal International Journal of Computer Vision archive\nScience, 9 pages . vol . 60 , 2 , 91-110 .\nInternational Preliminary Report on Patentability in related PCT Lu , Z.L. , Liu , J. , and Dosher, B.A. ( 2010 ) Modeling mechanisms of\nApplication No. PCT/US2014 /039162 filed May 22 , 2014 , dated perceptual learning with augmented Hebbian re -weighting. Vision\nNov. 24 , 2015 , 7 pages. Research , 50 ( 4 ) . 375-390 .\nInternational Preliminary Report on Patentability in related PCT Mahadevan, S. Proto -value functions: Developmental reinforce\nApplication No. PCT/US2014 /039239 filed May 22 , 2014 , dated ment learning. In Proceedings of the 22nd international conference\nNov. 24 , 2015 , 8 pages. on Machine learning, pp . 553-560 . ACM , 2005 .\nInternational Preliminary Report on Patentability dated Nov. 8 , Meuth , J.R. and Wunsch , D.C. ( 2007 ) A Survey of Neural Compu\n2016 from International Application No. PCT/US2015 /029438, 7 tation on Graphics Processing Hardware. 22nd IEEE International\npages. Symposium on Intelligent Control, Part of IEEE Multi-conference\nInternational Search Report and Written Opinion dated Feb. 18 , on Systems and Control, Singapore, Oct. 1-3 , 2007 , 5 pages .\n2015 from International Application No. PCT /US2014 /039162, 12 Mishkin M , Ungerleider LG . ( 1982 ) . \" Contribution of striate inputs\npages. to the visuospatial functions of parieto -preoccipital cortex in mon\nInternational Search Report and Written Opinion dated Feb. 23 , keys,\u201d Behav Brain Res, 6 ( 1 ) : 57-77 .\n2016 from International Application No. PCT /US2015 /029438 , 11 Mnih , Volodymyr, Kavukcuoglu , Koray, Silver, David , Rusu , Andrei\npages . A , Veness, Joel , Bellemare, Marc G , Graves, Alex , Riedmiller,\nInternational Search Report and Written Opinion dated Jul. 6 , 2017 Martin , Fidjeland, Andreas K , Ostrovski, Georg , et al . Human -level\nfrom International Application No. PCT/US2017 /029866 , 12 pages . control through deep reinforcement learning. Nature, 518 (7540 ):529\nInternational Search Report and Written Opinion dated Nov. 26 , 533 , Feb. 25 , 2015 .\n2014 from International Application No. PCT /US2014 /039239, 14 Moore , Andrew W and Atkeson, Christopher G. Prioritized sweep\npages. ing : Reinforcement learning with less data and less time . Machine\nInternational Search Report and Written Opinion dated Sep. 15 , Learning, 13 ( 1 ) : 103-130 , 1993 .\n2015 from International Application No. PCT /US2015 /021492 , 9 Najemnik, J. , and Geisler, W. ( 2009 ) . Simple summation rule for\npages. optimal fixation selection in visual search . Vision Research . 49 ,\nItti, L. , and Koch , C. ( 2001 ) . Computational modelling of visual 1286-1294 .\nattention . Nature Reviews Neuroscience, 2 ( 3 ) , 194-203 . Notice of Allowance dated Jul. 27 , 2016 from U.S. Appl. No.\nItti, L. , Koch , C. , and Niebur, E. ( 1998 ) . A Model of Saliency - Based 14/662,657 .\nVisual Attention for Rapid Scene Analysis, 1-6 . Notice of Allowance dated Dec. 16 , 2016 from U.S. Appl. No.\nJarrett, K. , Kavukcuoglu , K. , Ranzato , M. A. , & LeCun , Y. ( Sep. 14/662,657 .\n2009 ) . What is the best multi - stage architecture for object recog Oh , K.-S. , and Jung, K. ( 2004 ) . GPU implementation of neural\nnition ?. In Computer Vision , 2009 IEEE 12th International Confer networks. Pattern Recognition 37 , pp . 1311-1314 .\nence on (pp . 2146-2153 ) . IEEE . Oja, E. ( 1982 ) . Simplified neuron model as a principal component\nKipfer, P., Segal , M. , and Westermann , R. ( 2004 ) . UberFlow : A analyzer. Journal of Mathematical Biology 15 ( 3 ) , 267-273 .\nGPU - Based Particle Engine . In Proceedings of the SIGGRAPH/ Partial Supplementary European Search Report dated Jul. 4 , 2017\nEurographics Workshop on Graphics Hardware 2004 , pp . 115-122 . from European Application No. 14800348.6 , 13 pages .\nKolb , A. , L. Latta , and C. Rezk - Salama. ( 2004 ) . \u201c Hardware - Based Raijmakers, M.E.J. , and Molenaar, P. ( 1997 ) . Exact ART: A com\nSimulation and Collision Detection for Large Particle Systems.\u201d in plete implementation of an ART network Neural networks 10 (4 ) ,\nProceedings of the SIGGRAPH /Eurographics Workshop on Graph 649-669 .\nics Hardware 2004 , pp . 123-131 . Ranzato , M. A. , Huang, F. J. , Boureau , Y. L. , & Lecun , Y. ( Jun .\nKompella, Varun Raj, Luciw , Matthew , and Schmidhuber, J\u00fcrgen. 2007 ) . Unsupervised learning of invariant feature hierarchies with\nIncremental slow feature analysis: Adaptive low - complexity slow applications to object recognition . In Computer Vision and Pattern\nfeature updating from high -dimensional input streams. Neural Com Recognition, 2007. CVPR\u201907 . IEEE Conference on (pp . 1-8 ) . IEEE .\nputation , 24 ( 11 ) : 2994-3024 , 2012 . Raudies, F. , Eldridge, S. , Joshi , A. , and Versace, M. ( Aug. 20 , 2014 ) .\nKowler, E. ( 2011 ) . Eye movements: The past 25years. Vision Learning to navigate in a virtual world using optic flow and stereo\nResearch , 51 ( 13 ) , 1457-1483 . doi : 10.1016 / j.visres.2010.12.014 . disparity signals. Artificial Life and Robotics, DOI 10.1007/ s10015\nLarochelle H. , & Hinton G. ( 2012 ) . Learning to combine foveal 014-0153-1 .\nglimpses with a third - order Boltzmann machine . NIPS 2010,1243 Riesenhuber, M. , & Poggio , T. ( 1999 ) . Hierarchical models of\n1251 . object recognition in cortex . Nature Neuroscience , 2 ( 11 ) , 1019\nLeCun , Y., Kavukcuoglu, K. , & Farabet, C. (May 2010 ) . Convolu 1025 .\ntional networks and applications in vision . In Circuits and Systems Riesenhuber, M. , & Poggio , T. ( 2000 ) . Models of object recogni\n( ISCAS ) , Proceedings of 2010 IEEE International Symposium on tion . Nature neuroscience , 3 , 1199-1204 .\n( pp . 253-256 ) . IEEE . Rolfes, T. ( 2004 ) . Artificial Neural Networks on Programmable\nLee , D. D. and Seung, H. S. ( 1999 ) . Learning the parts of objects Graphics Hardware . In Game Programming Gems 4 , A. Kirmse, Ed .\nby non -negative matrix factorization. Nature, 401 ( 6755 ) : 788-791 . Charles River Media , Hingham , MA , pp . 373-378 .\nLee , D. D. , and Seung , H. S. ( 1997 ) . \u201c Unsupervised learning by Rublee , E. , Rabaud, V., Konolige , K. , & Bradski, G. ( 2011 ) . ORB :\nconvex and conic coding .\u201d Advances in Neural Information Pro An efficient alternative to SIFT or SURF. In IEEE International\ncessing Systems , 9 . Conference on Computer Vision ( ICCV ) 2011 , 2564-2571 .\nLegenstein , R. , Wilbert, N. , and Wiskott, L. Reinforcement learning Ruesch, J. et al . ( 2008 ) . Multimodal Saliency -Based Bottom -Up\non slow features of high - dimensional input streams. PLoS Compu Attention a Framework for the Humanoid Robot iCub . 2008 IEEE\ntational Biology, 6 ( 8 ) , 2010. ISSN 1553-734X . International Conference on Robotics and Automation , pp . 962-965 .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 6 of 22\n\n\n US RE48,438 E\n Page 5\n\n( 56 ) References Cited Tong , F. , Ze -Nian Li , ( 1995 ) . Reciprocal -wedge transform for\n space - variant sensing, Pattern Analysis and Machine Intelligence,\n OTHER PUBLICATIONS IEEE Transactions on , vol . 17 , No. 5 , pp . 500-51 . doi: 10.1109 /\n 34.391393 .\nRumelhart D. , Hinton G. , and Williams, R. ( 1986 ) . Learning inter Torralba, A. , Oliva , A. , Castelhano, M.S. , Henderson, J.M. ( 2006 ) .\nnal representations by error propagation. In Parallel distributed Contextual guidance of eye movements and attention in real -world\nprocessing : explorations in the microstructure of cognition , vol . 1 , scenes: the role of global features in object search . Psychological\nMIT Press . Review , 113 ( 4 ) .766-786 .\nRumpf, M. and Strzodka, R. Graphics processor units : New pros Van Hasselt , Hado, Guez, Arthur, and Silver, David . Deep rein\npects for parallel computing. In Are Magnus Bruaset and Aslak forcement learning with double q - learning. arXiv preprint arXiv :\n 1509.06461 , Sep. 22 , 2015 .\nTveito , editors , Numerical Solution of Partial Differential Equations Versace, M. ( 2006 ) From spikes to interareal synchrony: how\non Parallel Computers, vol . 51 of Lecture Notes in Computational attentive matching and resonance control learning and information\nScience and Engineering , pp . 89-134 . Springer, 2005 . processing by laminar thalamocortical circuits . NSF Science of\nSchaul, Tom , Quan , John , Antonoglou, Ioannis, and Silver, David . Learning Centers PI Meeting, Washington , DC , USA .\nPrioritized experience replay. arXiv preprint arXiv: 1511.05952 , Versace, M. , ( 2010 ) Open - source software for computational neu\nNov. 18 , 2015 . roscience: Bridging the gap between models and behavior. In\nSchmidhuber, J. ( 2010 ) . Formal theory of creativity, fun , and Horizons in Computer Science Research ,vol. 3 .\nintrinsic motivation ( 1990-2010 ) . Autonomous Mental Develop Versace, M. , Ames, H. , L\u00e9veill\u00e9, J. , Fortenberry, B. , and Gorchetchnikov,\n A. ( 2008 ) KinNeSS : A modular framework for computational\nment, IEEE Transactions on , 2 ( 3 ) , 230-247 . neuroscience. Neuroinformatics, 2008 Winter ; 6 (4 ) : 291-309 . Epub\nSchmidhuber, J\u00fcrgen . Curious model -building control systems. In Aug 10 , 2008 .\nNeural Networks, 1991. 1991 IEEE International Joint Conference Versace, M. , and Chandler, B. ( 2010 ) MONETA : A Mind Made from\non , pp . 1458-1463 . IEEE , 1991 . Memristors. IEEE Spectrum , Dec. 2010 .\nSeibert, M. , & Waxman , A.M. ( 1992 ) . Adaptive 3 - D Object Rec Webster, Bachevalier, Ungerleider ( 1994 ) . Connections of IT areas\nognition from Multiple Views. IEEE Transactions on Pattern Analy TEO and TE with parietal and frontal cortex in macaque monkeys.\nsis and Machine Intelligence , 14 ( 2 ) , 107-124 . Cerebal Cortex, 4 ( 5 ) , 470-483 .\nSherbakov, L. and Versace, M. ( 2014 ) Computational principles for Wiskott, Laurenz and Sejnowski, Terrence . Slow feature analysis:\nan autonomous active vision system . Ph.D., Boston University,\n Unsupervised learning of invariances. Neural Computation , 14 (4 ):715\n 770 , 2002 .\nhttp://search.proquest.com/docview/1558856407. Livitz G. , Versace M. , Gorchetchnikov A. , Vasilkoski Z. , Ames H. ,\nSherbakov, L. , Livitz , G. , Sohail , A. , Gorchetchnikov, A. , Mingolla , Chandler B. , Leveille J. and Mingolla E. ( 2011 ) Adaptive, brain -like\nE. , Ames, H. , and Versace, M. ( 2013a) CogEye: An online active systems give robots complex behaviors, The Neuromorphic Engi\nvision system that disambiguates and recognizes objects. NeuComp neer, : 10.2417 / 1201101.003500 Feb. 2011. 3 pages .\n2013 . Salakhutdinov, R. , & Hinton, G. E. ( 2009 ) . Deep boltzmann machines.\nSherbakov, L. , Livitz , G. , Sohail , A. , Gorchetchnikov, A. , Mingolla , In International Conference on Artificial Intelligence and Statistics\nE. , Ames, H. , and Versace, M ( 2013b) A computational model of the (pp . 448-455 ).\nrole of eye -movements in object disambiguation . Cosyne, Feb. Sherbakov, L. et al . 2012. CogEye : from active vision to context\n28 -Mar. 3 , 2013. Salt Lake City, UT, USA . identification, youtube, retrieved from the Internet on Oct. 10 , 2017 :\nSmolensky, P. ( 1986 ) . Information processing in dynamical sys URL : //www.youtube.com/watch ? v = i5PQk962B1k, 1 page .\ntems : Foundations of harmony theory. In D. E. Sherbakov, L. et al. 2013. CogEye: system diagram module brain\nSpratling, M. W. ( 2008 ) . Predictive coding as a model of biased area function algorithm approx # neurons, retrieved from the\ncompetition in visual attention . Vision Research, 48 ( 12 ) : 1391-1408 . Internet on Oct. 12 , 2017 : URL : //http ://www-labsticc.univ-ubs.\nSpratling, M. W. ( 2012 ) . Unsupervised learning of generative and fr / ~ coussy /neucomp2013 / index_fichiers /material / posters /\ndiscriminative weights encoding elementary image components in a NeuComp2013_final56x36.pdf, 1 page .\npredictive coding model of cortical function . Neural Computation , Snider, Greg, et al . \u201c From synapses to circuitry : Using memristive\n24 ( 1 ) : 60-103 . memory to explore the electronic brain . \u201d IEEE computer, vol . 44 ( 2 ) .\nSpratling, M. W. , De Meyer, K. , and Kompass , R. ( 2009 ) . Unsu (2011 ) : 21-28 .\npervised learning of overlapping image components using divisive Versace, TEDx Fulbright, Invited talk, Washington DC , Apr. 5 ,\ninput modulation . Computational intelligence and neuroscience . 2014. 30 pages .\nSprekeler, H. On the relation of slow feature analysis and laplacian Versace , Brain - inspired computing. Invited keynote address, Bionet\neigenmaps. Neural Computation, pp . 1-16 , 2011 . ics 2010 , Boston , MA , USA . 1 page .\nSutton , Richard S and Barto , Andrew G. Reinforcement learning:\nAn introduction. MIT Press , 1998 . * cited by examiner\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 7 of 22\n\n\nU.S. Patent Feb.i6, 2021 Sheet 1 of 5 US RE48,438 E\n\n\n\n\n 14? *\n\n\n\n\n 130\n ????????????\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 8 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 2 of 5 US RE48,438 E\n\n\n\n\n I\n\n\n 220\n\n\n I\n\n\n\n\n RAM\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 9 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 3 of 5 US RE48,438 E\n\n\n\n 13:25 326 330 FIG\n 3\n .\n 320\n\n\n ECXPANRSIDON CONTROLER INITIALIZATION\n 180\n EIXNTEPRUNATL TFEXRTUOREMS RTOEXATUMRE MBEAMNORKY POPULATION SBIHNADRIERS FRAMTO SMHEAMDOERRY BANK COMPUTATION )\n 4\n .\n (\n SEE\n FIG\n\n\n\n\n COMPUTATIONAL 3ST0RE3AM\n 309\n DATA\n 310\n PSDAHRASDTEAR GAENERADTOR\n 316\n BATA\n 314\n DATA\n 350\n\n 304\n IPANRPSUETR TEXTURE GENERATOR POPULATION COMPILER IPANRPSUETR TGENXETRAUTROER ODUATPUAT ACUMULATION INRAM\n 315\n\n 306 LAST T?ERATION YES\n 311 312\n\n 300\n UINTSERAECTIRON 3ST0RE2AM 307 GUINRTSAEPREFHAIRCE 305 SIMULATION INITIALIZATION PROGRES MONITOR RESULTS DISPLAY NO\n 390\n START GURSAPEHIRC INTERFACE INITIALIZATION INTERACTION USER\n 313\n END\n\n\n 308 317 318\n CPU120\n DOUATTPUAT 3ST0RE1AM\n O\n /\n I\n DISK\n\n\n INITIALIZATION DOUATPTUAT TODISK\n NO\n LAST I?TERATION YES\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 10 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 4 of 5 US RE48,438 E\n\n\n Expansion Card 180\n\n\n\n Isoss foxxos ir\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 11 of 22\n\n\nU.S. Patent Feb. 16 , 2021 Sheet 5 of 5 US RE48,438 E\n\n\n\n\n Blsthepofdiwfc.amnicpxklotuniolndwahrstg 5\n .\n FIG\n\n\n\n\n 2\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 12 of 22\n\n\n US RE48,438 E\n 1 2\n GRAPHIC PROCESSOR BASED for input /output should be designed so that it provides the\n ACCELERATOR SYSTEM AND METHOD synchronization with computation .\n In the case of GPGPU , the computation itself is performed\n outside\nMatter enclosed in heavy brackets [ ] appears in the 5 \u201c peripheral of the CPU , so the complete system comprises three\noriginal patent but forms no part of this reissue specifica hardware, and\u201d components: user interactive hardware , disk\ntion ; matter printed in italics indicates the additions ing unit ( CPUcomputational\n ) establishes\n hardware. The central process\n communication and synchroni\nmade by reissue ; a claim printed with strikethrough zation between peripherals. Each of the peripherals is pref\nindicates that the claim was canceled, disclaimed, or held erably controlled by a dedicated thread that is executed in\ninvalid by a prior post- patent action or proceeding . 10 parallel with minimal interactions and dependencies on the\n other threads .\n RELATED APPLICATIONS A GPU on a conventional video card is usually controlled\n through OpenGL , DirectX , or similar graphic application\n The present application is a broadening reissue applica programming\ntion of U.S. Pat. No. 9,189,828, filed Jan. 3 , 2014, which 15 context of graphic interfaces (APIs ). Such APIs establish the\nclaims a priority benefit, under 35 U.S.C. $ 120 , as a con GPU are made . This operations, within which all calls to the\ntinuation of U.S. application Ser. No. 11 / 860,254 , now U.S. within the same threadcontext\n of\n only works when initialized\n execution that uses it . As a result,\nPat . No. 8,648,867 B2 , filed Sep. 24 , 2007 , entitled \u201c Graphic in a preferred embodiment, the context is initialized within\nProcessor Based Accelerator System and Method , \u201d which in\nturn claims the priority benefit, under 35 U.S.C. 8119 (e ) , of 20 aevercomputational thread. This creates complications , how\n , in the interaction between the user interface thread that\nU.S. Application No. 60/ 826,892 , filed Sep. 25 , 2006. Each\nof the above - identified applications is incorporated herein by changes parameters of simulations and the computational\nreference in its entirety. More than one reissue application thread that uses these parameters.\nhas been filed for the reissue of U.S. Pat. No. 9,189,828 , A solution as proposed here is an implementation of the\nincluding this application and a reissue continuation appli- 25 computational stream of execution in hardware, so that\ncation filed Dec. 29, 2020. thread and context initialization are replaced by hardware\n initialization . This hardware implementation includes an\n BACKGROUND expansion card comprising a printed circuit board having ( a )\n one or more graphics processing units , (b ) two or more\n Graphics Processing Units (GPUs) are found in video 30 associated memory banks that are logically or physically\nadapters ( graphic cards) of most personal computers ( PCs ) , partitioned, (c ) a specialized controller, and (d) a local bus\nvideo game consoles , workstations, etc. and are considered providing signal coupling compatible with the PCI industry\nhighly parallel processors dedicated to fast computation of standards ( this includes but is not limited to PCI -Express ,\ngraphical content. With the advances of the computer and PCI - X , USB 2.0 , or functionally similar technologies ). The\nconsole gaming industries, the need for efficient manipula- 35 controller handles most of the primitive operations needed to\ntion and display of 3D graphics has accelerated the devel- set up and control GPU computation . As a result, the CPU\nopment of GPUs . is freed from this function and is dedicated to other tasks . In\n In addition , manufacturers of GPUs have included general this case a few controls ( simulation start and stop signals\npurpose programmability into the GPU architecture leading from the CPU and the simulation completion signal back to\nto the increased popularity of using GPUs for highly paral- 40 CPU) , GPU programs and input/output data are the infor\nlelizable and computationally expensive algorithms outside mation exchanged between CPU and the expansion card .\nof the computer graphics domain . When implemented on Moreover, since on every time step of the simulation the\nconventional video card architectures, these general purpose results from the previous time step are used but not changed ,\nGPU ( GPGPU) applications are not able to achieve optimal the results are preferably transferred back to CPU in parallel\nperformance , however. There is overhead for graphics- 45 with the computation .\nrelated features and algorithms that are not necessary for In general, according to one aspect , the invention features\nthese non-video applications. a computer system . This system comprises a central pro\n cessing unit , main memory accessed by the central process\n SUMMARY ing unit , and a video system for driving a video monitor in\n 50 response to the central processing unit as is common . The\n Numerical simulations, e.g. , finite element analysis , of computer system further comprises an accelerator that uses\nlarge systems of similar elements ( e.g. neural networks, input data from and provides output data to the central\ngenetic algorithms, particle systems, mechanical systems) processing unit. This accelerator comprises at least one\nare one example of an application that can benefit from graphics processing unit, accelerator memory for the graphic\nGPGPU computation . During numerical simulations, disk 55 processing unit, and an accelerator controller that moves the\nand user input /output can be performed independently of input data into the at least one graphics processing unit and\ncomputation because these two processes require interac- the accelerator memory to generate the output data .\ntions with peripheral hardware ( disk , screen , keyboard , In the preferred , the central processing unit transfers the\nmouse , etc ) and put relatively low load on the central input data for a simulation to the accelerator, after which the\nprocessing unit/system (CPU) . Complete independence is 60 accelerator executes simulation computations to generate\nnot desirable , however; user input might affect how the the output data, which is transferred to the central processing\ncomputation is performed and even interrupt it if necessary. unit. Preferably, the accelerator controller dictates an order\nFurthermore, the user output and the disk output are depen- of execution of instructions to the at least one graphics\ndent on the results of the computation. A reasonable solution processing unit . The use of the separate controller enables\nwould be to separate input/output into threads, so that it is 65 data transfer during execution such that the accelerator\ninteracting with hardware occurs in parallel with the com- controller transfers output data from the accelerator memory\nputation . In this case whatever CPU processing is required to main memory of the central processing unit .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 13 of 22\n\n\n US RE48,438 E\n 3 4\n In the preferred embodiment, the accelerator controller not limited to , workstations, server computers, supercom\ncomprises an interface controller that enables the accelerator puters, notebook computers, hand -held electronic devices\nto communicate over a bus of the computer system with the such as cell phones, mp3 players, or personal digital assis\ncentral processing unit. tants ( PDAs ) , multiprocessor systems, programmable con\n In general according to another aspect , the invention also 5 sumer electronics, networks of any of the above -mentioned\nfeatures an accelerator system for a computer system , which computing devices, and distributed computing environments\ncomprises at least one graphics processing unit , accelerator that including any of the above -mentioned computing\nmemory for the graphic processing unit and an accelerator devices.\ncontroller for moving data between the at least one graphics 10 In one implementation the GPU accelerator is imple\nprocessing unit and the accelerator memory .\n In general according to another aspect , the invention also mented as an expansion card 180 includes connections with\nfeatures a method for performing numerical simulations in a are installed along110with\n the motherboard , on which the one or more CPU's 120\n main , or system memory 130 and\ncomputer system . This method comprises a central process mass / non volatile data storage 140 , such as hard drive or\ning unit loading input data into an accelerator system from redundant array of independent drives (RAID ) array, for the\nmain memory of the central processing unit and an accel- 15 computer system 100. In the current example\nerator controller transferring the input data to a graphics card 180 communicates to the motherboard ,110 the expansion\nprocessing unit with instructions to be performed on the bus 190. This local bus 190 could be PCI , PCIviaExpress a local\ninput data . The accelerator controller then transfers output PCI - X , or any other functionally similar technology (de,\ndata generated by the graphic processing unit to the central 20 pending upon the availability on the motherboard 110 ) . An\nprocessing unit as output data .\n The above and other features of the invention including external version GPU accelerator is also a possible imple\nvarious novel details of construction and combinations of mentation . In this example, the external GPU accelerator is\nparts, and other advantages, will now be more particularly connected to the motherboard 110 through USB - 2.0 , IEEE\ndescribed with reference to the accompanying drawings and 1394 (Firewire ), or similar external /peripheral device inter\npointed out in the claims . It will be understood that the 25 face .\nparticular method and device embodying the invention are The CPU 120 and the system memory 130 on the moth\nshown by way of illustration and not as a limitation of the erboard 110 and the mass data storage system 140 are\ninvention . The principles and features of this invention may preferably independent of the expansion card 180 and only\nbe employed in various and numerous embodiments without communicate with each other and the expansion card 180\ndeparting from the scope of the invention . 30 through the system bus 200 located in the motherboard 110 .\n A system bus 200 in current generations of computers have\n BRIEF DESCRIPTION OF THE DRAWINGS bandwidths from 3.2 GB / s (Pentium 4 with AGTL + , Athlon\n XP with EVO ) to around 15 GB / s (Xeon Woodcrest with\n In the accompanying drawings, reference characters refer AGTL + , Athlon 64 /Opteron with Hypertransport), while the\nto the same parts throughout the different views. The draw- 35 local bus has maximal peak data transfer rates of 4 GB / s\nings are not necessarily to scale ; emphasis has instead been (PCI Express 16 ) or 2 GB / s ( PCI -X 2.0 ) . Thus the local bus\nplaced upon illustrating the principles of the invention . Of 190 becomes a bottleneck in the information exchange\nthe drawings: between the system bus 200 and the expansion card 180. The\n FIG . 1 is a schematic diagram illustrating a computer design of the expansion card and methods proposed herein\nsystem including the GPU accelerator according to an 40 minimizes the data transfer through the local bus 190 to\nembodiment of the present invention ; reduce the effect of this bottleneck .\n FIG . 2 is block diagram illustrating the architecture for the The system memory 130 is referred to as the main\nGPU accelerator according to an embodiment of the present random - access memory (RAM ) in the description herein .\ninvention; However, this is not intended to limit the system memory\n FIG . 3 is a block / flow diagram illustrating an exemplary 45 130 to only RAM technology. Other possible computer\nimplementation of the top level control of the GPU accel- storage media include , but are not limited to ROM ,\nerator system ; EEPROM , flash memory, or any other memory technology.\n FIG . 4 is a flow diagram illustrating an exemplary imple- In the illustrated example, the GPU accelerator system is\nmentation of the bottom level control of the GPU accelerator implemented on an expansion card 180 on which the one or\nsystem that is used to execute the target computation ; and 50 more GPU's 240 are mounted . It should be noted that the\n FIG . 5 is an example population of nine computational GPU accelerator system GPU 240 is separate from and\nelements arranged in a 3x3 square and a potential packing independent of any GPU on the standard video card 150 or\nscheme for texture pixels , according to an implementation of other video driving hardware such as integrated graphics\nthe present invention . systems. Thus the computations performed on the expansion\n 55 card 180 do not interfere with graphics display ( including\n DETAILED DESCRIPTION but not limited to manipulation and rendering of images ) .\n Various brand of GPU are relevant. Under current tech\n FIG . 1 shows a computer system 100 that has been nology, GPU's based on the GeForce series from NVIDIA\nconstructed according to the principles of the present inven- Corporation or the Catalyst series from ATI/ Advanced\ntion . 60 Micro Devices, Inc.\n In more detail, the computer system 100 in one example The output to a video monitor 170 is preferably through\nis a standard personal computer ( PC ) . However, this only the video card 150 and not the GPU accelerator system 180 .\nserves as an example environment as computing environ- The video card 150 is dedicated to the transfer of graphical\nment 100 does not necessarily depend on or require any information and connects to the motherboard 110 through a\ncombination of the components that are illustrated and 65 local bus 160 that is sometimes physically separate from the\ndescribed herein . In fact, there are many other suitable local bus 190 that connects the expansion card 180 to the\ncomputing environments for this invention, including, but motherboard 110 .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 14 of 22\n\n\n US RE48,438 E\n 5 6\n FIG . 2 is a block diagram illustrating the general archi- not require hardware implementation. Also the partitioning\ntecture of the GPU accelerator system and specifically the scheme is also altered based on new designs or needs of the\nexpansion card 180 in which at least one GPU 240 and algorithms being employed. The reason for this partitioning\nassociated memories 210 and 250 are mounted . Electrical is further explained in the Data Organization section, below .\n( signal) and mechanical coupling with a local bus 190 5 A local bus interface 230 on the controller 220 serves as\nprovides signal coupling compatible with the PCI industry a driver that allows the controller 220 to communicate\nstandards ( this includes but is not limited to PCI , PCI -X , PCI through the local bus 190 with the system bus 200 and thus\nExpress, or functionally similar technology ). the CPU 120 and RAM 130. This local bus interface 230 is\n The GPU accelerator further preferably comprises one not intended to be limited to PCI related technology. Other\nspecifically designed accelerator controller 220. Depending 10 drivers can be used to interface with comparable technology\nupon the implementation , the accelerator controller 220 is as a local bus 190 .\nfield programmable gate array ( FPGA ) logic , or custom built Data Organization\napplication - specific (ASIC ) chip mounted in the expansion Each computational element discussed above has output\ncard 180 , and in mechanical and signal coupling with the variables that affect the rest of the system . For example in\nGPU 240 and the associated memories 210 and 250. During 15 the case of a neural network it is the output of a neuron . A\ninitial design , a controller can be partially or even fully computational element also usually has several internal\nimplemented in software, in one example. variables that are used to compute output variables , but are\n The controller 220 commands the storage and retrieval of not exposed to the rest of the system , not even to other\narrays of data ( on a conventional video card the arrays of elements of the same population, typically. Each of these\ndata are represented as textures, hence the term ' texture' in 20 variables is represented as a texture . The important differ\nthis document refers to a data array unless specified other- ence between output variables and internal variables is their\nwise and each element of the texture is a pixel of color access .\ninformation ), execution of GPU programs (on a conven- Output variables are usually accessed by any element in\ntional video card these programs are called shaders , hence the system during every time step . The value of the output\nthe term \u201c shader ' in this document refers to a GPU program 25 variable that is accessed by other elements of the system\nunless specified otherwise ), and data transfer between the corresponds to the value computed on the previous, not the\nsystem bus 200 and the expansion card 180 through the local current, time step . This is realized by dedicating two textures\nbus 190 which allows communication between the main to output variables \u2014 one holds the value computed during\nCPU 120 , RAM 130 , and disk 140 . the previous time step and is accessible to all computational\n Two memory banks 210 and 250 are mounted on the 30 elements during the current time step , another is not acces\nexpansion card 180. In some example, these memory banks sible to other elements and is used to accumulate new values\nseparated in the hardware, as shown, or alternatively imple- for the variable computed during the current time step .\nmented as a single , logically partitioned memory compo- In -between time steps these tw te res are switched , so\nnent. that newly accumulated values serve as accessible input\n The reason to separate the memory into two partitions 210 35 during the next time step , while the old input is replaced with\n250 stems from the nature of the computations to which the new values of the variable. This switch is implemented by\nGPU accelerator system is applied . The elements of com- swapping the address pointers to respective textures as\nputation ( computational elements ) are characterized by a described in the System and Framework section .\nsingle output variable . Such computational elements often Internal variables are computed and used within the same\ninclude one or more equations. Computational elements are 40 computational element. There is no chance of a race con\nsame or similar within a large population and are computed dition in which the value is used before it is computed or\nin parallel. An example of such a population is a layer of after it has already changed on the next time step because\nneurons in an artificial neural network (ANN ), where all within an element the processing is sequential. Therefore, it\nneurons are described by the same equation. As a result, is possible to render the new value of internal variable into\nsome data and most of the algorithms are common to all 45 the same texture where the old was read from in the texture\ncomputational elements within population , while most of the memory bank . Rendering to more than one texture from a\ndata and some algorithms are specific for each equation . single shader is not implemented in current GPU architec\nThus, one memory , the shader memory bank 210 , is used to tures, so computational elements that track internal variables\nstore the shaders needed for the execution of the required would have to have one shader per variable . These shaders\ncomputations and the parameters that are common for all 50 can be executed in order with internal variables computed\ncomputational elements and is coupled with the controller first, followed by output variables .\n220 only. The second memory, the texture memory bank Further savings of texture memory is achieved through\n250 , is used to store all the necessary data that are specific using multiple color components per pixel ( texture element)\nfor every computational element (including, but not limited to hold data . Textures can have up to four color components\nto , input data , output data , intermediate results, and param- 55 that are all processed in parallel on a GPU . Thus, to\neters ) and is coupled with both the controller 220 and the maximize the use of GPU architecture it is desirable to pack\nGPU 240 . the data in such a way that all four components are used by\n The texture memory bank 250 is preferably further par- the algorithm . Even though each computational element can\ntitioned into four sections . The first partition 250 a is have multiple variables, designating one texture pixel per\ndesigned to hold the external input data patterns. The second 60 element is ineffective because internal variables require one\npartition 250b is designed to hold the data textures repre- texture and output variables require two textures . Further\nsenting internal variables. The third partition 250c is more , different element types have different numbers of\ndesigned to hold the data textures used as input at a variables and unless this number is precisely a multiple of\nparticular computation step on the GPU 240. The fourth four, texture memory can be wasted .\npartition 250d holds the data textures used to accommodate 65 A more reasonable packing scheme would be to pack four\nthe output of a particular computational step on the GPU computational elements into a pixel and have separate\n240. This partitioning scheme can be done logically , does textures for every variable associated with each computa\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 15 of 22\n\n\n US RE48,438 E\n 7 8\ntional element. In this case the packing scheme is identical The crucial feature of the interaction between the User\nfor all textures, and therefore can be accessed using the same Interaction Stream 302 and the Computational Stream 303 is\nalgorithm . Several ways to approach this packing scheme the shift of priorities. Outside of the simulation , the system\nare outlined here. An example population of nine computa- 100 is driven by the user input, thus the User Interaction\ntional elements arranged in a 3x3 square (FIG . 5a ) can be 5 Stream 302 has the priority and controls the data exchange\npacked by element (FIG . 5b ) , by row (FIG . 5c ) , or by square 304 between streams. After the user starts the simulation , the\n( FIG . 5d) . Computational Stream 303 takes the priority and controls\n Packing by element ( FIG . 5b ) means that elements 1,2,3,4 the data exchange between streams until the simulation is\n finished or interrupted 350 .\ngo into first pixel ; 5,6,7,8 go into second pixel ; 9 goes into\nthird pixel . This is the most compact scheme , but not 10 an The user starts 300 the framework through the means of\nconvenient because the geometrical relationship is not pre theoperating system and interacts with the software through\nserved during packing and its extraction depends on the size 306user interaction section 305 of the graphic user interface\n executed on the CPU 120. The start 300 of the imple\nof the population .\n Packing by row ( column; FIG . 5c ) means that elements 15 initializationbegins\n mentation with a user action that causes a GUI\n 307 , Disk input /output initialization 308 on the\n1,2,3 go into pixel ( 1,1 ) ; 3,4,5 go into pixel (2,1 ) , 7,8,9 go CPU 120 , and controller initialization 320 of the GPU\ninto pixel ( 3,1 ) . With this scheme the element\u2019s y coordinate accelerator on the expansion card 180. GUI initialization\nin the population is the pixel's y coordinate, while the includes opening of the main application window and setting\nelement\u2019s x coordinate in the population is the pixel's x the interface tools that allow the user to control the frame\ncoordinate times four plus the index of color component. 20 work . Disk I/O initialization can be performed at the start of\nFive by five populations in this case will use 2x5 texture, or the framework , or at the start of each individual simulation .\n10 pixels . Five of these pixels will only use one out of four The user interaction 305 controls the setting and editing of\ncomponents , so it wastes 37.5 % of this texture. 25x1 popu- the computational elements, parameters, and sources of\nlation will use 6x1 texture ( six pixels ) and will waste 12.5 % external inputs. It specifies which equations should have\nof it . 25 their output saved to disk and / or displayed on the screen . It\n Packing by square ( FIG . 5d ) means that elements 1,2,4,5 allows the user to start and stop the simulation . And it\ngo into pixel ( 1,1 ) ; 3,6 go into pixel ( 1,2 ) ; 7,8 go into pixel performs standard interface functions such as file loading\n( 2,1 ) , and 9 goes into pixel (2,2 ) . Both the row and the and saving , interactive help , general preferences and others.\ncolumn of the element are determined from the row (col- The user interaction 305 directs the CPU 120 to acquire\numn ) of the pixel times two plus the second ( first) bit of the 30 the new external input textures needed (this includes but is\ncolor component index . Five by five populations in this case not limited to loading from disk 140 or receiving them in\nwill use 3x3 texture , or 9 pixels . Four of these pixels will real time from a recording device ), parses them if necessary\nonly use out of four components, and one will only use 309 , and initializes their transfer the expansion card 180 ,\none component, so it wastes 34.4 % of this texture . This is where they are stored 325 in the texture memory bank 250\nmore advantageous than packing by row , since the texture is 35 by the controller 220. The user interaction 305 also directs\nsmaller and the waste is also lower. 25x1 population on the the CPU 120 to parse populations of elements that will be\nother hand will use 13x1 texture ( thirteen pixels ) and waste used in the simulation, convert them to GPU programs\n> 50 % of it , which is much worse than packing by row . ( shaders ) , compile them 310 , and initializes their transfer to\n In order to eliminate waste altogether the population the expansion card 180 , where they are stored 326 in the\nshould have even dimensions in the square packing, and it 40 shader memory bank 210 by the controller 220. This opera\nshould have a number of columns divisible by four in row tion is accompanied by the upload 309 of the initial data into\npacking. Theoretically, the chances are approximately the input partition of the texture memory bank 250 , and\nequivalent for both of these cases to occur, so the particular stores the shader order of execution in the controller 220 .\ntask and data sizes should determine which packing scheme The user can perform operations 309 and 310 as many times\nis preferable in each individual case . 45 as necessary prior to starting the simulation or between\n The System and Framework simulations .\n FIG . 3 shows an exemplary implementation of the top The editing of the system between simulations is difficult\nlevel system and method that is used to control the compu- to accomplish without the hardware implementation of the\ntation . It is a representation of one of several ways in which computational thread suggested herein . The system of equa\na system and method for processing numerical techniques 50 tions ( computational elements) is represented by textures\ncan be implemented in the invention described herein and so that track variables plus shaders that define processing\nthe implementation is not intended to be limited to the algorithms. As mentioned above , textures, shaders and other\nfollowing description and accompanying figure . graphics related constructs can only be initialized within the\n The method presented herein includes two execution rendering context, which is thread specific . Therefore tex\nstreams that run on the CPU 120 - User Interaction Stream 55 tures and shaders can only be initialized in the computa\n302 and Data Output Stream 301. These two streams pref- tional thread .\nerably do not interact directly, but depend on the same data Network editing is a user - interactive process, which\naccumulated during simulations . They can be implemented according to the scheme suggested above happens in the\nas separate threads with shared memory access and executed User Interaction Stream 302. The simulation software thus\non different CPUs in the case of multi -CPU computing 60 has to take the new parameters from the User Interaction\nenvironment. The third execution stream \u2014 Computational Stream 302 , communicate them to the Computational\nStream 303 runs on the GPU accelerator of the expansion Stream 303 and regenerate the necessary shaders and tex\ncard 180 and interacts with the User Interaction Stream 302 tures . This is hard to accomplish without a hardware imple\nthrough initialization routines and data exchange in between mentation of the Computational Stream 303. The Compu\nsimulations. The Computational Stream 303 interacts with 65 tational Stream 303 is forked from the User Interaction\nthe User Interaction Stream and the Data Output Stream Stream and it can access the memory of the parent thread,\nthrough synchronization procedures during simulations. but the reverse communication is harder to achieve. The\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 16 of 22\n\n\n US RE48,438 E\n 9 10\ncontroller 220 allows operations 309 and 310 to be per- timestep ( ) , TSimulator::outfileInterval ( ), and TSimulator::\nformed as many times as necessary by providing the nec- outmode ( ) , the application can set the time step of the\nessary communication to the User Interaction Stream 302 . simulation , the time step of disk output, and the mode of the\n After execution of the input parser texture generation 309 5 disk output. The external input pattern should be packed into\nand population parser shader generator and compiler 310 are a TPattern object and bound to the simulation object through\nperformed at least once , the user has the option to initialize TSimulator:: resetInputs( ) . method . TSimulator::\nthe simulation 311. During this initialization the main con simLength ( ) sets the length of the simulation .\ntrol of the framework is transferred to the GPU accelerator The second step is to create at least one population of\nsystem's accelerator controller 220 and computation 330 is equations ( Tpopulation object ). Population holds one equa\nstarted\ninterrupt(see\n the FIG . 4; 420, change\n simulation ). The the\n userinput\n retains\n , or the abilitythe\n to change to 10 tion object TEquation. This object contains only a formula\ndisplay properties of the framework , but these interactions the and does not hold element- specific data , so all elements of\nare queued to be performed at times determined by the population can share single TEquation .\ncontroller - driven data exchange 314 and 316 to avoid the before execution The TEquation object is converted to a GPU program\ncorruption of the data . 15 . GPU programs have to be executed within\n The progress monitor 312 is not necessary for perfor creates this context , within\n a graphical context which is stream specific . TSimulator\nmance , but adds convenience. It displays the percentage of fore all programs and dataa arrays Computational Stream , there\ncompleted time steps of the simulation and allows the user computation have to be initialized that within\n are necessary for\n Computational\nto plan the schedule using the estimates of the simulation Stream . Constructor of TPopulation is called from User\nwall clock times . Controller - driven data exchange 314 20 Interaction\nupdates the display of the results 313. Online screen output tialized in this constructor .\n Stream , so no GPU - related objects can be ini\nfor the user selected population allows the user to monitor\nthe activity and evaluate the qualitative behavior of the to TPopulation overcome\n :: fillElements ( ) is a virtual method designed\n this difficulty. It is called from within the\nnetwork . Simulations with unsatisfactory behavior can be Computational Stream after TSimulator ::user\n networkCreate ( ) is\nterminated early to change parameters and restart . Control- 25 called in the User Interaction Stream\nler - driven data exchange 314 also drives the output of the TPopulation :: fillElements ( ) to create TEquation and other\n . A has to override\nresults to disk 317. Data output to disk for convenience can computation related objects both element independent and\nbe done on an element per file basis . A suggested file format element-specific. Element independent objects include sub\nincludes a leftmost column that displays a simulated time for\neach of the simulation steps and subsequent columns that 30 components of TEquation and\n handle interdependencies objectsvariables\n between that describe how to\n implemented\ndisplay variable values during this time step in all elements through derivatives of TGate class .\nwith identical equations (e.g. all neurons in a layer of a Element - specific data is held in TElement objects. These\nneural network ).\n Controller -driven data exchange or input parser texture objects. Therereferences\n objects hold to TEquation and a set of TGate\n is one TElement per population, but the size\ngenerator 316 allows the user to change input that is gen- 35 of data arrays within this object corresponds to population\nerated on the fly during the simulation . This allows the size . All TElement objects have to be added to the TSimu\nframework monitoring of the input that is coming from a lator list of elements by calling TSimulator:: addUnit ( )\nrecording device ( video camera, microphone, cell recording method from TPopulation :: fillElements ( ).\nelectrode, etc ) in real time . Similar to the initial input parser Finally, TPopulation :: fillElements ( ) should contain a set\n309\ndata, itarray\n preprocesses\n suitable the\n for input\n textureintogeneration\n a universaland\n format of the 40 of TElement:: add* Dependency ( ) calls for each element.\n generates\ntextures . Unlike the initial parser 309 , here the textures are every Each of these calls sets a corresponding dependency for\ntransferred to hardware not whenever ready but upon the pendentTGatepart object. Here TGate object holds element inde\n of dependency and TElement::\nrequest of the controller 220 . add * Dependency sets element-specific details.\n The controller 220 also drives the conditional testing 315 45 System provided TPopulation handles the output of com\nand 318 informs the CPU - bound streams whether the simu putational elements, both when they need to exchange the\nlation is finished . If so , the control returns to the User data and when they need to output it to disk . User imple\nInteraction Stream . The user then can change parameters or\ninputs ( 309 and 310 ) , restart the simulation (311 ) or quit the mentation of TPopulation derivative can add screen output.\n Listing 1 is an example code of the user program that uses\nframework (390 ) . 50\n a recurrent competitive field (RCF ) equation:\n SANNDRA ( Synchronous Artificial Neuronal Network\nDistributed Runtime Algorithm ; http://www.kinness.net/\nDocs /SANNDRA /html) was developed to accelerate and Listing 1\noptimize processing of numerical integration of large non\nhomogenous systems of differential equations. This library 55 uint16_t static floatw m_compet\n = 3 , h = 3 ; = 0.5 ;\nis fully reworked in its version 2.x.x to support multiple static float m_persist = 1.0 ;\ncomputational backends including those based on multicore class TCablePopRCF : public TPopulation\nCPUs , GPUs and other processing systems . GPU based {TEq_RCF * m_equation ;\nbackend for SANNDRA -2.x.x can serve as an example TGate * m_gatel;\npractical software implementation of the method and archi- 60 TGate * m_gate2;\ntecture described above and pictorially represented in FIG . void createGatingStructure( )\n3. {\n To use SANNDRA, the application should create a m_gate2 m_gatel = new TGate ( 0 );\nTSimulator object either directly or through inheritance . } ; = new TGate ( 1 ) ;\nThis object will handle global simulation properties and 65 void createUnitStructure ( TBasicUnit* u )\ncontrol the User Interaction Stream , Data Output Stream , {\nand Computational Stream . Through TSimulator ::\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 17 of 22\n\n\n US RE48,438 E\n 11 12\n -continued expensive operation , computationally, but since the textures\n are referred to by texture IDs ( pointers ), swapping these\n Listing 1 pointers for input and output textures after each time step\nu-> addO20PInputDependency (m_gatel, O. , 0. , 0.004 , 0. , 0 , 0 ) ; achieves the same result at a much lesser cost .\n 5\nu- > addFullDependency (m_gate2, population ( ) ); In the hardware solution suggested herein , ID swapping is\n} equivalent to swapping the base memory address for two\npublic : TCablePopRCF ( ) : TPopulation ( \" compCPU RCF \u201d , w , h , true) { } ;\n-TCablePopRCFO ) { if (m_equation ) delete m_equation ; partitions of the texture memory bank 250. They are\n if (m_gatel) delete m_gatel;\n if (m_gate2) delete m_gate2 ; } ;\n swapped 485 during synchronization ( 485 , 430 , and 455 ) so\nbool fillElements( TSimulator * sim ) ; 10 that data transfer 445 and the computation 435-487 proceeds\n}; immediately and in parallel with data transfer as shown in\nbool TCablePopRCF :: fillElements ( TSimulatior* sim) FIG . 4. A hardware solution allows this parallelism through\n{ access of the controller 220 to the onboard texture memory\nm_equation = new TEQ_RCF (this, m_compet, m_persist );\ncreateGatingStructure ( ) ; bank 250 .\nfor( size_t i = 0 ; i < xSize ( ) ; ++ i ) 15\n The main computation and data exchange are executed by\n for(size_t j = 0 ; j < ySize ( ) ; ++ i)\n{ the controller 220. It runs three parallel substreams of\nTElement * u = new TCPUElement(this , m_equation, i , j ) ; execution : Computational Substream 403 , Data Output Sub\nsim-> addUnit( u ); stream 402 , and Data Input Substream 404. These streams\ncreateUnitStructure ( u );\n} 20 are synchronized with each other during the swap of pointers\nReturn true ; 485 to the input and output texture memory partitions of the\n} texture memory bank 250 and the check for the last iteration\nint\nmain ( ) 487. Algorithmically, these two operations are a single\n{ atomic operation, but the block diagram shows them as two\n// Input pattern generation ( 309 in FIG.3 ) 25 separate blocks for clarity.\nuint32_t * pat = new uint32_t [ w * h ];\nTRandom < float > randGen (0 ) ; The Computational Substream 403 performs a computa\nfor (uint32_t I = 0 ; I < w * h ; ++ i )\npat [i] = randGen.random ( ) ; tional cycle including a sequential execution of all shaders\nTpattern * p = new Tpattern (pat, w, h ) ; that were stored in the shader memory bank 210 using the\n// Setting up the simulation\n 30\n appropriate input and output textures. To begin the simula\nTSimulator * cableSim = new TSimulator ( \"data \" ) ; // ( 308 and 320 in tion the controller 220 initializes three execution substreams\nFIG. 3)\ncableSim- >timestep ( 0.05 ) ; // (320 in FIG . 3 ) 403 , 402 , and 404. On every simulation step , the Compu\ncableSim-> resetInputs (p ); // (325 in FIG . 3 ) tational Substream 403 determines which textures the GPU\ncableSim- > outfileInterval(0.1 ); // (308 in FIG . 3 ) 240 will need to perform the computations and initiates the\ncableSim- > outmode (SANNDRA ::timefunc ); // ( 308 in FIG . 3 )\ncableSim- > simLength (60.0 ); // (320 in FIG . 3 ) 35 upload 435 of them onto the GPU 240. The GPU 240 can\n// Preparing the population communicate directly with the texture memory bank 250 to\nTPopulation * cablePop new TCablePopRCFO ); // (310 in FIG . 3 ) upload the appropriate texture to perform the computations .\ncableSim- > networkCreate ( ); // (326 in FIG . 3 ) The controller 220 also pulls the first shader (known by the\nuint16_t user = 1 ;\nwhile (user) stored order ) from the shader memory bank 210 and uploads\n{ 40 450 it onto the GPU 240 .\nif (! cableSim-> simulationStart ( true, 1 ) ) // ( 311 in FIG . 3 )\nexit ( 1 ) ; The GPU 240 executes the following operations in this\nstd ::cout << \" Repeat ? \\ n \" ; // (305 in FIG . 3 )\nstd :: cin >> user; // ( 305 in FIG . 3 )\n order : performs the computation ( execution of the shader )\nif (user 1 ) 470 ; tells the controller 220 that it is done with the compu\ncableSim- > networkReset ( ); // ( 305 in FIG . 3 ) tations for the current shader; and after all shaders for this\n 45 particular equation are executed sends 480 the output tex\n{\nIf (cableSim ) tures to the output portion of the texture memory bank 250 .\nDelete cableSim ; // Also deletes cablePop and its internals This cycle continues through all of the equations based on\nexit (0 ) ;\n}; the branching step 482 .\n 50 An example shader that performs fourth order Runge\n FIG . 4 is a detailed flow diagram illustrating a part of an Kutta numerical integration is shown in Listing 2 using\nexemplary implementation of the bottom level system and GLSL notation ;\nmethod performed during the computation on the GPU\naccelerator of the expansion card 180 and is a more detailed Listing 2\nview of the computational box 330 in FIG . 3. FIG . 4 is a 55\nrepresentation of one of several ways in which a system and uniform sampler2DRect Variable ;\nmethod for processing numerical techniques can be imple uniform float integration_step ;\n float halfstep = integration_step * 0.5 ;\nmented .\n With systems of equations that have complex interdepen float fl_6step = integration_step / 6.0 ;\n vec4 output = texture2DRect (Variable, gl_TexCoord [0 ] .st );\ndencies it is likely that the variable in some equation from 60 // define equation here\na previous time step has to be used by some other equation vec4 rungekutta4 ( vec4 x )\nafter the new values of this variable are already computed {\n const vec4 kl = equation ( x );\nfor new time step . To avoid data confusion , the new values const vec4 k2 equation ( x + halfstep * kl ) ;\nof variables should be rendered in a separate texture. After const vec4 k3 equation ( x + halfstep * k2 );\nthe time step is completed for all equations, these new values 65 const vec4 k4 = equation ( x + integration step * k3 ) ;\nshould be copied over old values so that they are used as return fl_6step * (kl + 2.0 * (k2 + k3 ) + k4 );\ninput during the next time step . Copying textures is an\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 18 of 22\n\n\n US RE48,438 E\n 13 14\n -continued CONCLUSION\n Listing 2 This GPU accelerator system offers the following poten\n }\n tial advantages:\n Void main (void ) 5 1. Limited computations on the CPU 120. The CPU 120\n { is only used for user input, sending information to the\n output + = rungekutta4 (output );\n gl_FragColor = output;\n controller 220 , receiving output after each computational\n } cycle ( or less frequently as defined by the user ), writing this\n output to disk 140 , and displaying this output on the monitor\n 10 170. This frees the CPU 120 to execute other applications\n The shader in Listing 2 can be executed on conventional and allows the expansion card to run at its full capacity\nvideo card . Using the controller 220 this code can be further without being slowed down by extensive interactions with\noptimized , however. Since the integration step does not the CPU 120 .\nchange during the simulation , the step itself as well as the 2. Minimizing data transfer between the expansion card\nhalfstep and % of the step can be computed once per 15 180 and the system bus 200. All of the information needed\nsimulation , and updated in all shaders by a shader update to perform the simulations will be stored on the expansion\nprocedures 310 , 326 discussed above . card 180 and all simulations will take place on it . Further\n After all of the equations in the computational cycle are more , whatever data transfer remains necessary will take\ncomputed the main execution substream 403 on the control place in parallel with the computation , thus reducing the\nler 220 can switch 485 the reference pointers of the input and 20 impact\n 3. New of this\n waytransfer on the\n to execute GPUperformance\n programs (. shaders ). Previ\noutput portions of the texture memory bank 250 . ously, the CPU 120 had full control over the order of\n The two other substreams of execution on the controller shader's execution and was required to produce specific\n220 are waiting ( blocks 430 and 455 , respectively) for this commands on every cycle to tell the GPU 240 which shader\nswitch to begin their execution . The Data Input Substream 25 to use . With the invention disclosed herein , shaders will\n404 is controlling 440 the input of additional data from the initially be stored on the shader memory bank 210 on the\nCPU 120. This is necessary in cases where the simulation is expansion card 180 and will be sent to the GPU 240 for\nmonitoring the changing input, for example input from a execution by the general purpose controller 220 located on\nvideo camera or other recording device in the real time . This the expansion card .\nsubstream uploads new external input from the CPU 120 to 30 4. Multiple parallelisms . The GPU 240 is inherently\nthe texture memory bank 250 so it can be used by the main parallel and is well suited to perform parallel computations.\ncomputational substream 403 on the next computational step In parallel with the GPU 240 performing the next calcula\nand waits for the next iteration 475. The Data Output tion , the controller 220 is uploading the data from the\nSubstream 445 controls the output of simulation results to previous calculation into main memory 130. Furthermore,\nthe CPU 120 if requested by the user . This substream 35 the CPU 120 at the same time uses uploaded previous results\nuploads the results of the previous step to the main RAM to save them onto disk 140 and to display them on the screen\n130 so that the CPU 120 can save them on disk 140 or show through the system bus 200 .\nthem on the results display 313 and waits for the next 5. Reuse of existing and affordable technology. All hard\niteration 460 . ware used in the invention and mentioned here - in are based\n Since the Computational Substream 403 determines the 40 on currently available and reliable components. Further\ntiming of input 440 and output 445 data transfers, these data advance of these components will provide straightforward\ntransfers are driven by the controller 220. To further reduce improvements of the invention .\nthe data transfer overhead ( and disk 140 overhead also ) the While this invention has been particularly shown and\ncontroller 220 initiates transfer only after selected compu- described with references to preferred embodiments thereof,\ntational steps . For example, if the experimental data that is 45 it will be understood by those skilled in the art that various\nsimulated was recorded every 10 milliseconds (msec ) and changes in form and details may be made therein without\nthe simulation for better precision was computed every 1 departing from the scope of the invention encompassed by\nmsec , then only every tenth result has to be transferred to the appended claims .\nmatch the experimental frequency.\n This solution stores two copies of output data , one in the 50 What is claimed is :\nexpansion card texture memory bank 250 and another in the 1. \u00c0 computer system , comprising:\nsystem RAM 130. The copy in the system RAM 130 is a central processing unit to receive input data ;\naccessed twice : for disk I/O and screen visualization 313. An main memory , operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a unit via a bus , to store the input data received by the\ndirect read access to the onboard texture memory bank 250 55 central processing unit;\nby mapping the memory of the hardware onto a global an accelerator, operably coupled to the central processing\nmemory space . The alternative solution will double the unit and the [first] main memory via the bus , to receive\ncommunication through the local bus 190. Since the goal at least a portion of the input data from the main\ndiscussed herein is reducing the information transfer through memory , the accelerator comprising:\nthe local bus 190 , the former solution is favored . 60 at least one graphics processing unit to perform a\n The main substream 403 determines if this is the last sequence of computations on the at least a portion of\niteration 487. If it is the last iteration , the controller 220 the input data so as to generate output data , the\nwaits for the all of the execution substreams to finish 490 sequence of computations representing an artificial\nand then returns the control to the CPU 120 , otherwise it neural network, intermediate computations in the\nbegins the next computational cycle . 65 sequence of computations representing respective\n This repeats through all of the computational cycles of the layers of the artificial neural network and yielding\nsimulation . intermediate results; and\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 19 of 22\n\n\n US RE48,438 E\n 15 16\n accelerator memory , operably coupled to the [ graphic ] system comprising a central processing unit (CPU) , a main\n at least one graphics processing unit, to store the memory operably coupled to the central processing unit via\n results of the [ plurality of sequential] sequence of a bus , an accelerator operably coupled to the CPU and the\n computations; and main memory via the bus , the accelerator comprising a\n a controller, operably coupled to the at least one graphics 5 graphics processing unit (GPU) and an accelerator memory,\n processing unit and the accelerator memory, to initial- the method comprising:\n ize textures and shaders in the accelerator memory for ( A ) performing, by the GPU , the sequence of computa\n performing the sequence of computations, to control tions on a first portion of [ the] input data so as to\n performance of the sequence of computations by the at generate a first portion of [the] output data , the first\n least one graphics processing unit, to transfer the at 10 portion of the output data representing an output of a\n least a portion of the input data into the accelerator neuron in a first layer of the artificial neural network,\n memory during performance of the intermediate com intermediate computations in the sequence of compu\n putations in the sequence of computations by the at tations yielding intermediate results , wherein perform\n least one graphics processing unit, and to transfer at ing the sequence of computations on the first portion of\n least a portion of the output data from the accelerator 15 the input data comprises ( i ) assigning an output vari\n memory to the main memory during performance of the\n intermediate computations in the sequence of compu able to a first texture and a second texture , the output\n tations by the at least one [ graphic ] graphics processing variable being included in a first computational ele\n unit . ment of a plurality of computational elements, the\n 2. The computer system of claim 1 , wherein the central 20 plurality of computational elements representing the\nprocessing unit is configured to receive the input data in sequence of computations and ( ii ) accumulating a first\nresponse to a user interaction . value for the output variable in the first texture during\n 3. The computer system of claim 1 , wherein : a first time step ;\n the central processing unit is configured to receive the ( B ) in parallel with performing the sequence of compu\n input data at a first rate ; and 25 tations by the GPU in ( A ), transferring a second portion\n the at least one graphics processing unit is configured to of the input data from the main memory to the accel\n perform the sequence of computations at a second rate erator via the bus ; [ and]\n different than the first rate . ( C ) in parallel with performing the sequence of compu\n 4. The computer system of claim 1 , wherein the main tations by the GPU in ( A ), transferring a second portion\nmemory is configured to store a copy of the output data 30 of the output data from the accelerator memory to the\nstored in the accelerator memory . main memory via the bus, the second portion of the\n 5. The computer system of claim 1 , wherein an output of output data representing an output of a neuron in a\nat least one computation in the sequence of computations second layer in the artificial neural network ; and\nrepresents an output of at least one neuron in an artificial ( D ) performing, by the GPU , the sequence of computa\nneural network . 35 tions on the second portion of the input data , wherein\n 6. The computer system of claim 1 , wherein accelerator performing the sequence of computationson the second\nmemory comprises: portion of the input data comprises ( i ) accumulating a\n a first memory bank to store parameters common to all of second value for the output variable in the second\n the computations in the sequence of computations; and texture during a second time step and ( ii) making the\n a second memory bank to store data specific to at least one 40 first value of the output variable in the first texture\n computation in the sequence of computations. accessible to other computational elements in the plu\n 7. The computer system of claim 1 , wherein the controller rality of computational elements during the second\nis configured to transfer the output data from the accelerator time step.\nmemory to the main memory without transferring any of the 13. The method of claim 12 , further comprising:\nintermediate results from the accelerator memory to the 45 storing the input data in the main memory in response to\nmain memory so as to reduce data transfer via the bus . a user interaction .\n 8. The computer system of claim 1 , wherein the controller 14. The method of claim 12 , further comprising:\nis configured to transfer at least a portion of the output data receiving the input data at a first rate; and\nfrom the accelerator memory to the main memory after the wherein ( A ) comprises performing the sequence of com\nat least one graphics processing unit has begun to perform 50 putations at a second rate different than the first rate .\nanother sequence of computations. [ 15. The method of claim 12 , wherein ( A ) comprises:\n 9. The computer system of claim 8 , wherein the controller generating an output representative of an output of at least\nis configured to initiate transfer of the at least a portion of the one neuron in an artificial neural network .]\ninput data and to transfer the at least a portion of the output 16. The method of claim 12 , wherein (C ) comprises:\ndata in parallel with performance of at least one computation 55 transferring the second portion of the output data from the\nin the other sequence of computations by the at least one accelerator memory to the main memory without trans\ngraphics processing unit . ferring any of the intermediate results of the plurality of\n 10. The computer system of claim 1 , wherein the con sequential computations from the accelerator memory\ntroller is configured to control execution of the sequence of to the main memory so as to reduce data transfer via the\ncomputations by the at least one graphics processing unit . 60 bus .\n 11. The computer system of claim 1 , further comprising: 17. The method of claim 12 , wherein ( C ) comprises :\n at least one of a video camera, a microphone, or a cell transferring the second portion of the output data from the\n recording electrode, operably coupled to the central accelerator memory to the main memory after the GPU\n [ processor) processing unit , to acquire the input data in has begun to perform another sequence of computa\n real time . 65 tions.\n 12. A method of performing a sequence of computations 18. The method of claim 17 , wherein (C ) further com\nrepresenting an artificial neural network on a computer prises:\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 20 of 22\n\n\n US RE48,438 E\n 17 18\n initiating transfer of the second portion of the output data 27. The method of claim 26 , further comprising :\n in parallel with performance of at least one computa- storing, in a second memory partition of the memory, data\n tion in the other sequence of computations. specific to the first computation in the sequence of\n 19. The method of claim 12 , further comprising : computations.\n acquiring the input data in real time with at least one of 5 28. The method of claim 27, further comprising :\n a video camera , a microphone, or a cell recording storing, in the second memory partition , external input\n electrode operably coupled to the CPU . data patterns, representations of internal variables, an\n 20. The method of claim 12 , further comprising : input of the computation in the sequence of computa\n storing parameters common to all of the computations in 10 tions , and the output of the computation in the sequence\n the sequence of computations in a first memory bank in of computations.\n the accelerator memory ; and 29. The method of claim 21 , wherein storing the first\n storing data specific to at least one computation in the output data comprises :\n sequence of computations in a second memory bank in accumulating, in the memory, outputs of computational\n the accelerator memory. 15 elements executed by the GPU in performing the first\n 21. A method of performing a sequence of computations computation in the sequence of computations.\nrepresenting an artificial neural network, the method com- 30. The method of claim 21 , further comprising:\nprising : storing, in the memory, an output of a previous compu\n receiving, at a central processing unit ( CPU ), first input tation in the sequence of computations; and\n data acquired from an external system in real time; 20 accessing, by the GPU , the output of the previous com\n initializing, by a controller operably coupled to a graph- putation during performance of the computation in the\n ics processing unit (GPU ), textures and shaders in a sequence of computations.\n memory opera coupled to the GPU ; 31. The method of claim 21 , wherein performing the first\n transferring the first input data received by the CPU to the computation comprises executing a plurality of computa\n memory operably coupled to the GPU : 25 tional elements representing a layer of neurons in an arti\n performing, by the graphics processing unit (GPU ), a first ficial neural network.\n computation in the sequence of computations on the 32. The method of claim 31 , wherein all neurons in the\n first input data based on the textures and shaders to layer of neurons are described by the same equation .\n generate first output data , computations in the 30 33. The method of claim 21 , further comprising :\n acquiring the second input data with at least one of a\n sequence of computations representing respective lay\n ers of neurons in the artificial neural network , an video camera , a microphone, or a cell recording elec\n trode .\n output of the first computation in the sequence of 34. The method of claim 21 , further comprising :\n computations representing an output of a first neuron in loading the second input data from disk .\n a first layer in the artificial neural network ; 35 35. A system for performing a sequence of computations,\n storing , in the memory operably coupled to the GPU , the the system comprising:\n first input data and the first output data ; and a camera to generate input data in real time ;\n transferring second input data acquired from the external a first memory partition ;\n system in real time into the memory operably coupled a second memory partition operably coupled to the first\n to the GPU after the GPU starts the first computation 40 memory partition ; and\n and before the GPU starts a second computation of the a processing unit , operably coupled to the camera , the\n sequence of computations, an output of the second first memory partition , and the second memory parti\n computation in the sequence of computations repre tion , to perform the sequence of computations on a first\n senting an output of a second neuron in a second layer portion of the input data so as to generate a first\n in the artificial neural network . 45 portion of output data , intermediate computations in\n 22. The method of claim 21 , wherein transferring the the sequence of computations yielding intermediate\nsecond input data comprises transferring the second input results, the first portion of the output data representing\ndata via a bus operably coupled to the CPU . an output of an artificial neural network,\n 23. The method of claim 21 , further comprising: wherein the first memory partition is configured to trans\n transferring the first output data from the memory to 50 fer a second portion of the input data to the second\n another memory during the second computation in the memory partition in parallel with performance the\n sequence of computations. sequence of computations by the processing unit ,\n 24. The method of claim 23, further comprising: wherein the second memory partition is configured to\n storing intermediate results of the sequence of computa 55 transfer a second portion of the output data to the first\n tions in the memory, and memory partition in parallel with performance the\n wherein transferring the first output data from the sequence of computations by the processing unit, and\n wherein the sequence of computations represents the\n memory to the other memory occurs without transfer artificial neural network , each neuron in the artificial\n ring the intermediate results of the sequence of com neural network has an output variable assigned to a\n putations. 60 first texture and a second texture in the memory , the first\n 25. The method of claim 23 , wherein transferring the texture holds a first value of the output variable com\nsecond input data and transferring the first output data puted during a previous time step of the sequence of\noccurs in parallel . computations and accessible to other neurons in the\n 26. The method of claim 21 ,further comprising: neural network during a current time step of the\n storing , in a first memory partition of the memory, param- 65 sequence of computations and the second texture accu\n eters common to all of the computations in the mulates a second value of the output variable computed\n sequence of computations. during the current time step .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 21 of 22\n\n\n US RE48,438 E\n 19 20\n 36. The system of claim 35, wherein the first memory in the sequence of computations representing layers\npartition and the second memory partition are logical par of the neural network and yielding intermediate\ntitions . results ; and\n 37. The system of claim 35, wherein the processing unit is accelerator memory, operably coupled to the at least\ncomprises a graphics processing unit ( GPU ) . 5\n one processing unit, to store the results of the\n 38. The system of claim 35, wherein the processing unit is sequence of computations; and\nconfigured to receive the input data at a first rate and to a controller, operably coupled to the at least one\nperform the sequence of computations at a second rate is processing unit and the accelerator memory, to con\ndifferent than the first rate. trol transfer of the at least a portion of the input data\n 39. The system of claim 35, wherein the second memory 10 into the accelerator memory during performance of\npartition is configured to transfer the second portion of the the intermediate computations in the sequence of\noutput data to the first memory partition without transfer computations by the at least one processing unit, to\nring any of the intermediate results to the first memory control transfer at least a portion of the output data\npartition .\n 40. A system for executing an artificial neural network, 15 from the accelerator memory to the main memory\nthe system comprising: during performance of the intermediate computa\n a central processing unit (CPU ) to provide first input tions in the sequence of computations by the at least\n data ; one processing unit, and to control performance of\n a memory, operably coupled to the CPU , to store the first the sequence of computations by the at least one\n input data in a first partition, referenced by a first 20 processing unit.\n pointer, before computing a first layer of neurons of the 45. The computer system of claim 44, wherein the central\n artificial neural network ; processing unit is configured to receive the input data in\n a processing unit, operably coupled to the memory, to response to a user interaction .\n perform , during computation of the first layer of neu- 46. The computer system of claim 44, wherein :\n rons, at least one calculation on the first input data so 25 the central processing unit is configured to receive the\n as to generate first output data , the first output data input data at a first rate ; and\n representing an output of at least one neuron in the first the at least one processing unit is configured to perform\n layer of neurons ; and the sequence of computations at a second rate different\n a controller, operably coupled to the processing unit and than the first rate.\n the memory, to : 30 47. The computer system of claim 44, wherein the main\n store the first output data in a second partition of the memory is configured to store a copy of the output data\n memory, the second partition referenced by a second stored in the accelerator memory.\n pointer, and to swap the firstpointer with the second 48. The computer system of claim wherein an output\n pointer at the end of the computation of the first layer of at least one computation in the sequence of computations\n of neurons, such that the firstoutput data becomes an 35 represents an output of at least one neuron in an artificial\n input for a second layer of neurons of the artificial neural network .\n neural network, 49. The computer system of claim 44, wherein accelerator\n transfer the first output data to another memory during memory comprises:\n computation of the second layer of neurons, and a first memory partition to store parameters common to\n dictate an order of execution of instructions to the 40 all of the computations in the sequence of computa\n processing unit to perform the computation of the tions ; and\n first layer of neurons . a second memory partition to store data specific to at least\n 41. The system of claim 40 , wherein the processing unit one computation in the sequence of computations.\ncomprises a graphics processing unit . 50. The computer system of claim 44, wherein the con\n 42. The system of claim 40 , wherein the controller is 45 troller is configured to transfer the output data from the\nconfigured to send instructions for performing the at least accelerator memory to the main memory without transfer\none calculation to the processing unit. ring any of the intermediate results from the accelerator\n 43. The system of claim 40 , wherein the memory further memory to the main memory so as to reduce data transfer\ncomprises : via the bus.\n a third partition to store internal variables; and 50 51. The computer system of claim 44, wherein the con\n a fourth partition to store data used as input at a troller is configured to transfer at least a portion of the\n particular layer of neurons of the artificial neural output data from the accelerator memory to the main\n network. memory after the at least one processing unit has begun to\n 44. A computer system , comprising: perform another sequence of computations.\n a central processing unit to receive input data acquired 55 52. The computer system of claim 51 , wherein the con\n from an external system ; troller is configured to initiate transfer of the at least a\n main memory, operably coupled to the central processing portion of the input data and to transfer the at least a portion\n unit via a bus, to store the input data received by the of the output data in parallel with performance of at least\n central processing unit; one computation in the other sequence of computations by\n an accelerator, operably coupled to the central processing 60 the at least one processing unit .\n unit and the main memory via the bus, to receive at 53. The computer system of claim 44, wherein the con\n least a portion of the input data from the main memory , troller is configured to control execution of the sequence of\n the accelerator comprising : computations by the at least one processing unit .\n at least one processing unit to perform a sequence of 54. The computer system of claim 44, further comprising :\n computations representing an artificial neural net- 65 at least one of a video camera , a microphone, or a cell\n work on the at least a portion of the input data so as recording electrode, operably coupled to the central\n to generate output data, intermediate computations processing unit, to acquire the input data in real time .\n\f Case 7:26-cv-00318 Document 1-4 Filed 08/17/26 Page 22 of 22\n\n\n US RE48,438 E\n 21 22\n 55. The computer system of claim 1 , wherein the control\nler is configured to inform the central processing unit that\nthe sequence of computations is finished .\n 56. The computer system of claim 1 , wherein the control\nler is configured to reduce a processing load on the central 5\nprocessing unit.\n 57. The computer system of claim 1 , wherein the control\nler is configured to reduce interactions between the central\nprocessing unit and the accelerator.\n 10\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:21.450912-07:00","document_number":"1","attachment_number":4,"pacer_doc_id":"181037209214","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 3","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294675/","id":490294675,"tags":[],"absolute_url":"/docket/74659430/1/5/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.082136-07:00","date_modified":"2026-08-21T19:39:07.745965-07:00","sha1":"c81d412905e75a927b55cc09a47e2fe8d624c94d","page_count":20,"file_size":1057834,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.5.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.5.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 1 of 20\n\n\n\n\n EXHIBIT\n\n 4\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 2 of 20\n I 1111111111111111\n 1111111111111111\n IIIIIIIIIIIII1111111111111111\n Illlllll llll\n US00RE49461E\nc19) United States\nc12) Reissued Patent (10) Patent Number: US RE49,461 E\n Gorchetchnikov et al. (45) Date of Reissued Patent: Mar. 14, 2023\n\n\n(54) GRAPHIC PROCESSOR BASED (56) References Cited\n ACCELERATOR SYSTEM AND METHOD\n U.S. PATENT DOCUMENTS\n(71) Applicant: Neurala, Inc., Boston, MA (US)\n 5,063,603 A 11/1991 Burt\n(72) Inventors: Anatoli Gorchetchnikov, Newton, MA 5,136,687 A 8/1992 Edelman et al.\n (US); Heather Marie Ames, Milton, (Continued)\n MA (US); Massimiliano Versace,\n Milton, MA (US); Fabrizio Santini, FOREIGN PATENT DOCUMENTS\n Jamaica Plain, MA (US) EP 1224622 Bl 11/2004\n WO 190208 11/2014\n(73) Assignee: Neurala, Inc., Boston, MA (US)\n (Continued)\n(21) Appl. No.: 17/136,343\n(22) Filed: Dec. 29, 2020 OTHER PUBLICATIONS\n Related U.S. Patent Documents Hodgkin, A. L., and Huxley, A. F. 1952. Quantitative description of\nReissue of: membrane current and its application to conduction and excitation\n(64) Patent No.: 9,189,828 m nerve. J Physiol 117, pp. 500-544.\n Issued: Nov. 17, 2015 (Continued)\n Appl. No.: 14/147,015\n Primary Examiner - William H. Wood\n Filed: Jan.3, 2014\n (74) Attorney, Agent, or Firm - Smith Baluch LLP\nU.S. Applications:\n(63) Continuation of application No. 15/808,201, filed on (57) ABSTRACT\n Nov. 9, 2017, now Pat. No. Re. 48,438, which is an An accelerator system is implemented on an expansion card\n (Continued) comprising a printed circuit board having (a) one or more\n graphics processing units (GPUs), (b) two or more associ-\n(51) Int. Cl. ated memory banks (logically or physically partitioned), (c)\n G06T 1160 (2006.01) a specialized controller, and (d) a local bus providing signal\n coupling compatible with the PCI industry standards. The\n G06F 9/50 (2006.01) controller handles most of the primitive operations to set up\n (Continued) and control GPU computation. Thus, the computer's central\n(52) U.S. Cl. processing unit (CPU) can be dedicated to other tasks. In this\n CPC .............. G06T 1120 (2013.01); G06F 9/5027 case a few controls (simulation start and stop signals from\n the CPU and the simulation completion signal back to CPU),\n (2013.01); G06T 1160 (2013.01); GPU programs and input/output data are exchanged between\n (Continued) CPU and the expansion card. Moreover, since on every time\n(58) Field of Classification Search step of the simulation the results from the previous time step\n CPC ... G06F 9/5027; G06F 2209/509; G06T 1/20; are used but not changed, the results are preferably trans-\n ferred back to CPU in parallel with the computation.\n G06T 1/60; G06N 3/00; G06N 3/02;\n (Continued) 21 Claims, 5 Drawing Sheets\n\n Expansion Card mo\n\n\n\n\n I ............\n\n\n 41 r:;-\n\n\n r~\u00b7--\u00b7 .tll)\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 3 of 20\n\n\n US RE49,461 E\n Page 2\n\n\n Related U.S. Application Data 2011/0004341 Al 1/2011 Sarvadevabhatla et al.\n 2011/0173015 Al 7/2011 Chapman et al.\n application for the reissue of Pat. No. 9,189,828, 2011/0279682 Al 11/2011 Li et al.\n which is a continuation of application No. 11/860, 2012/0072215 Al 3/2012 Yu et al.\n 2012/0089552 Al 4/2012 Chang et al.\n 254, filed on Sep. 24, 2007, now Pat. No. 8,648,867. 2012/0197596 Al 8/2012 Comi\n(60) Provisional application No. 60/826,892, filed on Sep. 2012/0316786 Al 12/2012 Liu et al.\n 2013/0126703 Al 5/2013 Caulfield\n 25, 2006. 2013/0131985 Al 5/2013 Weiland et al.\n 2014/0019392 Al 1/2014 Buibas et al.\n(51) Int. Cl. 2014/0032461 Al 1/2014 Weng\n G06T 1120 (2006.01) 2014/0052679 Al 2/2014 Sinyavskiy et al.\n G06N 3/063 (2023.01) 2014/0089232 Al 3/2014 Buibas et al.\n 2015/0127149 Al 5/2015 Sinyavskiy et al.\n G06N 20/00 (2019.01) 2015/0134232 Al 5/2015 Robinson\n(52) U.S. Cl. 2015/0224648 Al 8/2015 Lee et al.\n CPC ........ G06F 2209/509 (2013.01); G06N 3/063 2016/0075017 Al 3/2016 Laurent et al.\n (2013.01); G06N 20/00 (2019.01) 2016/0082597 Al 3/2016 Gorshechnikov et al.\n 2016/0096270 Al 4/2016 Gabardos et al.\n(58) Field of Classification Search 2016/0198000 Al 7/2016 Gorshechnikov et al.\n CPC ............ G06N 3/04; G06N 3/06; G06N 3/063; 2017 /0024877 Al 1/2017 Versace et al.\n G06N 3/08; G06N 3/1 O; G06N 5/00; 2017/0076194 Al 3/2017 Versace et al.\n G06N 7/00; G06N 7/02; G06N 7/04; 2017/0193298 Al 7/2017 Versace et al.\n G06N 7/046; G06N 7/06; G06N 20/00\n See application file for complete search history. FOREIGN PATENT DOCUMENTS\n\n WO 2014204615 A2 12/2014\n(56) References Cited WO 2015143173 A2 9/2015\n WO 2016014137 A2 1/2016\n U.S. PATENT DOCUMENTS\n\n 5,142,665\n A 8/ 1992 Bigus OTHER PUBLICATIONS\n 5,172,253\n A 12/1992 Lynne\n 5,388,206\n A 2/1995 Poulton et al. Hopfield, J. 1982. Neural networks and physical systems with\n 6,018,696\n A 1/2000 Matsuoka et al.\n emergent collective computational abilities. In Proc Natl Acad Sci\n 6,336,051\n Bl 1/2002 Pangels et al.\n 6,647,508\n B2 11/2003 Zalewski et al. USA, vol. 79, pp. 2554-2558.\n 7,119,810\n B2 10/2006 Sumanaweera et al. Ilie, A. 2002. Optical character recognition on graphics hardware.\n 7,219,085\n B2 * 5/2007 Buck . G06V 10/955 Tech. Rep. integrative paper, UNCCH, Department of Computer\n 706/12 Science, 9 pages.\n 7,477,256 Bl* 1/2009 Johnson .................... G06F 3/14 International Preliminary Report on Patentability in related PCT\n 345/506 Application No. PCT/US2014/039162 filed May 22, 2014, dated\n 7,525,547 Bl* 4/2009 Diard ........................ G06T 1/20\n Nov. 24, 2015, 7 pages.\n 345/522\n 7,765,029 B2 7/2010 Fleischer et al. International Preliminary Report on Patentability in related PCT\n 7,861,060 Bl* 12/2010 Nickolls et al. .. ... ... .... ... . 712/22 Application No. PCT/US2014/039239 filed May 22, 2014, dated\n 7,873,650 Bl 1/2011 Chapman et al. Nov. 24, 2015, 8 pages.\n 8,392,346 B2 3/2013 Ueda et al. International Preliminary Report on Patentability dated Nov. 8,\n 8,510,244 B2 8/2013 Carson et al. 2016 from International Application No. PCT/US2015/029438, 7\n 8,583,286 B2 11/2013 Fleischer et al. pages.\n 8,648,867 B2 * 2/2014 Gorchetchnikov et al ... 345/501\n International Search Report and Written Opinion dated Feb. 18,\n 9,031,692 B2 5/2015 Zhu\n 9,177,246 B2 11/2015 Bui bas et al. 2015 from International Application No. PCT/US2014/039162, 12\n 9,189,828 B2 11/2015 Gorchetchnikov et al. pages.\n 9,626,566 B2 4/2017 Versace et al. International Search Report and Written Opinion dated Feb. 23,\n 10,083,523 B2 9/2018 Versace et al. 2016 from International Application No. PCT/US2015/029438, 11\n RE48,438 E * 2/2021 Gorchetchnikov ... G06F 9/5027 pages.\n 2001/0010034 Al 7/2001 Burton International Search Report and Written Opinion dated Jul. 6, 2017\n 2002/0046271 Al 4/2002 Huang\n from International Application No.PCT/US2017/029866, 12 pages.\n 2002/0050518 Al 5/2002 Roustaei\n 2002/0064314 Al 5/2002 Comaniciu et al. International Search Report and Written Opinion dated Nov. 26,\n 2002/0168100 Al 11/2002 Woodall 2014 from International Application No. PCT/US2014/039239, 14\n 2003/0026588 Al 2/2003 Elder et al. pages.\n 2003/00787 54 Al 4/2003 Hamza International Search Report and Written Opinion dated Sep. 15,\n 2004/0015334 Al 1/2004 Ditlow et al. 2015 from International Application No. PCT/US2015/021492, 9\n 2005/0166042 Al 7/2005 Evans pages.\n 2006/0129506 Al* 6/2006 Edelman. G05D 1/0088\n Itti, L., and Koch, C. (2001). Computational modelling of visual\n 706/12\n 2006/0184273 Al 8/2006 Sawada et al. attention. Nature Reviews Neuroscience, 2 (3), 194-203.\n 2007/0052713 Al 3/2007 Chung et al. Itti, L., Koch, C., and Niebur, E. (1998). A Model of Saliency-Based\n 2007/0198222 Al 8/2007 Schuster et al. Visual Attention for Rapid Scene Analysis, 1-6.\n 2007/0279429 Al 12/2007 Ganzer Jarrett, K., Kavukcuoglu, K., Ranzato, M. A., & LeCun, Y. (Sep.\n 2008/0033897 Al 2/2008 Lloyd 2009). What is the best multi-stage architecture tor object recogni-\n 2008/0066065 Al 3/2008 Kim et al. tion?. In Computer Vision, 2009 IEEE 12th International Confer-\n 2008/0258880 Al 10/2008 Smith et al.\n ence on (pp. 2146-2153) IEEE.\n 2009/0080695 Al 3/2009 Yang\n 2009/0089030 Al 4/2009 Sturrock et al. Khaligh-Razavi, S.-M et al., Deep Supervised, but Not Unsuper-\n 2009/0116688 Al 5/2009 Monacos et al. vised, Models May Explain IT Cortical Representation, PLoS\n 2010/0048242 Al 2/2010 Rhoads et al. Computational Biology, vol. 10, Issue 11, 29 pages (Nov. 2014).\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 4 of 20\n\n\n US RE49,461 E\n Page 3\n\n\n(56) References Cited Minih, Volodymyr, Kavukcuoglu, Koray, Silver, David, Rusu, Andrei\n A, Veness, Joel, Bellemare, Marc G, Graves, Alex, Riedmiller,\n OTHER PUBLICATIONS Martin, Fidjeland, Andreas K, Ostrovski, Georg, et al. Human-level\n control through deep reinforcement learning. Nature, 518(7540):529-\nKim, S., Novel approaches to clustering, biclustering and algo- 533, Feb. 25, 2015.\nrithms based on adaptive resonance theory and Intelligent control, Montrym et al., The GeForce 6800, in IEEE Micro, vol. 25, No. 2,\nDoctoral Dissertations, Missouri University of Science and Tech- pp. 41-51, March-Apr. 2005.\nnology, 125 pages (2016). Moore, Andrew W and Atkeson, Christopher G. Prioritized sweep-\nKipfer, P., Segal, M., and Westermann, R. 2004. UberFlow: A ing: Reinforcement learning with less data and less time. Machine\nGPU-Based Particle Engine. In Proceedings of the SIGGRAPH/ Learning, 13(1): 103-130,1993.\nEurographics Workshop on Graphics Hardware 2004, pp. 115-122. Najernnik, J., and Geisler, W. (2009). Simple sununation rule for\nKolb, A., L. Latta, and C. RF7K-SALAMA. 2004. \"Hardware- optimal fixation selection in visual search. Vision Research. 49,\nBased Simulation and Collision Detection for Large Particle Sys- 1286-1294.\ntems.\" In Proceedings of the SIGGRAPH/Eurographics Workshop Non-Final Office Action dated Jan. 4, 2018 from U.S. Appl. No.\non Graphics Hardware 2004, pp. 123-131. 15/262,637, 23 pages.\nKompella, Varun Raj, Luciw, Matthew, and Schmidhuber, Jurgen. Non-Final Office Action dated May 31, 2018 from U.S. Appl. No.\nIncremental slow feature analysis: Adaptive low-complexity slow 14/947,516, 16 pages.\nfeature updating from high-dimensional input streams Neural Com- Notice of Alllowance dated May 22, 2018 from U.S. Appl. No.\nputation, 24( 11):2994-3024, 2012. 15/262,637, 6 pages.\nKowler, E. (2011). Eye movements: The past 25years. Vision Notice of Allowance dated Jul. 27, 2016 from U.S. Appl. No.\nResearch, 51(13), 1457-1483. doi:10.1016/j.visres.2010.12.014. 14/662,657.\nLarochelle H., & Hinton G. (2012). Learning to combine foveal Notice of Allowance dated Dec. 16, 2016 from U.S. Appl. No.\nglimpses with a third-order Boltzmann machine. NIPS 2010,1243- 14/662,657.\n1251. Oh, K.-S., and Jung, K. 2004. GPU implementation of neural\nLeCun, Y., Kavukcuoglu, K., & Farabet, C. (May 2010). Convolu- networks. Pattern Recognition 37, pp. 1311-1314.\ntional networks and applications in vision. In Circuits and Systems Oja, E. (1982). Simplified neuron model as a principal component\n(ISCAS), Proceedings of 2010 IEEE International Symposium on analyzer. Journal of Mathematical Biology 15(3), 267-273.\n(pp. 253-256). IEEE. Partial Supplementary European Search Report dated Jul. 4, 2017\nLee, D. D. and Seung, H. S. (1999). Learning the parts of objects from European Application No. 14800348.6, 13 pages.\nby non-negative matrix factorization. Nature, 401(6755):788-791. Perumalla, \"Discrete-event execution alternatives on general pur-\nLee, D. D., and Seung, H. S. (1997). \"Unsupervised learning by pose graphical processing units (GPGPU s).\" Proceedings of the\nconvex and conic coding.\" Advances in Neural Information Pro- 20th Workshop on Principles of Advanced and Distributed Simu-\ncessing Systems, 9. lation. IEEE Computer Society, 2006.8 pages.\nLegenstein, R., Wilbert, N., and Wiskott, L. Reinforcement learning Raijmakers, M.E.J., and Molenaar, P. (1997). Exact Art: A complete\non slow features of high-dimensional input streams. PLoS Compu- implementation of an ART network Neural networks 10 (4), 649-\ntational Biology, 6(8), 2010. ISSN 1553-734X. 13 pages. 669.\nLeveille, J., Ames, H., Chandler, B., Gorchetchnikov, A., Mingolla, Ranzato, M.A., Huang, F. J., Boureau, Y. L., & Lecun, Y. (2007,\nE., Patrick, S., and Versace, M. (2010) Learning in a distributed June). Unsupervised learning of invariant feature hierarchies with\nsoftware architecture for large-scale neural modeling. BIONET- applications to object recognition. In Computer Vision and Pattern\nICS 10, Boston, MA, USA. 8 pages. Recognition, 2007. CVPR'07. IEEE Conference on (pp. 1-8). IEEE.\nLivitz G., Versace M., Gorchetchnikov A., Vasilkoski Z., Ames H., Raudies, F., Eldridge, S., Joshi, A., and Versace, M. (Aug. 20, 2014).\nChandler B., Leveille J. andMingolla E. (2011) Adaptive, brain-like Learning to navigate in a virtual world using optic flow and stereo\nsystems give robots complex behaviors, The Neuromorphic Engi- disparity signals. Artificial Life and Robotics, DOI 10.1007/10015-\nneer,: 10.2417/1201101.003500 Feb. 2011. 3 pages. 014-0153-l. 15 pages.\nLivitz, G., Versace, M., Gorchetchnikov, A., Vasilkoski, Z., Ames, Adelson, E. H, Anderson, C. H, Bergen, JR., Burt, P. J, & Ogden,\nH., Chandler, B., Leveille, J., Mingolla, E., Snider, G., Amerson, R., J. M (1984) Pyramid methods in image processing. RCA engineer,\nCarter, D., Abdalla, H., and Qureshi, S. (2011) Visually-Guided 29(6), 33-41.\nAdaptive Robot (ViGuAR). Proceedings of the International Joint Aggarwal, Charu C, Hinneburg, Alexander, and Keim, Daniel A. On\nConference on Neural Networks (IJCNN) 2011, San Jose, CA, the surprising behavior of distance metrics in high dimensional\nUSA. 9 pages. space. Springer, 2001. 15 pages.\n Al-Kaysi, A. M. et al., A Multichannel Deep Belief Network for the\nLowe, D.G.(2004). Distinctive Image Features from Scale-Invariant\n Classification of EEG Data, from Ontology-based Information\nKeypoints. Journal International Journal of Computer Vision archive\n Extraction for Residential Land Use Suitability: A Case Study of the\nvol. 60, 2, 91-110.\n City of Regina, Canada, DOI 10.1007/978-3-319-26561-2_5, 8\nLu, Z.L., Liu, J., and Dosher, B.A.(2010) Modeling mechanisms of pages (Nov. 2015).\nperceptual learning with augmented Hebbian re-areighting Vision Ames, H, Versace, M., Gorchetchnikov, A., Chandler, B., Livitz, G.,\nResearch, 50(4). 375-390. Leveille, J., Mingolla, E., Carter, D., Abdalla, H., and Snider, G.\nLuo et al., \"Ailificial neural network computation on graphic (2012) Persuading computers to act more like brains. In Advances\nprocess unit.\" Neural Networks, 2005. IJCNN'05. Proceedings in Neuromorphic Mernristor Science and Applications, Kozma,\n2005 IEEE International Joint Conference on vol. 1 IEEE, 2005 pp. R.Pino,R., and Pazienza, G. (eds), Springer Verlag. 25 pages.\n622-626. Ames, H. Mingolla, E., Sohail, A., Chandler, B., Gorchetchnikov,\nMahadevan, S. Proto-value functions: Developmental reinforce- A., Leveille, J., Livitz, G. and Versace, M. (2012) The Animat. IEEE\nment learning. In Proceedings of the 22nd international conference Pulse, Feb. 2012, 3(1), 47-50.\non Machine learning, pp. 553-560. ACM, 2005. Apolloni, B. et al., Training a network of mobile neurons, Proceed-\nMeuth, J.R. and Wunsch, D.C. (2007) A Survey of Neural Compu- ings of International Joint Conference on Neural Networks, San\ntation On Graphics Processing Hardware. 22nd IEEE International Jose, CA, doi: 10.1109/IJCNN.2011.6033427, pp. 1683-1691 (Jul.\nSymposium on Intelligent Control, Part of IEEE Multi-conference 31-Aug. 5, 2011).\non Systems and Control, Singapore, Oct. 1-3, 2007, 5 pages. Artificial Intelligence as a Service. Invited talk, Defrag, Broomfield,\nMishkin M, Ungerleider LG. (1982). \"Contribution of striate inputs CO, Nov. 4-6, 2013. 22 pages.\nto the visuospatial functions of parieto-preoccipital cortex in mon- Aryananda, L. 2006. Attending to learn and learning to attend for a\nkeys,\" Behav Brain Res, 6 (1): 57-77. social robot. Humanoids 06, pp. 618-623.\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 5 of 20\n\n\n US RE49,461 E\n Page 4\n\n\n(56) References Cited Ellias, S. A., and Grossberg, S. 1975. Pattern formation, contrast\n control and oscillations in the short term memory of shunting\n OTHER PUBLICATIONS on-center off-surround networks Biol Cybern 20, pp. 69-98.\n Extended European Search Report and Written Opinion dated Jun.\nBaraldi, A. and Alpaydin, E. ( 1998). Simplified Art: A new class of 1, 2017 from European Application No. 14813864.7, 10 pages.\nArt algorithms. International Computer Science Institute, Berkeley, Extended European Search Report and Written Opinion dated Oct.\nCA, TR-98-004, 1998. 42 pages. 12, 2017 from European Application No. 14800348.6, 12 pages.\nBaraldi, A. and Alpaydin, E. (2002). Constructive feedforward Art Extended European Search Report and Written Opinion Oct. 23,\nclustering networks-Part I. IEEE Transactions on Neural Net- 2017 from European Application No. 15765396.5, 3 pages.\nworks 13(3), 645-661. Faz!, A., Grossberg, S., and Mingolla, E. (2009). View-invariant\nBaraldi, A. and Parmiggiani, F. (1997). Fuzzy combination of object category learning, recognition, and search: How spatial and\nKohonen's and ART neural network models to detect statistical object attention are coordinated using surface-based attentional\nregularities in a random sequence of multi-valued input patterns. In shrouds. Cognitive Psychology 58, 1-48.\nInternational Conference on Neural Networks, IEEE. 6 pages. Fiildiak, P. (1990). Forming sparse representations by local anti-\nBaraldi, Andrea and Alpaydin, Ethem. Constructive feedforward Hebbian learning, Biological Cybernetics, vol. 64, pp. 165-170.\nART clustering networks-part II IEEE Transactions on Neural Friston K., Adams R., Perrinet L., & Breakspear M. (2012). Per-\nNetworks, 13(3):662-677, May 2002. ISSN 1045-9227. doi: 10.1109/ ceptions as hypotheses: saccades as experiments. Frontiers in Psy-\ntnn.2002.1000131. URL http://dx.doi.org/l 0.1109/tnn.2002. chology, 3 (151), 1-20.\n1000131. Galbraith, B.V, Guenther, F.H., and Versace, M. (2015) A neural\nBengio, Y., Courville, A., & Vincent, P. Representation learning: A network-based exploratory learning and motor planning system for\nreview and new perspectives, IEEE Transactions on Pattern Analy- co-robots Frontiers in Neuroscience, in press. 10 pages.\nsis and Machine Intelligence, vol. 35 Issue 8, Aug. 2013. pp. George, D. and Hawkins, J. (2009). Towards a mathematical theory\n1798-1828. of cortical micro-circuits. PLoS Computational Biology 5( 10), 1-26.\nBerenson, D et al., A robot path planning framework that learns Georgii, J., and Westermann, R. 2005. Mass-spring systems on the\nfrom experience, 2012 International Conference on Robotics and GPU. Simulation Modelling Practice and Theory 13, pp. 693-702.\nAutomation, 2012, 9 pages [retrieved from the internet] URL:http:// Gorchetchnikov A., Hasselmo M. E. (2005). A biophysical imple-\nusers. wpi .edu/-dberenson/lightning. pdf. mentation of a bidirectional graph search algorithm to solve mul-\nBernhard, F., and Keriven, R. 2005. Spiking Neurons on GPUs. tiple goal navigation tasks. Connection Science, 17(1-2), pp. 145-\nTech. Rep. 05-15, Ecole Nationale des Ponts et Chauss'es, 8 pages. 166.\nBes!, P. J., & Jain, R. C. (1985). Three-dimensional object recog- Gorchetchnikov A., Hasselmo M. E. (2005). A simple rule for\nnition. ACM Computing Surveys (CSUR), 17(1), 75-145. spike-timing-dependent plasticity: local influence of AHP current.\nBoddapati, V., Classifying Environmental Sounds with Image Net- Neurocomputing, 65-66, pp. 885-890.\nworks, Thesis, Faculty of Computing Blekinge Institute of Tech- Gorchetchnikov A., Versace M., Hasselmo M. E. (2005). A Model\nnology, 37 pages (Feb. 2017). of STDP Based on Spatially and Temporally Local Information:\nBohn, C.-A. Kohonen. 1998. Feature Mapping Through Graphics Derivation and Combination with Gated Decay. Neural Networks,\nHardware. In Proceedings of 3rd Int. Conference on Computational 18, pp. 458-466.\nIntelligence and Neurosciences, 4 pages. Gorchetchnikov A., Versace M., Hasselmo M. E. (2005). Spatially\nBradski, G., & Grossberg, S. (1995). Fast-learning Viewnet archi- and temporally local spike-timing-dependent plasticity rule. In:\ntectures for recognizing three-dimensional objects from multiple Proceedings of the International Joint Conference on Neural Net-\ntwo-dimensional views. Neural Networks, 8 (7-8), 1053-1080. works, No. 1568 in IEEE CD-ROM Catalog No. 05CH37662C, pp.\nCanny, J.A. (1986). Computational Approach To Edge Detection, 390-396.\nIEEE Trans. Pattern Analysis and Machine Intelligence, 8(6):679- Gorcheichnikov, A. 2017. An Approach to a Biologically Realistic\n698. Simulation of Natural Memory. Master's thesis, Middle Tennessee\nCarpenter, G.A. and Grossberg, S. (1987). A massively parallel State University, Murfreesboro, TN, 70 pages.\narchitecture for a self-organizing neural pattern recognition machine. Grossberg, S. (1973). Contour enhancement, short-term memory,\nComputer Vision, Graphics, and Image Processing 37, 54-115. and constancies in reverberating neural networks. Studies in Applied\nCarpenter, G.A., and Grossberg, S. (1995). Adaptive resonance Mathematics 52, 213-257.\ntheory (ART). In M. Arbib (Ed.), The handbook of brain theory and Grossberg, S., and Huang, T.R. (2009). Artscene: A neural system\nneural networks, (pp. 79-82). Cambridge, M.A.: MIT press. for natural scene classification. Journal of Mision, 9 (4), 6.1-19.\nCarpenter, G.A., Grossberg, S. and Rosen, D.B. (1991). Fuzzy Art: doi: 10.1167/9.4.6.\nFast stable learning and categorization of analog patterns by an Grossberg, S., and Versace, M. (2008) Spikes, synchrony, and\nadaptive resonance system Neural Networks 4, 759-771. attentive learning by laminar thalamocortical circuits. Brain Research,\nCarpenter, Gail A and Grossberg, Stephen. The art of adaptive 1218C, 278-312 [Authors listed alphabetically].\npattern recognition by a self-organizing neural network. Computer, Hagen, T. R., Hjelmervik, J., Lie, K.-A., Natvig, J., and Ofstad\n21(3):77-88, 1988. Henriksen, M. 2005. Visual simulation of shallow-water waves.\nCoifman, R.R. and Maggioni, M. Diffusion wavelets. Applied and Simulation Modelling Practice and Theory 13, pp. 716-726.\nComputational Harmonic Analysis, 21(1):53-94, 2006. Hasselt, Hado Van. Double q-learning. In Advances in Neural\nCoifman, R.R., Lafon, S., Lee, A.B., Maggioni, M., Nadler, B., Information Processing Systems, pp. 2613-2621,2010.\nWarner, F., and Zucker, S.W. Geometric diffusions as a tool for Hinton, G. E., Osindero, S., and Teh, Y. (2006). A fast learning\nharmonic analysis and structure definition of data: Diffusion maps. algorithm for deep belief nets. Neural Computation, 18, 1527-1554.\nProceedings of the National Academy of Sciences of the United Ren, Y et al., Ensemble Classification and Regression-Recent\nStates of America, 102(21):7426, 2005. 21 pages. Developments, Applications and Future Directions, in IEEE Com-\nCornwall et al., Automatically translating a general purpose C++ putational Intelligence Magazine, 10.1109/MCI.2015.2471235, 14\nimage processing library for GPUs. Proceedings 20th IEEE Inter- pages (2016).\nnational Parallel & Distributed Processing Symposium, 2006, 8 Riesenhuber, M., & Poggio, T. (1999). Hierarchical models of\npages. object recognition in cortex. Nature Neuroscience, 2 (11), 1019-\nDavis, C. E. 2005. Graphic Processing Unit Computation of Neural 1025.\nNetworks. Master's thesis, University of New Mexico, Albuquer- Riesenhuber, M., & Poggio, T. (2000). Models of object recogni-\nque, NM, 121 pages. tion. Nature neuroscience, 3, 1199-1204.\nDosher, B.A., and Lu, Z.L. (2010). Mechanisms of perceptual Rolfes, T. 2004. Artificial Neural Networks on Programmable\nattention in preening of location. Vision Res., 40(10-12). 1269- Graphics Hardware. In Game Programming Gems 4, A. Kirmse, Ed.\n1292. Charles River Media, Hingham, MA, pp. 373-378.\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 6 of 20\n\n\n US RE49,461 E\n Page 5\n\n\n(56) References Cited Snider, Greg, et al. \"From synapses to circuitry: Using mernristive\n memory to explore the electronic brain.\" IEEE Computer, vol.\n OTHER PUBLICATIONS 44(2). (2011): 21-28.\n Spratling, M. W. (2008). Predictive coding as a model of biased\nRublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011). ORB: competition in visual attention. Vision Research, 48(12): 1391-1408.\nAn efficient alternative to SIFT or SURF. In IEEE International Spratling, M. W. (2012). Unsupervised learning of generative and\nConference on Computer Vision (ICCV) 2011, 2564-2571. discriminative weights encoding elementary image components in a\nRuesch, J et al. 2008. Multimodal Saliency-Based Bottom-Up predictive coding model of cortical function. Neural Computation,\nAttention a Framework for the Humanoid Robot iCub. 2008 IEEE 24(1):60-103.\nInternational Conference on Robotics and Automation, pp. 962-965. Spratling, M. W., De Meyer, K., and Kompass, R. (2009). Unsu-\nRumelhart D., Hinton G., and Williams, R. (1986). Learning inter- pervised learning of overlapping image components using divisive\nnal representations by error propagation. In Parallel distributed input modulation. Computational intelligence and neuroscience. 20\nprocessing: explorations in the microstructure of cognition, vol. 1, pages.\nMIT Press. 45 pages. Sprekeler, H. On the relation of slow feature analysis and laplacian\nRumpf, M. and Strzodka, R. Graphics processor units: New pros- eigenmaps. Neural Computation, pp. 1-16, 2011.\npects for parallel computing. In Are Magnus Bruaset and Aslak Sun, Z et al., Recognition of SAR target based on multilayer\nTveito, editors, Numerical Solution of Partial Differential Equations auto-encoder and SNN, International Journal of Innovative Com-\non Parallel Computers, vol. 51 of Lecture Notes in Computational puting, Information and Control, vol. 9, No. 11, pp. 4331-4341,\nScience and Engineering, pp. 89-134. Springer, 2005. Nov. 2013.\nSalakhutdinov, R., & Hinton, G. E. (2009). Deep boltzmann machines. Sutton, R. S., and Barto, A.G. (1998). Reinforcement learning: An\nIn International Conference on Artificial Intelligence and Statistics introduction(vol. 1, No. 1). Cambridge: MIT press. 10 pages.\n(pp. 448-455). Tong, F., Ze-Nian Li, (1995). Reciprocal-wedge transform for\nSchaul, Tom, Quan, John, Antonoglou, Ioannis, and Silver, David. space-variant sensing, Pattern Analysis and Machine Intelligence,\nPrioritized experience replay. arXiv preprint arXiv: 1511.05952, IEEE Transactions on , vol. 17, No. 5, pp. 500-551 doi: 10\nNov. 18, 2015. 21 pages. 1109/34.391393.\nSchmidhuber, J. (2010). Formal theory of creativity, fun, and Torralba, A., Oliva, A., Castelhano, M.S., Henderson, J.M. (2006).\nintrinsic motivation (1990-2010). Autonomous Mental Develop- Contextual guidance of eye movements and attention in real-world\nment, IEEE Transactions on, 2(3), 230-247. scenes: the role of global features in object search Psychological\nSchmidhuber, Jurgen. Curious model-building control systems. In Review, 113(4).766-786.\nNeural Networks, 1991. 1991 IEEE International Joint Conference Van Hasselt, Hado, Guez, Arthur, and Silver, David. Deep rein-\non, pp. 1458-1463. IEEE, 1991. forcement learning with double q-learning. arXiv preprint arXiv:\nSeibert, M., & Waxman, A.M. (1992). Adaptive 3-D Object Rec- 1509.06461, Sep. 22, 2015. 7 pages.\nognition from Multiple Views. IEEE Transactions on Pattern Analy- Versace, Brain-inspired computing. Invited keynote address, Bionet-\nsis and Machine Intelligence, 14 (2), 107-124. ics 2010, Boston, MA, USA. 1 page.\nSetoain et al., \"Parallel hyperspectral image processing on com- Versace, M. (2006) From spikes to interareal synchrony: how\nmodity graphics hardware.\" Parallel Processing Workshops, 2006. attentive matching and resonance control learning and Information\nICPP 2006 Workshops 2006 International Conference on. IEEE, processing by laminar thalamocortical circuits. NSF Science of\n2006. 8 pages. Learning Centers PI Meeting, Washington, DC, USA. 1 page.\nSherbakov, L. and Versace, M. (2014) Computational principles for Versace, M., (2010) Open-source software for computational neu-\nan autonomous active vision system. Ph.D., Boston University, roscience: Bridging the gap between models and behavior. In\nhttp://search.proquest.com/docview/l 558856407. 194 pages. Horizons in Computer Science Research, vol. 3 43 pages.\nSherbakov, L. et al. 2012. CogEye: from active vision to context Versace,M., Ames, H., Leveille, J., Fortenberry,B., and Gorchetchnikov,\nidentification, youtube, retrieved from the Internet an Oct. 10, 2017: A. (2008) KlnNeSS: A modular framework for computational\nURL://www.youtube.com/watch?v~i5PQk962Blk, 1 page. neuroscience Neuroinforrnatics, 2008 Winter; 6(4):291-309. Epub\nSherbakov, L. et al. 2013. CogEye: system diagram module brain Aug. 10, 2008.\narea function algorithm approx # neurons, retrieved from the Versace, M., and Chandler, B. (2010) MoNeta: A Mind Made from\nInternet on Oct. 12, 2017: URL://http://www-labsticc.univ-ubs.fr/ Mernristors. IEEE Spectrum, Dec. 2010. 8 pages.\n--coussy/neucomp2013/index_fichiers/material/posters/NeuCornp2013_ Versace, TEDx Fulbright, Invited talk, Washington DC, Apr. 5,\nfinal56x36.pdf, 1 page. 2014. 30 pages.\nSherbakov, L., Livitz, G., Sohail, A., Gorchetchnikov, A., Mingolla, Webster, Bachevalier, Ungerleider (1994). Connections of IT areas\nE., Ames, H., and Versace, M (2013b) A computational model of the TEO and TE with parietal and frontal cortex in macaque monkeys.\nrole of eye-movements in object disambiguation. Cosyne, Feb. Cerebal Cortex, 4(5), 470-483.\n28-Mar. 3, 2013. Salt Lake City, UT, USA. 2 pages. Wiskott, Laurenz and Sejnowski, Terrence. Slow feature analysis:\nSherbakov, L., Livitz, G., Sohail, A., Gorchetchnikov, A., Mingolla, Unsupervised learning ofinvariances. Neural Computation, 14(4):715-\nE., Ames, H., and Versace, M. (2013a) CogEye: An online active 770, 2002.\nvision system that disambiguates and recognizes objects NeuComp Wu, Yan & J. Cai, H. (2010). A Simulation Study of Deep Belief\n2013.2 pages. Network Combined with the Self-Organizing Mechanism of Adap-\nSmolensky, Paul. Information processing in dynamical systems: tive Resonance Theory. 10.1109/CISE.2010.56//265, 4 pages.\nFoundations of harmony theory. No. CU-CS-321-86. Colorado\nUniv At Boulder Dept of Computer Science, 1986. 88 pages. * cited by examiner\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 7 of 20\n\n\nU.S. Patent Mar.14,2023 Sheet 1 of 5 US RE49,461 E\n\n\n\n\n ,/\n ........... \u00b7\u00b7\u00b7\u00b7\u00b7t--\u00b7.c:\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 8 of 20\n\n\nU.S. Patent Mar.14,2023 Sheet 2 of 5 US RE49,461 E\n\n\n\n\n w\n rt: ~ X:\n \\,,,\n ;- 0\n :) 2;\n\n ..., X\n < ~\n ~\n d w w 00 :t, ..-. -0\n 0\n i,t:<\n N\n \\{)\n (~,\n ..1 r- :fil <.(}\n t>J\n ':\")\n ;-b\n (~\n\n\n\n /\n\n\n CJ\n \"<......\n N\n\n ~~ >-\n ~\n 0\n 0\n N\n .,\n <\".\\!\n\n\n -\n u !!\n w\n ..J\n V '?\n\n /\n z.U ,.J.\n 0 :)\n l.\n :Ez\n ~\n \u00b5 tt; I~ a..\n f.!: \u00b7'\u00a2(\n wm '(--,\n }\u00ab\n ~\n 'r-r {;)\u00b7\n 0 0\n >.::t \u00b7,AWMWM ,A~ \u2022M\u00b7,A \u2022, 'OW\n\n\n %:\n w - Mi2J.1ma11:)d\n _.,,,~..,\n ~\n 0\n (.'>)\n N\n '\"I ..,, 'T\n\n\n - -\n\n\n ,,.,\n 0\n -.:.:::-:\n ..... --\u00b7\n f~\n .......\n ~\n ,~\\;,j\u2022t: r.?-!.if:~/? ~/:f::\n \u00b7\u00b7, :.~.....'!'\u00b7\";\u2022,;-\u00b7.:\u00b7~:\u00b7:\u00b7~ .:-:;::~\n lh.~-:;_/~ ~/~//~\n 435\n 40;: 403\n !: ~rsp.H~ t~xr:1r:=:-~~-ff>i:=:.\n t>:::-:-:.~u:\n ::-:\u2022\n ~n>:::-~n:.ny\n I\n .j t;-;:u:}; ?:-;:GPU\n .\n Sl i~:~t1-:::-r~;:\n fi t)fn\n ~b:=:::!Jt:;::\u00b7\n r'f}f.::f::K::ry\n b::~nK:r:-GPU\n\n ;\u2666------------------\n :r\n\n\n\n\n 480,----~--~\n i Outp~:ttf:<_:,;\u00b7:sp<::,\n i vpk-.:;;:r!to ::';ix.~:Jrfs\n j r?>:3 r\":::(\u00b7!~).. (3;~t: k ,......................\n 440\n\n\n i::;:~~=~:,:~:;:::;t~:~::~\n f\n 4~}fj\n\n V'.J::i~t\n fr_;r:~w::tp i fr:=!\u00b7\n \\:\\'\u00b0::~~t f)W{)p\n r,..\u00b7\n \u2713 y5\n O{ lnpcl!i\u25catl!f,rn\n t:::\u2022:-:.b.Jr<::\n 44~=,\n ........\n p:_);ntr~f::'>\n 1............. t..l~~;~~;~r\u2022:::;~t',J\n N.:.'::..................\n L:~::-(t-.....\n .........,\n --------~~..........\n tt-~~ratK>11?,,.\"\"r\u00b7-.:--\n ..... ........_ ......... \u2713\n\n\n Y<::-s\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 11 of 20\n\n\n\n\n e\n \u2022\n 00\n \u2022\n ~\n ~\n ~\n ~\n\n =\n ~\n\n\n\n 3 .i 2 3 4 1 2 3 -\u00b7\u00b7.. -\u00b7\u00b7 ...\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\n .............-...... ....\n 1 2 3\n ~--\u2022-\u2022.\u2022\n -~\u00b7\u00b7\u00b7'\u00b7\u00b7\u00b7\u00b7'\n 6 5 6 7 8 4 5 6 4 5 6 ~\n ~\n\n7!8 9 \u00b7g 7 8 9 7 8 9 :-:\n ' \" ....\n ...\n ~\n N\n 0\n ::a) b) ........,.,.,\n ,, ...... ' .... -\u00b7\u00b7\n )1 \u00b7\u00b7\u00b7\u00b7\u00b7\u00b7=\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7\u00b7--\u2022\u2022'\u00b7\u2022\u00b7\u00b7\n N\n ~\n\n\n\n\n Bo1dlines shmv the packing of data into pixels ',Nlthfour color components,\n rJJ\n =-\n ('D\n ....\n ('D\n\n Ul\n\n ....\n 0\n\n Ul\n\n\n\n\n FIG\u00bb5 d\n rJl.\n\n ~\n ~\n \\0\n ~\n 0--,\n\n \"'\"\"'\n ~\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 12 of 20\n\n\n US RE49,461 E\n 1 2\n GRAPHIC PROCESSOR BASED processing unit/system (CPU). Complete independence is\n ACCELERATOR SYSTEM AND METHOD not desirable, however; user input might affect how the\n computation is performed and even interrupt it if necessary.\n Furthermore, the user output and the disk output are depen-\n Matter enclosed in heavy brackets [ ] appears in the 5 dent on the results of the computation. A reasonable solution\n original patent but forms no part of this reissue specifica- would be to separate input/output into threads, so that it is\n tion; matter printed in italics indicates the additions interacting with hardware occurs in parallel with the com-\n made by reissue; a claim printed with strikethrough putation. In this case whatever CPU processing is required\n indicates that the claim was canceled, disclaimed, or held for input/output should be designed so that it provides the\n invalid by a prior post-patent action or proceeding. 10 synchronization with computation.\n In the case of GPGPU, the computation itself is performed\n RELATED APPLICATIONS outside of the CPU, so the complete system comprises three\n \"peripheral\" components: user interactive hardware, disk\n The present application is a reissue continuation appli- hardware, and computational hardware. The central process-\n cation of U.S. application Ser. No. 15/808,201, which was 15 ing unit (CPU) establishes communication and synchroni-\nfiled on Nov. 9, 2017 as a broadening reissue application of zation between peripherals. Each of the peripherals is pref-\n U.S. Pat. No. 9,189,828, filed Jan. 3, 2014, which claims a erably controlled by a dedicated thread that is executed in\n priority benefit, under 35 U.S.C. \u00a7120, as a continuation of parallel with minimal interactions and dependencies on the\n U.S. application Ser. No. 11/860,254, now U.S. Pat. No. other threads.\n 8,648,867 B2, filed Sep. 24, 2007, entitled \"Graphic Pro- 20 A GPU on a conventional video card is usually controlled\n cessor Based Accelerator System and Method,\" which in through OpenGL, DirectX, or similar graphic application\n turn claims the priority benefit, under 35 U.S.C. \u00a7119(e), of programming interfaces (APis ). Such APis establish the\n U.S. Application No. 60/826,892, filed Sep. 25, 2006. The context of graphic operations, within which all calls to the\npresent application is also a broadening reissue application GPU are made. This context only works when initialized\n of U.S. Pat. No. 9,189,828, filed Jan. 3, 2014, which is a 25 within the same thread of execution that uses it. As a result,\n continuation of U.S. application Ser. No. 11/860,254, now in a preferred embodiment, the context is initialized within\n U.S. Pat. No. 8,648,867 B2, filed Sep. 24, 2007, which in a computational thread. This creates complications, how-\n turn claims the priority benefit, under 35 U.S. C. \u00a7 119(e), of ever, in the interaction between the user interface thread that\n U.S. Application No. 60/826,892, filed Sep. 25, 2006. Each changes parameters of simulations and the computational\n of the above-identified applications is incorporated herein by 30 thread that uses these parameters.\n reference in its entirety. More than one reissue application A solution as proposed here is an implementation of the\n has been filed for the reissue of U.S. Pat. No. 9,189,828, computational stream of execution in hardware, so that\n including this application and U.S. application Ser. No. thread and context initialization are replaced by hardware\n 15/808,201. initialization. This hardware implementation includes an\n 35 expansion card comprising a printed circuit board having (a)\n BACKGROUND one or more graphics processing units, (b) two or more\n associated memory banks that are logically or physically\n Graphics Processing Units (GPUs) are found in video partitioned, (c) a specialized controller, and (d) a local bus\n adapters (graphic cards) of most personal computers (PCs), providing signal coupling compatible with the PCI industry\n video game consoles, workstations, etc. and are considered 40 standards (this includes but is not limited to PCI-Express,\n highly parallel processors dedicated to fast computation of PCI-X, USB 2.0, or functionally similar technologies). The\n graphical content. With the advances of the computer and controller handles most of the primitive operations needed to\n console gaming industries, the need for efficient manipula- set up and control GPU computation. As a result, the CPU\n tion and display of 3D graphics has accelerated the devel- is freed from this function and is dedicated to other tasks. In\n opment of GPUs. 45 this case a few controls (simulation start and stop signals\n In addition, manufacturers of GPU shave included general from the CPU and the simulation completion signal back to\n purpose programmability into the GPU architecture leading CPU), GPU programs and input/output data are the infor-\n to the increased popularity of using GPU s for highly paral- mation exchanged between CPU and the expansion card.\n lelizable and computationally expensive algorithms outside Moreover, since on every time step of the simulation the\n of the computer graphics domain. When implemented on 50 results from the previous time step are used but not changed,\n conventional video card architectures, these general purpose the results are preferably transferred back to CPU in parallel\n GPU (GPGPU) applications are not able to achieve optimal with the computation.\n performance, however. There is overhead for graphics- In general, according to one aspect, the invention features\n related features and algorithms that are not necessary for a computer system. This system comprises a central pro-\n these non-video applications. 55 cessing unit, main memory accessed by the central process-\n ing unit, and a video system for driving a video monitor in\n SUMMARY response to the central processing unit as is common. The\n computer system further comprises an accelerator that uses\n Numerical simulations, e.g., finite element analysis, of input data from and provides output data to the central\n large systems of similar elements (e.g. neural networks, 60 processing unit. This accelerator comprises at least one\n genetic algorithms, particle systems, mechanical systems) graphics processing unit, accelerator memory for the graphic\n are one example of an application that can benefit from processing unit, and an accelerator controller that moves the\n GPGPU computation. During numerical simulations, disk input data into the at least one graphics processing unit and\n and user input/output can be performed independently of the accelerator memory to generate the output data.\n computation because these two processes require interac- 65 In the preferred, the central processing unit transfers the\n tions with peripheral hardware (disk, screen, keyboard, input data for a simulation to the accelerator, after which the\n mouse, etc) and put relatively low load on the central accelerator executes simulation computations to generate\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 13 of 20\n\n\n US RE49,461 E\n 3 4\nthe output data, which is transferred to the central processing In more detail, the computer system 100 in one example\nunit. Preferably, the accelerator controller dictates an order is a standard personal computer (PC). However, this only\nof execution of instructions to the at least one graphics serves as an example environment as computing environ-\nprocessing unit. The use of the separate controller enables ment 100 does not necessarily depend on or require any\ndata transfer during execution such that the accelerator 5 combination of the components that are illustrated and\ncontroller transfers output data from the accelerator memory described herein. In fact, there are many other suitable\nto main memory of the central processing unit. computing environments for this invention, including, but\n In the preferred embodiment, the accelerator controller not limited to, workstations, server computers, supercom-\ncomprises an interface controller that enables the accelerator puters, notebook computers, hand-held electronic devices\nto communicate over a bus of the computer system with the 10 such as cell phones, mp3 players, or personal digital assis-\ncentral processing unit. tants (PDAs), multiprocessor systems, progranimable con-\n In general according to another aspect, the invention also sumer electronics, networks of any of the above-mentioned\nfeatures an accelerator system for a computer system, which computing devices, and distributed computing environments\ncomprises at least one graphics processing unit, accelerator\n that including any of the above-mentioned computing\nmemory for the graphic processing unit and an accelerator 15\n devices.\ncontroller for moving data between the at least one graphics\n In one implementation the GPU accelerator is imple-\nprocessing unit and the accelerator memory.\n In general according to another aspect, the invention also mented as an expansion card 180 includes connections with\nfeatures a method for performing numerical simulations in a the motherboard 110, on which the one or more CPU's 120\ncomputer system. This method comprises a central process- 20 are installed along with main, or system memory 130 and\ning unit loading input data into an accelerator system from mass/non volatile data storage 140, such as hard drive or\nmain memory of the central processing unit and an accel- redundant array of independent drives (RAID) array, for the\nerator controller transferring the input data to a graphics computer system 100. In the current example, the expansion\nprocessing unit with instructions to be performed on the card 180 communicates to the motherboard 110 via a local\ninput data. The accelerator controller then transfers output 25 bus 190. This local bus 190 could be PCI, PCI Express,\ndata generated by the graphic processing unit to the central PCI-X, or any other functionally similar technology (de-\nprocessing unit as output data. pending upon the availability on the motherboard 110). An\n The above and other features of the invention including external version GPU accelerator is also a possible imple-\nvarious novel details of construction and combinations of mentation. In this example, the external GPU accelerator is\nparts, and other advantages, will now be more particularly 30 connected to the motherboard 110 through USB-2.0, IEEE\ndescribed with reference to the accompanying drawings and 1394 (Firewire), or similar external/peripheral device inter-\npointed out in the claims. It will be understood that the face.\nparticular method and device embodying the invention are The CPU 120 and the system memory 130 on the moth-\nshown by way of illustration and not as a limitation of the erboard 110 and the mass data storage system 140 are\ninvention. The principles and features of this invention may 35 preferably independent of the expansion card 180 and only\nbe employed in various and numerous embodiments without communicate with each other and the expansion card 180\ndeparting from the scope of the invention. through the system bus 200 located in the motherboard 110.\n A system bus 200 in current generations of computers have\n BRIEF DESCRIPTION OF THE DRAWINGS bandwidths from 3.2 GB/s (Pentium 4 withAGTL+, Athlon\n 40 XP with EV6) to around 15 GB/s (Xeon Woodcrest with\n In the accompanying drawings, reference characters refer AGTL+, Athlon 64/Opteron with Hypertransport), while the\nto the same parts throughout the different views. The draw- local bus has maximal peak data transfer rates of 4 GB/s\nings are not necessarily to scale; emphasis has instead been (PCI Express 16) or 2 GB/s (PCI-X 2.0). Thus the local bus\nplaced upon illustrating the principles of the invention. Of 190 becomes a bottleneck in the information exchange\nthe drawings: 45 between the system bus 200 and the expansion card 180. The\n FIG. 1 is a schematic diagram illustrating a computer design of the expansion card and methods proposed herein\nsystem including the GPU accelerator according to an minimizes the data transfer through the local bus 190 to\nembodiment of the present invention; reduce the effect of this bottleneck.\n FIG. 2 is block diagram illustrating the architecture for the The system memory 130 is referred to as the main\nGPU accelerator according to an embodiment of the present 50 random-access memory (RAM) in the description herein.\ninvention; However, this is not intended to limit the system memory\n FIG. 3 is a block/flow diagram illustrating an exemplary 130 to only RAM technology. Other possible computer\nimplementation of the top level control of the GPU accel- storage media include, but are not limited to ROM,\nerator system; EEPROM, flash memory, or any other memory technology.\n FIG. 4 is a flow diagram illustrating an exemplary imple- 55 In the illustrated example, the GPU accelerator system is\nmentation of the bottom level control of the GPU accelerator implemented on an expansion card 180 on which the one or\nsystem that is used to execute the target computation; and more GPU's 240 are mounted. It should be noted that the\n FIG. 5 is an example population of nine computational GPU accelerator system GPU 240 is separate from and\nelements arranged in a 3x3 square and a potential packing independent of any GPU on the standard video card 150 or\nscheme for texture pixels, according to an implementation of 60 other video driving hardware such as integrated graphics\nthe present invention. systems. Thus the computations performed on the expansion\n card 180 do not interfere with graphics display (including\n DETAILED DESCRIPTION but not limited to manipulation and rendering of images).\n Various brand of GPU are relevant. Under current tech-\n FIG. 1 shows a computer system 100 that has been 65 nology, GPU's based on the GeForce series from NVIDIA\nconstructed according to the principles of the present inven- Corporation or the Catalyst series from ATI/Advanced\ntion. Micro Devices, Inc.\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 14 of 20\n\n\n US RE49,461 E\n 5 6\n The output to a video monitor 170 is preferably through partition 250b is designed to hold the data textures repre-\nthe video card 150 and not the GPU accelerator system 180. senting internal variables. The third partition 250c is\nThe video card 150 is dedicated to the transfer of graphical designed to hold the data textures used as input at a\ninformation and connects to the motherboard 110 through a particular computation step on the GPU 240. The fourth\nlocal bus 160 that is sometimes physically separate from the 5 partition 250d holds the data textures used to accommodate\nlocal bus 190 that connects the expansion card 180 to the the output of a particular computational step on the GPU\nmotherboard 110. 240. This partitioning scheme can be done logically, does\n FIG. 2 is a block diagram illustrating the general archi- not require hardware implementation. Also the partitioning\ntecture of the GPU accelerator system and specifically the scheme is also altered based on new designs or needs of the\nexpansion card 180 in which at least one GPU 240 and 10 algorithms being employed. The reason for this partitioning\nassociated memories 210 and 250 are mounted. Electrical is further explained in the Data Organization section, below.\n(signal) and mechanical coupling with a local bus 190 A local bus interface 230 on the controller 220 serves as\nprovides signal coupling compatible with the PCI industry a driver that allows the controller 220 to communicate\nstandards (this includes but is not limited to PCI, PCI-X, PCI through the local bus 190 with the system bus 200 and thus\nExpress, or functionally similar technology). 15 the CPU 120 and RAM 130. This local bus interface 230 is\n The GPU accelerator further preferably comprises one not intended to be limited to PCI related technology. Other\nspecifically designed accelerator controller 220. Depending drivers can be used to interface with comparable technology\nupon the implementation, the accelerator controller 220 is as a local bus 190.\nfield programmable gate array (FPGA) logic, or custom built Data Organization\napplication-specific (ASIC) chip mounted in the expansion 20 Each computational element discussed above has output\ncard 180, and in mechanical and signal coupling with the variables that affect the rest of the system. For example in\nGPU 240 and the associated memories 210 and 250. During the case of a neural network it is the output of a neuron. A\ninitial design, a controller can be partially or even fully computational element also usually has several internal\nimplemented in software, in one example. variables that are used to compute output variables, but are\n The controller 220 commands the storage and retrieval of 25 not exposed to the rest of the system, not even to other\narrays of data (on a conventional video card the arrays of elements of the same population, typically. Each of these\ndata are represented as textures, hence the term 'texture' in variables is represented as a texture. The important differ-\nthis document refers to a data array unless specified other- ence between output variables and internal variables is their\nwise and each element of the texture is a pixel of color access.\ninformation), execution of GPU programs (on a conven- 30 Output variables are usually accessed by any element in\ntional video card these programs are called shaders, hence the system during every time step. The value of the output\nthe term 'shader' in this document refers to a GPU program variable that is accessed by other elements of the system\nunless specified otherwise), and data transfer between the corresponds to the value computed on the previous, not the\nsystem bus 200 and the expansion card 180 through the local current, time step. This is realized by dedicating two textures\nbus 190 which allows communication between the main 35 to output variables----one holds the value computed during\nCPU 120, RAM 130, and disk 140. the previous time step and is accessible to all computational\n Two memory banks 210 and 250 are mounted on the elements during the current time step, another is not acces-\nexpansion card 180. In some example, these memory banks sible to other elements and is used to accumulate new values\nseparated in the hardware, as shown, or alternatively imple- for the variable computed during the current time step.\nmented as a single, logically partitioned memory compo- 40 In-between time steps these two textures are switched, so\nnent. that newly accumulated values serve as accessible input\n The reason to separate the memory into two partitions 210 during the next time step, while the old input is replaced with\n250 stems from the nature of the computations to which the new values of the variable. This switch is implemented by\nGPU accelerator system is applied. The elements of com- swapping the address pointers to respective textures as\nputation (computational elements) are characterized by a 45 described in the System and Framework section.\nsingle output variable. Such computational elements often Internal variables are computed and used within the same\ninclude one or more equations. Computational elements are computational element. There is no chance of a race con-\nsame or similar within a large population and are computed dition in which the value is used before it is computed or\nin parallel. An example of such a population is a layer of after it has already changed on the next time step because\nneurons in an artificial neural network (ANN), where all 50 within an element the processing is sequential. Therefore, it\nneurons are described by the same equation. As a result, is possible to render the new value of internal variable into\nsome data and most of the algorithms are common to all the same texture where the old was read from in the texture\ncomputational elements within population, while most of the memory bank. Rendering to more than one texture from a\ndata and some algorithms are specific for each equation. single shader is not implemented in current GPU architec-\nThus, one memory, the shader memory bank 210, is used to 55 tures, so computational elements that track internal variables\nstore the shaders needed for the execution of the required would have to have one shader per variable. These shaders\ncomputations and the parameters that are common for all can be executed in order with internal variables computed\ncomputational elements and is coupled with the controller first, followed by output variables.\n220 only. The second memory, the texture memory bank Further savings of texture memory is achieved through\n250, is used to store all the necessary data that are specific 60 using multiple color components per pixel (texture element)\nfor every computational element (including, but not limited to hold data. Textures can have up to four color components\nto, input data, output data, intermediate results, and param- that are all processed in parallel on a GPU. Thus, to\neters) and is coupled with both the controller 220 and the maximize the use of GPU architecture it is desirable to pack\nGPU 240. the data in such a way that all four components are used by\n The texture memory bank 250 is preferably further par- 65 the algorithm. Even though each computational element can\ntitioned into four sections. The first partition 250 a is have multiple variables, designating one texture pixel per\ndesigned to hold the external input data patterns. The second element is ineffective because internal variables require one\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 15 of 20\n\n\n US RE49,461 E\n 7 8\ntexture and output variables require two textures. Further- Stream 303-runs on the GPU accelerator of the expansion\nmore, different element types have different numbers of card 180 and interacts with the User Interaction Stream 302\nvariables and unless this number is precisely a multiple of through initialization routines and data exchange in between\nfour, texture memory can be wasted. simulations. The Computational Stream 303 interacts with\n Amore reasonable packing scheme would be to pack four 5 the User Interaction Stream and the Data Output Stream\ncomputational elements into a pixel and have separate through synchronization procedures during simulations.\ntextures for every variable associated with each computa- The crucial feature of the interaction between the User\ntional element. In this case the packing scheme is identical Interaction Stream 302 and the Computational Stream 303 is\nfor all textures, and therefore can be accessed using the same the shift of priorities. Outside of the simulation, the system\nalgorithm. Several ways to approach this packing scheme 10 100 is driven by the user input, thus the User Interaction\nare outlined here. An example population of nine computa- Stream 302 has the priority and controls the data exchange\ntional elements arranged in a 3x3 square (FIG. Sa) can be 304 between streams. After the user starts the simulation, the\npacked by element (FIG. Sb), by row (FIG. Sc), or by square Computational Stream 303 takes the priority and controls\n(FIG. Sd). the data exchange between streams until the simulation is\n Packing by element (FIG. Sb) means that elements 1,2,3,4 15 finished or interrupted 350.\ngo into first pixel; 5,6,7,8 go into second pixel; 9 goes into The user starts 300 the framework through the means of\nthird pixel. This is the most compact scheme, but not an operating system and interacts with the software through\nconvenient because the geometrical relationship is not pre- the user interaction section 305 of the graphic user interface\nserved during packing and its extraction depends on the size 306 executed on the CPU 120. The start 300 of the imple-\nof the population. 20 mentation begins with a user action that causes a GUI\n Packing by row (colunm; FIG. Sc) means that elements initialization 307, Disk input/output initialization 308 on the\n1,2,3 go into pixel (1,1); 3,4,5 go into pixel (2,1), 7,8,9 go CPU 120, and controller initialization 320 of the GPU\ninto pixel (3,1). With this scheme the element's y coordinate accelerator on the expansion card 180. GUI initialization\nin the population is the pixel's y coordinate, while the includes opening of the main application window and setting\nelement's x coordinate in the population is the pixel's x 25 the interface tools that allow the user to control the frame-\ncoordinate times four plus the index of color component. work. Disk I/O initialization can be performed at the start of\nFive by five populations in this case will use 2x5 texture, or the framework, or at the start of each individual simulation.\n10 pixels. Five of these pixels will only use one out of four The user interaction 305 controls the setting and editing of\ncomponents, so it wastes 37.5% of this texture. 25xl popu- the computational elements, parameters, and sources of\nlation will use 6xl texture (six pixels) and will waste 12.5% 30 external inputs. It specifies which equations should have\nof it. their output saved to disk and/or displayed on the screen. It\n Packing by square (FIG. Sd) means that elements 1,2,4,5 allows the user to start and stop the simulation. And it\ngo into pixel (1,1); 3,6 go into pixel (1,2); 7,8 go into pixel performs standard interface functions such as file loading\n(2,1), and 9 goes into pixel (2,2). Both the row and the and saving, interactive help, general preferences and others.\ncolunm of the element are determined from the row (col- 35 The user interaction 305 directs the CPU 120 to acquire\nunm) of the pixel times two plus the second (first) bit of the the new external input textures needed (this includes but is\ncolor component index. Five by five populations in this case not limited to loading from disk 140 or receiving them in\nwill use 3x3 texture, or 9 pixels. Four of these pixels will real time from a recording device), parses them if necessary\nonly use two out of four components, and one will only use 309, and initializes their transfer to the expansion card 180,\none component, so it wastes 34.4% of this texture. This is 40 where they are stored 325 in the texture memory bank 250\nmore advantageous than packing by row, since the texture is by the controller 220. The user interaction 305 also directs\nsmaller and the waste is also lower. 25xl population on the the CPU 120 to parse populations of elements that will be\nother hand will use 13xl texture (thirteen pixels) and waste used in the simulation, convert them to GPU programs\n>50% of it, which is much worse than packing by row. (shade rs), compile them 310, and initializes their transfer to\n In order to eliminate waste altogether the population 45 the expansion card 180, where they are stored 326 in the\nshould have even dimensions in the square packing, and it shader memory bank 210 by the controller 220. This opera-\nshould have a number of columns divisible by four in row tion is accompanied by the upload 309 of the initial data into\npacking. Theoretically, the chances are approximately the input partition of the texture memory bank 250, and\nequivalent for both of these cases to occur, so the particular stores the shader order of execution in the controller 220.\ntask and data sizes should determine which packing scheme 50 The user can perform operations 309 and 310 as many times\nis preferable in each individual case. as necessary prior to starting the simulation or between\n The System and Framework simulations.\n FIG. 3 shows an exemplary implementation of the top The editing of the system between simulations is difficult\nlevel system and method that is used to control the compu- to accomplish without the hardware implementation of the\ntation. It is a representation of one of several ways in which 55 computational thread suggested herein. The system of equa-\na system and method for processing numerical techniques tions (computational elements) is represented by textures\ncan be implemented in the invention described herein and so that track variables plus shaders that define processing\nthe implementation is not intended to be limited to the algorithms. As mentioned above, textures, shaders and other\nfollowing description and accompanying figure. graphics related constructs can only be initialized within the\n The method presented herein includes two execution 60 rendering context, which is thread specific. Therefore tex-\nstreams that run on the CPU 120-User Interaction Stream tures and shaders can only be initialized in the computa-\n302 and Data Output Stream 301. These two streams pref- tional thread.\nerably do not interact directly, but depend on the same data Network editing is a user-interactive process, which\naccumulated during simulations. They can be implemented according to the scheme suggested above happens in the\nas separate threads with shared memory access and executed 65 User Interaction Stream 302. The simulation software thus\non different CPUs in the case of multi-CPU computing has to take the new parameters from the User Interaction\nenvironment. The third execution stream-Computational Stream 302, communicate them to the Computational\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 16 of 20\n\n\n US RE49,461 E\n 9 10\nStream 303 and regenerate the necessary shaders and tex- practical software implementation of the method and archi-\ntures. This is hard to accomplish without a hardware imple- tecture described above and pictorially represented in FIG.\nmentation of the Computational Stream 303. The Compu- 3.\ntational Stream 303 is forked from the User Interaction To use SANNDRA, the application should create a\nStream and it can access the memory of the parent thread, 5 TSimulator object either directly or through inheritance.\nbut the reverse communication is harder to achieve. The This object will handle global simulation properties and\ncontroller 220 allows operations 309 and 310 to be per- control the User Interaction Stream, Data Output Stream,\nformed as many times as necessary by providing the nec- and Computational Stream. Through TSimulator: :time-\nessary communication to the User Interaction Stream 302. step( ) TSimulator: :outfileinterval( ), and TSimulator: :out-\n After execution of the input parser texture generation 309 lO mode( ), the application can set the time step of the simu-\nand population parser shader generator and compiler 310 are lation, the time step of disk output, and the mode of the disk\nperformed at least once, the user has the option to initialize output. The external input pattern should be packed into a\nthe simulation 311. During this initialization the main con- TPattem object and bound to the simulation object through\ntrol of the framework is transferred to the GPU accelerator 15 TSimulator::resetinputs( ) method. TSimulator::sim-\nsystem's accelerator controller 220 and computation 330 is Length( ) sets the length of the simulation.\nstarted (see FIG. 4; 420). The user retains the ability to The second step is to create at least one population of\ninterrupt the simulation, change the input, or to change the equations (Tpopulation object). Population holds one equa-\n tion object TEquation. This object contains only a formula\ndisplay properties of the framework, but these interactions\n and does not hold element-specific data, so all elements of\nare queued to be performed at times determined by the 20\n the population can share single TEquation.\ncontroller-driven data exchange 314 and 316 to avoid the\n The TEquation object is converted to a GPU program\ncorruption of the data. before execution. GPU programs have to be executed within\n The progress monitor 312 is not necessary for perfor- a graphical context, which is stream specific. TSimulator\nmance, but adds convenience. It displays the percentage of creates this context within a Computational Stream, there-\ncompleted time steps of the simulation and allows the user 25 fore all programs and data arrays that are necessary for\nto plan the schedule using the estimates of the simulation computation have to be initialized within Computational\nwall clock times. Controller-driven data exchange 314 Stream. Constructor of TPopulation is called from User\nupdates the display of the results 313. Online screen output Interaction Stream, so no GPU-related objects can be ini-\nfor the user selected population allows the user to monitor tialized in this constructor.\nthe activity and evaluate the qualitative behavior of the 30 TPopulation: :fillElements( ) is a virtual method designed\nnetwork. Simulations with unsatisfactory behavior can be to overcome this difficulty. It is called from within the\nterminated early to change parameters and restart. Control- Computational Stream after TSimulator: :networkCreate( ) is\nler-driven data exchange 314 also drives the output of the called in the User Interaction Stream. A user has to override\nresults to disk 317. Data output to disk for convenience can 35 TPopulation::fillElements( ) to create TEquation and other\nbe done on an element per file basis. A suggested file format computation related objects both element independent and\nincludes a leftmost colunm that displays a simulated time for element-specific. Element independent objects include sub-\neach of the simulation steps and subsequent colunms that components of TEquation and objects that describe how to\ndisplay variable values during this time step in all elements handle interdependencies between variables implemented\n through derivatives of TGate class.\nwith identical equations (e.g. all neurons in a layer of a 40\n Element-specific data is held in TElement objects. These\nneural network).\n objects hold references to TEquation and a set of TGate\n Controller-driven data exchange or input parser texture\n objects. There is one TElement per population, but the size\ngenerator 316 allows the user to change input that is gen- of data arrays within this object corresponds to population\nerated on the fly during the simulation. This allows the size. All TElement objects have to be added to the TSimu-\nframework monitoring of the input that is coming from a 45 lator list of elements by calling TSimulator::addUnit( )\nrecording device (video camera, microphone, cell recording method from TPopulation: :fillElements( ).\nelectrode, etc) in real time. Similar to the initial input parser Finally, TPopulation::fillElements() should contain a set\n309, it preprocesses the input into a universal format of the of TElement::add*Dependency( ) calls for each element.\ndata array suitable for texture generation and generates Each of these calls sets a corresponding dependency for\ntextures. Unlike the initial parser 309, here the textures are 50 every TGate object. Here TGate object holds element inde-\ntransferred to hardware not whenever ready but upon the pendent part of dependency and TElement::\nrequest of the controller 220. add*Dependency( ) sets element-specific details.\n The controller 220 also drives the conditional testing 315 System provided TPopulation handles the output of com-\nand 318 informs the CPU-bound streams whether the simu-\n putational elements, both when they need to exchange the\nlation is finished. If so, the control returns to the User 55 data and when they need to output it to disk. User imple-\nInteraction Stream. The user then can change parameters or mentation of TPopulation derivative can add screen output.\ninputs (309 and 310), restart the simulation (311) or quit the Listing 1 is an example code of the user program that uses\nframework (390). a recurrent competitive field (RCF) equation:\n SANNDRA (Synchronous Artificial Neuronal Network\nDistributed Runtime Algorithm; http://www.kinness.net/ 60\n LISTING 1\nDocs/SANNDRA/html) was developed to accelerate and\noptimize processing of numerical integration of large non- uint16_t w - 3, h - 3;\nhomogenous systems of differential equations. This library static float m_compet = 0.5;\n static float m_persist = 1.0;\nis fully reworked in its version 2.x.x to support multiple class TCablePopRCF : public TPopulation\ncomputational backends including those based on multicore 65 {\nCPUs, GPUs and other processing systems. GPU based TEq_RCF* m_equation;\nbackend for SANNDRA-2.x.x can serve as an example\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 17 of 20\n\n\n US RE49,461 E\n 11 12\n LISTING I-continued for new time step. To avoid data confusion, the new values\n of variables should be rendered in a separate texture. After\nTGate* m_gatel; the time step is completed for all equations, these new values\nTGate* m_gate2;\nvoid createGatingStructure( )\n should be copied over old values so that they are used as\n{ 5 input during the next time step. Copying textures is an\nm_gatel - new TGate(0); expensive operation, computationally, but since the textures\nm_gate2 - new TGate(l); are referred to by texture IDs (pointers), swapping these\n}; pointers for input and output textures after each time step\nvoid createUnitStructure(TBasicUnit* u)\n achieves the same result at a much lesser cost.\n{\nu->addO2OPlnputDependency(m_gatel, 0., 0., 0.004, 0., 0, 0); 10\n In the hardware solution suggested herein, ID swapping is\nu->addFullDependency(m_gate2, population()); equivalent to swapping the base memory address for two\n} partitions of the texture memory bank 250. They are\npublic: TCablePopRCF() : TPopulation(\"compCPU RCF\", w, h, true) { }; swapped 485 during synchronization (485, 430, and 455) so\n~TCablePopRCF() {if(m_equation) delete m_equation;\n that data transfer 445 and the computation 435-487 proceeds\n if(m_gatel) delete m_gatel;\n if(m_gate2) delete m_gate2;}; immediately and in parallel with data transfer as shown in\n 15 FIG. 4. A hardware solution allows this parallelism through\nboo! fillElements(TSimulator* sim);\n}; access of the controller 220 to the onboard texture memory\nboo! TCablePopRCF::fillElements(TSimulatior* sim) bank 250.\n{ The main computation and data exchange are executed by\nm_equation - new TEq_RCF(this, m_compet, m_persist);\ncreateGatingStructure( );\n the controller 220. It runs three parallel substreams of\nfor(size_t i - 0; i < xSize( ); ++i) 20 execution: Computational Substream 403, Data Output Sub-\nfor(size_t j - 0; j < ySize( ); ++j) stream 402, and Data Input Substream 404. These streams\n{ are synchronized with each other during the swap of pointers\nTElement* u - new TCPUElement(this, m_equation, i, j); 485 to the input and output texture memory partitions of the\nsim->addUnit(u);\ncreate U nitStructure( u);\n texture memory bank 250 and the check for the last iteration\n} 25 487. Algorithmically, these two operations are a single\nReturn true; atomic operation, but the block diagram shows them as two\n} separate blocks for clarity.\nint The Computational Substream 403 performs a computa-\nmain() tional cycle including a sequential execution of all shaders\n{\n// Input pattern generation (309 in FIG.3) that were stored in the shader memory bank 210 using the\n 30\nuint32_t* pat - new uint32_t[w*h]; appropriate input and output textures. To begin the simula-\nTRandom randGen (0); tion the controller 220 initializes three execution sub streams\nfor(uint32_t I - 0; I < w*h; ++i) 403, 402, and 404. On every simulation step, the Compu-\npat[i] - randGen.random( ); tational Substream 403 determines which textures the GPU\nTpattern* p - new Tpattern(pat, w, h);\n// Setting up the simulation\n 240 will need to perform the computations and initiates the\n 35 upload 435 of them onto the GPU 240. The GPU 240 can\nTSimulator* cableSim - new TSimulator(\"data\"); //(308 and 320 in\nFIG. 3) communicate directly with the texture memory bank 250 to\ncableSim->timestep(0.05); //(320 in FIG. 3) upload the appropriate texture to perform the computations.\ncableSim->resetlnputs(p); //(325 in FIG. 3) The controller 220 also pulls the first shader (known by the\ncableSim->outfileinterval(0.1); //(308 in FIG. 3)\ncableSim->outmode(SANNDRA::timefunc); //(308 in FIG. 3)\n stored order) from the shader memory bank 210 and uploads\ncableSim->simLength(60.0); //(320 in FIG. 3) 40 450 it onto the GPU 240.\n// Preparing the population The GPU 240 executes the following operations in this\nTPopulation* cablePop - new TCablePopRCF( ); //(310 in FIG. 3) order: performs the computation (execution of the shader)\ncableSim->networkCreate( ); //(326 in FIG. 3) 470; tells the controller 220 that it is done with the compu-\nuintl 6_t user= 1;\n tations for the current shader; and after all shaders for this\nwhile(user)\n{ 45 particular equation are executed sends 480 the output tex-\nif(! cableSim->simulationStart(true, 1)) //(311 in FIG. 3) tures to the output portion of the texture memory bank 250.\nexit(!); This cycle continues through all of the equations based on\nstd::cout<<\"Repeat?ln\"; //(305 in FIG. 3) the branching step 482.\nstd::cin>>user; //(305 in FIG. 3) An example shader that performs fourth order Runge-\nif(user -- 1)\ncableSim->networkReset( ); //(305 in FIG. 3)\n Kutta numerical integration is shown in Listing 2 using\n 50\n{ GLSL notation;\nIf(cableSim)\nDelete cableSim; //Also deletes cablePop and its internals LISTING 2\nexit(0);\n}; uniform sarnpler2DRect Variable;\n 55 uniform float integration_step;\n float halfstep - integration_step*0.5;\n FIG. 4 is a detailed flow diagram illustrating a part of an float fl_6step - integration_step/6.0;\nexemplary implementation of the bottom level system and vec4 output - texture2DRect(Variable, gl_TexCoord[0].st);\n // define equation( ) here\nmethod performed during the computation on the GPU\n vec4 rungekutta4(vec4 x)\naccelerator of the expansion card 180 and is a more detailed {\nview of the computational box 330 in FIG. 3. FIG. 4 is a 60 canst vec4 kl - equation(x);\nrepresentation of one of several ways in which a system and canst vec4 k2 - equation(x + halfstep*kl);\nmethod for processing numerical techniques can be imple- canst vec4 k3 - equation(x + halfstep*k2);\n canst vec4 k4 - equation(x + integration step*k3);\nmented. return fl_6step*(kl + 2.0*(k2 + k3) + k4);\n With systems of equations that have complex interdepen- }\ndencies it is likely that the variable in some equation from 65 Void main(void)\na previous time step has to be used by some other equation {\nafter the new values of this variable are already computed\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 18 of 20\n\n\n US RE49,461 E\n 13 14\n LISTING 2-continued cycle (or less frequently as defined by the user), writing this\n output to disk 140, and displaying this output on the monitor\n output +- rungekutta4( output);\n 170. This frees the CPU 120 to execute other applications\n gl_FragColor - output;\n }\n and allows the expansion card to run at its full capacity\n 5 without being slowed down by extensive interactions with\n the CPU 120.\n The shader in Listing 2 can be executed on conventional 2. Minimizing data transfer between the expansion card\nvideo card. Using the controller 220 this code can be further 180 and the system bus 200. All of the information needed\noptimized, however. Since the integration step does not to perform the simulations will be stored on the expansion\nchange during the simulation, the step itself as well as the 10 card 180 and all simulations will take place on it. Further-\nhalfstep and 1/4 of the step can be computed once per more, whatever data transfer remains necessary will take\nsimulation, and updated in all shaders by a shader update place in parallel with the computation, thus reducing the\nprocedures 310, 326 discussed above. impact of this transfer on the performance.\n After all of the equations in the computational cycle are 3. New way to execute GPU programs (shaders). Previ-\ncomputed the main execution substream 403 on the control- 15 ously, the CPU 120 had full control over the order of\n!er 220 can switch 485 the reference pointers of the input and shader's execution and was required to produce specific\noutput portions of the texture memory bank 250. commands on every cycle to tell the GPU 240 which shader\n The two other substreams of execution on the controller to use. With the invention disclosed herein, shaders will\n220 are waiting (blocks 430 and 455, respectively) for this initially be stored on the shader memory bank 210 on the\nswitch to begin their execution. The Data Input Substream\n 20 expansion card 180 and will be sent to the GPU 240 for\n404 is controlling 440 the input of additional data from the execution by the general purpose controller 220 located on\nCPU 120. This is necessary in cases where the simulation is the expansion card.\nmonitoring the changing input, for example input from a 4. Multiple parallelisms. The GPU 240 is inherently\nvideo camera or other recording device in the real time. This\n parallel and is well suited to perform parallel computations.\nsubstream uploads new external input from the CPU 120 to\n In parallel with the GPU 240 performing the next calcula-\nthe texture memory bank 250 so it can be used by the main 25\ncomputational sub stream 403 on the next computational step tion, the controller 220 is uploading the data from the\nand waits for the next iteration 475. The Data Output previous calculation into main memory 130. Furthermore,\nSubstream 445 controls the output of simulation results to the CPU 120 at the same time uses uploaded previous results\nthe CPU 120 if requested by the user. This substream to save them onto disk 140 and to display them on the screen\nuploads the results of the previous step to the main RAM 30 through the system bus 200.\n130 so that the CPU 120 can save them on disk 140 or show 5. Reuse of existing and affordable technology. All hard-\nthem on the results display 313 and waits for the next ware used in the invention and mentioned here-in are based\niteration 460. on currently available and reliable components. Further\n Since the Computational Substream 403 determines the advance of these components will provide straightforward\ntiming of input 440 and output 445 data transfers, these data 35 improvements of the invention.\ntransfers are driven by the controller 220. To further reduce While this invention has been particularly shown and\nthe data transfer overhead (and disk 140 overhead also) the described with references to preferred embodiments thereof,\ncontroller 220 initiates transfer only after selected compu- it will be understood by those skilled in the art that various\ntational steps. For example, if the experimental data that is changes in form and details may be made therein without\nsimulated was recorded every 10 milliseconds (msec) and\n 40 departing from the scope of the invention encompassed by\nthe simulation for better precision was computed every 1 the appended claims.\nmsec, then only every tenth result has to be transferred to\nmatch the experimental frequency.\n This solution stores two copies of output data, one in the What is claimed is:\nexpansion card texture memory bank 250 and another in the [1. A computer system, comprising:\nsystem RAM 130. The copy in the system RAM 130 is 45 a central processing unit to receive input data;\naccessed twice: for disk I/O and screen visualization 313. An main memory, operably coupled to the central processing\nalternative solution would be to provide CPU 120 with a unit via a bus, to store the input data received by the\ndirect read access to the onboard texture memory bank 250 central processing unit;\nby mapping the memory of the hardware onto a global an accelerator, operably coupled to the central processing\nmemory space. The alternative solution will double the 50 unit and the first memory via the bus, to receive at least\ncommunication through the local bus 190. Since the goal a portion of the input data from the main memory, the\ndiscussed herein is reducing the information transfer through accelerator comprising:\nthe local bus 190, the former solution is favored. at least one graphics processing unit to perform a\n The main substream 403 determines if this is the last sequence of computations on the at least a portion of\niteration 487. If it is the last iteration, the controller 220 55 the input data so as to generate output data, inter-\nwaits for the all of the execution substreams to finish 490 mediate computations in the sequence of computa-\nand then returns the control to the CPU 120, otherwise it tions yielding intermediate results; and\nbegins the next computational cycle.\n accelerator memory, operably coupled to the graphic\n This repeats through all of the computational cycles of the\n processing unit, to store the results of the plurality of\nsimulation.\n 60 sequential computations; and\n CONCLUSION a controller, operably coupled to the at least one graphics\n processing unit and the accelerator memory, to transfer\n This GPU accelerator system offers the following poten- the at least a portion of the input data into the accel-\ntial advantages: erator memory, and to transfer at least a portion of the\n 1. Limited computations on the CPU 120. The CPU 120 65 output data from the accelerator memory to the main\nis only used for user input, sending information to the memory during performance of the sequence of com-\ncontroller 220, receiving output after each computational putations by the at least one graphic processing unit.]\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 19 of 20\n\n\n US RE49,461 E\n 15 16\n [2. The computer system of claim 1, wherein the central [13. The method of claim 12, further comprising:\nprocessing unit is configured to receive the input data in storing the input data in the main memory in response to\nresponse to a user interaction.] a user interaction.]\n [3. The computer system of claim 1, wherein: [14. The method of claim 12, further comprising:\n the central processing unit is configured to receive the 5 receiving the input data at a first rate; and\n input data at a first rate; and wherein (A) comprises performing the sequence of com-\n the at least one graphics processing unit is configured to putations at a second rate different than the first rate.]\n perform the sequence of computations at a second rate [15. The method of claim 12, wherein (A) comprises:\n different than the first rate.] generating an output representative of an output of at least\n 10\n [4. The computer system of claim 1, wherein the main one neuron in an artificial neural network.]\nmemory is configured to store a copy of the output data [16. The method of claim 12, wherein (C) comprises:\nstored in the accelerator memory.] transferring the second portion of the output data from the\n [5. The computer system of claim 1, wherein an output of accelerator memory to the main memory without trans-\nat least one computation in the sequence of computations 15 ferring any of the intermediate results of the plurality of\nrepresents an output of at least one neuron in an artificial sequential computations from the accelerator memory\nneural network.] to the main memory so as to reduce data transfer via the\n [6. The computer system of claim 1, wherein accelerator bus.]\nmemory comprises: [17. The method of claim 12, wherein (C) comprises:\n a first memory bank to store parameters common to all of 20 transferring the second portion of the output data from the\n the computations in the sequence of computations; and accelerator memory to the main memory after the GPU\n a second memory bank to store data specific to at least one has begun to perform another sequence of computa-\n computation in the sequence of computations.] tions.]\n [7. The computer system of claim 1, wherein the control- [18. The method of claim 17, wherein (C) further com-\nler is configured to transfer the output data from the accel- 25 prises:\nerator memory to the main memory without transferring any initiating transfer of the second portion of the output data\nof the intermediate results from the accelerator memory to in parallel with performance of at least one computa-\nthe main memory so as to reduce data transfer via the bus.] tion in the other sequence of computations.]\n [8. The computer system of claim 1, wherein the control- [19. The method of claim 12, further comprising:\nler is configured to transfer at least a portion of the output 30 acquiring the input data in real time with at least one of\ndata from the accelerator memory to the main memory after a video camera, a microphone, or a cell recording\nthe at least one graphics processing unit has begun to electrode operably coupled to the CPU.]\nperform another sequence of computations.] [20. The method of claim 12, further comprising:\n [9. The computer system of claim 8, wherein the control- 35 storing parameters common to all of the computations in\n!er is configured to initiate transfer of the at least a portion the sequence of computations in a first memory bank in\nof the input data and to transfer the at least a portion of the the accelerator memory; and\noutput data in parallel with performance of at least one storing data specific to at least one computation in the\ncomputation in the other sequence of computations by the at sequence of computations in a second memory bank in\nleast one graphics processing unit.] 40 the accelerator memory.]\n [10. The computer system of claim 1, wherein the con- 21. A method of executing computations representing an\ntroller is configured to control execution of the sequence of artificial neural network on a computer system comprising\ncomputations by the at least one graphics processing unit.] at least one central processing unit (CPU), a processing\n [11. The computer system of claim 1, further comprising: unit, a first memory partition, and a second memory parti-\n at least one of a video camera, a microphone, or a cell 45 tion, the method comprising:\n recording electrode, operably coupled to the central executing, by the at least one CPU, a user interaction\n processor unit, to acquire the input data in real time.] stream, the user interaction stream controlling transfer\n [12. A method of performing a sequence of computations of inputs to the artificial neural network to the first\non a computer system comprising a central processing unit memory partition and the second memory partition;\n(CPU), a main memory operably coupled to the central 50 executing, by the processing unit, a computational stream,\nprocessing unit via a bus, an accelerator operably coupled to the computational stream controlling data exchange\nthe CPU and the main memory via the bus, the accelerator between the user interaction stream and the computa-\ncomprising a graphics processing unit (GPU) and an accel- tional stream during execution of the computations\nerator memory, the method comprising: representing the artificial neural network;\n (A) performing, by the GPU, the sequence of computa- 55 shifting control of a data exchange between the user\n tions on a first portion of the input data so as to generate interaction stream and the computational stream to the\n a first portion of the output data, intermediate compu- computational stream in response to starting execution\n tations in the sequence of computations yielding inter- of the computations representing the artificial neural\n mediate results; network;\n (B) in parallel with performing the sequence of compu- 60 shifting control of the data exchange between the user\n tations by the GPU in (A), transferring a second portion interaction stream and the computational stream to the\n of the input data from the main memory to the accel- user interaction stream in response to completion or\n erator via the bus; and interruption of the computations representing the arti-\n (C) in parallel with performing the sequence of compu- ficial neural network;\n tations by the GPU in (A), transferring a second portion 65 queueing a user command received by the user interaction\n of the output data from the accelerator memory to the stream during execution of the computations represent-\n main memory via the bus.] ing the artificial neural network; and\n\f Case 7:26-cv-00318 Document 1-5 Filed 08/17/26 Page 20 of 20\n\n\n US RE49,461 E\n 17 18\n executing the user command during execution of the a second memory partition;\n computations representing the artificial neural network at least one central processing unit (CPU), operably\n at times determined by the computational stream. coupled to the camera, the first memory partition, and\n 22. The method of claim 21, wherein the user interaction the second memory partition, to execute a user inter-\nstream controls the data exchange between the user inter- 5 action stream, the user interaction stream controlling\naction stream and the computational stream outside of transfer of the input data acquired by the camera to the\nexecution of the computations representing the artificial first memory partition and the second memory partition\nneural network. during execution of the computations representing the\n 23. The method of claim 21, wherein executing the user artificial neural network;\ninteraction stream comprises: 10\n a processing unit, operably coupled to the first memory\n controlling setting and editing of computational elements\n partition, the second memory partition, and the at least\n of the computations representing the artificial neural\n one CPU, to execute a computational stream, the\n network\n 24. The method of claim 21, wherein executing the user computational stream controlling transfer of the input\ninteraction stream comprises: 15\n data from the first memory partition and the second\n controlling setting and editing of parameters of the com- memory partition during execution of the computations\n putations representing the artificial neural network. representing the artificial neural network, the execution\n 25. The method of claim 21, wherein executing the user of the computations representing the artificial neural\ninteraction stream comprises: network occurring while the camera is acquiring the\n controlling setting and editing of parameters of the inputs 20 input data; and\n to the artificial neural network. a controller, operably coupled to the at least one CPU and\n 26. The method of claim 21, wherein executing the user the processing unit, to queue user interactions received\ninteraction stream comprises: by the user interaction stream during the execution of\n specifying an output to be saved to disk and/or displayed the computations representing the artificial neural net-\n on a screen. 25\n work for performance at times selected to avoid data\n 27. The method of claim 21, wherein executing the user corruption.\ninteraction stream comprises: 34. The system of claim 33, wherein the user interactions\n parsing elements to be used in the computations repre- cause interruption of the computations representing the\n senting the artificial neural network. artificial neural network.\n 28. The method of claim 27, wherein the processing unit 30 35. The system of claim 33, wherein the user interactions\ncomprises a graphics processing unit ( GPU) and executing cause a change in inputs to the artificial neural network.\nthe user interaction stream further comprises: 36. The system of claim 33, wherein the user interactions\n converting the elements into GPU programs. cause a change in display properties of an output of the\n 29. The method of claim 28, wherein executing the user computations representing the artificial neural network.\ninteraction stream comprises: 35\n 3 7. The system of claim 33, wherein the controller is\n compiling the GPU programs. configured to request the input data during the execution of\n 30. The method of claim 29, wherein executing the user the computations representing the artificial neural network.\ninteraction stream comprises: 38. The method of claim 21, wherein the user command\n transferring the GPU programs to the second memory causes interruption of the computations representing the\n partition. 40\n artificial neural network.\n 31. The method of claim 21, further comprising: 39. The method of claim 21, wherein the user command\n executing, by the at least one CPU, a data output stream, causes a change in the inputs to the artificial neural net-\n the data output stream controlling transfer of outputs of work.\n the computations representing the artificial neural net- 40. The method of claim 21, wherein the user command\n work to disk. 45\n causes a change in display properties of an output of the\n 32. The method of claim 21, further comprising: computations representing the artificial neural network.\n generating the inputs with a video camera during execu- 41. The method of claim 21, wherein, during execution of\n tion of the computations. the computations representing the artificial neural network,\n 33. A system for executing computations representing an the computational stream controls the data exchange\nartificial neural network, the system comprising: 50\n between the user interaction stream and the computational\n a camera to acquire input data for the artificial neural stream by requesting the inputs to the artificial neural\n network; network.\n a first memory partition; * * * * *\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:29.221892-07:00","document_number":"1","attachment_number":5,"pacer_doc_id":"181037209215","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 4","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294676/","id":490294676,"tags":[],"absolute_url":"/docket/74659430/1/6/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.102112-07:00","date_modified":"2026-08-21T19:46:03.286747-07:00","sha1":"139712393608679690482efc26c4a12579583e89","page_count":27,"file_size":3012612,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.6.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.6.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 1 of 27\n\n\n\n\n EXHIBIT\n\n 5\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 2 of 27\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n\n Plaintiff, Civil Action No. 7:24-cv-00221-ADA-DTG\n\n v. JURY TRIAL DEMANDED\n\nNVIDIA CORPORATION,\n\n Defendant.\n\n\n PLAINTIFF NEURAL AI, LLC\u2019S NOTICE OF SERVICE\n OF SUBPOENA TO TESLA, INC.\n\n PLEASE TAKE NOTICE that Plaintiff Neural AI, LLC (\u201cNeural AI\u201d) will serve (1) a\n\nSubpoena to Produce Documents, Information, or Objects on Tesla, Inc. (\u201cTesla\u201d), attached hereto\n\nas Attachment 1, and (2) a Subpoena for Testimony on Tesla, attached as Attachment 2.\n\nDated: June 24, 2026\n\n\n Respectfully submitted,\n\n /s/ Emily Portuguese\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n\fCase 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 3 of 27\n\n\n\n\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 5125174\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 4 of 27\n\n\n\n\n CERTIFICATE OF SERVICE\n\n\n I certify that on June 24, 2026 true and correct copies of Neural AI\u2019s Subpoena to\n\nProduce Documents, Information, or Objects and Subpoena for Testimony on Tesla has been\n\nserved all counsel of record electronically.\n\n\n /s/ Emily Portuguese\n Emily Portuguese\n\fCase 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 5 of 27\n\n\n\n\n Attachment 1\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 6 of 27\nAO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action\n\n\n UNITED STATES DISTRICT COURT\n for the\n __________\n Western District of __________\n Texas\n\n Neural Al, LLC )\n Plaintiff )\n v. ) Civil Action No. 7:24-cv-00221-ADA-DTG\n NVIDIA Corporation )\n ) (If the action is pending in another district, state where:\n Defendant ) __________ District of __________ )\n\n SUBPOENA TO PRODUCE DOCUMENTS, INFORMATION, OR OBJECTS\n OR TO PERMIT INSPECTION OF PREMISES IN A CIVIL ACTION\n Tesla,\nTo: Google Inc.\n LLC\n\n\n\n -\n c/o C T Corporation\n clo Corporation System, 1999\n Service Company BryanLawyers\n OBA CSC- St., Suite\n Inc,900, Dallas,\n 211 E. TX 75201\n 7th Street, Suite 620, Austin, TX 78701\n\n Production: YOU ARE COMMANDED to produce at the time, date, and place set forth below the following\ndocuments, electronically stored information, or objects, and permit their inspection, copying, testing, or sampling of the\nmaterial: See Exhibit A\n\n\n\n Place: Planet Depos- Downtown Austin clo Lexitas Legal, Date and Time:\n 100 Congress Ave, Ste. 2000, Austin, TX 78701\n 0711412026 11 :59 pm\n\n\n 0 Inspection of Premises: YOU ARE COMMANDED to permit entry onto the designated premises, land, or\nother property possessed or controlled by you at the time, date, and location set forth below, so that the requesting party\nmay inspect, measure, survey, photograph, test, or sample the property or any designated object or operation on it.\n\n Place: Date and Time:\n\n\n\n\n The provisions of Fed. R. Civ. P. 45(c), relating to your protection as a person subject to a subpoena, and Rule\n45 (d) and (e), relating to your duty to respond to this subpoena and the potential consequences of not doing so, are\nattached.\n\n\nDate: 06/24/2026\n\n CLERK OF COURT\n OR\n Isl Emily Portuguese\n Signature of Clerk or Deputy Clerk Attorney\u2019s signature\n\n\nThe name, address, e-mail, and telephone number of the attorney representing (name of party)\nNeural Al LLC , who issues or requests this subpoena, are:\nEmily Portuguese, Susman Godfrey LLP, One Manhattan West, 50th Floor, New York, New York 10001\neportuguese@susmangodfrey.com, 212-729-2082\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 7 of 27\nAO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action (Page 2)\n\nCivil Action No. 7:24-cv-00221-ADA-DTG\n\n PROOF OF SERVICE\n (This section should not be filed with the court unless required by Fed. R. Civ. P. 45.)\n\n This subpoena for (name of individual and title, if any)\nwas received by me on (date) .\n\n 0 I served the subpoena by delivering a copy to the named person as follows:\n\n\n on (date) ; or\n\n 0 I returned the subpoena unexecuted because:\n .\n\n Unless the subpoena was issued on behalf of the United States, or one of its officers or agents, I have also\n tendered to the witness fees for one day\u2019s attendance, and the mileage allowed by law, in the amount of\n $ .\n\nMy fees are $ for travel and $ for services, for a total of $ 0.00 .\n\n\n I declare under penalty of perjury that this information is true.\n\n\nDate:\n Server\u2019s signature\n\n\n\n Printed name and title\n\n\n\n\n Server\u2019s address\n\nAdditional information regarding attempted service, etc:\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 8 of 27\n AO 88B (Rev. 06/09) Subpoena to Produce Documents, Information, or Objects or to Permit Inspection of Premises in a Civil Action(Page 3)\n\n\n\n\n Federal Rule of Civil Procedure 45 (c), (d), and (e) (Effective 12/1/07)\n(c) Protecting a Person Subject to a Subpoena. (d) Duties in Responding to a Subpoena.\n (1) Avoiding Undue Burden or Expense; Sanctions. A party or (1) Producing Documents or Electronically Stored Information.\nattorney responsible for issuing and serving a subpoena must take These procedures apply to producing documents or electronically\nreasonable steps to avoid imposing undue burden or expense on a stored information:\nperson subject to the subpoena. The issuing court must enforce this (A) Documents. A person responding to a subpoena to produce\nduty and impose an appropriate sanction \u2014 which may include lost documents must produce them as they are kept in the ordinary\nearnings and reasonable attorney\u2019s fees \u2014 on a party or attorney course of business or must organize and label them to correspond to\nwho fails to comply. the categories in the demand.\n (2) Command to Produce Materials or Permit Inspection. (B) Form for Producing Electronically Stored Information Not\n (A) Appearance Not Required. A person commanded to produce Specified. If a subpoena does not specify a form for producing\ndocuments, electronically stored information, or tangible things, or electronically stored information, the person responding must\nto permit the inspection of premises, need not appear in person at the produce it in a form or forms in which it is ordinarily maintained or\nplace of production or inspection unless also commanded to appear in a reasonably usable form or forms.\nfor a deposition, hearing, or trial. (C) Electronically Stored Information Produced in Only One\n (B) Objections. A person commanded to produce documents or Form. The person responding need not produce the same\ntangible things or to permit inspection may serve on the party or electronically stored information in more than one form.\nattorney designated in the subpoena a written objection to (D) Inaccessible Electronically Stored Information. The person\ninspecting, copying, testing or sampling any or all of the materials or responding need not provide discovery of electronically stored\nto inspecting the premises \u2014 or to producing electronically stored information from sources that the person identifies as not reasonably\ninformation in the form or forms requested. The objection must be accessible because of undue burden or cost. On motion to compel\nserved before the earlier of the time specified for compliance or 14 discovery or for a protective order, the person responding must show\ndays after the subpoena is served. If an objection is made, the that the information is not reasonably accessible because of undue\nfollowing rules apply: burden or cost. If that showing is made, the court may nonetheless\n (i) At any time, on notice to the commanded person, the serving order discovery from such sources if the requesting party shows\nparty may move the issuing court for an order compelling production good cause, considering the limitations of Rule 26(b)(2)(C). The\nor inspection. court may specify conditions for the discovery.\n (ii) These acts may be required only as directed in the order, and (2) Claiming Privilege or Protection.\nthe order must protect a person who is neither a party nor a party\u2019s (A) Information Withheld. A person withholding subpoenaed\nofficer from significant expense resulting from compliance. information under a claim that it is privileged or subject to\n (3) Quashing or Modifying a Subpoena. protection as trial-preparation material must:\n (A) When Required. On timely motion, the issuing court must (i) expressly make the claim; and\nquash or modify a subpoena that: (ii) describe the nature of the withheld documents,\n (i) fails to allow a reasonable time to comply; communications, or tangible things in a manner that, without\n (ii) requires a person who is neither a party nor a party\u2019s officer revealing information itself privileged or protected, will enable the\nto travel more than 100 miles from where that person resides, is parties to assess the claim.\nemployed, or regularly transacts business in person \u2014 except that, (B) Information Produced. If information produced in response to a\nsubject to Rule 45(c)(3)(B)(iii), the person may be commanded to subpoena is subject to a claim of privilege or of protection as trial-\nattend a trial by traveling from any such place within the state where preparation material, the person making the claim may notify any\nthe trial is held; party that received the information of the claim and the basis for it.\n (iii) requires disclosure of privileged or other protected matter, if After being notified, a party must promptly return, sequester, or\nno exception or waiver applies; or destroy the specified information and any copies it has; must not use\n (iv) subjects a person to undue burden. or disclose the information until the claim is resolved; must take\n (B) When Permitted. To protect a person subject to or affected by reasonable steps to retrieve the information if the party disclosed it\na subpoena, the issuing court may, on motion, quash or modify the before being notified; and may promptly present the information to\nsubpoena if it requires: the court under seal for a determination of the claim. The person\n (i) disclosing a trade secret or other confidential research, who produced the information must preserve the information until\ndevelopment, or commercial information; the claim is resolved.\n (ii) disclosing an unretained expert\u2019s opinion or information that\ndoes not describe specific occurrences in dispute and results from (e) Contempt. The issuing court may hold in contempt a person\nthe expert\u2019s study that was not requested by a party; or who, having been served, fails without adequate excuse to obey the\n (iii) a person who is neither a party nor a party\u2019s officer to incur subpoena. A nonparty\u2019s failure to obey must be excused if the\nsubstantial expense to travel more than 100 miles to attend trial. subpoena purports to require the nonparty to attend or produce at a\n (C) Specifying Conditions as an Alternative. In the circumstances place outside the limits of Rule 45(c)(3)(A)(ii).\ndescribed in Rule 45(c)(3)(B), the court may, instead of quashing or\nmodifying a subpoena, order appearance or production under\nspecified conditions if the serving party:\n (i) shows a substantial need for the testimony or material that\ncannot be otherwise met without undue hardship; and\n (ii) ensures that the subpoenaed person will be reasonably\ncompensated.\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 9 of 27\n\n\n\n\n EXHIBIT A\n\n DEFINITIONS AND INSTRUCTIONS\n\n 1. The term \u201cNVIDIA GPUs\u201d means the Hopper, Ada Lovelace, Ampere, Turing,\n\nVolta, Pascal, Maxwell, Jetson, and Blackwell architectures of NVIDIA graphics processing units.\n\nFor avoidance of doubt, those architectures include the following devices: DGX line of\n\nsupercomputers and servers (including at least DGX B300, DGX B200, DGX GB200, DGX\n\nGB300, DGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200,\n\nDGX H200, DGX H100, DGX BasePOD, DGX SuperPOD with H200, DGX A100), HGX line of\n\nsupercomputers and servers (including at least HGX B300, HGX B200, HGX H100, HGX H200,\n\nEos SuperPOD), OVX line of supercomputers and servers (including at least OVX L40S), EGX\n\nline of supercomputers and servers (including at least EGX Server with Quadro RTX A6000, EGX\n\nServer with A40, EGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000),\n\nGB300 NVL72, GB200 NVL72; Nvidia\u2019s GPU accelerators and superchips, including those with\n\nNVIDIA\u2019s Blackwell, Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, and Maxwell GPU\n\narchitectures, including at least, RTX PRO 6000 Server Edition, RTX PRO 6000 Workstation,\n\nRTX PRO 6000 Max-Q Workstation, RTX PRO 5000, RTX PRO 4500, RTX PRO 4000, RTX\n\nPRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, GB300, GB200, H100, H200,\n\nGH200, GH100, L40, L40S, L4, RTX 6000, RTX 6000 Ada, RTX 5000, Ada, RTX 4500 Ada,\n\nRTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3000, RTX 2000, RTX 1000, RTX 500, RTX\n\n4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060 Ti,\n\nand RTX 4060, GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce RTX\n\n4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU, A100,\n\nA40, A30, A16, A10, A2, A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 10 of 27\n\n\n\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500, GeForce RTX\n\n3090 Ti, GeForce RTX 3090, GeForce RTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti,\n\nGeForce RTX 3070, GeForce RTX 3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB),\n\nGeForce RTX 3050 (6 GB), GeForce RTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop\n\nGPU, GeForce RTX 3070 Ti Laptop GPU, GeForce RTX 3070 Laptop GPU, GeForce RTX 3060\n\nLaptop GPU, GeForce RTX 3050 Ti Laptop GPU, GeForce RTX 3050 Laptop GPU, GeForce\n\nMX570 Laptop GPU, Tesla T4 GPUs, Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 8000,\n\nQuadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000,\n\nT1000 8GB, T1200, Quadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB, Titan\n\nRTX, GeForce RTX 2080 Ti, GeForce RTX 2080, Super, GeForce RTX 2080, GeForce RTX 2070\n\nSuper, GeForce RTX 2070, GeForce RTX 2060 Super, GeForce RTX 2060, GeForce RTX 2500,\n\nGeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630, GeForce MX550, GeForce MX450, GeForce MX430, Tesla V100,\n\nQuadro GV100, Titan V GPU, Tesla P100, P40, P4, Quadro GP100, Quadro P6000, Quadro\n\nP5200, Quadro P5000, Quadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro\n\nP2200, Quadro P2000, Quadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500,\n\nQuadro P400, Titan Xp, Titan X, GeForce GTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070\n\nTi, GeForce GTX 1070, GeForce GTX 1060, GeForce GTX 1050 Ti, GeForce GTX 1050,\n\nGeForce MX300, GeForce MX200, GeForce MX150, Tesla M60, M40, M10, Quadro M6000\n\n24GB, Quadro M6000 (12GB), Quadro M5000, Quadro M5000M, Quadro M5500, Quadro\n\nM4000, Quadro M4000M, Quadro M3000M, Quadro M2200, Quadro M2000, Quadro M2000M,\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 11 of 27\n\n\n\n\nQuadro M1200, Quadro M1000M, Quadro M620, Quadro M600M, Quadro M520, Quadro\n\nM500M, NVS 810, Tesla M6, GTX Titan X, GeForce GTX 980Ti, GeForce GTX 980, GeForce\n\nGTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M,\n\nGeForce GTX 960M, GeForce GTX 950M, GeForce GTX 750 Ti, GeForce GTX 750, GeForce\n\nMX130, and GeForce MX110; and Jetson modules, including at least the Jetson Thor Series,\n\nJetson Thor, Jetson T5000, Jetson T4000, Jetson AGX Orin Series, Jetson AGX Orin Developer\n\nKit, Jetson AGX Orin 64GB, Jetson AGX Orin Industrial, Jetson AGX Orin 32GB, Jetson Orin\n\nNX Series, Jetson Orin NX 16GB, Jetson Orin NX 8GB, Jetson Orin Nano Series, Jetson Orin\n\nNano Super Developer Kit, Jetson Orin Nano 8GB, Jetson Orin Nano 4GB, Jetson AGX Xavier\n\nSeries, Jetson AGX Xavier Industrial, Jetson AGX Xavier 64GB, Jetson AGX Xavier 32GB,\n\nJetson Xavier NX Series, Jetson Xavier NX 16GB, Jetson Xavier NX 8GB, Jetson TX2 Series,\n\nJetson TX2i, Jetson TX2, Jetson TX2 4GB, Jetson TX2 NX, Jetson Nano, any and all variations\n\nof the aforementioned products (including at least products having different options for number of\n\nGPUs).\n\n 2. The terms \u201cand\u201d and \u201cor\u201d are not intended to be read disjunctively but rather\n\nconjunctively unless the context of a particular request clearly indicates otherwise. \u201cOr\u201d should be\n\nunderstood to include and encompass \u201cand\u201d; and \u201cand\u201d should be understood to include and\n\nencompass \u201cor.\u201d\n\n 3. The terms \u201cany\u201d or \u201ceach\u201d should be understood to include and encompass \u201call.\u201d\n\n 4. The terms \u201cconcerning,\u201d \u201crelated to\u201d or \u201crelating to\u201d, and \u201cregarding\u201d and any\n\nvariation of these terms mean analyzing, alluding to, concerning, considering, commenting on,\n\nconsulting, comprising, containing, contradicting, describing, dealing with, discussing,\n\nestablishing, evidencing, identifying, involving, noting, recording, reporting on, relating to,\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 12 of 27\n\n\n\n\nreflecting, referring to, regarding, stating, showing, studying, mentioning, memorializing, or\n\npertaining to, directly or indirectly, in whole or in part.\n\n 5. The term \u201cCPU(s)\u201d means Central Processing Unit(s).\n\n 6. The term \u201cDocument(s)\u201d shall have the broadest meaning possible under Federal\n\nRules 26 and 34 and shall include without limitation: documents, Electronically Stored\n\nInformation, communications in written, electronic, and recorded form, and tangible things. A\n\ndraft or non-identical copy of a document shall be considered a separate document within the\n\nmeaning of the term \u201cdocument.\u201d Any comment, notation, or other marking shall be sufficient to\n\ndistinguish documents that are otherwise similar in appearance and to make them separate\n\ndocuments for purposes of your response. Any preliminary form, intermediate form, superseded\n\nversion, or amendment of any document is to be considered a separate document.\n\n 7. The term \u201cGPU(s)\u201d means Graphics Processing Unit(s).\n\n 8. The terms \u201cinclude\u201d and \u201cincluding\u201d mean including without limitation.\n\n 9. The term \u201cNVIDIA\u201d means Defendant NVIDIA Corporation, its predecessors,\n\npresent and former directors, officers, accountants, affiliates, attorneys, partners, managers, agents,\n\nemployees, representatives, in-house and outside counsel, and any other person or entity acting on\n\nbehalf of or under control of Defendant NVIDIA Corporation.\n\n 10. The term \u201cperson(s)\u201d means and includes natural persons and formal or informal\n\nentities and organizations, including public and private corporations, partnerships, professional\n\ncorporations, limited liability companies, business trusts, banking institutions, associations, firms,\n\njoint ventures, commissions, bureaus, departments, and any other legal entity, including\n\nany divisions, subsidiaries, departments, and other units thereof.\n\n 11. The term \u201cSource Code\u201d means human-readable instructions written in a\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 13 of 27\n\n\n\n\nprogramming language, including all comments, annotations, declarations, functions, classes, and\n\nother components used to define the behavior of a software program. For purposes of these\n\nrequests, \u201cSource Code\u201d includes all associated files necessary to understand, compile, and execute\n\nthe code, such as scripts, header files, makefiles, configuration files, and documentation. Unless\n\notherwise stated, \u201cSource Code\u201d includes all versions and revisions relevant to the time periods\n\nand subject matter described in each interrogatory.\n\n 12. The terms \u201cYou\u201d or \u201cYour\u201d refer to Tesla, Inc., including but not limited to its\n\npredecessors, successors, parents, subsidiaries, divisions, affiliates, and all past or present\n\ndirectors, officers, partners, managers, employees, contractors, agents, representatives,\n\naccountants, consultants, in-house and outside counsel, and any other person or entity acting or\n\npurporting to act on its behalf or subject to its control. This definition expressly includes, without\n\nlimitation, any Tesla parent, subsidiary, affiliate, or other related entity that has used, licensed,\n\ndeployed, evaluated, or integrated NVIDIA GPUs or software.\n\n 13. The use of the singular form of any word includes the plural and vice versa.\n\n 14. These Requests seek the production of all documents, electronically stored\n\ninformation (\u201cESI\u201d), and tangible things in your possession, custody, or control, as that phrase is\n\nused in Federal Rule of Civil Procedure 34, as of the date of compliance with this subpoena and\n\nthat come into your possession, custody, or control at any time prior to production. This includes\n\nmaterials held by You directly, as well as by Your affiliates, subsidiaries, agents, representatives,\n\nor any other person or entity acting on Your behalf.\n\n 15. If You are aware of the existence (past or present) of any responsive documents,\n\nESI, or tangible items that are not in Your current possession, custody, or control, You must\n\nidentify such materials and provide:\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 14 of 27\n\n\n\n\n a. A description of the item(s);\n\n b. The name and contact information of the person or entity currently believed\n\n to have possession, custody, or control; and\n\n c. The reason You are unable to produce the material(s).\n\n 16. If You believe that no responsive documents, ESI, or tangible things exist in\n\nresponse to a particular request, You must state so in writing with respect to that request.\n\n 17. If You withhold any document, ESI, or portion thereof based on a claim of attorney-\n\nclient privilege, work product doctrine, or any other legal protection, You must produce a privilege\n\nlog that complies with Federal Rule of Civil Procedure 26(b)(5).\n\n 18. You must produce all documents and ESI: as they are kept in the usual course of\n\nbusiness or organized and labeled to correspond to the categories in this subpoena, as required\n\nunder FRCP 34(b)(2)(E); in their native electronic format (with original metadata intact) wherever\n\npossible, or as searchable, OCR-scanned PDFs with corresponding load files; with complete\n\nfamily groupings (e.g., attachments must be produced with their parent emails/documents); and in\n\nthe same folders or directories in which they were maintained, preserving original file structures\n\nand naming conventions.\n\n 19. You must maintain and produce a record of the source of each document or ESI\n\nitem produced, including:\n\n a. The file path or directory location;\n\n b. The name of the custodian (individual, team, or department) from whose\n files the document was collected;\n\n c. The system or platform (e.g., email server, shared drive, cloud service) from\n which the document was obtained.\n\nThis information may be provided in metadata load files, a source log, or other mutually agreed\n\nformat. This instruction is consistent with Federal Rule of Civil Procedure 34(b)(2)(E) and\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 15 of 27\n\n\n\n\nproportional discovery principles under Rule 26(b)(1). If any of the above information is not\n\nreasonably available or unduly burdensome to collect, You must so state and explain the basis for\n\nthat assertion.\n\n 20. For all ESI, You must preserve and produce standard metadata fields, including but\n\nnot limited to: filename, filepath, author, date created, date last modified, recipients, sender,\n\nsubject line (for emails), and document type. You must not degrade or alter metadata through\n\nprocessing or production.\n\n 21. You are under a continuing obligation to supplement or correct Your production if\n\nYou discover or obtain additional responsive materials prior to the close of discovery or the\n\nresolution of this matter.\n\n 22. Unless otherwise stated, the relevant period is from September 13, 2018 to the\n\npresent.\n\n 23. Unless otherwise stated, all requests herein are limited to documents, electronically\n\nstored information, and tangible things that relate to NVIDIA GPU-Acceleration Hardware or\n\nNVIDIA GPU-Acceleration Software that were:\n\n a. purchased, acquired, licensed, used, implemented, deployed, tested, or\n evaluated within the United States, or\n\n b. purchased, acquired, licensed, or used for the purpose of supporting,\n enabling, or operating any facility, system, team, data center, personnel,\n product, service, customer, or business activity located in or directed toward\n the United States.\n\nThis instruction is intended to encompass both U.S.-based activity and non-U.S. activity that\n\ndirectly supports or enables U.S. operations or usage.\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 16 of 27\n\n\n\n\n REQUESTS FOR PRODUCTION OF DOCUMENTS\n\n1. Documents sufficient to identify all software, frameworks, libraries, APIs, scripts, Source\n\n Code, configuration files, and custom code You use to perform computations on NVIDIA\n\n GPUs.\n\n2. Documents sufficient to show whether You use NVIDIA\u2019s Aerial, Clara Parabricks, cuBLAS,\n\n cuDNN, cuFFT, cuQuantum, cuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac,\n\n Isaac Lab, Maxine, Memory Map, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo,\n\n PyTorch, RAPIDS, Riva, Runtime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream),\n\n or any other NVIDIA software as part of computations You perform using NVIDIA GPUs.\n\n3. Documents sufficient to show whether You use sample Source Code provided by NVIDIA as\n\n part of computations You perform using NVIDIA GPUs.\n\n4. Documents sufficient to show whether and how any software You use to perform\n\n computations on NVIDIA GPUs calls, invokes, interfaces with, wraps, depends on, sits on top\n\n of, modifies, extends, or implements functionality provided by CUDA, cuDNN, TensorRT,\n\n CUDA libraries, CUDA drivers, CUDA runtime, CUDA applications or frameworks or any\n\n other NVIDIA software.\n\n5. Documents sufficient to show the architecture, design, data flow, control flow, and execution\n\n flow of any system in which You use NVIDIA GPUs to perform computations, including\n\n diagrams, technical specifications, design documents, Powerpoints, slide decks, internal and\n\n external presentations, Source Code, configuration files, build files, deployment files, runtime\n\n logs, and profiler traces.\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 17 of 27\n\n\n\n\n6. Documents sufficient to show whether computations You performed using NVIDIA GPUs\n\n involved artificial neural networks, neural-network computational layers or computations with\n\n outputs as inputs for other neurons or layers.\n\n7. Documents sufficient to show whether You use a pointer to data stored in memory (e.g.\n\n memory bank or partition), using as an input to a subsequent computational layer the pointer\n\n to output data from a GPU computation, using pointers in neural network computations,\n\n swapping an input pointer with the pointer to data output from a GPU computation, pointer\n\n swapping, pointer rotation, buffer swapping, ping-pong buffers, double or triple buffering,\n\n alternating input/output buffers, or any other technique in which output data from one\n\n computation, layer, iteration, time step, or cycle becomes input data for a later computation,\n\n layer, iteration, time step, or cycle.\n\n8. Documents sufficient to show whether You store input data, output data, intermediate results,\n\n tensors, activations, weights, parameters, internal variables, GPU programs, kernels, textures,\n\n shaders, or other GPU-computation-related data in separate, partitioned, logical, physical,\n\n first/second, input/output, texture, shader, shared, global, device, host, pinned, GPU RAM,\n\n GPU cache(s), or unified memory regions (shared by CPU and GPU) when performing\n\n computations using NVIDIA GPUs.\n\n9. Documents sufficient to show how input data is received, acquired, stored, transferred, copied,\n\n streamed, prefetched, staged, queued, or loaded from CPU memory, host memory, system\n\n memory, storage, sensors, cameras, or other input sources to NVIDIA GPU memory\u2014\n\n including GPU RAM (e.g. GPU HBM, GDDR) and/or GPU cache(s)\u2014before, during, or in\n\n parallel with computations You perform using NVIDIA GPUs.\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 18 of 27\n\n\n\n\n10. Documents sufficient to show how output data from a GPU computation(s), intermediate\n\n results of GPU computations, tensors, buffers, activations, variables, or other computation\n\n results are stored, transferred, copied, streamed, written back, returned, accumulated, reused,\n\n or made available including asynchronously from NVIDIA GPU memory to CPU memory,\n\n host memory, system memory, storage, display, network, or another memory location before,\n\n during, or in parallel with computations You perform using NVIDIA GPUs\u2014and also\n\n including in the opposite direction, copying data from CPU or host or other memory to a queue\n\n for GPU computation while other GPU computations are occurring.\n\n11. Documents sufficient to show how computations You perform using NVIDIA GPUs are\n\n scheduled, ordered, controlled, queued, synchronized, parallelized, launched, interrupted,\n\n resumed, or executed, including through kernels, CUDA streams, CUDA graphs, events,\n\n threads, controllers, schedulers, compilers, runtimes, inference engines, run lists, run engines,\n\n or custom software.\n\n12. Documents sufficient to show whether and how user inputs, user commands, configuration\n\n changes, parameter changes, model changes, computational-element changes, input changes,\n\n interruptions, or display/output changes affect computations You perform using NVIDIA\n\n GPUs and/or queue them for GPU computation.\n\fCase 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 19 of 27\n\n\n\n\n Attachment 2\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 20 of 27\n AO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action\n\n\n UNITED STATES DISTRICT COURT\n for the\n Western District of __________\n __________ Texas\n\n NeuralAl,\n Nerual AI,LLC\n LLC )\n Plaintiff )\n v. ) Civil Action No. 7:24-cv-00221-ADA-DTG\n NVIDIA Corporation )\n )\n Defendant )\n\n SUBPOENA TO TESTIFY AT A DEPOSITION IN A CIVIL ACTION\n\n To: Tesla, Inc. Google LLC\n c/o C\n c/o T Corporation\n Corporation System,\n Service 1999 Bryan\n Company St., Suite\n DBA CSC- 900, Inc,\n Lawyers 211 TX\n Dallas, 75201\n E. 7th Street, Suite 620, Austin, TX 78701\n (Name of person to whom this subpoena is directed)\n\n Testimony: YOU ARE COMMANDED to appear at the time, date, and place set forth below to testify at a\n deposition to be taken in this civil action. If you are an organization, you must designate one or more officers, directors,\n or managing agents, or designate other persons who consent to testify on your behalf about the following matters, or\n those set forth in an attachment:\n See Exhibit A.\n\n\n Place: Planet Depos - Downtown Austin c/o Lexitas Legal, Date and Time:\n 100 Congress Ave, Ste. 2000, Austin, TX 78701 07/21/2026 9:00 am\n\n\n The deposition will be recorded by this method: audio, video, and stenographic means\n\n 0 Production: You, or your representatives, must also bring with you to the deposition the following documents,\n electronically stored information, or objects, and must permit inspection, copying, testing, or sampling of the\n material:\n\n\n\n\n The following provisions of Fed. R. Civ. P. 45 are attached \u2013 Rule 45(c), relating to the place of compliance;\n Rule 45(d), relating to your protection as a person subject to a subpoena; and Rule 45(e) and (g), relating to your duty to\n respond to this subpoena and the potential consequences of not doing so.\n\n Date: 06/24/2026\n CLERK OF COURT\n OR\n /s/ Emily Portuguese\n Signature of Clerk or Deputy Clerk Attorney\u2019s signature\n\n The name, address, e-mail address, and telephone number of the attorney representing (name of party)\n Neural Al, LLC , who issues or requests this subpoena, are:\n Emily Portuguese, Susman Godfrey LLP, One Manhattan West, 50th Floor, New York, New York 10001\neportuguese\u00aesusmangodfi'ey.com, 212-729-2082\n Notice to the person who issues or requests this subpoena\n If this subpoena commands the production of documents, electronically stored information, or tangible things before\n trial, a notice and a copy of the subpoena must be served on each party in this case before it is served on the person to\n whom it is directed. Fed. R. Civ. P. 45(a)(4).\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 21 of 27\nAO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action (Page 2)\n\nCivil Action No. 7:24-cv-00221-ADA-DTG\n\n PROOF OF SERVICE\n (This section should not be filed with the court unless required by Fed. R. Civ. P. 45.)\n\n I received this subpoena for (name of individual and title, if any)\non (date) .\n\n 0 I served the subpoena by delivering a copy to the named individual as follows:\n\n\n on (date) ; or\n\n 0 I returned the subpoena unexecuted because:\n .\n\n Unless the subpoena was issued on behalf of the United States, or one of its officers or agents, I have also\n tendered to the witness the fees for one day\u2019s attendance, and the mileage allowed by law, in the amount of\n $ .\n\nMy fees are $ for travel and $ for services, for a total of $ 0.00 .\n\n\n I declare under penalty of perjury that this information is true.\n\n\nDate:\n Server\u2019s signature\n\n\n\n Printed name and title\n\n\n\n\n Server\u2019s address\n\nAdditional information regarding attempted service, etc.:\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 22 of 27\n\nAO 88A (Rev. 02/14) Subpoena to Testify at a Deposition in a Civil Action (Page 3)\n\n Federal Rule of Civil Procedure 45 (c), (d), (e), and (g) (Effective 12/1/13)\n(c) Place of Compliance. (i) disclosing a trade secret or other confidential research, development,\n or commercial information; or\n (1) For a Trial, Hearing, or Deposition. A subpoena may command a (ii) disclosing an unretained expert\u2019s opinion or information that does\nperson to attend a trial, hearing, or deposition only as follows: not describe specific occurrences in dispute and results from the expert\u2019s\n (A) within 100 miles of where the person resides, is employed, or study that was not requested by a party.\nregularly transacts business in person; or (C) Specifying Conditions as an Alternative. In the circumstances\n (B) within the state where the person resides, is employed, or regularly described in Rule 45(d)(3)(B), the court may, instead of quashing or\ntransacts business in person, if the person modifying a subpoena, order appearance or production under specified\n (i) is a party or a party\u2019s officer; or conditions if the serving party:\n (ii) is commanded to attend a trial and would not incur substantial (i) shows a substantial need for the testimony or material that cannot be\nexpense. otherwise met without undue hardship; and\n (ii) ensures that the subpoenaed person will be reasonably compensated.\n (2) For Other Discovery. A subpoena may command:\n (A) production of documents, electronically stored information, or (e) Duties in Responding to a Subpoena.\ntangible things at a place within 100 miles of where the person resides, is\nemployed, or regularly transacts business in person; and (1) Producing Documents or Electronically Stored Information. These\n (B) inspection of premises at the premises to be inspected. procedures apply to producing documents or electronically stored\n information:\n(d) Protecting a Person Subject to a Subpoena; Enforcement. (A) Documents. A person responding to a subpoena to produce documents\n must produce them as they are kept in the ordinary course of business or\n (1) Avoiding Undue Burden or Expense; Sanctions. A party or attorney must organize and label them to correspond to the categories in the demand.\nresponsible for issuing and serving a subpoena must take reasonable steps (B) Form for Producing Electronically Stored Information Not Specified.\nto avoid imposing undue burden or expense on a person subject to the If a subpoena does not specify a form for producing electronically stored\nsubpoena. The court for the district where compliance is required must information, the person responding must produce it in a form or forms in\nenforce this duty and impose an appropriate sanction\u2014which may include which it is ordinarily maintained or in a reasonably usable form or forms.\nlost earnings and reasonable attorney\u2019s fees\u2014on a party or attorney who (C) Electronically Stored Information Produced in Only One Form. The\nfails to comply. person responding need not produce the same electronically stored\n information in more than one form.\n (2) Command to Produce Materials or Permit Inspection. (D) Inaccessible Electronically Stored Information. The person\n (A) Appearance Not Required. A person commanded to produce responding need not provide discovery of electronically stored information\ndocuments, electronically stored information, or tangible things, or to from sources that the person identifies as not reasonably accessible because\npermit the inspection of premises, need not appear in person at the place of of undue burden or cost. On motion to compel discovery or for a protective\nproduction or inspection unless also commanded to appear for a deposition, order, the person responding must show that the information is not\nhearing, or trial. reasonably accessible because of undue burden or cost. If that showing is\n (B) Objections. A person commanded to produce documents or tangible made, the court may nonetheless order discovery from such sources if the\nthings or to permit inspection may serve on the party or attorney designated requesting party shows good cause, considering the limitations of Rule\nin the subpoena a written objection to inspecting, copying, testing, or 26(b)(2)(C). The court may specify conditions for the discovery.\nsampling any or all of the materials or to inspecting the premises\u2014or to\nproducing electronically stored information in the form or forms requested. (2) Claiming Privilege or Protection.\nThe objection must be served before the earlier of the time specified for (A) Information Withheld. A person withholding subpoenaed information\ncompliance or 14 days after the subpoena is served. If an objection is made, under a claim that it is privileged or subject to protection as trial-preparation\nthe following rules apply: material must:\n (i) At any time, on notice to the commanded person, the serving party (i) expressly make the claim; and\nmay move the court for the district where compliance is required for an (ii) describe the nature of the withheld documents, communications, or\norder compelling production or inspection. tangible things in a manner that, without revealing information itself\n (ii) These acts may be required only as directed in the order, and the privileged or protected, will enable the parties to assess the claim.\norder must protect a person who is neither a party nor a party\u2019s officer from (B) Information Produced. If information produced in response to a\nsignificant expense resulting from compliance. subpoena is subject to a claim of privilege or of protection as\n trial-preparation material, the person making the claim may notify any party\n (3) Quashing or Modifying a Subpoena. that received the information of the claim and the basis for it. After being\n notified, a party must promptly return, sequester, or destroy the specified\n (A) When Required. On timely motion, the court for the district where information and any copies it has; must not use or disclose the information\ncompliance is required must quash or modify a subpoena that: until the claim is resolved; must take reasonable steps to retrieve the\n information if the party disclosed it before being notified; and may promptly\n (i) fails to allow a reasonable time to comply; present the information under seal to the court for the district where\n (ii) requires a person to comply beyond the geographical limits compliance is required for a determination of the claim. The person who\nspecified in Rule 45(c); produced the information must preserve the information until the claim is\n (iii) requires disclosure of privileged or other protected matter, if no resolved.\nexception or waiver applies; or\n (iv) subjects a person to undue burden. (g) Contempt.\n (B) When Permitted. To protect a person subject to or affected by a The court for the district where compliance is required\u2014and also, after a\nsubpoena, the court for the district where compliance is required may, on motion is transferred, the issuing court\u2014may hold in contempt a person\nmotion, quash or modify the subpoena if it requires: who, having been served, fails without adequate excuse to obey the\n subpoena or an order related to it.\n\n\n For access to subpoena materials, see Fed. R. Civ. P. 45(a) Committee Note (2013).\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 23 of 27\n\n\n\n\n EXHIBIT A\n\n DEFINITIONS AND INSTRUCTIONS\n\n 24. The term \u201cNVIDIA GPUs\u201d means the Hopper, Ada Lovelace, Ampere, Turing,\n\nVolta, Pascal, Maxwell, Jetson, and Blackwell architectures of NVIDIA graphics processing units.\n\nFor avoidance of doubt, those architectures include the following devices: DGX line of\n\nsupercomputers and servers (including at least DGX B300, DGX B200, DGX GB200, DGX\n\nGB300, DGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200,\n\nDGX H200, DGX H100, DGX BasePOD, DGX SuperPOD with H200, DGX A100), HGX line of\n\nsupercomputers and servers (including at least HGX B300, HGX B200, HGX H100, HGX H200,\n\nEos SuperPOD), OVX line of supercomputers and servers (including at least OVX L40S), EGX\n\nline of supercomputers and servers (including at least EGX Server with Quadro RTX A6000, EGX\n\nServer with A40, EGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000),\n\nGB300 NVL72, GB200 NVL72; Nvidia\u2019s GPU accelerators and superchips, including those with\n\nNVIDIA\u2019s Blackwell, Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, and Maxwell GPU\n\narchitectures, including at least, RTX PRO 6000 Server Edition, RTX PRO 6000 Workstation,\n\nRTX PRO 6000 Max-Q Workstation, RTX PRO 5000, RTX PRO 4500, RTX PRO 4000, RTX\n\nPRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, GB300, GB200, H100, H200,\n\nGH200, GH100, L40, L40S, L4, RTX 6000, RTX 6000 Ada, RTX 5000, Ada, RTX 4500 Ada,\n\nRTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3000, RTX 2000, RTX 1000, RTX 500, RTX\n\n4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060 Ti,\n\nand RTX 4060, GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce RTX\n\n4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU, A100,\n\nA40, A30, A16, A10, A2, A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 24 of 27\n\n\n\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500, GeForce RTX\n\n3090 Ti, GeForce RTX 3090, GeForce RTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti,\n\nGeForce RTX 3070, GeForce RTX 3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB),\n\nGeForce RTX 3050 (6 GB), GeForce RTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop\n\nGPU, GeForce RTX 3070 Ti Laptop GPU, GeForce RTX 3070 Laptop GPU, GeForce RTX 3060\n\nLaptop GPU, GeForce RTX 3050 Ti Laptop GPU, GeForce RTX 3050 Laptop GPU, GeForce\n\nMX570 Laptop GPU, Tesla T4 GPUs, Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 8000,\n\nQuadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000,\n\nT1000 8GB, T1200, Quadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB, Titan\n\nRTX, GeForce RTX 2080 Ti, GeForce RTX 2080, Super, GeForce RTX 2080, GeForce RTX 2070\n\nSuper, GeForce RTX 2070, GeForce RTX 2060 Super, GeForce RTX 2060, GeForce RTX 2500,\n\nGeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630, GeForce MX550, GeForce MX450, GeForce MX430, Tesla V100,\n\nQuadro GV100, Titan V GPU, Tesla P100, P40, P4, Quadro GP100, Quadro P6000, Quadro\n\nP5200, Quadro P5000, Quadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro\n\nP2200, Quadro P2000, Quadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500,\n\nQuadro P400, Titan Xp, Titan X, GeForce GTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070\n\nTi, GeForce GTX 1070, GeForce GTX 1060, GeForce GTX 1050 Ti, GeForce GTX 1050,\n\nGeForce MX300, GeForce MX200, GeForce MX150, Tesla M60, M40, M10, Quadro M6000\n\n24GB, Quadro M6000 (12GB), Quadro M5000, Quadro M5000M, Quadro M5500, Quadro\n\nM4000, Quadro M4000M, Quadro M3000M, Quadro M2200, Quadro M2000, Quadro M2000M,\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 25 of 27\n\n\n\n\nQuadro M1200, Quadro M1000M, Quadro M620, Quadro M600M, Quadro M520, Quadro\n\nM500M, NVS 810, Tesla M6, GTX Titan X, GeForce GTX 980Ti, GeForce GTX 980, GeForce\n\nGTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M,\n\nGeForce GTX 960M, GeForce GTX 950M, GeForce GTX 750 Ti, GeForce GTX 750, GeForce\n\nMX130, and GeForce MX110; and Jetson modules, including at least the Jetson Thor Series,\n\nJetson Thor, Jetson T5000, Jetson T4000, Jetson AGX Orin Series, Jetson AGX Orin Developer\n\nKit, Jetson AGX Orin 64GB, Jetson AGX Orin Industrial, Jetson AGX Orin 32GB, Jetson Orin\n\nNX Series, Jetson Orin NX 16GB, Jetson Orin NX 8GB, Jetson Orin Nano Series, Jetson Orin\n\nNano Super Developer Kit, Jetson Orin Nano 8GB, Jetson Orin Nano 4GB, Jetson AGX Xavier\n\nSeries, Jetson AGX Xavier Industrial, Jetson AGX Xavier 64GB, Jetson AGX Xavier 32GB,\n\nJetson Xavier NX Series, Jetson Xavier NX 16GB, Jetson Xavier NX 8GB, Jetson TX2 Series,\n\nJetson TX2i, Jetson TX2, Jetson TX2 4GB, Jetson TX2 NX, Jetson Nano, any and all variations\n\nof the aforementioned products (including at least products having different options for number of\n\nGPUs).\n\n 25. The terms \u201cand\u201d and \u201cor\u201d are not intended to be read disjunctively but rather\n\nconjunctively unless the context of a particular request clearly indicates otherwise. \u201cOr\u201d should be\n\nunderstood to include and encompass \u201cand\u201d; and \u201cand\u201d should be understood to include and\n\nencompass \u201cor.\u201d\n\n 26. The terms \u201cany\u201d or \u201ceach\u201d should be understood to include and encompass \u201call.\u201d\n\n 27. The terms \u201cconcerning,\u201d \u201crelated to\u201d or \u201crelating to\u201d, and \u201cregarding\u201d and any\n\nvariation of these terms mean analyzing, alluding to, concerning, considering, commenting on,\n\nconsulting, comprising, containing, contradicting, describing, dealing with, discussing,\n\nestablishing, evidencing, identifying, involving, noting, recording, reporting on, relating to,\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 26 of 27\n\n\n\n\nreflecting, referring to, regarding, stating, showing, studying, mentioning, memorializing, or\n\npertaining to, directly or indirectly, in whole or in part.\n\n 28. The term \u201cCPU(s)\u201d means Central Processing Unit(s).\n\n 29. The term \u201cGPU(s)\u201d means Graphics Processing Unit(s).\n\n 30. The terms \u201cinclude\u201d and \u201cincluding\u201d mean including without limitation.\n\n 31. The term \u201cNVIDIA\u201d means Defendant NVIDIA Corporation, its predecessors,\n\npresent and former directors, officers, accountants, affiliates, attorneys, partners, managers, agents,\n\nemployees, representatives, in-house and outside counsel, and any other person or entity acting on\n\nbehalf of or under control of Defendant NVIDIA Corporation.\n\n 32. The term \u201cperson(s)\u201d means and includes natural persons and formal or informal\n\nentities and organizations, including public and private corporations, partnerships, professional\n\ncorporations, limited liability companies, business trusts, banking institutions, associations, firms,\n\njoint ventures, commissions, bureaus, departments, and any other legal entity, including\n\nany divisions, subsidiaries, departments, and other units thereof.\n\n 33. The term \u201cSource Code\u201d means human-readable instructions written in a\n\nprogramming language, including all comments, annotations, declarations, functions, classes, and\n\nother components used to define the behavior of a software program. For purposes of these topics,\n\n\u201cSource Code\u201d includes all associated files necessary to understand, compile, and execute the\n\ncode, such as scripts, header files, makefiles, configuration files, and documentation. Unless\n\notherwise stated, \u201cSource Code\u201d includes all versions and revisions relevant to the time periods\n\nand subject matter described in each interrogatory.\n\n 34. The terms \u201cYou,\u201d or \u201cYour\u201d refer to Tesla, Inc., including but not limited to its\n\npredecessors, successors, parents, subsidiaries, divisions, affiliates, and all past or present\n\f Case 7:26-cv-00318 Document 1-6 Filed 08/17/26 Page 27 of 27\n\n\n\n\ndirectors, officers, partners, managers, employees, contractors, agents, representatives,\n\naccountants, consultants, in-house and outside counsel, and any other person or entity acting or\n\npurporting to act on its behalf or subject to its control. This definition expressly includes, without\n\nlimitation, any Tesla parent, subsidiary, affiliate, or other related entity that has used, licensed,\n\ndeployed, evaluated, or integrated NVIDIA Hardware and/or NVIDIA Software.\n\n 35. The use of the singular form of any word includes the plural and vice versa.\n\n DEPOSITION TOPICS\n\n1. The NVIDIA software and libraries You use to perform computations, including but not\n\n limited to NVIDIA\u2019s Aerial, Clara Parabricks, cuBLAS, cuDNN, cuFFT, cuQuantum,\n\n cuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac, Isaac Lab, Maxine, Memory\n\n Map, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo, PyTorch, RAPIDS, Riva,\n\n Runtime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream).\n\n2. The NVIDIA sample Source Code You use, in whole or in part, to conduct computations.\n\n3. Your customizations and/or data inputs to NVIDIA software that alter the way in which\n\n NVIDIA software performs computations and/or a description of the data input to NVIDIA\n\n software on which computations are run.\n\n4. Identification of Your software that uses NVIDIA GPUs to perform computations.\n\n5. Using Your software, the ways in which output data from a GPU computation(s), including\n\n intermediate results of GPU computations are stored, referenced by a pointer, transferred,\n\n copied, streamed, written back, returned, accumulated, reused, or made available including\n\n asynchronously from NVIDIA GPU memory to CPU memory, host memory, system memory,\n\n storage, display, network, or another memory location before, during, or in parallel with\n\n computations performed using NVIDIA GPUs.\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:25.293463-07:00","document_number":"1","attachment_number":6,"pacer_doc_id":"181037209216","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 5","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294677/","id":490294677,"tags":[],"absolute_url":"/docket/74659430/1/7/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.117551-07:00","date_modified":"2026-08-23T05:23:21.334822-07:00","sha1":"12f2680ed6cd5449d794c8a268060e5cc510af6c","page_count":2,"file_size":67034,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.7.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.7.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-7 Filed 08/17/26 Page 1 of 2\n\n\n\n\n EXHIBIT\n\n 6\n\f Case 7:26-cv-00318 Document 1-7 Filed 08/17/26 Page 2 of 2\n\n\n\n\n AFFIDAVIT OF SERVICE\n UNITED STATES DISTRICT COURT\n Western District of Texas\nCase Number: 7:24-CV-00221-ADA-\nDTH\n\nPlaintiff:\nNeural Al, LLC\nVS.\n\nDefendant:\nNvidia Corporation\n\nFor:\nEmily Portuguese\n\nReceived by Anthony Collins on the 24th day of June, 2026 at 4:46 pm to be served on Tesla, Inc c/o CT\nCorporation System, 1999 Bryan St, Ste 900, Dallas, Dallas County, TX 75201.\n\nI, Anthony Collins, being duly sworn, depose and say that on the 25th day of June, 2026 at 2:55 pm, I:\n\nExecuted service by hand delivering a true copy of the Subpoena to: Al Johnson , an authorized\nacceptance agent employed by Registered Agent CT Corporation System, Inc., who is authorized to\naccept service of process for Tesla, Inc , at the address of: 1999 Bryan St, Ste 900, Dallas, Dallas\nCounty, TX 75201, and informed said person of the contents therein, in compliance with state statutes.\n\nDescription of Person Served: Age: 30s, Sex: M, Race/Skin Color: Black, Height: 5'11, Weight: 220, Hair:\nBald, Glasses: Y\n\n\"I certify that I am over the age of 18, have no interest in the above action. and authorized to serve\nprocess in the judicial circuit in which the process was served. I have personal knowledge of the facts set\nforth in this affidavit. I declare under the penalty of perjury that the fore ing i r e and correct.\n\n %\"`\"'\"'\u2022 ERIC JACOB HARRIS\n **\u2022:-g:allotary Public, State ofas\n :\u2022.\n COMM. Expires OS-12-2.7.2S\n \u00b0!;,* Notary ID 135527875\n Anthony Collins\n PSC-357 Expires 12/31/2027\nSubscribed and Sworn to before me on the 26th day\nof June, 2026 by the affiant who is personally known Lexitas\nto me. 1235 Broadway\n 2nd Floor\n New York, NY 10001\nNOTARY PUBLIC (718) 672-1117\n\n Our Job Serial Number: ONT-2026006077\n Ref: 27254654\n\n\n\n\n Copyright @ 1992-2026 DreamBuilt Software, LLC. - Process Server's Toolbox V9.Oe\n\n\n\n\n 1111111111111111111111111111111111\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:27.446052-07:00","document_number":"1","attachment_number":7,"pacer_doc_id":"181037209217","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 6","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294678/","id":490294678,"tags":[],"absolute_url":"/docket/74659430/1/8/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.135253-07:00","date_modified":"2026-08-21T19:38:57.835857-07:00","sha1":"81803991bc19b586ecd11e6debfa769ded8b1c46","page_count":18,"file_size":220151,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.8.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.8.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 1 of 18\n\n\n\n\n EXHIBIT\n\n 7\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 2 of 18\n\n\n\n\n UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n MIDLAND-ODESSA DIVISION\nNEURAL AI, LLC )\n )\n )\n Plaintiff, )\nv. ) Civil Action No. 7:24-cv-00221\n )\nNVIDIA CORPORATION )\n ) JURY TRIAL DEMANDED\n )\n Defendant. )\n\n\n PLAINTIFF\u2019S AMENDED DISCLOSURE OF ASSERTED CLAIMS AND FINAL\nINFRINGEMENT CONTENTIONS PURSUANT TO THE COURT\u2019S STANDING ORDER\n GOVERNING PATENT PROCEEDINGS\n\n Pursuant to the Court\u2019s December 2, 2025 Order (Dkt. 132), the First Amended Scheduling\n\n Order (Dkt. 135), and the Parties December 29, 2025 Joint Stipulation (Dkt. 136), Plaintiff Neural\n\n AI, LLC (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) provides the following Disclosure of Asserted Claims and\n\n Final Infringement Contentions (\u201cDisclosure\u201d) as to U.S. Patent Nos. 8,648,867 (\u201cthe \u2019867\n\n Patent\u201d), RE49,461 (\u201cthe \u2019461 Patent\u201d), and RE48,438 (\u201cthe \u2019438 Patent\u201d) (collectively, the\n\n \u201cAsserted Patents\u201d or \u201cPatents-in-Suit\u201d) against Defendant Nvidia Corporation (\u201cNvidia\u201d or\n\n \u201cDefendant\u201d). This Disclosure is made solely for the purpose of this action.\n\n Pursuant to the Court\u2019s December 2, 2025 Order (Dkt. 132), Plaintiff identifies source code\n\n for each accused instrumentality based on Nvidia\u2019s productions prior to November 30, 2025.\n\n Nvidia continues to produce source code on a rolling basis. Plaintiff therefore expressly reserves\n\n the right to serve additional or supplemental claim chart exhibits identifying newly produced or\n\n previously unavailable source code that satisfies the asserted claim elements, including after\n\n service of this Disclosure. Moreover, Plaintiff\u2019s investigation regarding infringement and\n\n additional potential grounds of infringement is ongoing. Nvidia\u2019s has not produced all necessary\n\n 1\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 3 of 18\n\n\n\ntechnical documents or source code sufficient to show the operation of the Accused Products, and\n\nkey categories of technical materials, including source code, remain outstanding. See, e.g., Dkt.\n\n103, 125. This Disclosure is therefore based upon information that Plaintiff has been able to obtain\n\nand review to date, together with its good-faith beliefs regarding the Accused Products and their\n\noperation, and is made without prejudice to Plaintiff\u2019s right to supplement or amend its Disclosure\n\nas additional facts are ascertained, discovery is conducted, code is reviewed, analysis is done, and\n\nresearch is completed. Plaintiff reserves the right to amend and/or supplement its infringement\n\ncontentions as additional information becomes available.\n\n For each Asserted Patent, Plaintiff identifies the following Accused Products of which it is\n\ncurrently aware. The identification of Accused Products is based on Plaintiff\u2019s research and\n\nanalysis to date, without the benefit of full discovery. Indeed, Nvidia has not yet complied with\n\nits obligation under the OGP to produce sufficient \u201ctechnical documents, including software where\n\napplicable, sufficient to show the operation of the accused product(s).\u201d See, e.g., Dkt. 103, 125.\n\nNvidia has continued to belatedly produce source code well into the fact discovery period\u2014\n\nincluding most recently on January 13, 2026\u2014and has not produced related technical documents,\n\nand Plaintiff has been required to repeatedly press Nvidia to obtain these late and piecemeal\n\nproductions. As a result, Plaintiff\u2019s current identification of Accused Products is necessarily based\n\non publicly available and otherwise accessible information. Accordingly, Plaintiff\u2019s current\n\nidentification is based on publicly available and otherwise accessible information. Plaintiff\n\nexpressly reserves the right to amend or supplement these contentions\u2014including by identifying\n\nadditional Accused Products or asserting additional bases for infringement\u2014under the applicable\n\nrules and any Court orders, including to reflect future productions by Nvidia.\n\n Accused Products. The Accused Products, as described in the accompanying Exhibits 1\u2013\n\n599 and any later-served amended or supplemental Exhibits, comprise integrated combinations of\n\n 2\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 4 of 18\n\n\n\nNvidia\u2019s software and hardware that together implement GPU-accelerated computing. These\n\ninclude Nvidia\u2019s GPU accelerators and superchips; Nvidia\u2019s computers, supercomputers, data\n\ncenters, servers, and workstations incorporating those GPUs; and the full Nvidia software stack\n\nthat operates on and controls that hardware to enable accelerated execution.\n\n Based on its present understanding of Nvidia\u2019s infringing software architecture and how\n\nthe accused functionality is implemented across Nvidia\u2019s integrated hardware and software stack,\n\nNeural AI has organized its infringement charts by Nvidia application in Exhibits 100-199. These\n\napplication charts further reference sub-charts identifying hardware described in Exhibits 1-99 as\n\nwell as software and libraries described in Exhibits 200-599. The sub charts are broken out as\n\nfollows: Exhibits 1\u201399 include hardware such as Nvidia\u2019s GPU accelerators and superchips, and\n\nNvidia\u2019s computers, supercomputers, data centers, servers, workstations that implement its GPU\n\naccelerators and superchips, coupling of some of this hardware to CPUs, as well as information on\n\nNvidia\u2019s infringing software products and CUDA code. Exhibits 100-199 include Nvidia\u2019s\n\napplication frameworks and platforms. Application frameworks and platforms processed by the\n\nhardware, in turn, call on lower-level Nvidia software. Exhibits 200\u2013299 include Nvidia\u2019s\n\nmachine-learning frameworks and inference platforms used by these applications, including\n\nPyTorch and TensorRT. Exhibits 300\u2013399 include Nvidia\u2019s neural-network primitive libraries,\n\nacceleration libraries, and functions, including cuDNN and related components. Exhibits 400\u2013499\n\ninclude Nvidia\u2019s mathematical libraries. Exhibits 500\u2013599 include low-level CUDA runtime,\n\ndriver, memory-management components, memory maps, and sub-component code used in the\n\nCUDA stack.\n\n This organization reflects Neural AI\u2019s present understanding of Nvidia\u2019s software\n\narchitecture. Nvidia has not produced a complete source-code production for any bucket or\n\ncategory, and Neural AI therefore expressly reserves the right to chart additional software,\n\n 3\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 5 of 18\n\n\n\nlibraries, or components, and to modify or supplement its chart organization and infringement\n\ntheories as discovery continues. Moreover, the charts provided within each category are\n\nrepresentative of other similar Nvidia software, libraries, and components within that same\n\ncategory. Neural AI has charted exemplar implementations rather than every produced file, and\n\nreserves the right to rely on other Nvidia software and libraries within the same category as\n\nadditional accused instrumentalities.\n\n In further details, the Accused Products include, without limitation, the following: Nvidia\u2019s\n\nGPU accelerators and superchips, including those with Nvidia\u2019s \u201cBlackwell,\u201d \u201cHopper,\u201d \u201cAda\n\nLovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d \u201cPascal,\u201d and \u201cMaxwell\u201d GPU architectures. These\n\nGPUs and superchips implement, and are specifically designed for, GPU-acceleration for artificial\n\nintelligence and neural networks.\n\n Nvidia\u2019s Blackwell GPUs include RTX PRO 6000 Server Edition, RTX PRO 6000\n\nWorkstation, RTX PRO 6000 Max-Q Workstation, RTX PRO 6000, RTX PRO 5000, RTX PRO\n\n4500, RTX PRO 4000, RTX PRO 3000, RTX PRO 2000, RTX PRO 1000, RTX PRO 500, RTX\n\n5090, RTX 5090 D, RTX 5080, RTX 5070 Ti, RTX 5070, RTX 5060, RTX 5060 Ti, RTX 5050,\n\nRTX 5080 Laptop, RTX 5090 Laptop, RTX 5070 Ti Laptop, RTX 5060 Laptop, RTX 5070\n\nLaptop, and RTX 5050 Laptop. In addition, Nvidia\u2019s superchips that implement GPU accelerators\n\ninclude the GB300 and GB200.\n\n Nvidia\u2019s Hopper GPUs include the H100 and H200 GPUs, including by not limited to\n\nPCle, SXM, and NVL models. In addition, Nvidia\u2019s superchips that implement GPU accelerators\n\ninclude the GH200, or Grace Hopper Superchip, which implements the Hopper-GPU architecture.\n\n Nvidia\u2019s Ada Lovelace (or Lovelace) GPUs include Nvidia Data Center GPUs, including\n\nL40, L40S, and L4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including RTX\n\nAda Generations series GPUs and Laptop GPUs (including RTX 6000, RTX 6000 Ada, RTX 5000\n\n 4\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 6 of 18\n\n\n\nAda, RTX 4500 Ada, RTX 4050, RTX 4000 Ada, RTX 4000 SFF, RTX 3500, RTX 3050, RTX\n\n3000, RTX 2000, RTX 1000, RTX 500); and GeForce RTX 40 series GPUs and Laptop GPUs\n\n(RTX 4090, RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4070 SUPER, RTX 4070, RTX 4060\n\nTi, and RTX 4060; GeForce RTX 4090 Laptop GPU, GeForce RTX 4080 Laptop GPU, GeForce\n\nRTX 4070 Laptop GPU, GeForce RTX 4060 Laptop GPU, GeForce RTX 4050 Laptop GPU).\n\n Nvidia\u2019s Ampere GPUs include Nvidia Data Center GPUs, including A100, A40, A30,\n\nA16, A10, and A2 GPUs; Nvidia Workstation and Professional Laptop GPUs, including RTX A\n\nseries GPUs and Laptop GPUs (A800 40GB Active, RTX A6000, RTX A5500, RTX A5000, RTX\n\nA4500, RTX A4000, RTX A2000, RTX A2000 12GB, RTX A1000, RTX A400, RTX A5500,\n\nRTX A4500, RTX A3000 12GB, RTX A2000 8GB, RTX A1000 6GB, RTX A500); GeForce\n\nRTX 30 series GPUs and Laptop GPUs (GeForce RTX 3090 Ti, GeForce RTX 3090, GeForce\n\nRTX 3080 Ti, GeForce RTX 3080, GeForce RTX 3070 Ti, GeForce RTX 3070, GeForce RTX\n\n3060 Ti, GeForce RTX 3060, GeForce RTX 3050 (8 GB), GeForce RTX 3050 (6 GB), GeForce\n\nRTX 3080 Ti Laptop GPU, GeForce RTX 3080 Laptop GPU, GeForce RTX 3070 Ti Laptop GPU,\n\nGeForce RTX 3070 Laptop GPU, GeForce RTX 3060 Laptop GPU, GeForce RTX 3050 Ti Laptop\n\nGPU, GeForce RTX 3050 Laptop GPU); and GeForce MX570 Laptop GPU.\n\n Nvidia\u2019s Turing GPUs include Nvidia Data Center GPUs, including Tesla T4 GPUs and\n\nQuadro RTX 8000 (passive) and Quadro RTX 6000 (passive) GPUs; Nvidia Workstation and\n\nProfessional Laptop GPUs, including T series GPUs and Laptop GPUs, Quadro T series Laptop\n\nGPUs, and Quadro RTX series GPUs and Laptop GPUs (Quadro RTX 8000, Quadro RTX 6000,\n\nQuadro RTX 5000, Quadro RTX 4000, Quadro RTX 3000, Quadro T2000, T1000 8GB, T1200,\n\nQuadrio T1000, T1000 (4GB), T600, T550, T500 T400, T400 4GB); Titan series Titan RTX GPU;\n\nGeForce RTX 20 series GPUs and Laptop GPUs (GeForce RTX 2080 Ti, GeForce RTX 2080\n\nSuper, GeForce RTX 2080, GeForce RTX 2070 Super, GeForce RTX 2070, GeForce RTX 2060\n\n 5\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 7 of 18\n\n\n\nSuper, GeForce RTX 2060, GeForce RTX 2500); GeForce GTX 16 series GPUs and Laptop GPUs\n\n(GeForce GTX 1660 Ti, GeForce GTX 1660 Super, GeForce GTX 1660, GeForce GTX 1650 Ti,\n\nGeForce GTX 1650 Super, GeForce GTX 1650 (G5), GeForce GTX 1650 (G6), GeForce GTX\n\n1650, GeForce GTX 1630); and GeForce MX550, MX450, and MX430 Laptop GPUs.\n\n Nvidia\u2019s Volta GPUs include Nvidia Data Center GPUs, including the Tesla V100 GPU;\n\nNvidia Workstation GPUs, including Quadro GV100; and Titan series Titan V GPU.\n\n Nvidia\u2019s Pascal GPUs include Nvidia Data Center GPUs, including Tesla P100, P40, and\n\nP4 GPUs; Nvidia Workstation and Professional Laptop GPUs, including the Quadro GP100 GPU\n\nand Quadro P series GPUs and Laptop GPUs (Quadro P6000, Quadro P5200, Quadro P5000,\n\nQuadro P4200, Quadro P4000, Quadro P3200, Quadro P3000, Quadro P2200, Quadro P2000,\n\nQuadro P1000, Quadro P620, Quadro P600, Quadro P520, Quadro P500, Quadro P400); Titan\n\nseries Titan Xp and Titan X GPUs; GeForce GTX 10 series GPUs and Laptop GPUs (GeForce\n\nGTX 1080 Ti, GeForce GTX 1080, GeForce GTX 1070 Ti, GeForce GTX 1070, GeForce GTX\n\n1060, GeForce GTX 1050 Ti, GeForce GTX 1050); and GeForce MX300 series, MX200 series,\n\nand MX150 Laptop GPUs.\n\n Nvidia\u2019s Maxwell GPUs include Nvidia Data Center GPUs, including Tesla M60, M40,\n\nand M10 GPUs; Nvidia Workstation and Professional Laptop GPUs, including Quadro M series\n\nGPUs and Laptop GPUs (Quadro M6000 24GB, Quadro M6000 (12GB), Quadro M5000, Quadro\n\nM5000M, Quadro M5500, Quadro M4000, Quadro M4000M, Quadro M3000M, Quadro M2200,\n\nQuadro M2000, Quadro M2000M, Quadro M1200, Quadro M1000M, Quadro M620, Quadro\n\nM600M, Quadro M520, Quadro M500M), the NVS 810 GPU, and Tesla M6 series Laptop GPUs;\n\nTitan series GTX Titan X GPU; GeForce GTX 900 series GPUs and Laptop GPUs (GeForce GTX\n\n980Ti, GeForce GTX 980, GeForce GTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce\n\nGTX 970M, GeForce GTX 965M, GeForce GTX 960M, GeForce GTX 950M); GeForce GTX\n\n 6\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 8 of 18\n\n\n\n700 series GPUs and Laptop GPUs (GeForce GTX 750 Ti, GeForce GTX 750); and GeForce\n\nMX130 series and MX110 Laptop GPUs.\n\n The Accused Products further include Nvidia\u2019s computers, supercomputers, data centers,\n\nservers, and workstations that implement its GPU accelerators and superchips. These computer\n\nhardware systems include: the DGX line of supercomputers, the HGX line of supercomputers, the\n\nOVX line of supercomputers, and the EGX line of servers for data centers and edge devices.\n\nNvidia\u2019s DGX supercomputers include the DGX B300, DGX B200, DGX GB200, DGX GB300,\n\nDGX Spark, DGX Station, DGX SuperPOD with GB300, DGX SuperPOD with GB200, DGX\n\nH200, DGX BasePOD, DGX A100, and DGX SuperPOD with DGX GB200. Nvidia\u2019s HGX\n\nincludes least the HGX B300, HGX B200, HGX H100, HGX H200, and EoS SuperPOD. And\n\nNvidia\u2019s EGX includes at least EGX Server with Quadro RTX A6000, EGX Server with A40,\n\nEGX Server with Quadro RTX 8000, EGX Server with Quadro RTX 6000), GB300 NVL72,\n\nGB200 NVL72.\n\n The Accused Products include Nvidia\u2019s software, platforms, libraries, and services for\n\naccelerated computing. These products include, without limitation, application frameworks,\n\nplatforms, and domains such as NVIDIA Drive, Isaac, Holoscan, RAPIDS, NVBlox, NeMo,\n\nMerlin, Modulus, MONAI, Morpheus, Riva, Maxine, Clara, Metropolis, Tokkio, Avatar, NIM\n\nMicroservices, Omniverse, Clara Train, TAO, DRIVE Sim, DLSS, PhysX, OptiX, Texture Tools,\n\nJetPack, DeepStream, DOCA, Magnum IO, Aerial, BioNeMo, CUDA-X HPC, Unified Compute\n\nFramework, AI Enterprise, the DGX Platform, NGC, and AI Foundation Models. Application\n\nframeworks and platforms processed by the hardware, in turn, call on lower-level Nvidia software.\n\nThe Accused Products further include machine-learning frameworks and inference platforms such\n\nas PyTorch, TensorRT, Triton, TensorFlow, Torch-TensorRT, JAX, and Spark. They also include\n\nneural-network, mathematical, and acceleration libraries such as cuDNN, cuFFT, cuDSS,\n\n 7\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 9 of 18\n\n\n\ncuSOLVER, cuRAND, CUTLASS, DALI, cuTensor, cuGraph, cuSPARSELt, NPP, NeuralVDB,\n\ncuNumeric, cuCIM, Sionna, cuBLAS, cuSPARSE, NCCL, Thrust, CUB, AmgX, and nvmath-\n\npython. The Accused Products further include low-level CUDA runtime, driver, and system\n\ncomponents, including CUDA, the CUDA Toolkit, CUDA Runtime and Driver components,\n\nCUDA Python, CUDA Quantum, Base Command, GPNVAPI, NVSHMEM, DCGM, Displaced\n\nMicro-Mesh, FLARE, GVDB Voxels, KickstartRT, Mesh Shading, Optical Flow, PTX, SASS,\n\nkernel implementations, firmware, scheduling logic, backend libraries, the CUDA Driver Internal\n\nLayer (CUI), CUDA Driver API, vGPU, GPUDirect, NVML, and other associated runtime, driver,\n\nand sub-component code. The Accused Products further include other Nvidia software, platforms,\n\nand services that implement similar accelerated computing functionality or operate using the same\n\nCUDA-based execution models, architectures, libraries, and runtime components, whether or not\n\nexpressly listed above. The Accused Products also encompass associated and underlying software\n\nlibraries and components that enable or support accelerated execution and any source code or sub-\n\ncomponent code necessary to understand or effect CUDA execution, whether or not separately\n\nanalyzed.\n\n The Accused Products further include (1) any additional products identified in the\n\naccompanying Exhibits 1\u2013599 attached hereto and any later-served amended or supplemental\n\nExhibits; (2) any products that include the same functionality or features described in the Exhibits;\n\nand (3) any prior or subsequent versions of the products identified in the Exhibits that include the\n\nsame features or functionality.\n\n The Accused Products infringe each of the Asserted Patents in a manner fully consistent\n\nwith the Court\u2019s Claim Construction Order (Dkt. 95). For example, as construed, the Accused\n\nProducts implement the claimed \u201caccelerator\u201d as hardware, software, or a combination thereof that\n\nneed not be physically separate from the CPU. The Accused Products likewise perform the claimed\n\n 8\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 10 of 18\n\n\n\nmethod steps in a pipelined and overlapping manner that satisfies the Court\u2019s ordering\n\nrequirements for the asserted claims of the \u2019867 and \u2019438 patents, including parallel execution\n\nwhere permitted and the specific sequencing constraints identified by the Court.\n\n On information and belief, the asserted claim elements charted for one chip architecture\n\nare evidenced by documentation and source code pertaining to the Accused Products for other chip\n\narchitectures, for which the Accused Products have capabilities and functionalities that are\n\nsubstantially the same for the asserted claim elements. Indeed, this is reflected by the different chip\n\narchitectures of the Accused Products sharing materially similar technical specifications and\n\noverlapping documentation as each Accused Product pertains to the accused claim elements.\n\nLikewise, the software implementations analyzed and charted are representative of other related\n\nNvidia software products that rely on the same CUDA-based execution models, libraries, runtime\n\ncomponents, and architectural design choices, and that therefore implement the asserted claim\n\nelements in substantially the same manner such as, for example, performing a math operation or\n\ncomputation in a neural network.\n\n The present infringement contentions also accuse Nvidia\u2019s newly announced Vera CPU\n\nand Rubin GPU products. Neural AI will supplement its infringement contentions as Nvidia\n\nproduces additional information and as these products become commercially available. The\n\npresent infringement contentions further accuse Nvidia\u2019s Jetson hardware products, which, based\n\non information obtained during ongoing fact and expert discovery, infringe one or more Asserted\n\nClaims. Neural AI will supplement its infringement contentions as Nvidia produces additional\n\ninformation regarding these products.\n\n These Final Infringement Contentions are based on public evidence and the limited\n\ndiscovery and source code made available by Nvidia to date, which largely consists of a single\n\nsoftware version for each accused application or library. Nvidia\u2019s has not produced all necessary\n\n 9\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 11 of 18\n\n\n\ntechnical documents or source code sufficient to show the operation of the Accused Products, and\n\nkey categories of technical materials and code remain outstanding. See, e.g., Dkt. 103, 125.\n\nDiscovery and expert analysis are ongoing and include, among other things, the multiple software\n\nversions and builds that may be used in combination, as well as additional libraries and software\n\nfunctions that infringe the asserted claims under the same theories reflected in the claim charts.\n\nThese include, for example, acceleration libraries and functions that infringe the claims in the same\n\nmanner as the cuDNN CTCLoss and RNNForward functions in the two cuDNN source-code\n\nversions produced by Nvidia. Accordingly, Plaintiff expressly reserves the right to amend,\n\nsupplement, or refine these contentions as additional facts are ascertained, discovery is conducted,\n\nanalysis is performed, and research is completed. Plaintiff further reserves the right to amend or\n\nsupplement its Disclosure under the applicable rules and any Court orders, including to reflect\n\nfuture productions by Nvidia.\n\n Defendant\u2019s Infringement. Based upon currently available information, Plaintiff\n\nidentifies the following asserted claims:\n\n \u2022 The \u2019867 Patent. Defendant has infringed and is infringing claims 16-19, literally and/or\n\n under the doctrine of equivalents. Defendant has infringed and is infringing these\n\n claims both directly and indirectly (by inducing infringement pursuant to 35 U.S.C. \u00a7\n\n 271(b) and/or by contributing to infringement pursuant to 35 U.S.C.\u00a7 271(c)).\n\n \u2022 The \u2019461 Patent. Defendant has infringed and is infringing claims 21-25, 27-30,\n\n literally and/or under the doctrine of equivalents. Defendant has infringed and is\n\n infringing these claims both directly and indirectly (by inducing infringement\n\n pursuant to 35 U.S.C. \u00a7 271(b) and/or by contributing to infringement pursuant to 35\n\n U.S.C.\u00a7 271(c)).\n\n \u2022 The \u2019438 Patent. Defendant has infringed and is infringing claims 1, 3-6, 8-9, 12, 14,\n 10\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 12 of 18\n\n\n\n 17-18, 21-23, 29-30, 32, 40, 43-44, 46, 48, 51-52, 55-57, literally and/or under the\n\n doctrine of equivalents. Defendant has infringed and is infringing these claims both\n\n directly and indirectly (by inducing infringement pursuant to 35 U.S.C. \u00a7 271(b) and/or\n\n by contributing to infringement pursuant to 35 U.S.C. \u00a7 271(c)).\n\n Based upon currently available information, Plaintiff asserts that Defendant has infringed\n\nand/or continues to infringe the patents and claims as identified and described in the infringement\n\ncharts for the Accused Products attached as the accompanying Exhibits 1\u2013599 and any later-served\n\namended or supplemental Exhibits. These exhibits contain illustrative examples of Defendant\u2019s\n\npresently known infringement of the Asserted Claims by evidencing the correspondence between\n\n(i) elements of the Asserted Claims and (ii) corresponding structures and/or functions of the\n\nAccused Products. Such examples are illustrative and not exhaustive, additional materials may\n\nevidence infringement, and additional bases of infringement may be present and uncovered during\n\ndiscovery. Plaintiff reserves the right to amend or supplement its Disclosure, including the\n\nattached claim charts, upon Nvidia\u2019s compliance with its discovery obligations.\n\n Each element of each asserted claim is presently alleged to be literally present. However,\n\nto the extent Defendant argues that a limitation is not literally present in the Accused Products, then\n\nDefendant still infringes under the doctrine of equivalents. Any differences alleged to exist\n\nbetween any of the Asserted Claims and any of the Accused Products are insubstantial, and\n\ntherefore each Accused Product also meets each limitation under the doctrine of equivalents, as the\n\nidentified features of the Accused Product perform substantially the same function in substantially\n\nthe same way to achieve substantially the same result as the corresponding claim limitations.\n\nPlaintiff reserves the right to supplement this Disclosure as discovery is conducted, Defendant\n\nprovides any alleged non-infringement positions, and claim construction is completed.\n\n Defendant directly infringes each of the asserted claims under 35 U.S.C. \u00a7271(a) at least\n\n 11\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 13 of 18\n\n\n\nby using, operating, testing, advertising, making, installing, maintaining, distributing, supporting,\n\nproviding instructions for, offering to sell, selling, and/or otherwise providing services including\n\nthe Accused Products\u2014or systems incorporating the Accused Products\u2014within the United States\n\nand/or importing the Accused Products into the United States. Defendant also directly infringes\n\neach of the claims at least by performing, or being responsible for the performance of (e.g., the\n\nacts are attributable to it), each of the claimed steps as set forth in the accompanying charts.\n\nDefendant\u2019s acts of direct infringement are further set forth in the accompanying Exhibits and any\n\nlater-served amended or supplemental Exhibits.\n\n Defendant also indirectly infringes the Asserted Claims by inducing infringement pursuant\n\nto 35 U.S.C. \u00a7 271(b)). Defendant has had knowledge of each of the asserted patents and of the\n\nspecific manner by which the Accused Products infringe each patent since at least September 2024,\n\nwhen Plaintiff filed and served its original complaint. Defendant knowingly induced one or more\n\nthird parties (e.g., business partners, customers, or others), to infringe the Asserted Claims by\n\nmaking the Accused Products available on Defendant\u2019s website, widely advertising the Accused\n\nProducts, providing applications that allow partners and users to access the Accused Products,\n\nproviding instructions for installing the Accused Products, and providing technical support to users\n\nand/or engaging in activities that aid and abet infringement of the Asserted Patents by end users\n\nwithin the United States, with knowledge and intent that performance of such actions would\n\ninfringe the Asserted Claims. Defendant committed these acts with knowledge or willful blindness\n\nthat such induced acts would constitute infringement of the Asserted Claims at least as of the filing\n\nof the original complaint. Defendant also has had actual or constructive notice of the technology\n\nclaimed in the Asserted Patents since at least 2007, when the inventors of the Asserted Patents first\n\ndiscussed their patented technologies with Mr. Sanford Russell, then the CTO of Nvidia. In\n\naddition, the inventors of the Asserted Patents held multiple discussions with Nvidia regarding a\n\n 12\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 14 of 18\n\n\n\npotential investment in or acquisition of their company, Neurala, Inc., and its assets, including the\n\npatent family that includes the Asserted Patents. Defendant knew or should have known that it\n\ninfringed the Asserted Patents based on its knowledge of the same. Alternatively or additionally,\n\nDefendant was willfully blind to the fact that it infringed the Asserted Patents despite its\n\nknowledge of the same, based on, for example, Defendant\u2019s having cited the application for the\n\n\u2019867 Patent on the face of its own patent since at least June 28, 2010, see Nvidia U.S. Patent No.\n\n8,922,566, and the similarity of the Accused Products to Plaintiff\u2019s patented technology.\n\nAdditional evidence of Defendant\u2019s inducement of infringement by others is set forth in each of\n\naccompanying Exhibits.\n\n Defendant also indirectly infringes the Asserted Claims by contributing to infringement\n\npursuant to 35 U.S.C. \u00a7 271(c). Each of the Accused Products is a material part of the claims, and\n\nDefendant knew each of the Accused Products is especially made or especially adapted for use in\n\nan infringement of the Asserted Claims. Further, the Accused Products have no substantial non-\n\ninfringing uses, as set forth in the example claim charts.\n\n Defendant also contributes to infringement by its customers and end users of the Accused\n\nProducts by offering to sell or selling within the United States or importing into the United States\n\nthe Accused Products, which are for use in practicing, and under normal operation practice,\n\nmethods claimed in the Asserted Patents, constituting a material part of the inventions claimed,\n\nand not a staple article or commodity of commerce suitable for substantial non-infringing use.\n\nIndeed, the Accused Products and the exemplary functionality identified in the accompanying\n\nExhibits have no substantial non-infringing uses but instead are specifically designed to practice\n\nthe Asserted Patents. Additional evidence of Defendant\u2019s contributory infringement is set forth in\n\neach of the Exhibits and any later-served amended or supplemental Exhibits.\n\n Priority Dates. Plaintiff presently identifies the following priority dates for the Asserted\n\n 13\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 15 of 18\n\n\n\nPatents:\n\n \u2022 All Asserted Claims of the \u2019867 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019867 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed invention, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n See NAI_0007731-NAI_0007923.\n\n \u2022 All Asserted Claims of the \u2019461 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019461 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed inventions, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n See NAI_0007731-NAI_0007923.\n\n \u2022 All Asserted Claims of the \u2019438 Patent are entitled to a priority date corresponding\n\n to the conception of the claimed inventions, which occurred no later than February\n\n 25, 2005. The conception of the inventions claimed in the \u2019438 Patent was followed\n\n by continuous diligence and actual reduction to practice of the claimed inventions\n\n no later than December 14, 2005. Following the actual reduction to practice of the\n\n claimed inventions, there was constructive reduction to practice corresponding to\n\n the filing of U.S. Provisional Application No. 60/826,892 on September 25, 2006.\n\n 14\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 16 of 18\n\n\n\n See NAI_0007731-NAI_0007923.\n\n Plaintiff\u2019s investigation and analysis is ongoing, and Plaintiff reserves the right to assert and\n\nrely on an earlier invention date in the event Defendant identifies alleged prior art that dated earlier\n\nthan the identified priority date corresponding to a date of conception followed by diligence and\n\nreduction to practice of the claimed inventions.\n\n\n\n DATED: January 20, 2026 Respectfully submitted,\n\n /s/ Tanner Laiche\n Max L. Tribble\n Texas State Bar 20213950\n Brian D. Melton\n Texas State Bar 24010620\n Rocco Magni\n Texas State Bar 24092745\n Samuel Drezdzon\n Texas State Bar 24117374\n SUSMAN GODFREY L.L.P.\n 1000 Louisiana\n Suite 5100\n Houston, TX 77002\n Telephone: (713) 651-9366\n Facsimile: (713) 654-6666\n mtribble@susmangodfrey.com\n bmelton@susmangodfrey.com\n rmagni@susmangodfrey.com\n sdrezdzon@susmangodfrey.com\n\n Tamar Lusztig\n NY State Bar 5125174\n Emily Portuguese\n NY State Bar 5920327\n One Manhattan West, 50th Floor\n New York, NY 10001\n tlusztig@susmangodfrey.com\n eportuguese@susmangodfrey.com\n\n Tanner Laiche\n WA State Bar 60450\n 401 Union Street, Suite 3000\n Seattle, WA 98101\n tlaiche@susmangodfrey.com\n 15\n\fCase 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 17 of 18\n\n\n\n\n Mark D. Siegmund\n Texas State Bar No. 24117055\n CHERRY JOHSON SIEGMUND\n JAMES PC\n Bridgeview Center\n 7901 Fish Pond Road, 2nd Floor\n Waco, Texas 76710\n msiegmund@cjsjlaw.com\n\n Max Ciccarelli\n Texas State Bar No. 00787242\n CICCARELLI LAW FIRM LLC\n 100 N. 6th Street, Suite 502\n Waco, Texas 76701\n Max@CiccarelliLawFirm.com\n\n Attorneys for Plaintiff Neural AI, LLC\n\n\n\n\n 16\n\f Case 7:26-cv-00318 Document 1-8 Filed 08/17/26 Page 18 of 18\n\n\n\n\n CERTIFICATE OF SERVICE\n\n The undersigned hereby certifies that a true and correct copy of the foregoing document has\n\nbeen served on January 20, 2026 to all counsel of record via electronic mail.\n\n\n\n /s/ Tanner Laiche\n Tanner Laiche\n\n\n\n\n 17\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:31.327334-07:00","document_number":"1","attachment_number":8,"pacer_doc_id":"181037209218","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 7","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294679/","id":490294679,"tags":[],"absolute_url":"/docket/74659430/1/9/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.151258-07:00","date_modified":"2026-08-21T18:44:25.655509-07:00","sha1":"df2986bf95401e2adbd3bdb84c41ee30a46aaabf","page_count":47,"file_size":458680,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.9.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.9.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-9 Filed 08/17/26 Page 1 of 47\n\n\n\n\n EXHIBIT\n\n 8\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 2 of 47\n 1 of 46\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n PUBLIC VERSION\nNEURAL AI, LLC\n Civil Action No. 7:24-cv-00221\n Plaintiff,\n\n v. JURY TRIAL DEMANDED\n\nNVIDIA CORPORATION.,\n\n Defendant.\n\n\n DEFENDANT\u2019S AMENDED ANSWER TO PLAINTIFF\u2019S AMENDED COMPLAINT\n\n Defendant NVIDIA Corporation (\u201cNVIDIA\u201d) hereby provides its amended answer to\n\nPlaintiff Neural AI, LLC\u2019s (\u201cPlaintiff\u201d) Amended Complaint for Patent Infringement (Dkt. 30)\n\n(\u201cComplaint\u201d). The headings and subheadings in Defendant\u2019s Answer are used solely for purposes\n\nof convenience and organization to mirror those appearing in the Complaint; to the extent that any\n\nheadings or other non-numbered statements in the Complaint contain or imply any allegations,\n\nDefendant denies each and every allegation therein. Except as expressly admitted, all allegations\n\nin the Complaint are denied.\n\n 1. Defendant admits that graphics processor units may be used for artificial\n\nintelligence, machine learning, or complex numerical simulation applications. Defendant admits\n\nthat GPU computing powers many of the most advanced and powerful forms of artificial\n\nintelligence over the past decade. Defendant denies the remaining allegations of Paragraph 1.\n\n 2. Defendant admits that graphics processor units may be used for complex numerical\n\nsimulation, machine learning, and training complex models. Defendant further admits that\n\ngraphics processor units typically have more computational processors than central processing\n\nunits and are capable of parallel processing. Defendant does not have knowledge or information\n\n\n\n\n 1\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 3 of 47\n 2 of 46\n\n\n\n\nsufficient to form a belief as to the truth of the remaining allegations contained in Paragraph 2 and\n\non that basis denies them.\n\n 3. Defendant denies the allegations in Paragraph 3 of the Complaint.\n\n 4. Defendant denies the allegations in Paragraph 4 of the Complaint.\n\n NATURE OF THE CASE\n\n 5. Defendant admits that Plaintiff has asserted claims for patent infringement arising\n\nunder 35 U.S.C. \u00a7 1, et seq., but denies all of Plaintiff\u2019s allegations of infringement. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 5 of the Complaint.\n\n 6. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 6 and on that basis denies them.\n\n 7. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 7 and on that basis denies them.\n\n 8. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 8 and on that basis denies them.\n\n 9. Defendant admits that Defendant is a Delaware corporation with its headquarters\n\nin Santa Clara, California. Defendant further admits that it is registered to conduct business in\n\nTexas. Defendant further admits that it has an office in Austin, Texas. Except as expressly\n\nadmitted, Defendant denies the remaining allegations in Paragraph 9 of the Complaint.\n\n JURISDICTION & VENUE\n\n 10. Defendant admits that Plaintiff purports to assert claims for patent infringement\n\narising under 35 U.S.C. \u00a7 1, et seq., but denies all claims of infringement by Defendant. Defendant\n\nadmits that the Court has subject matter jurisdiction pursuant to 28 U.S.C. \u00a7\u00a7 1331 and 1338(a).\n\n\n\n\n 2\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 4 of 47\n 3 of 46\n\n\n\n\nExcept as expressly admitted, Defendant denies the remaining allegations in Paragraph 10 of the\n\nComplaint.\n\n 11. Defendant admits that this Court has personal jurisdiction for the purpose of this\n\nparticular action and that it does business in Texas and in this District. Except as expressly\n\nadmitted, Defendant denies the remaining allegations in Paragraph 11 of the Complaint.\n\n 12. Defendant admits that Defendant has conducted business in within this District, but\n\ndenies all allegations of infringement. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 12 of the Complaint.\n\n 13. Defendant admits that this Court has personal jurisdiction for the purpose of this\n\nparticular action. To the extent the allegations of Paragraph 13 purport to quote from or\n\ncharacterize the contents of written documents, those documents speak for themselves. Defendant\n\ndenies the remaining allegations in Paragraph 13 of the complaint.\n\n 14. Defendant admits that venue is proper for this case but denies that it is a convenient\n\nforum for NVIDIA and its witnesses. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 14 of the complaint.\n\n 15. Defendant Nvidia Corporation is a registered business in Texas and has regular and\n\nestablished places of business Defendant admits that it is a registered business in Texas and\n\nmaintains an office located at 11001 Lakeline Blvd, Suite 100 Bldg. 2, Austin, Texas 78717. To\n\nthe extent the allegations of Paragraph 15 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 15 of the complaint.\n\n\n\n\n 3\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 5 of 47\n 4 of 46\n\n\n\n\n 16. Defendant admits that it has hundreds of employees in this District. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the remaining\n\nallegations contained in Paragraph 16 and on that basis denies them.\n\n 17. Defendant admits that it has open job postings for jobs that may be filled in a\n\nnumber of locations, including in this District. Defendant does not have knowledge or information\n\nsufficient to form a belief as to the truth of the remaining allegations contained in Paragraph 17\n\nand on that basis denies them.\n\n 18. Defendant admits that it engages it engages in business in this District. Defendant\n\nadmits that it has customer-facing personnel and operations in this District. Defendant admits that\n\nit provides technical support to partners and customers for its products in this District. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 18 of the Complaint.\n\n 19. Defendant denies the allegations in Paragraph 19 of the Complaint.\n\n 20. Defendant admits that it sells products and provides services in the State of Texas,\n\nbut denies that those products or services infringe the Asserted Patents. To the extent the\n\nallegations in Paragraph 20 of the Complaint relate to the knowledge or actions of third parties,\n\nDefendant does not have knowledge or information sufficient to form a belief as to the truth of\n\nthose allegations and on that basis denies them. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 20 of the Complaint.\n\n 21. Defendant admits that it sells products and provides services in the State of Texas,\n\nbut denies that the use of those products or services infringes the Asserted Patents. To the extent\n\nthe allegations in Paragraph 21 of the Complaint relate to the knowledge or actions of third parties,\n\nDefendant does not have knowledge or information sufficient to form a belief as to the truth of\n\n\n\n\n 4\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 6 of 47\n 5 of 46\n\n\n\n\nthose allegations and on that basis denies them. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 21 of the Complaint.\n\n 22. Defendant admits that it partners with resellers and managed service providers for\n\nthe sale or installation of certain NVIDIA products. To the extent the allegations of Paragraph 22\n\npurport to quote from or characterize the contents of websites, those documents speak for\n\nthemselves. Except as expressly admitted, Defendant denies the remaining allegations in\n\nParagraph 22 of the Complaint.\n\n 23. Defendant admits that it partners with data center providers. To the extent the\n\nallegations of Paragraph 23 purport to quote from or characterize the contents of written\n\ndocuments, those documents speak for themselves. To the extent the allegations in Paragraph 23\n\nof the Complaint relate to the knowledge or actions of third parties, Defendant does not have\n\nknowledge or information sufficient to form a belief as to the truth of those allegations and on that\n\nbasis denies them. Except as expressly admitted, Defendant denies the remaining allegations in\n\nParagraph 23 of the Complaint.\n\n 24. Defendant denies the allegations in Paragraph 24 of the Complaint.\n\n 25. To the extent the allegations of Paragraph 25 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 25 of the Complaint.\n\n 26. To the extent the allegations of Paragraph 26 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 26 of the Complaint.\n\n\n\n\n 5\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 7 of 47\n 6 of 46\n\n\n\n\n 27. To the extent the allegations of Paragraph 27 purport to quote from or characterize\n\nthe contents of websites, those websites speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 27 of the Complaint.\n\n 28. Defendant denies the allegations in Paragraph 28 of the Complaint.\n\n 29. To the extent the allegations of Paragraph 29 purport to quote from or characterize\n\nthe contents of written documents, those documents speak for themselves. Defendant denies the\n\nremaining allegations in Paragraph 29 of the Complaint.\n\n 30. Defendant denies the allegations in Paragraph 30 of the Complaint.\n\n PLAINTIFF\u2019S PATENTED INNOVATIONS\n\n 31. Defendant does not have knowledge or information sufficient to form a belief as to\n\nthe truth of the allegations contained in Paragraph 31 and on that basis denies them.\n\n The GPU-Based Acceleration Patents\n U.S. Patent Nos. 8,648,867, RE49,461, and RE48,438\n\n 32. Defendant admits that the \u2019461 Patent purports to be a continuation of the \u2019438\n\nPatent, which purports to be an application for reissue of U.S. Patent No. 9,189,828, which purports\n\nto be a continuation of the \u2019867 Patent. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 32 of the Complaint.\n\n 33. Defendant admits that Exhibit 1 to the Complaint appears to be a copy of the \u2019867\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nSeptember 24, 2007, and was issued on February 11, 2014. Defendant further admits that the \u2019867\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019867\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 33 of the Complaint.\n\n\n\n\n 6\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 8 of 47 of 46\n\n\n\n\n 34. Defendant admits that Exhibit 2 to the Complaint appears to be a copy of the \u2019438\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nNovember 9, 2017, and was issued on February 16, 2021. Defendant further admits that the \u2019438\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019438\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 34 of the Complaint.\n\n 35. Defendant admits that Exhibit 3 to the Complaint appears to be a copy of the \u2019461\n\nPatent, which is titled \u201cGraphic Processor Based Accelerator System and Method,\u201d was filed on\n\nDecember 29, 2020, and was issued on March 14, 2023. Defendant further admits that the \u2019461\n\nPatent purports to claim priority to U.S. Provisional App. No. 60/826,892 but denies that the \u2019461\n\nPatent is entitled to that claim of priority. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 35 of the Complaint.\n\n 36. To the extent the allegations of Paragraph 36 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 36 of the Complaint.\n\n 37. To the extent the allegations of Paragraph 37 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 37 of the Complaint.\n\n 38. To the extent the allegations of Paragraph 38 purport to quote from or characterize\n\nthe contents of the \u2019867 Patent, that document speaks for itself. Defendant denies the remaining\n\nallegations in Paragraph 38 of the Complaint.\n\n\n\n\n 7\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 9 of 47\n 8 of 46\n\n\n\n\n 39. To the extent the allegations of Paragraph 39 purport to quote from or characterize\n\nthe contents of the \u2019461 and \u2019438 Patents, those documents speak for themselves. Defendant denies\n\nthe remaining allegations in Paragraph 39 of the Complaint.\n\n 40. To the extent the allegations of Paragraph 40 purport to quote from or characterize\n\nthe contents of the Asserted Patents, those documents speak for themselves. Defendant denies the\n\nremaining allegations in Paragraph 40 of the Complaint.\n\n ACCUSED PRODUCTS\n\n 41. Defendant admits that it offers GPUs and various hardware and software products,\n\nbut specifically denies that those products infringe the Asserted Patents. To the extent the\n\nallegations of Paragraph 41 purport to quote from or characterize the contents of websites, those\n\ndocuments speak for themselves. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 41 of the Complaint.\n\n 42. Defendant admits that \u201cHopper,\u201d \u201cAda Lovelace,\u201d \u201cAmpere,\u201d \u201cTuring,\u201d \u201cVolta,\u201d\n\n\u201cPascal,\u201d and \u201cMaxwell\u201d are architectures of Defendant\u2019s GPUs but specifically denies that those\n\nproducts infringe the Asserted Patents. To the extent the allegations of Paragraph 42 purport to\n\nquote from or characterize the contents of websites, those documents speak for themselves. Except\n\nas expressly admitted, Defendant denies the remaining allegations in Paragraph 42 of the\n\nComplaint.\n\n 43. Defendant admits that its products include the H100 and H200 GPUs, but\n\nspecifically denies that those products infringe the Asserted Patents. Defendant further admits that\n\nit offers the GH200 \u201cGrace Hopper Superchip,\u201d but likewise specifically denies that this product\n\ninfringes the Asserted Patents. To the extent the allegations of Paragraph 43 purport to quote from\n\n\n\n\n 8\n\f CaseCase\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page 10\n Page\n of 47\n 9 of 46\n\n\n\n\nor characterize the contents of websites, those documents speak for themselves. Except as\n\nexpressly admitted, Defendant denies the remaining allegations in Paragraph 43 of the Complaint.\n\n 44. Defendant admits that its products include GPUs with the Ada Lovelace\n\narchitecture, but specifically denies that those products infringe the Asserted Patents. To the extent\n\nthe allegations of Paragraph 44 purport to quote from or characterize the contents of websites,\n\nthose documents speak for themselves. Except as expressly admitted, Defendant denies the\n\nremaining allegations in Paragraph 44 of the Complaint.\n\n 45. Defendant admits that its products include GPUs with the Ampere architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 45 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 45 of the Complaint.\n\n 46. Defendant admits that its products include GPUs with the Turing architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 46 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 46 of the Complaint.\n\n 47. Defendant admits that its products include GPUs with the Volta architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 47 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 47 of the Complaint.\n\n\n\n\n 9\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 11 of 47\n 10 of 46\n\n\n\n\n 48. Defendant admits that its products include GPUs with the Pascal architecture, but\n\nspecifically denies that those products infringe the Asserted Patents. To the extent the allegations\n\nof Paragraph 48 purport to quote from or characterize the contents of websites, those documents\n\nspeak for themselves. Except as expressly admitted, Defendant denies the remaining allegations\n\nin Paragraph 48 of the Complaint.\n\n 49. Defendant admits that its products include GPUs with the Maxwell architecture,\n\nbut specifically denies that those products infringe the Asserted Patents. To the extent the\n\nallegations of Paragraph 49 purport to quote from or characterize the contents of websites, those\n\ndocuments speak for themselves. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 49 of the Complaint.\n\n 50. Defendant admits that certain of Defendant\u2019s GPU architectures support the CUDA\n\nplatform. To the extent the allegations of Paragraph 50 purport to quote from or characterize the\n\ncontents of websites, those documents speak for themselves. Except as expressly admitted,\n\nDefendant denies the remaining allegations in Paragraph 50 of the Complaint.\n\n 51. Defendant admits that it has marketed products under the EGX, HGX, DGX, and\n\nOVX product names, but specifically denies that those products infringe the Asserted Patents. To\n\nthe extent the allegations of Paragraph 51 purport to quote from or characterize the contents of\n\nwebsites, those documents speak for themselves. Except as expressly admitted, Defendant denies\n\nthe remaining allegations in Paragraph 51 of the Complaint.\n\n 52. To the extent the allegations of Paragraph 52 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 52 of the Complaint.\n\n\n\n\n 10\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 12 of 47\n 11 of 46\n\n\n\n\n 53. Paragraph 53 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 54. To the extent the allegations of Paragraph 54 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 54 of the Complaint.\n\n 55. To the extent the allegations of Paragraph 55 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 55 of the Complaint.\n\n 56. To the extent the allegations of Paragraph 56 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 56 of the Complaint.\n\n 57. To the extent the allegations of Paragraph 57 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 57 of the Complaint.\n\n 58. To the extent the allegations of Paragraph 58 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 58 of the Complaint.\n\n 59. To the extent the allegations of Paragraph 59 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 59 of the Complaint.\n\n 60. To the extent the allegations of Paragraph 60 purport to quote from or characterize\n\nthe contents of websites, those documents speak for themselves. Defendant denies the remaining\n\nallegations in Paragraph 60 of the Complaint.\n\n\n\n\n 11\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 13 of 47\n 12 of 46\n\n\n\n\n FIRST CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019867 PATENT)\n\n 61. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 62. Defendant denies the allegations in Paragraph 62 of the Complaint.\n\n 63. Paragraph 63 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 64. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 64 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 64 of the Complaint.\n\n 65. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 65 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 65 of the Complaint.\n\n 66. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 66 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 66 of the Complaint.\n\n 67. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 67 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 67 of the Complaint.\n\n\n\n\n 12\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 14 of 47\n 13 of 46\n\n\n\n\n 68. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 68 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 68 of the Complaint.\n\n 69. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 69 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 69 of the Complaint.\n\n 70. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 70 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 70 of the Complaint.\n\n 71. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 71 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 71 of the Complaint.\n\n 72. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 72 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 72 of the Complaint.\n\n 73. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 73 purport to quote from or characterize the contents of\n\n\n\n\n 13\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 15 of 47\n 14 of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 73 of the Complaint.\n\n 74. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 74 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 74 of the Complaint.\n\n 75. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 75 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 75 of the Complaint.\n\n 76. Defendant denies the allegations in Paragraph 76 of the Complaint.\n\n 77. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 77 of the Complaint.\n\n 78. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 78 of the Complaint.\n\n\n\n\n 14\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 16 of 47\n 15 of 46\n\n\n\n\n 79. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 79 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 79 of the Complaint.\n\n 80. Defendant denies the allegations in Paragraph 80 of the Complaint.\n\n 81. Defendant denies the allegations in Paragraph 81 of the Complaint.\n\n 82. Defendant denies the allegations in Paragraph 82 of the Complaint.\n\n 83. Defendant denies the allegations in Paragraph 83 of the Complaint.\n\n 84. Defendant denies the allegations in Paragraph 84 of the Complaint.\n\n 85. Defendant denies the allegations in Paragraph 85 of the Complaint.\n\n 86. Defendant denies the allegations in Paragraph 86 of the Complaint.\n\n 87. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n87 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 87 of the Complaint.\n\n 88. Defendant denies the allegations in Paragraph 88 of the Complaint.\n\n 89. Defendant denies the allegations in Paragraph 89 of the Complaint.\n\n 90. Defendant denies the allegations in Paragraph 90 of the Complaint.\n\n 91. Defendant denies the allegations in Paragraph 91 of the Complaint.\n\n 92. Defendant denies the allegations in Paragraph 92 of the Complaint.\n\n 93. Defendant denies the allegations in Paragraph 93 of the Complaint.\n\n\n\n\n 15\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 17 of 47\n 16 of 46\n\n\n\n\n SECOND CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019438 PATENT)\n\n 94. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 95. Defendant denies the allegations in Paragraph 95 of the Complaint.\n\n 96. Paragraph 96 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 97. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 97 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 97 of the Complaint.\n\n 98. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 98 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 98 of the Complaint.\n\n 99. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 99 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 99 of the Complaint.\n\n 100. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 100 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 100 of the Complaint.\n\n\n\n\n 16\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 18 of 47\n 17 of 46\n\n\n\n\n 101. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 101 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 101 of the Complaint.\n\n 102. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 102 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 102 of the Complaint.\n\n 103. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 103 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 103 of the Complaint.\n\n 104. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 104 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 104 of the Complaint.\n\n 105. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 105 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 105 of the Complaint.\n\n 106. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 106 purport to quote from or characterize the contents of\n\n\n\n\n 17\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 19 of 47\n 18 of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 106 of the Complaint.\n\n 107. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 107 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 107 of the Complaint.\n\n 108. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 108 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 108 of the Complaint.\n\n 109. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 109 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 109 of the Complaint.\n\n 110. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 110 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 110 of the Complaint.\n\n 111. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 111 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 111 of the Complaint.\n\n\n\n\n 18\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 20 of 47\n 19 of 46\n\n\n\n\n 112. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 112 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 112 of the Complaint.\n\n 113. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 113 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 113 of the Complaint.\n\n 114. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 114 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 114 of the Complaint.\n\n 115. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 115 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 115 of the Complaint.\n\n 116. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 116 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 116 of the Complaint.\n\n 117. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 117 purport to quote from or characterize the contents of\n\n\n\n\n 19\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 21 of 47\n 20 of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 117 of the Complaint.\n\n 118. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 118 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 118 of the Complaint.\n\n 119. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 119 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 119 of the Complaint.\n\n 120. Defendant denies the allegations in Paragraph 120 of the Complaint.\n\n 121. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Defendant admits that it became aware of the \u2019438 Patent since at\n\nleast the filing of this Complaint. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 121 of the Complaint.\n\n 122. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 122 of the Complaint.\n\n\n\n\n 20\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 22 of 47\n 21 of 46\n\n\n\n\n 123. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 123 purport to quote from or characterize the contents\n\nof websites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 123 of the Complaint.\n\n 124. Defendant denies the allegations in Paragraph 124 of the Complaint.\n\n 125. Defendant denies the allegations in Paragraph 125 of the Complaint.\n\n 126. Defendant denies the allegations in Paragraph 126 of the Complaint.\n\n 127. Defendant denies the allegations in Paragraph 127 of the Complaint.\n\n 128. Defendant denies the allegations in Paragraph 128 of the Complaint.\n\n 129. Defendant denies the allegations in Paragraph 129 of the Complaint.\n\n 130. Defendant denies the allegations in Paragraph 130 of the Complaint.\n\n 131. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n131 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 131 of the Complaint.\n\n 132. Defendant denies the allegations in Paragraph 132 of the Complaint.\n\n 133. Defendant denies the allegations in Paragraph 133 of the Complaint.\n\n 134. Defendant denies the allegations in Paragraph 134 of the Complaint.\n\n 135. Defendant denies the allegations in Paragraph 135 of the Complaint.\n\n 136. Defendant denies the allegations in Paragraph 136 of the Complaint.\n\n 137. Defendant denies the allegations in Paragraph 137 of the Complaint.\n\n\n\n\n 21\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 23 of 47\n 22 of 46\n\n\n\n\n THIRD CAUSE OF ACTION\n (INFRINGEMENT OF THE \u2019461 PATENT)\n\n 138. Defendant restates and incorporates by reference its answers to the preceding\n\nparagraphs of the Complaint.\n\n 139. Defendant denies the allegations in Paragraph 139 of the Complaint.\n\n 140. Paragraph 140 of the Complaint does not contain any allegation which requires a\n\nresponse.\n\n 141. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 141 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 141 of the Complaint.\n\n 142. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 142 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 142 of the Complaint.\n\n 143. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 143 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 143 of the Complaint.\n\n 144. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 144 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 144 of the Complaint.\n\n\n\n\n 22\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 24 of 47\n 23 of 46\n\n\n\n\n 145. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 145 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 145 of the Complaint.\n\n 146. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 146 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 146 of the Complaint.\n\n 147. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 147 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 147 of the Complaint.\n\n 148. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 148 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 148 of the Complaint.\n\n 149. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 149 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 149 of the Complaint.\n\n 150. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 150 purport to quote from or characterize the contents of\n\n\n\n\n 23\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 25 of 47\n 24 of 46\n\n\n\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 150 of the Complaint.\n\n 151. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 151 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 151 of the Complaint.\n\n 152. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 152 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 152 of the Complaint.\n\n 153. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 153 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 153 of the Complaint.\n\n 154. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 154 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 154 of the Complaint.\n\n 155. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 155 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 155 of the Complaint.\n\n\n\n\n 24\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 26 of 47\n 25 of 46\n\n\n\n\n 156. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 156 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 156 of the Complaint.\n\n 157. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 157 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 157 of the Complaint.\n\n 158. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 158 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 158 of the Complaint.\n\n 159. Defendant specifically denies that it has infringed any of the Asserted Patents. To\n\nthe extent the allegations of Paragraph 159 purport to quote from or characterize the contents of\n\nwebsites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 159 of the Complaint.\n\n 160. Defendant denies the allegations in Paragraph 160 of the Complaint.\n\n 161. Defendant admits that Mr. Sanford Russell was employed by NVIDIA in 2007.\n\nDefendant denies that Mr. Russell was, at any point in time, the CTO of NVIDIA. Defendant does\n\nnot have knowledge or information sufficient to form a belief as to the truth of the allegations\n\nregarding the alleged communications between NVIDIA employees and Neurala made in 2007\n\nand on that basis denies them. Defendant admits that it became aware of the \u2019461 Patent since at\n\n\n\n\n 25\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 27 of 47\n 26 of 46\n\n\n\n\nleast the filing of this Complaint. Except as expressly admitted, Defendant denies the remaining\n\nallegations in Paragraph 161 of the Complaint.\n\n 162. Defendant admits that, in or around 2016, Defendant had discussions with Neurala,\n\nInc. and that at least Mr. Alvin Lin and/or Mr. Jeff Herbst from Defendant were involved.\n\nDefendant also admits that, in 2016, Defendant had discussions with at least one of the inventors\n\nregarding potential investments in Neurala, Inc. Defendant denies the remaining allegations in\n\nParagraph 162 of the Complaint.\n\n 163. Defendant admits that it hosted its GPU Technology Conference in May of 2017.\n\nTo the extent the allegations of Paragraph 161 purport to quote from or characterize the contents\n\nof websites, those websites speak for themselves. Defendant denies the remaining allegations in\n\nParagraph 163 of the Complaint.\n\n 164. Defendant denies the allegations in Paragraph 164 of the Complaint.\n\n 165. Defendant denies the allegations in Paragraph 165 of the Complaint.\n\n 166. Defendant denies the allegations in Paragraph 166 of the Complaint.\n\n 167. Defendant denies the allegations in Paragraph 167 of the Complaint.\n\n 168. Defendant denies the allegations in Paragraph 168 of the Complaint.\n\n 169. Defendant denies the allegations in Paragraph 169 of the Complaint.\n\n 170. Defendant denies the allegations in Paragraph 170 of the Complaint.\n\n 171. Defendant admits that it sells and has sold its products and provides certain\n\ntechnical support to its customers for those products. To the extent the allegations of Paragraph\n\n171 purport to quote from or characterize the contents of websites, those websites speak for\n\nthemselves. Defendant denies the remaining allegations in Paragraph 171 of the Complaint.\n\n 172. Defendant denies the allegations in Paragraph 172 of the Complaint.\n\n\n\n\n 26\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 28 of 47\n 27 of 46\n\n\n\n\n 173. Defendant denies the allegations in Paragraph 173 of the Complaint.\n\n 174. Defendant denies the allegations in Paragraph 174 of the Complaint.\n\n 175. Defendant denies the allegations in Paragraph 175 of the Complaint.\n\n 176. Defendant denies the allegations in Paragraph 176 of the Complaint.\n\n 177. Defendant denies the allegations in Paragraph 177 of the Complaint.\n\n PRAYER FOR RELIEF\n\n Defendant denies any factual assertions contained in Plaintiff\u2019s Prayer for Relief.\n\nDefendant further denies that Plaintiff is entitled to any relief whatsoever, including but not limited\n\nto the relief sought in Paragraphs A-G of the Complaint.\n\n DEMAND FOR JURY TRIAL\n\n A response is not required to Plaintiff\u2019s demand for a jury trial.\n\n DEFENSES\n\n Defendant repeats and re-alleges the allegations of the preceding Paragraphs as if fully set\n\nforth herein. Defendant asserts the following defenses to Plaintiff\u2019s Complaint, without admitting\n\nor acknowledging that Defendant bears the burden of proof as to any of them or that any must be\n\npleaded as defenses. Defendant specifically reserves all rights to allege additional defenses that\n\nbecome known through the course of discovery.\n\n FIRST DEFENSE\n (Non-Infringement)\n\n Defendant has not and does not infringe, either literally or under the doctrine of\n\nequivalents, any valid and enforceable claim of any Asserted Patent, whether directly, indirectly,\n\ncontributorily, by inducement, individually, jointly, willfully, or otherwise. Additionally, with\n\nrespect to Plaintiff\u2019s allegations of indirect, joint, and willful infringement, Defendant lacks the\n\nrequisite mens rea.\n\n\n\n 27\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 29 of 47\n 28 of 46\n\n\n\n\n SECOND DEFENSE\n (Invalidity & Ineligibility)\n\n The Asserted Claims are invalid under at least 35 U.S.C. \u00a7\u00a7 101, 102, 103, and/or 112.\n\nDefendant incorporates by reference its forthcoming invalidity contentions and all amendments\n\nthereto.\n\n THIRD DEFENSE\n (No Willfulness, Enhanced Damages, or Attorneys\u2019 Fees)\n\n Plaintiff is not entitled to enhanced damages under 35 U.S.C. \u00a7 284, at least because\n\nPlaintiff has failed to show, and cannot show, that any infringement has been willful and/or\n\nknowing. Plaintiff is not entitled to an award of attorney\u2019s fees under 35 U.S.C. \u00a7 285, at least\n\nbecause Plaintiff has failed to show, and cannot show, that this case is \u201cexceptional\u201d in Plaintiff\u2019s\n\nfavor as would be required by the statute.\n\n FOURTH DEFENSE\n (Statutory Limitation on Damages)\n\n Plaintiff\u2019s claims for relief are statutorily limited in whole or in part by 35 U.S.C. \u00a7\u00a7 286\n\nand/or 287. In addition, to the extent Plaintiff seeks damages for allegedly infringing acts\n\ncommitted more than six years prior to the filing of the Complaint in this action, it is barred from\n\nrecovery of such damages.\n\n Additionally, to the extent Plaintiff or any licensee of the Asserted Patent failed to properly\n\nmark any of their relevant products as required by 35 U.S.C. \u00a7 287 or otherwise failed to give\n\nproper notice that Defendant\u2019s actions allegedly infringed any Asserted Claim, Defendant is not\n\nliable to Plaintiff for the acts alleged to have been performed before Defendant received actual\n\nnotice of infringement.\n\n FIFTH DEFENSE\n (License, Exhaustion, Waiver, and Estoppel)\n\n Plaintiff\u2019s claims are barred, in whole or in part, by license, exhaustion, and/or the\n\n\n 28\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 30 of 47\n 29 of 46\n\n\n\n\ndoctrines of waiver and/or equitable estoppel.\n\n SIXTH DEFENSE\n (Inexcusable Delay)\n\n Plaintiff is barred from enforcing the Asserted Patents due to inexcusable delay in reviving\n\nthe \u2019867 patent after it was abandoned.\n\n SEVENTH DEFENSE\n (Intervening Rights)\n\n Plaintiff\u2019s claims are barred by the doctrine of absolute intervening rights and the doctrine\n\nof equitable intervening rights with respect to any Accused Product or technology that predates\n\nthe date of the revival of the \u2019867 patent and/or the date of the reissue of the \u2019461 and \u2019438 patents.\n\n35 U.S.C. \u00a7 252.\n\n EIGHTH DEFENSE\n (Improper Reissue)\n\n The claims of the \u2019438 and \u2019461 patents are invalid pursuant to 35 U.S.C. \u00a7 251 because\n\nthey enlarge the scope of the claims of the original patent and/or because they improperly recapture\n\nsubject matter that the patentee intentionally surrendered to obtain a valid patent.\n\n NINTH DEFENSE\n (28 U.S.C. \u00a7 1498)\n\n On information and belief, Plaintiff\u2019s claims against NVIDIA for patent infringement are\n\nbarred, in whole or in part, by 28 U.S.C. \u00a7 1498.\n\n TENTH DEFENSE\n (Ensnarement)\n\n Plaintiff is barred by the doctrine of ensnarement from contending that any Asserted Claim\n\ncovers any product, service, or method practiced, manufactured, used, sold, or offered for sale by\n\nDefendant in any manner that would ensnare the prior art.\n\n\n\n\n 29\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 31 of 47\n 30 of 46\n\n\n\n\n ELEVENTH DEFENSE\n (Territoriality)\n\n Plaintiff is not entitled to damages arising from any purported indirect infringement by\n\nDefendant that is premised on direct infringement by end-users occurring outside of the United\n\nStates under 35 U.S.C. \u00a7 271.\n\n TWELFTH DEFENSE\n (No Standing)\n\n Plaintiff\u2019s claims are barred because Plaintiff lacks standing to bring this suit. Specifically,\n\nPlaintiff cannot prove that it is the rightful owner of the Asserted Patents.\n\n THIRTEENTH DEFENSE\n (Failure to State a Claim)\n\n The Complaint fails to state a claim upon which relief may be granted.\n\n FOURTEENTH DEFENSE\n (Inconvenient Venue)\n\n For the convenience of parties and witnesses, venue for this action is not convenient\n\nin this district and would be more appropriate in another district. 28 U.S.C. \u00a7 1404.\n\n FIFTEENTH DEFENSE\n (Reservation of Defenses)\n\n Defendant reserves all affirmative defenses under Rule 8(c) of the Federal Rules of\n\nCivil Procedure, as well as any other defenses at law or in equity that may exist now or that may\n\nbe available in the future.\n\n SIXTEENTH DEFENSE\n (Unenforceability Due to Inequitable Conduct)\n\n 1. Each of the claims of the Asserted Patents is unenforceable due to inequitable\n\nconduct committed by prior assignee Neurala, one or more of the named inventors (Anatoli\n\nGorchetchnikov, Heather Marie Ames, Massimiliano Versace, and Fabrizio Santini), and/or\n\n\n\n\n 30\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 32 of 47\n 31 of 46\n\n\n\n\nprosecution counsel for prior assignee Neurala or current assignee NAI (including, but not limited\n\nto, Christopher Max Colice).\n\n 2. This case arises from a pattern of concealment and misrepresentation surrounding\n\na purported improvement to general-purpose computing on graphics processing units (\u201cGPGPU\u201d).\n\nLong before Neurala\u2019s initial patent application that led to the Asserted patents was filed, NVIDIA\n\nhad pioneered GPGPU computing, developing both the hardware and software foundations for\n\nexecuting general-purpose numerical computations on GPUs. NVIDIA\u2019s engineers\u2014through\n\ntechnologies such as BrookGPU and well-known books such as GPU Gems 2\u2014publicly disclosed\n\nthe very concepts later claimed by Neurala. These were not obscure academic papers; they were\n\nwell-known, widely cited works intended to teach the industry how to harness GPUs for scientific\n\ncomputing. Neurala and the named inventors were well aware of not only NVIDIA\u2019s role in\n\nGPGPU development, but of NVIDIA\u2019s specific teachings in GPU Gems 2 and other technologies.\n\n 3. Against this backdrop, the Asserted Patents claim a narrow and incremental\n\npurported improvement to GPGPU\u2014such as merely offloading certain setup and control functions\n\nfrom the host CPU to an \u201caccelerator controller.\u201d This supposed improvement did not create a\n\nnew GPGPU paradigm; it merely repeated well understood ideas from NVIDIA\u2019s prior work and\n\ncontributions to the field. Yet Neurala\u2019s inventors and attorneys withheld NVIDIA\u2019s key patents\n\nand publications\u2014including GPU Gems 2\u2014from the Patent Office while advancing their own\n\napplication.\n\n 4. There are two independent bases for an inequitable conduct finding, either of which,\n\nstanding alone, renders the Asserted Patents unenforceable. Together, they demonstrate a\n\ncoordinated pattern of misleading conduct intended to misdirect the Patent Office about the true\n\n\n\n\n 31\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 33 of 47\n 32 of 46\n\n\n\n\nstate of the art, the inventors\u2019 knowledge of it, and Neurala\u2019s desperate attempt to get patents on\n\ntechnology it did not invent.\n\n 5. Withholding of Material Prior Art. The claims of the Asserted Patents are\n\nunenforceable due to Neurala\u2019s and/or one or more of the named inventors\u2019 intentional withholding\n\nof material prior art references during prosecution of the Asserted Patents with an intent to deceive\n\nthe Patent Office.\n\n 6. False Declaration Regarding Abandonment. The claims of the Asserted Patents\n\nare also unenforceable due to Neurala\u2019s filing of a false declaration regarding abandonment of the\n\napplication that issued as the \u2019867 patent. Without that false declaration, none of the Asserted\n\nPatents would have issued.\n\n 7. Collectively, these acts form a coherent pattern of inequitable conduct\u2014a deliberate\n\neffort to obscure NVIDIA\u2019s pioneering role in GPGPU computing and to mislead the Patent Office\n\ninto granting patents on technology NVIDIA and others had already disclosed to the world.\n\nWithholding Material Prior Art\n\n 8. Neurala, each of the named inventors, and their counsel involved in the prosecution\n\nof the Asserted Patents had a duty of candor and good faith in dealing with the Patent Office, as\n\nrequired by 37 C.F.R. \u00a7 1.56. Their individual and collective failure to disclose known material\n\nprior art was done with specific intent to mislead or deceive the Patent Office into issuing each of\n\nthe Asserted Patents. As a result, all of the Asserted Patents are unenforceable due to inequitable\n\nconduct.\n\n\n\n\n 32\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 34 of 47\n 33 of 46\n\n\n\n\nKnowledge and Materiality of GPU Gems 2\n\n 9. Neurala, one or more of the named inventors, and/or Neurala\u2019s prosecution counsel\n\nintentionally withheld GPU Gems 2: Programming Techniques for High-Performance Graphics\n\nand General-Purpose Computation (\u201cGPU Gems 2\u201d) (March 2005) from the Patent Office.\n\n 10. NVIDIA published a series of books that provided practical guidance and\n\ntechniques for using GPUs in general-purpose applications, helping developers harness the parallel\n\nprocessing power of GPUs for a wide range of fields. One of those books was GPU Gems 2, which\n\nNVIDIA published on the Internet in March 2005 and made it available for download for free to\n\nanyone that wanted to download it. See Ex. 1 (April 1, 2025 Invalidity Contentions Ex. A11), Ex.\n\n2 (August 29, 2025 Supplemental Invalidity Contentions Supp. Ex. A11), Ex. 3 (April 1, 2025\n\nInvalidity Contentions App\u2019x B), Ex. 4 (April 1, 2025 Invalidity Contentions Ex. C11). At least\n\none of the named inventors\n\n\n\n . Despite extensive knowledge of GPU Gems 2 and its direct relevance for teaching\n\ntechniques for using GPUs in general-purpose applications, that inventor and Neurala withheld\n\nGPU Gems 2 from the Patent Office for all seven years that the \u2019867 patent was pending and every\n\nyear since.\n\n 11. not only was intimately familiar with GPU Gems 2 prior to the\n\nfiling of the provisional patent application from which the Asserted Patents claim priority but also\n\n\n\n\n 33\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 35 of 47\n 34 of 46\n\n\n\n\n 12. and one or more of the other named inventors met regularly to\n\ndiscuss the implementation of the project and the filing of the provisional patent application from\n\nwhich the Asserted Patents claim priority. See, e.g.,\n\nOn information and belief, the other named inventors and/or prosecution counsel were also aware\n\nof GPU Gems 2 due to the close nature of their working relationship with .\n\n 13. As NVIDIA detailed in its Invalidity Contentions served April 1, 2025, August 29,\n\n2025, and October 30, 2025, GPU Gems 2 is a prior art reference that is material to each of\n\nthe \u2019867, \u2019461, and \u2019438 patents. GPU Gems 2 anticipates the \u2019867 and \u2019438 patents and discloses\n\nkey elements of the \u2019461 patent claims. When combined with other unconsidered prior art,\n\nincluding NVIDIA\u2019s own patents, GPU Gems 2 renders obvious all of the Asserted Claims of the\n\nAsserted Patents. See Exs. 1, 2, 3, 4; see also Ex. 14 (April 1, 2025, Preliminary Invalidity\n\nContentions Cover Pleading), Ex. 15 (August 29, 2025 Supplemental Invalidity Contentions Cover\n\nPleading), Ex. 16 (October 30, 2025 Second Supplemental Invalidity Contentions Cover Pleading).\n\nThe Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but for the\n\nwithholding of GPU Gems 2. Therasense, Inc. v. Becton, Dickinson & Co., 649 F.3d 1276, 1290\u2013\n\n91 (Fed. Cir. 2011). For example, as NVIDIA details in its Invalidity Contentions with respect to\n\nclaim 16 of the \u2019867 patent, GPU Gems 2 teaches \u201can accelerator controller, operably coupled to\n\nthe accelerator memory and the central processing unit.\u201d See Ex. 1 at 15\u201344; Ex. 2; Ex. 17 (\u2019867\n\npatent file history) at 43. GPU Gems 2 teaches that the accelerator controller that \u201ctransfer[s] the\n\n\n\n\n 34\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 36 of 47\n 35 of 46\n\n\n\n\nat least the portion of the input data into the accelerator memory before the first computational\n\ncycle.\u201d See Ex. 1 at 27\u201344; Ex. 2; Ex. 17 at 43. GPU Gems 2 teaches that the accelerator controller\n\nthat \u201ctransfer[s] the first output data from the accelerator memory to the main memory during the\n\nsecond computational cycle\u201d and \u201cdirect[s] the second output data into the second partition during\n\nthe second computational cycle.\u201d See Ex. 1 at 54\u201369; Ex. 2; Ex. 17 at 43. GPU Gems 2 teaches\n\nthat the accelerator controller \u201cswap[s] the first pointer and the second pointer at the conclusion of\n\nthe second computational cycle such that the second output data becomes an input for a third\n\ncomputational cycle of the plurality of computational cycles.\u201d See Ex. 1 at 69\u201378; Ex. 2; Ex. 17\n\nat 43.\n\nKnowledge and Materiality of SANNDRA/KInNeSS\n\n 14. Mr. Gorchetchnikov and Mr. Massimiliano Versace\u2014named inventors of the\n\nAsserted Patents\u2014developed Synchronous Artificial Neuronal Networks Distributed Runtime\n\nAlgorithm (SANNDRA) and its implementation on KDE Integrated NeuroSimulation Software\n\n(KInNeSS) (\u201cSANNDRA/KInNeSS\u201d) (March 2005). SANNDRA version 1.1.x and KInNeSS\n\n0.3.3 (on which SANNDRA was implemented) were publicly available and in use by March 2005\n\nbased at least on the following information: KInNeSS: A new software environment for\n\nsimulations of neuronal activity; 9th International conference on Cognitive and Neural Systems\n\n(Boston, MA, 2004); https://web.archive.org/web/20051030032020/http://www.kinness.net/\n\n(KInNeSS documentation);\n\nhttps://web.archive.org/web/20080828055305fw /http://symphony.bu.edu/\n\nDocs/SANNDRA/html/index.html (SANNDRA API documentation); see also Ex. 5 (April 1, 2025\n\nInvalidity Contentions Ex. A7), Ex. 6 (April 1, 2025 Invalidity Contentions Ex. B7), Ex. 7 (April\n\n1, 2025 Invalidity Contentions Ex. C7).\n\n\n\n\n 35\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 37 of 47\n 36 of 46\n\n\n\n\n 15. Furthermore, the specification of the Asserted Patents acknowledges that\n\nSANNDRA \u201cwas developed to accelerate and optimize processing of numerical integration of\n\nlarge non-homogenous systems of differential equations.\u201d \u02bc867 patent at 9:25\u201235. And although\n\nthe Asserted Patents reference version 2.x.x of SANNDRA as \u201can example practical software\n\nimplementation of the method and architecture described above and pictorially represented in FIG.\n\n3,\u201d the applicant failed to identify previous, publicly available versions of SANNDRA or KInNeSS\n\nas relevant prior art to the Patent Office. Id.;\n\n\n\n Neurala, the named\n\ninventors, and their prosecution counsel further failed to fully disclose the relevance and\n\nmateriality of their own software to the Patent Office, despite being the ones in the best position\n\nto do so. NVIDIA expects further discovery, including complete production of the documents that\n\nNVIDIA requested from Neurala on July 30, 2025, to shed further light on Mr. Gorchetchnikov\u2019s\n\nand Mr. Versace\u2019s concealment of earlier versions of SANNDRA/KInNeSS.\n\n 16. SANNDRA 1.1.x and earlier versions, as implemented on KInNeSS, together with\n\nother undisclosed references (such as Nickolls and Kirk, among others) renders obvious all of the\n\nAsserted Claims of the Asserted Patents as shown by Exs. 5, 6, 7, and 16. The Patent Office would\n\nnot have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but for the withholding of\n\nSANNDRA/KInNeSS. Therasense, 649 F.3d at 1290\u201391.\n\nKnowledge and Materiality of Cg\n\n 17. The C for Graphics (Cg) language (2003) is a high-level shading language created\n\nby NVIDIA in collaboration with Microsoft to program graphics shaders on GPUs. Cg was made\n\navailable as an open-source release and in public use by 2003 based at least on the following\n\n\n\n\n 36\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 38 of 47\n 37 of 46\n\n\n\n\ninformation: The Cg Tutorial: The Definitive Guide to Programmable Real-Time Graphics (2003);\n\nhttps://web.archive.org/web/20041205090713/http://developer.nvidia.com:80/object/cg_toolkit.h\n\ntml (\u201cCg Toolkit\u201d); see also Ex. 8 (April 1, 2025 Invalidity Contentions Ex. A4), Ex. 9 (April 1,\n\n2025 Invalidity Contentions Ex. B4), Ex. 10 (April 1, 2025 Invalidity Contentions Ex. C4).\n\n 18.\n\n\n\n 19. The Cg language, together with other references (such as Nickolls and Kirk, among\n\nothers) renders obvious all of the Asserted Claims of the Asserted Patents as shown by Exs. 8, 9,\n\n10, and 16. The Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to issue but\n\nfor the withholding of Cg. Therasense, 649 F.3d at 1290\u201391.\n\nKnowledge and Materiality of BrookGPU\n\n 20. The BrookGPU programming language (2004) is an early system developed at\n\nStanford University to enable general-purpose computing on graphics processing units (GPGPU).\n\nBrookGPU was publicly available and in use by 2004 based at least on the following information:\n\nBuck et al., Brook for GPUs: Stream Computing on Graphics Hardware, ACM, 2004 (\u201cBrook for\n\nGPU 2004\u201d); https://web.archive.org/web/20041205061111/http://graphics.stanford.edu/projects/\n\nbrookgpu/start.html (BrookGPU documentation); see also Ex. 11 (April 1, 2025 Invalidity\n\nContentions Ex. A5), Ex. 12 (April 1, 2025 Invalidity Contentions Ex. B5), Ex. 13 (April 1, 2025\n\nInvalidity Contentions Ex. C5).\n\n 21.\n\n\n\n 22. The BrookGPU language, together with other references (such as Nickolls and Kirk,\n\namong others) renders obvious all of the Asserted Claims of the Asserted Patents, as shown by Exs.\n\n\n\n\n 37\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 39 of 47\n 38 of 46\n\n\n\n\n11, 12, 13, and 16. The Patent Office would not have allowed the \u2019867, \u2019461, or \u2019438 patents to\n\nissue but for the withholding of BrookGPU. Therasense, 649 F.3d at 1290\u201391.\n\nWithholding of Material References\n\n 23. Despite their awareness of numerous references relevant to the technology of the\n\nAsserted Patents, the named inventors did not provide any prior art to the Patent Office during\n\nprosecution of the \u02bc867 patent. Only four references are disclosed on the face of the \u2019867 patent\n\nas having been considered during prosecution, and all four were identified by the examiner in a\n\nNotice of References Cited. Although each of the four references cited by the examiner relates to\n\ngraphics rendering, none of the references provide the practical guidance and techniques for using\n\nGPUs in general-purpose applications that GPU Gems 2 does, or the relevant applied examples\n\nand implementations that system art, such as SANNDRA/KInNeSS, Cg, or BrookGPU provides.\n\nFor each of the two reissue patents, the applicant took the opposite approach and submitted\n\nhundreds of references, none of which was GPU Gems 2, SANNDRA/KInNeSS, Cg, or\n\nBrookGPU, and none of which provide the relevant applied examples and implementations that\n\nGPU Gems 2 does.\n\n 24. None of GPU Gems 2, SANNDRA/KInNeSS, Cg, or BrookGPU is cumulative of\n\nthe information already on record. Unlike the four graphics-rendering references cited by the\n\nexaminer, these materials disclose practical architectures, applied examples, and implementation-\n\nlevel guidance applicable to GPGPU\u2014the very subject matter of the Asserted Patents. But for\n\ntheir withholding, the Patent Office would not have allowed any of the \u2019867, \u2019461, or \u2019438 patents\n\nto issue. The deliberate withholding of these NVIDIA and other GPGPU-related references\n\ndeprived the examiner of the most relevant prior art and materially misled the Patent Office about\n\nthe true state of the art.\n\n\n\n\n 38\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 40 of 47\n 39 of 46\n\n\n\n\nIntent to Deceive\n\n 25. \u2014and by extension, Neurala\u2014knew of these material prior art\n\nreferences. On information and belief, the other named inventors and their counsel involved in\n\nprosecution were similarly aware of these references and knew of their materiality to the Asserted\n\nPatents. and others were aware of GPU Gems 2, SANNDRA/KInNeSS, Cg,\n\nor BrookGPU and made the conscious decision to withhold it from the Patent Office.\n\n 26. The inventors did not file their provisional application until September 25, 2006,\n\nover a year after first reviewed GPU Gems 2. And the inventors did not\n\ndisclose GPU Gems 2 to the Patent Office at any time during nearly seven years of prosecution of\n\nthe application that led to the \u2019867 patent.\n\n\n\n\n Each of these references was material and not\n\ncumulative of the bare record before the examiner during prosecution of the \u2019867 patent. It is\n\nsimply not credible that the named inventors did not think that any prior art was material to\n\nprosecution. These facts demonstrate an intent to deceive the Patent Office by not providing any\n\nprior art for its consideration. Thus, for the \u02bc867 patent, by withholding all known references, the\n\ninventors may have aimed to create a misleading impression of the uniqueness and inventiveness\n\nof their claims. For the two reissue patents, the applicant attempted to flood the Patent Office with\n\nreferences to distract from the key prior art that was omitted: GPU Gems 2, SANNDRA/KInNeSS,\n\nCg, and BrookGPU. These facts, and those yet to be ascertained through discovery demonstrate\n\nthat the most reasonable inference to draw is that the named inventors intended to deceive the\n\nPatent Office by withholding references during prosecution.\n\n\n\n\n 39\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 41 of 47\n 40 of 46\n\n\n\n\n 27. On information and belief, Neurala and the named inventors also intentionally\n\nchose not to disclose GPU Gems 2 to its attorneys responsible for prosecution of the patent\n\napplications at the Patent Office, knowing that such attorneys also owed a duty of candor to the\n\nPatent Office, and would disclose the reference to the Patent Office and the examiner if the\n\nattorneys were to become aware of it. In either case, Neurala and the named inventors violated\n\nthe duty of candor that each of them owed to the Patent Office.\n\n 28. A pattern of deceiving the Patent Office continues. For example, current assignee\n\nNeural AI recently paid the 11-year maintenance fee for the \u2019867 patent and did so with a\n\nrepresentation that it was entitled to small entity status, even though it knew it was no longer\n\nentitled to small entity status due to\n\n 29. These facts, when taken together with the evidence of intent presented for the other\n\nbasis of inequitable conduct, demonstrate a pattern of conduct that shows a continuing intent to\n\ndeceive the Patent Office.\n\nFiling a False Declaration Regarding Abandonment of an Application\n\n 30. In addition to failing to disclose material prior art references during the prosecution\n\nof the \u2019867 patent that, if cited, would have precluded the claims in that patent from issuing,\n\nNeurala and its prosecution counsel also affirmatively misled the Patent Office when it revived the\n\nabandoned application that issued as the \u2019867 patent. But for its misrepresentation, none of the\n\nAsserted Patents would have issued because the patent application from which all of those patents\n\nstem would have remained abandoned. This affirmative misrepresentation is part of Neurala\u2019s\n\ncontinued pattern of inequitable conduct in front of the Patent Office to obtain the Asserted Patents.\n\n 31. Specifically, Neurala and its prosecution counsel allowed the application that issued\n\nas the \u2019867 patent (U.S. Patent Appl. No. 11/860,254) (\u201cthe \u2019254 application\u201d) to become\n\n\n\n\n 40\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 42 of 47\n 41 of 46\n\n\n\n\nabandoned for more than two years before belatedly filing a false, generic declaration alleging that\n\nthe \u2019254 application had been unintentionally abandoned in an attempt to revive it. The claims of\n\neach of the Asserted Patents are therefore also unenforceable due to a false declaration regarding\n\nabandonment of the application that issued as the \u2019867 patent.\n\n 32. The application that issued as the \u2019867 patent (U.S. Patent Appl. No. 11/860,254)\n\n(\u201cthe \u2019254 application\u201d) was filed on September 24, 2007. The Patent Office issued a non-final\n\nOffice Action on September 16, 2010, with a three month non-statutory time period for reply. A\n\nresponse to the Office Action was due on December 16, 2010 without payment of extension fees,\n\nbut the applicant neither filed a response nor requested an extension of time under the provisions\n\nof 37 C.F.R. \u00a7 1.136(a), and the \u2019254 application became abandoned on December 17, 2010, the\n\nday after the expiration of the shortened non-statutory deadline established in the non-final Office\n\nAction.\n\n 33. On April 12, 2011, the Patent Office mailed a notice of abandonment to the\n\napplicant. It was not until July 31, 2013\u2014more than two years later\u2014that the applicant filed a\n\npetition to revive the \u2019254 application. The petition was signed by Christopher Max Colice of\n\nFoley & Lardner LLP and included the statement that \u201c[t]he entire delay in filing the required reply\n\nfrom the due date for the required reply until the filing of a grantable petition under 37 CFR 1.137(b)\n\nwas unintentional.\u201d Ex. 17 at 64\u201388. At the time, no power of attorney had been filed listing Mr.\n\nColice as Neurala\u2019s attorney of record. There is no indication in the statement what investigation\n\nMr. Colice undertook or how he determined that abandonment was \u201cunintentional.\u201d Furthermore,\n\nthe attorney advisor reviewing the petition noted that it was \u201cnot apparent whether the person\n\nsigning the statement of unintentional delay was in a position to have firsthand or direct knowledge\n\n\n\n\n 41\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 43 of 47\n 42 of 46\n\n\n\n\nof the facts and circumstances of the delay at issue,\u201d and that there was \u201cno indication that the\n\npetition [wa]s signed by a registered patent attorney or patent agent of record. Id. at 62.\n\n 34. Materiality. The material prong is met \u201c[w]hen the patentee has engaged in\n\naffirmative acts of egregious misconduct, such as the filing of an unmistakably false affidavit.\u201d\n\nTherasense, 649 F.3d at 1292; see also Rohm & Haas Co. v. Crystal Chem. Co., 722 F.3d 1556,\n\n1571 (Fed. Cir. 1983) (\u201cthere is no room to argue that submission of false affidavits is not\n\nmaterial\u201d); Intellect Wireless, Inc. v. HTC Corp., 732 F.3d 1339, 1342 (Fed. Cir. 2013); Apotex,\n\nInc. v. UCB, Inc., 763 F.3d 1354 (Fed. Cir. 2014). An affirmative act of egregious misconduct is\n\ninherently material. Therasense, 649 F.3d at 1292. The filing of a false revival petition under 37\n\nCFR \u00a7 1.137(a) is an affirmative act of egregious misconduct. In re Rembrandt Techs. LP Patent\n\nLitig., 899 F.3d 1254, 1272\u201374 (Fed. Cir. 2018). Because the \u02bc867 patent (which issued from the\n\n\u02bc254 application) was the first patent in the family, but for the false statement in the Petition for\n\nRevival, all Asserted Patents would not have issued and thus would no longer be in force. The\n\nfalse statement to the Patent Office is therefore material to patentability. See, e.g., 3D Med.\n\nImaging Sys. LLC v. Visage Imaging Inc., 228 F. Supp. 3d 1331, 1338\u201339 (N.D. Ga. 2017).\n\n 35. Intent. The intent of Mr. Colice and/or the named inventors to deceive the Patent\n\nOffice is evidenced at least by the length of time that elapsed between the dates of the Office\n\nAction (September 16, 2010), when the application became abandoned (December 18, 2010), the\n\nNotice of Abandonment (April 12, 2011), and the applicant\u2019s Petition for Revival (July 31, 2013).\n\nFor example, there is no explanation why it took almost three full years from when the Office\n\nAction was issued for the applicant to respond to the Office Action. Even after Neurala LLC was\n\nnotified of the abandonment, it took more than two years to file a Petition for Revival, accompanied\n\nby only a generic statement that the entire delay was unintentional. The most reasonable inference\n\n\n\n\n 42\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 44 of 47\n 43 of 46\n\n\n\n\nto draw is that there was an intent to allow the application to become abandoned, and that Mr.\n\nColice and/or the named inventors intended to deceive the Patent Office by submitting a false\n\ndeclaration stating that the abandonment of the \u2019254 application was unintentional.\n\n 36. The \u02bc254 application was the first application in the chain of applications for the\n\nAsserted Patents (it issued as the asserted \u02bc867 patent). If the \u02bc254 application had been deemed\n\nabandoned and no false declaration had been filed, no patents in the asserted patent family would\n\nhave issued. Therefore, all Asserted Patents should be rendered unenforceable due to the filing of\n\na false declaration in the prosecution of the \u02bc867 patent.\n\n 37. These facts, when taken together with the evidence of intent presented for the other\n\nbasis of inequitable conduct, demonstrate a pattern of conduct that shows a continuing intent to\n\ndeceive the Patent Office.\n\nInfectious Unenforceability\n\n 38. The \u02bc254 application\u2014from which the \u2019867 patent issued\u2014was the first application\n\nin the chain of applications for the Asserted Patents. Each of the \u2019461 and the \u2019438 patents is a\n\nchild of the \u2019867 patent, and the pattern of blatantly inequitable conduct that pervades the\n\nprosecution of this patent family renders the claims of each of the Asserted Patents unenforceable.\n\n 39. Here, but for the false declaration filed in support of the revival of the \u2019254\n\napplication, no patents in the asserted patent family would have ever issued. Lumenyte Intern.\n\nCorp. v. Cable Lite Corp., 96-1011, 1996 U.S. App. LEXIS 16400 (Fed. Cir. July 9, 1996) (a false\n\naffidavit filed to revive an abandoned patent results in the unenforceability of later-filed, related\n\npatents).\n\n 40. Additionally, but for the withholding of material prior art\u2014including GPU Gems\n\n2, Cg, SANNDRA/KInNeSS, and BrookGPU, the Patent Office would not have allowed any of\n\n\n\n\n 43\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 45 of 47\n 44 of 46\n\n\n\n\nthe Asserted Patents to issue. All of the Asserted Patents were procured through a pattern of\n\ninequitable conduct as discussed herein.\n\n 41. From withholding material prior art to falsely claiming unintentional abandonment\n\nto obtain a patent, Neurala, the named inventors, and prosecution counsel have exhibited a pattern\n\nof intentional deception of the Patent Office. The remedy for the repeated and egregious instances\n\nof inequitable conduct is to render each of the Asserted Patents unenforceable.\n\n PRAYER FOR RELIEF\n\n Wherefore, Defendant respectfully requests judgment in its favor with the following relief:\n\n a) The Complaint be dismissed with prejudice.\n\n b) Judgment that Defendant has not infringed and is not infringing, either directly or\n\n indirectly, any of the claims the \u2019867 Patent, the \u2019461 Patent, and the\u2019 438 Patent,\n\n in violation of 35 U.S.C. \u00a7 271.\n\n c) Judgment that the claims of the \u2019867 Patent, the \u2019461 Patent, and the\u2019 438 Patent\n\n are invalid.\n\n d) Judgment that the \u2019867 Patent, the \u2019461 Patent and the \u2019438 Patent, including all of\n\n their claims, are unenforceable due to inequitable conduct.\n\n e) Judgment and determination that this case is exceptional under 35 U.S.C. \u00a7 285 and\n\n that Defendant is entitled to its attorneys\u2019 fees, costs, and expenses in defending\n\n this action.\n\n f) Such other relief, including other monetary and equitable relief, as this Court deems\n\n just and proper.\n\n DEMAND FOR JURY TRIAL\n\n Defendant demands a jury trial on all issues so triable.\n\n\n\n\n 44\n\f Case Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 46 of 47\n 45 of 46\n\n\n\n\nDated: November 18, 2025\n\n /s/ L. Kieran Kieckhefer\n L. Kieran Kieckhefer (pro hac vice)\n Jaysen S. Chung (pro hac vice)\n GIBSON, DUNN & CRUTCHER LLP\n One Embarcadero Center, Suite 2600\n San Francisco, CA 94111\n (415) 393-8200\n kkieckhefer@gibsondunn.com\n jschung@gibsondunn.com\n\n Brian Rosenthal\n Ahmed ElDessouki (pro hac vice)\n GIBSON, DUNN & CRUTCHER LLP\n 200 Park Ave.\n New York, NY 10166\n (212) 351-4000\n brosenthal@gibsondunn.com\n aeldessouki@gibsondunn.com\n\n Lillian J. Mao (pro hac vice)\n GIBSON DUNN & CRUTCHER LLP\n 1881 Page Mill Road\n Palto Alto, CA 94301-1211\n (650) 849-5307\n lmao@gibsondunn.com\n\n Barry K. Shelton (Texas State Bar No. 24055029)\n SHELTON COBURN LLP\n 311 RR 620 S, Suite 205\n Austin, TX 78734\n (512) 263 2165\n bshelton@sheltoncoburn.com\n\n Counsel for Defendant NVIDIA Corporation\n\n\n\n\n 45\n\fCase Case\n 7:24-cv-00221-ADA-DTG\n 7:26-cv-00318 Document\n Document\n 1-9 130\n Filed 08/17/26\n Filed 11/25/25\n Page Page\n 47 of 47\n 46 of 46\n\n\n\n\n CERTIFICATE OF SERVICE\n\n I hereby certify that all counsel of record are being served with a copy of the foregoing\n\n documents via electronic mail on November 18, 2025.\n\n /s/ L. Kieran Kieckhefer\n L. Kieran Kieckhefer\n\n\n\n\n 46\n\f","ocr_status":1,"date_upload":"2026-08-17T14:36:29.556194-07:00","document_number":"1","attachment_number":9,"pacer_doc_id":"181037209219","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 8","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294680/","id":490294680,"tags":[],"absolute_url":"/docket/74659430/1/10/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.166405-07:00","date_modified":"2026-08-21T19:48:13.092700-07:00","sha1":"4694cb55b058ceaa7c031ca3d3639cdca62d8ea6","page_count":47,"file_size":264031,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.10.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.10.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 1 of 47\n\n\n\n\n EXHIBIT\n\n 9\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 2 of 47\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\nNEURAL AI, LLC,\n Case No. 7:24-cv-00221-ADA-DTG\n Plaintiff,\n JURY TRIAL DEMANDED\n v.\n\nNVIDIA CORPORATION,\n\n Defendants.\n\n\n\n\n NON-PARTY TESLA, INC.\u2019S OBJECTIONS AND RESPONSES TO PLAINTIFF\n NEURAL AI, LLC\u2019S SUBPOENA\n Non-Party Tesla, Inc. (\u201cTesla\u201d) hereby serves the following objections and responses\n(\u201cResponses\u201d) to Plaintiff Neural AI, LLC\u2019s (\u201cNeural AI\u201d or \u201cPlaintiff\u201d) (1) Subpoena to\nProduce Documents, Information, or Objects, and (2) Subpoena for Testimony.\n\n PRELIMINARY STATEMENT\n 1. Tesla\u2019s objections and responses to the Requests are made to the best of its\ncurrent knowledge, information, belief, and understanding of the Requests. Tesla reserves the\nright to supplement or amend any responses should future investigation indicate that such\nsupplementation or amendment is necessary.\n 2. Tesla\u2019s responses to the Requests are made solely for the purpose of and in\nrelation to the above-captioned action. Each response is given subject to all appropriate\nobjections (including, but not limited to, objections concerning privilege, competency,\nrelevancy, materiality, propriety, and admissibility). All objections are reserved and may be\ninterposed at any time.\n 3. Tesla\u2019s responses include only information that is within Tesla\u2019s possession,\ncustody, or control.\n\n\n\n\n 1\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 3 of 47\n\n\n\n\n 4. Tesla incorporates by reference each and every general objection set forth\nbelow into each and every specific response. From time to time, a specific response may repeat\na general objection for emphasis or some other reason. The failure to include any general\nobjection in any specific response shall not be interpreted as a waiver of any general objection\nto that response.\n 5. Nothing contained in these Responses and Objections or provided in response\nto the Requests consists of, or should be construed as, an admission relating to the accuracy,\nrelevance, existence, or nonexistence of any alleged facts or information referenced in any\nRequest.\n GENERAL OBJECTIONS\n 1. Tesla objects to the Subpoena to the extent it seeks the disclosure of Tesla's\nhighly confidential, proprietary, or trade secret technical and business information, including\nbut not limited to internal software architectures, source code, AI/ML model designs, GPU\ncomputing infrastructure, data flow and execution flow diagrams, and engineering\nspecifications. As a non-party to this litigation, the burden on Tesla to review and produce\nsuch highly sensitive competitive information outweighs the potential relevance to the\nunderlying action.\n 2. Tesla generally objects to each Request, including the Definitions and\nInstructions, on the grounds and to the extent that it purports to impose obligations beyond\nthose imposed or authorized by the Federal Rules, the Federal Rules of Evidence, the Local\nRules, the Court\u2019s standing orders, any other applicable federal or state law, and any\nagreements between the parties. Tesla will construe and respond to the Requests in accordance\nwith the requirements of the Federal Rules and other applicable rules or laws.\n 3. Tesla generally objects to Defendant\u2019s Requests on the grounds that they are\noverbroad, oppressive, unduly burdensome, and disproportionate to the needs of the case,\nparticularly given that Tesla is not a party to this litigation. Tesla objects that the overbreadth\n\n\n\n\n 2\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 4 of 47\n\n\n\n\nof this Subpoena subjects it to undue burden and expense in both searching for and producing\nthe documents called for by this Subpoena.\n 4. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent the Request seeks documents and information that are irrelevant to\nthe claims in, or defenses to, this action, is disproportionate to the needs of the case, and/or is\nof such marginal relevance that its probative value is outweighed by the burden imposed on\nTesla in having to provide such information, including any Request that seeks information for\nany time period outside that which is relevant to the claims and defenses asserted in this action\nparticularly given that Tesla is not a party to this litigation.\n 5. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent the Request is vague, ambiguous, unreasonably cumulative, or\nduplicative, including to the extent it seeks documents or communications that are otherwise\nresponsive to other specific Requests.\n 6. Tesla generally objects to the Requests, including the Definitions and\nInstructions, on the basis that they specify an overbroad and unduly burdensome time period,\nor a time period when the asserted patents in the underlying action were not in force, and seek\ndocuments and things outside of the time period relevant to the claims and defenses asserted\nin this action.\n 7. Tesla generally objects to the Requests, including the Definitions and\nInstructions, to the extent that they are overbroad and/or unduly burdensome, including to the\nextent that they call for the production of \u201cAny,\u201d \u201cany\u201d or \u201call\u201d documents or communications\nconcerning the subject matter referenced therein.\n 8. Tesla generally objects to each Request, including the Definitions and\nInstructions, on the grounds and to the extent the Request purports to request the identification\nand disclosure of any information, communication(s), or document(s) that were prepared in\nanticipation of litigation or in connection with any internal investigation conducted at the\ndirection of counsel, constitute attorney work product, reveal privileged attorney-client\n\n 3\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 5 of 47\n\n\n\n\ncommunications, are covered under the common interest privilege and/or joint defense\nprivilege, or are otherwise protected or immune from disclosure under any applicable\nprivilege(s), law(s), or rule(s). Tesla hereby asserts all such applicable privileges and\nprotections and excludes privileged and protected information from its responses to each\nRequest. See generally Fed. R. Evid. 502.\n 9. Tesla generally objects to any Request to the extent it seeks production of\ninformation and/or documents that comprise or contain confidential information of a third\nparty to whom Tesla believes it owes a duty of confidentiality or otherwise protected from\ndisclosure by agreements between Tesla and other parties.\n 10. Tesla generally objects to any Request to the extent that it seeks to require Tesla\nto provide any information beyond what is available to Tesla at the present time after\nreasonable search of its own records and a reasonable inquiry of its present employees. For\nexample, Telsa objects to any request that seeks information that is not within Tesla\u2019s\npossession, custody or control. Tesla also objects to any Request that seeks to impose a duty\non Tesla to create materials that Tesla does not create or maintain in the ordinary course of\nbusiness.\n 11. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent that it requests information that is confidential, proprietary, or\ncompetitively sensitive.\n 12. Tesla generally objects to each Request to the extent that the information sought\nis more appropriately pursued through another discovery tool.\n 13. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent it is argumentative, lacks foundation, or incorporates allegations and\nassertions that are disputed or erroneous. In furnishing the responses herein, Tesla does not\nconcede the truth of any factual assertion or implication contained in any Request, Definition,\nor Instruction.\n\n\n\n 4\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 6 of 47\n\n\n\n\n 14. Tesla generally objects to any Request to the extent that it seeks information or\nmaterials that are publicly available, already in Defendant\u2019s possession, custody, or control,\nor are equally available to Defendant from another less burdensome source, such as parties to\nthe litigation.\n 15. Tesla generally objects to any Request to the extent that it fails to describe the\ninformation requested with reasonable particularity, is indefinite as to time and scope, seeks\ninformation that is not relevant to the claims or defenses of the parties in this case, and/or is\nnot proportional to the needs of the case.\n 16. Tesla generally objects to any Request to the extent that it requires Tesla to\ndraw legal conclusions.\n 17. Tesla generally objects to each Request, including the Definitions and\nInstructions, to the extent that it purports to impose an obligation to conduct anything beyond\na reasonable and diligent search of reasonably accessible files (including electronic files)\nwhere responsive documents reasonably would be expected to be found. Any Requests that\nseek to require Tesla to go beyond such a search are overbroad and unduly burdensome.\n 18. Tesla generally objects to the Subpoena to the extent it seeks production of\nTesla's proprietary source code, internal software, custom code, configuration files, build files,\ndeployment files, runtime logs, profiler traces, or other engineering artifacts. Such materials\nconstitute Tesla's core intellectual property and trade secrets, and their production to a non-\nparty in a dispute between Neural AI and NVIDIA is disproportionate, unduly burdensome,\nand risks competitive harm.\n 19. Tesla\u2019s willingness to provide any document or information in response to a\nRequest shall not be interpreted as an admission that such document or information exists, that\nit is relevant to a claim or defense in this action, or that it is admissible for any purpose. Tesla\ndoes not waive its right to object to the admissibility of any document or information produced\nby any party on any ground.\n\n\n\n 5\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 7 of 47\n\n\n\n\n OBJECTIONS TO THE DEFINITIONS AND INSTRUCTIONS\n 1. Tesla objects to Plaintiff's definition of \"NVIDIA GPUs\" as overbroad, unduly\nburdensome, and disproportionate to the needs of the case. The definition encompasses\nvirtually every NVIDIA GPU product ever manufactured across eight architecture generations,\nincluding hundreds of individual product SKUs spanning consumer, enterprise, data center,\nand embedded platforms. This sweeping definition, combined with the document requests,\nwould require Tesla to search for, collect, and review documents relating to any and all\nNVIDIA hardware it has ever used, regardless of relevance to the patents-in-suit. Further,\nsuch information can be obtained through other, less burdensome and more appropriate means,\nincluding from parties to the litigation.\n 2. Tesla objects to Plaintiff's definitions of \"and\" and \"or\" as overbroad, unduly\nburdensome, impermissibly vague, and not proportional to the needs of the case, to the extent\nthey purport to change the customary and usual meaning of these terms and alter the meaning\nof a phrase to impose requirements in excess of those under the Federal Rules of Civil\nProcedure and the Local Rules of this Court.\n 3. Tesla objects to Plaintiff's definitions of \"any\" and \"each\" as overbroad, unduly\nburdensome, impermissibly vague, and not proportional to the needs of the case, to the extent\nit purports to seek information that is unrelated to the present case, to the extent would capture\ndocuments of no evidentiary value and impose an undue burden on a non-party, and to the\nextent it exceeds the obligations imposed by the Federal Rules, the local rules of this Court,\nand any orders this Court entered in this case.\n 4. Tesla objects to Plaintiff's definitions of \"concerning,\" \"related to,\" \"relating\nto,\" and \"regarding\" as overbroad, vague, and disproportionate to the needs of the case. The\ndefinitions encompass over twenty verbs including \"alluding to,\" \"contradicting,\"\n\"mentioning,\" and \"memorializing,\" which would capture documents of no evidentiary value\nand impose an undue burden on a non-party.\n\n\n\n\n 6\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 8 of 47\n\n\n\n\n 5. Tesla objects to Plaintiff\u2019s definition of \u201cdocument(s)\u201d as overly broad and\nunduly burdensome, to the extent it purports to seek information that is unrelated to the present\ncase, and to the extent it exceeds the obligations imposed by the Federal Rules, the local rules\nof this Court, and any orders this Court entered in this case. Tesla recognizes that Plaintiff\u2019s\ndefinition of \u201cdocument\u201d does not include \u201cSource Code\u201d and Plaintiff has provided a separate\ndefinition for \u201cSource Code.\u201d\n 6. Tesla objects to Plaintiff's definition of \"persons\" as overbroad, unduly\nburdensome, vague, ambiguous, and not proportional to the needs of the case to the extent it\npurports to include \"formal or informal entities and organizations\" and extends to \"public and\nprivate corporations, partnerships, professional corporations, limited liability companies,\nbusiness trusts, banking institutions, associations, firms, joint ventures, commissions, bureaus,\ndepartments, and any other legal entity, including any divisions, subsidiaries, departments, and\nother units thereof\" regardless of relevance to any claim or defense in this case. Tesla further\nobjects to the inclusion of \"informal entities and organizations\" and \"any other legal entity\" as\nunbounded in scope and undefined, as these terms could be interpreted to encompass virtually\nany grouping of individuals or organizational unit without limitation. The breadth of this\ndefinition, when applied across the Requests, would impose an undue burden on Tesla\u2014a non-\nparty\u2014by requiring it to search for and identify documents involving an unlimited universe of\npersons and entities with no meaningful nexus to the claims or defenses in this action. Tesla\nwill interpret the term \"persons\" according to its customary usage in the context of a particular\nRequest and limit any response accordingly.\n 7. Tesla objects to Plaintiff's definition of \"Source Code\" as overbroad and unduly\nburdensome and as calling for information not relevant to the case. The definition encompasses\n\"all versions and revisions,\" \"all associated files,\" \"scripts, header files, makefiles,\nconfiguration files, and documentation\"\u2014effectively requiring Tesla to potentially produce\nentire software development repositories relating to any system that touches an NVIDIA GPU.\nTesla objects to this definition as it would require the disclosure of sensitive Tesla trade secrets\n\n 7\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 9 of 47\n\n\n\n\nand intellectual property in a dispute to which Tesla is not a party and without making any\nshowing of relevance or that the request is proportional to the needs of the case.\n 8. Tesla objects to Plaintiff\u2019s definition of \u201cYou,\u201d and \u201cYour\u201d and its inclusion of\n\u201cbut not limited to its predecessors, successors, parents, subsidiaries, divisions, affiliates, and\nall past or present directors, officers, partners, managers, employees, contractors, agents,\nrepresentatives, accountants, consultants, in-house and outside counsel\u201d as overbroad and\nunduly burdensome to the extent the terms are meant to include any individual(s), entit(ies),\nor any other person(s) over which Tesla exercises no control and to the extent Defendant\npurports to use the terms to impose obligations on Tesla that go beyond the requirements of\nthe Federal Rules and the Local Rules. Tesla further objects to these definitions to the extent\nthat Defendant purports to use these defined terms to seek documents that are not relevant to\nthe claims and defenses in this action, including seeking documents from Tesla\u2019s subsidiaries\nand with respect to products not at issue in this litigation. Tesla will construe the terms \u201cYou\u201d\nand \u201cYour\u201d so as to include only the following: Tesla, Inc. and its employees. Tesla further\nobjects to the definition and its inclusion of \u201cin-house and outside counsel\u201d to the extent that\nit seeks information protected by attorney-client or work product privilege.\n 9. Tesla objects to Plaintiff's Instructions to the extent they impose obligations that\ngo beyond the requirements of the Federal Rules and the Local Rules.\n 10. Tesla objects to Plaintiff's instruction regarding production format, metadata,\nnative format, and load file specifications as overbroad, unduly burdensome, and\ndisproportionate to the needs of this case.\n 11. Tesla objects to Plaintiff's instruction that Tesla identify materials not in its\npossession, custody, or control as overbroad, vague, and disproportionate to the needs of this\ncase to the extent it purports to impose interrogatory-style narrative obligations on Tesla to\nexplain the absence of documents or to speculate regarding the existence, destruction, or\nlocation of documents.\n\n\n\n 8\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 10 of 47\n\n\n\n\n 12. Tesla objects to Plaintiff's instruction regarding continuing obligations to\nsupplement to the extent it purports to impose obligations beyond those required by the Federal\nRules for non-parties.\n 13. Tesla objects to Plaintiff's Requests to the extent they fail to impose a\nmeaningful geographic scope limitation, and further objects to the extent it purports to\nencompass \"non-U.S. activity that directly supports or enables U.S. operations or usage.\" This\nlanguage is undefined, unbounded, and would require Tesla, a non-party, to make subjective\nlegal and factual determinations regarding the geographic nexus of its global computing\noperations. As written, the Requests could be construed to require Tesla to search for, collect,\nand produce documents from any facility, system, or operation worldwide that has any\narguable connection to the United States, imposing a burden and expense on a non-party that\nis grossly disproportionate to the needs of this case. Tesla will construe the Requests as limited\nto activity occurring within the United States.\n 14. Tesla objects to Plaintiff\u2019s \u201cInstructions\u201d to the extent they impose obligations\nthat go beyond the requirements of the Federal Rules and the Local Rules.\n 15. Tesla objects to Plaintiff\u2019s instruction regarding production format and load file\nspecifications as overbroad, unduly burdensome, and disproportionate to the needs of this case.\nTesla further objects to the extent that this instruction purports to impose obligations that go\nbeyond the requirements of the Federal Rules and the Local Rules, as applicable. Tesla will\nconstrue this instruction in accordance with the Federal Rules and the Local Rules.\n 16. Tesla objects to Plaintiff\u2019s instruction regarding the production of documents in\nnative format as overbroad, vague, and disproportionate to the needs of this case to the extent\nit purports to unilaterally define the criteria for native productions that differ from standard\nindustry practice.\n 17. Tesla objects to Plaintiff\u2019s instruction regarding metadata load file requirements\nas overbroad, unduly burdensome, and disproportionate to the needs of this case to the extent\nit purports to unilaterally dictate specific metadata that may not be reasonably available,\n\n 9\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 11 of 47\n\n\n\n\nautomatically generated, or maintained in the ordinary course of business without undue\nburden or expense. Tesla further objects to the extent it purports to unilaterally impose rigid\nmetadata reporting requirements, which as previously stated, may or may not be technically\nfeasible.\n 18. All Tesla\u2019s responses herein and/or related documents will be provided subject\nto any protective order entered in this case by the Court, or, if no protective order is entered,\nas HIGHLY CONFIDENTIAL ATTORNEYS\u2019 EYES ONLY.\n 19. The foregoing general reservations and objections are incorporated into each of\nthe responses and objections to the specific Request set forth below.\n\n\n\n\n 10\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 12 of 47\n\n\n\n\n OBJECTIONS AND RESPONSES TO DOCUMENT REQUESTS\nDOCUMENT REQUEST NO. 1:\n Documents sufficient to identify all software, frameworks, libraries, APIs, scripts,\nSource Code, configuration files, and custom code You use to perform computations on\nNVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 1:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla or third-party software, frameworks, libraries, APIs, scripts, Source Code, configuration\nfiles, and custom code.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case, particularly given that\nTesla is not a party to this litigation. The Request seeks identification of \"all software,\nframeworks, libraries, APIs, scripts, Source Code, configuration files, and custom code\" used\nto perform computations on NVIDIA GPUs, a scope that could potentially encompass virtually\nany software system Tesla operates. Tesla further objects to the terms \"all\" and \"Source Code\"\nas overbroad and unduly burdensome as they would require an exhaustive identification effort\nthat is disproportionate to the discovery need in a dispute to which Tesla is not a party.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\n\n 11\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 13 of 47\n\n\n\n\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla confidential trade secrets and proprietary technical\ninformation, internal software architectures, and custom-built AI/ML frameworks that\nconstitute core intellectual property and are not relevant to this case. Production of such\nmaterials, particularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to this Request as vague and ambiguous as to what\n\"computations\" are relevant to the underlying litigation, and as to the meaning and scope of\nthe terms \u201csoftware, frameworks, libraries, APIs, scripts, Source Code, configuration files, and\ncustom code.\u201d\n Tesla further objects to this Request to the extent it seeks information beyond the use\nof NVIDIA products.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to the extent this\nRequest seeks to impose a duty on Tesla to create materials or compile information that Tesla\ndoes not create or maintain in the ordinary course of business. Tesla further objects on the\ngrounds that information regarding NVIDIA's software, frameworks, and libraries is more\nreadily available from NVIDIA, the Defendant in this action.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that\nit calls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 12\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 14 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 2:\n Documents sufficient to show whether You use NVIDIA's Aerial, Clara Parabricks,\ncuBLAS, cuDNN, cuFFT, cuQuantum, cuSOLVER, cuSPARSE, Drive, DriveWorks,\nHoloscan, Isaac, Isaac Lab, Maxine, Memory Map, Merlin, Metropolis, Modulus, Monai,\nMorpheus, NeMo, PyTorch, RAPIDS, Riva, Runtime Driver, TensorFlow, TensorRT, Triton,\nVSS (Deepstream), or any other NVIDIA software as part of computations You perform using\nNVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 2:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla\u2019s use of software specifically listed in this Request or \u201cany other NVIDIA software.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The Request identifies over thirty\nspecific NVIDIA software products and then adds the catch-all phrase \"or any other NVIDIA\nsoftware,\" rendering the Request virtually unlimited in scope.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentationrepresenting\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's highly confidential and proprietary information\n\n\n 13\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 15 of 47\n\n\n\n\nregarding its internal software stack, computing infrastructure, and technology choices, which\nconstitute competitively sensitive business information. Disclosure of which specific NVIDIA\ntools Tesla does or does not use would reveal Tesla's internal technology strategy and\ncompetitive posture.\n Tesla further objects on the grounds that information regarding NVIDIA's software\nproducts and their usage by customers is more appropriately sought from NVIDIA, the\nDefendant in this action, which possesses licensing records, telemetry data, and customer\nusage information.\n Tesla objects that the catch-all phrase \"any other NVIDIA software\" renders the\nRequest vague, ambiguous, and boundless in scope, as it effectively extends the demand to\nevery NVIDIA product, tool, or library, without temporal or functional limitation.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects on the grounds\nthat the information sought is more readily obtainable from the parties to this litigation or from\nother less burdensome sources, thus disproportionate to the needs of the case.\n Tesla further objects to the extent that this Request is duplicative of Request No. 1.\nTesla objects to this Request on the grounds that it is unduly burdensome and oppressive to\nthe extent that it seeks information and documents that are equally available to the parties in\nthis litigation.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n 14\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 16 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 3:\n\n Documents sufficient to show whether You use sample Source Code provided by\nNVIDIA as part of computations You perform using NVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 3:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The term \"sample Source Code\" is\nvague and ambiguous, as it is unclear what specific sample code is being referenced, what\nuniverse of NVIDIA sample code is at issue, or how Tesla would be expected to identify\nwhether any portion of its codebase derives from, incorporates, or was inspired by NVIDIA\nsample code. Tesla objects to this Request to the extent that compliance would require Tesla\nto conduct an exhaustive audit of its entire codebase against an undefined body of NVIDIA\nsample code, a task that is extraordinarily burdensome and disproportionate for a non-party.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's proprietary\nsource code and internal software, which constitute trade secrets and core intellectual property.\nTesla further objects on the grounds that NVIDIA sample source code is publicly available\nand/or in NVIDIA's possession, custody, or control, and information regarding its distribution\nto customers is more appropriately sought from the Defendant. Tesla further objects to this\nRequest to the extent it seeks documents containing confidential, proprietary or trade secret\ninformation without making any showing of relevance or that the request is proportional to the\nneeds of the case.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case. Tesla objects to this Request to the extent it seeks\ninformation protected by the attorney-client privilege or the work-product doctrine. Tesla\n\n 15\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 17 of 47\n\n\n\n\nfurther objects to the extent this Request seeks to impose a duty on Tesla to create materials\nor compile analyses that Tesla does not create or maintain in the ordinary course of business.\nTesla further objects to the extent that this Request is duplicative of Requests Nos. 1 and 2.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 4:\n Documents sufficient to show whether and how any software You use to perform\ncomputations on NVIDIA GPUs calls, invokes, interfaces with, wraps, depends on, sits on top\nof, modifies, extends, or implements functionality provided by CUDA, cuDNN, TensorRT,\nCUDA libraries, CUDA drivers, CUDA runtime, CUDA applications or frameworks or any\nother NVIDIA software.\nRESPONSE TO DOCUMENT REQUEST NO. 4:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether and how Tesla software \u201cperform[s] computations on NVIDIA GPUs calls, invokes,\ninterfaces with, wraps, depends on, sits on top of, modifies, extends, or implements\nfunctionality provided by CUDA, cuDNN, TensorRT, CUDA libraries, CUDA drivers, CUDA\nruntime, CUDA applications or frameworks or any other NVIDIA software.\u201d\n\n 16\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 18 of 47\n\n\n\n\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to this\nRequest as being vague and ambiguous to the extent it seeks documents showing \"whether and\nhow\" Tesla's software \"calls, invokes, interfaces with, wraps, depends on, sits on top of,\nmodifies, extends, or implements\" NVIDIA functionality.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\nlitigation, rendering meaningful compliance impossible without speculation.\n Tesla objects to this Request to the extent it implicates a scope that would require Tesla\nto map and document every interaction between its proprietary software systems and\nNVIDIA's computing stack across its entire business. The enumeration of nine distinct verbs\ndescribing software interaction, combined with the catch-all phrase \"or any other NVIDIA\nsoftware,\" renders the Request virtually unlimited in scope.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary technical\ninformation, including internal software architectures, custom integrations, dependency\nstructures, and engineering designs that constitute core intellectual property.\n Tesla objects to this Request as compound and vague. The Request combines at least\nnine distinct NVIDIA technologies with at least nine distinct functional relationships, creating\na matrix of discrete inquiries posed as a single request.\n Tesla further objects to the terms \"sits on top of,\" \"wraps,\" and \"interfaces with\" as\nvague, ambiguous, and susceptible to multiple technical interpretations that render meaningful\ncompliance impossible without speculation. The catch-all phrase \"any other NVIDIA\n\n 17\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 19 of 47\n\n\n\n\nsoftware\" renders the Request boundless in scope, extending the demand to every NVIDIA\nproduct, tool, or library, without temporal or functional limitation.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as architectural mappings, dependency analyses, or software interaction\ndiagrams, that Tesla does not create or maintain in the ordinary course of business. Tesla\nobjects to this Request to the extent it seeks information protected by the attorney-client\nprivilege or the work-product doctrine. Tesla further objects to the extent that this Request is\nduplicative of Requests Nos. 1, 2, and 3.\n Tesla further objects on the grounds that information regarding NVIDIA's CUDA\nplatform, software libraries, and their interfaces is extensively documented in NVIDIA's\npublic developer resources and is more appropriately sought from the Defendant. Tesla further\nobjections to this Request to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 5:\n Documents sufficient to show the architecture, design, data flow, control flow, and\nexecution flow of any system in which You use NVIDIA GPUs to perform computations,\nincluding diagrams, technical specifications, design documents, Powerpoints, slide decks,\ninternal and external presentations, Source Code, configuration files, build files, deployment\nfiles, runtime logs, and profiler traces.\n\n\n 18\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 20 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 5:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201cthe architecture, design, data flow, control flow, and execution flow of any system in which\n[Tesla] use[s] NVIDIA GPUs to perform computations, including diagrams, technical\nspecifications, design documents, Powerpoints, slide decks, internal and external presentations,\nSource Code, configuration files, build files, deployment files, runtime logs, and profiler\ntraces.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. The Request\nseeks the \"architecture, design, data flow, control flow, and execution flow\" of every system\nin which Tesla uses NVIDIA GPUs, along with virtually every category of engineering artifact,\n\"diagrams, technical specifications, design documents, Powerpoints, slide decks, internal and\nexternal presentations, Source Code, configuration files, build files, deployment files, runtime\nlogs, and profiler traces.\" Tesla objects to this Request to the extent that this catch-all request\nwould potentially require Tesla to produce an unknowable quantum of engineering\ndocumentation and source code representing Tesla's valuable intellectual property. The\nRequest fails to identify with sufficient particularity what specific implementation or\napplication of these techniques is at issue in the underlying litigation, rendering meaningful\ncompliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's highly confidential trade secrets and proprietary\n\n 19\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 21 of 47\n\n\n\n\ntechnical information, including proprietary source code, internal software architectures, and\ncustom-built AI/ML frameworks that constitute core intellectual property. Production of such\nmaterials, particularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to this Request to the extent it seeks information beyond the use\nof NVIDIA products.\n Tesla further objects to the extent that the scope of this Request is overbroad and bears\nno reasonable relationship to the claims or defenses in the underlying action. Tesla further\nobjects to the extent this Request seeks to impose a duty on Tesla to create materials that Tesla\ndoes not create or maintain in the ordinary course of business. Tesla objects to this Request to\nthe extent it seeks information protected by the attorney-client privilege or the work-product\ndoctrine. Tesla further objects to the extent that this Request is duplicative of prior Requests,\nincluding Requests Nos. 1, 2, 3, and 4.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla\nwill not produce any documents in response to this Request.\n\n\nDOCUMENT REQUEST NO. 6:\n Documents sufficient to show whether computations You performed using NVIDIA\nGPUs involved artificial neural networks, neural-network computational layers or\ncomputations with outputs as inputs for other neurons or layers.\nRESPONSE TO DOCUMENT REQUEST NO. 6:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n\n 20\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 22 of 47\n\n\n\n\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether computations Tesla performed using NVIDIA GPUs involved artificial neural\nnetworks, neural-network computational layers or computations with outputs as inputs for\nother neurons or layers.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. Tesla further objects to this Request\nto the extent it seeks documents containing confidential, proprietary or trade secret information\nwithout making any showing of relevance or that the request is proportional to the needs of\nthe case.\n Tesla further objects to the terms \"artificial neural networks, neural-network\ncomputational layers or computations with outputs as inputs for other neurons or layers\" as\nvague and ambiguous to the extent they could encompass virtually any computation Tesla\nperforms on NVIDIA hardware.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case.\n Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to the extent this\nRequest seeks to impose a duty on Tesla to create materials that Tesla does not create or\nmaintain in the ordinary course of business. Tesla further objects on the grounds that\ninformation regarding NVIDIA GPUs' neural network capabilities is publicly available and\nmore appropriately sought from NVIDIA. Tesla further objects to the extent that this Request\nis duplicative of prior Requests, including Requests Nos. 1, 4, and 5.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\n\n\n\n 21\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 23 of 47\n\n\n\n\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 7:\n Documents sufficient to show whether You use a pointer to data stored in memory (e.g.\nmemory bank or partition), using as an input to a subsequent computational layer the pointer to\noutput data from a GPU computation, using pointers in neural network computations, swapping\nan input pointer with the pointer to data output from a GPU computation, pointer swapping,\npointer rotation, buffer swapping, ping-pong buffers, double or triple buffering, alternating\ninput/output buffers, or any other technique in which output data from one computation, layer,\niteration, time step, or cycle becomes input data for a later computation, layer, iteration, time\nstep, or cycle.\nRESPONSE TO DOCUMENT REQUEST NO. 7:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, whether\nTesla uses a pointer to data stored in memory.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to this\nRequest to the extent that it identifies a sweeping range of fundamental computing techniques,\npointer usage, pointer swapping, pointer rotation, buffer swapping, ping-pong buffers, double\nor triple buffering, alternating input/output buffers, and then adds the catch-all phrase \"or any\n\n\n 22\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 24 of 47\n\n\n\n\nother technique in which output data from one computation, layer, iteration, time step, or cycle\nbecomes input data for a later computation, layer, iteration, time step, or cycle.\"\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary technical information, including internal memory management\nstrategies, GPU optimization techniques, custom buffer management implementations, and\nlow-level engineering designs that constitute core intellectual property. Tesla further objects to\nthis Request to the extent it seeks documents containing confidential, proprietary or trade secret\ninformation without making any showing of relevance or that the request is proportional to the\nneeds of the case.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as memory architecture diagrams, pointer flow analyses, or buffer management\ndocumentation, that Tesla does not create or maintain in the ordinary course of business.\n Tesla further objects on the grounds that the information sought is more readily\nobtainable from the parties to this litigation or from other less burdensome sources, thus\ndisproportionate to the needs of the case.\n Tesla objects to this Request to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla objects to this Request to the extent it seeks\ninformation protected by the attorney-client privilege or the work-product doctrine.\n Tesla further objects on the grounds to the extent that the techniques described, pointer\nusage, buffer swapping, double buffering, ping-pong buffers, are well-known, standard\ncomputing techniques whose operation is extensively and publicly documented in computer\nscience literature, NVIDIA's developer guides, and CUDA programming documentation, and\nare more appropriately explored through NVIDIA's own materials.\n Tesla objects to this Request to the extent that this catch-all request would potentially\nrequire Tesla to produce an unknowable quantum of engineering documentation representing\nTesla's valuable intellectual property. The Request fails to identify with sufficient particularity\nwhat specific implementation or application of these techniques is at issue in the underlying\n\n 23\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 25 of 47\n\n\n\n\nlitigation, rendering meaningful compliance impossible without speculation. Tesla objects to\nthis Request to the extent it uses generic technical terminology to describe fundamental and\nubiquitous computing operations that are not unique to any particular proprietary technology,\npatented method, or party to this litigation. As drafted, this Request is overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla further objects to the extent that this Request is duplicative of prior Requests,\nincluding Requests Nos. 1, 4, 5, and 6.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 8:\n Documents sufficient to show whether You store input data, output data, intermediate\nresults, tensors, activations, weights, parameters, internal variables, GPU programs, kernels,\ntextures, shaders, or other GPU-computation-related data in separate, partitioned, logical,\nphysical, first/second, input/output, texture, shader, shared, global, device, host, pinned, GPU\nRAM, GPU cache(s), or unified memory regions (shared by CPU and GPU) when performing\ncomputations using NVIDIA GPUs.\nRESPONSE TO DOCUMENT REQUEST NO. 8:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n\n 24\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 26 of 47\n\n\n\n\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nwhether Tesla stores certain GPU-computation-related data in certain memory regions.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case to the extent that\nit would potentially require Tesla to produce an unknowable quantum of engineering\ndocumentation representing Tesla's valuable intellectual property. The Request fails to identify\nwith sufficient particularity what specific implementation or application of these techniques is\nat issue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary intellectual property, including GPU memory allocation\nstrategies, custom memory optimization techniques, tensor management implementations, and\ninternal computing architectures. Tesla further objects to the catch-all phrase \"or other GPU-\ncomputation-related data\" as vague, ambiguous, and unbounded.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any\nparticular proprietary technology, patented method, or party to this litigation. The Request\nfails to identify with sufficient particularity what specific implementation or application of\nthese techniques is at issue in the underlying litigation, rendering meaningful compliance\nimpossible without speculation. As drafted, this Request is therefore facially overbroad,\nunduly burdensome, and disproportionate to the needs of this case, particularly as directed to\na non-party.\n\n\n\n 25\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 27 of 47\n\n\n\n\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as memory allocation maps, data storage analyses, or memory partition\ndocumentation, that Tesla does not create or maintain in the ordinary course of business. Tesla\nobjects to this Request to the extent it seeks information protected by the attorney-client\nprivilege or the work-product doctrine.\n Tesla further objects on the grounds that NVIDIA's own documentation, developer\nguides, CUDA programming manuals, and hardware specifications describe in detail the\nmemory architecture, memory types, and memory management capabilities of NVIDIA GPUs\nand are publicly available and more appropriately sought from the Defendant. Tesla further\nobjects to the extent that this Request is duplicative of prior Requests, including Requests Nos.\n1, 4, 5, and 7.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 9:\n Documents sufficient to show how input data is received, acquired, stored, transferred,\ncopied, streamed, prefetched, staged, queued, or loaded from CPU memory, host memory,\nsystem memory, storage, sensors, cameras, or other input sources to NVIDIA GPU memory\nincluding GPU RAM (e.g. GPU HBM, GDDR) and/or GPU cache(s) before, during, or in\nparallel with computations You perform using NVIDIA GPUs.\n\n\n\n\n 26\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 28 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 9:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nhow Tesla handles input data related to GPU computations.\n Tesla objects to this Request to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to the\nRequest to the extent that it seeks documents showing \"how\" input data is \"received, acquired,\nstored, transferred, copied, streamed, prefetched, staged, queued, or loaded\" from an extensive\nand open-ended list of sources, \"CPU memory, host memory, system memory, storage, sensors,\ncameras, or other input sources, \" to NVIDIA GPU memory, \"before, during, or in parallel with\"\ncomputations. This scope encompasses virtually every data pipeline and data ingestion pathway\nacross Tesla's entire computing infrastructure that touches any NVIDIA hardware.\n Tesla objects to this Request to the extent it seeks the disclosure of Tesla's confidential\ntrade secrets and proprietary technical information, including internal data pipeline\narchitectures, sensor fusion systems, data preprocessing workflows, custom data loading and\nprefetching implementations, and GPU optimization strategies that constitute core intellectual\nproperty. Tesla specifically objects to the reference to \"sensors, cameras, or other input sources\"\nas a transparent attempt to compel disclosure of Tesla's most valuable and competitively\nsensitive proprietary intellectual property. Tesla further objects to this Request to the extent it\nseeks documents containing confidential, proprietary or trade secret information without\nmaking any showing of relevance or that the request is proportional to the needs of the case.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to identify\n\n 27\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 29 of 47\n\n\n\n\nwith sufficient particularity what specific implementation or application of these techniques is\nat issue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation. As drafted, this Request is therefore overbroad, unduly burdensome, and\ndisproportionate to the needs of this case, particularly as directed to a non-party.\n Tesla further objects to the catch-all phrase \"or other input sources\" as vague, ambiguous,\nand unbounded. Tesla further objects to the extent this Request seeks to impose a duty on Tesla\nto create materials, such as data flow diagrams, pipeline architecture documents, or data transfer\nanalyses, that Tesla does not create or maintain in the ordinary course of business. Tesla objects\nto this Request to the extent it seeks information protected by the attorney-client privilege or\nthe work-product doctrine.\n Tesla further objects on the grounds that the data transfer and memory management\nmechanisms described\u2014CPU-to-GPU transfers, memory staging, prefetching, streaming\u2014are\nstandard computing operations whose architecture is publicly documented by NVIDIA in its\nCUDA programming guides and developer documentation and are more appropriately sought\nfrom the Defendant. Tesla further objects to the extent that this Request is duplicative of prior\nRequests, including Requests Nos. 4, 5, 7, and 8.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 28\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 30 of 47\n\n\n\n\nDOCUMENT REQUEST NO. 10:\n Documents sufficient to show how output data from a GPU computation(s),\nintermediate results of GPU computations, tensors, buffers, activations, variables, or other\ncomputation results are stored, transferred, copied, streamed, written back, returned,\naccumulated, reused, or made available including asynchronously from NVIDIA GPU\nmemory to CPU memory, host memory, system memory, storage, display, network, or\nanother memory location before, during, or in parallel with computations You perform using\nNVIDIA GPUs and also including in the opposite direction, copying data from CPU or host or\nother memory to a queue for GPU computation while other GPU computations are occurring.\nRESPONSE TO DOCUMENT REQUEST NO. 10:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nhow Tesla handles output data related to GPU computations.\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. The Request\nseeks documents showing how output data, intermediate results, tensors, buffers, activations,\nvariables, and \"other computation results\" are \"stored, transferred, copied, streamed, written\nback, returned, accumulated, reused, or made available (including asynchronously)\" of\nunbounded scope, including different memory domains such as \"GPU memory to CPU memory,\nhost memory, system memory, storage, display, network, or another memory location\" and then\nextends the Request to encompass data movement \"in the opposite direction\" as well. Tesla\nobjects to the catch-all phrases \"or other computation results\" and \"or another memory location\"\nas vague, ambiguous, and unbounded, rendering the Request limitless in scope. Tesla objects\nto this Request to the extent that this catch-all request would potentially require Tesla to produce\nan unknowable quantum of engineering documentation representing Tesla's valuable\n\n 29\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 31 of 47\n\n\n\n\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\nproperty, including GPU-CPU data transfer strategies, asynchronous computing pipelines,\ncustom memory management implementations, and real-time inference data flow architectures.\nProduction of such materials in a dispute between Neural AI and NVIDIA would risk\ncompetitive harm to Tesla, a non-party.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any\nparticular proprietary technology, patented method, or party to this litigation. The Request\nfails to identify with sufficient particularity what specific implementation or application of\nthese techniques is at issue in the underlying litigation, rendering meaningful compliance\nimpossible without speculation. As drafted, this Request is therefore facially overbroad,\nunduly burdensome, and disproportionate to the needs of this case, particularly as directed to\na non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as bidirectional data flow diagrams, memory transfer analyses, or\nasynchronous pipeline documentation, that Tesla does not create or maintain in the ordinary\ncourse of business. Tesla objects to this Request to the extent it seeks information protected\nby the attorney-client privilege or the work-product doctrine.\n Tesla further objects on the grounds that the data transfer mechanisms described, GPU-\nto-CPU transfers, asynchronous memory operations, streaming, write-back, are standard\ncomputing operations extensively documented in NVIDIA's public CUDA programming\n\n 30\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 32 of 47\n\n\n\n\nguides and developer resources and are more appropriately sought from the Defendant. Tesla\nfurther objects to the extent that this Request is substantially duplicative of prior Requests,\nincluding Requests Nos. 4, 5, 7, 8, and 9.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 11:\n Documents sufficient to show how computations You perform using NVIDIA GPUs\nare scheduled, ordered, controlled, queued, synchronized, parallelized, launched, interrupted,\nresumed, or executed, including through kernels, CUDA streams, CUDA graphs, events,\nthreads, controllers, schedulers, compilers, runtimes, inference engines, run lists, run engines,\nor custom software.\nRESPONSE TO DOCUMENT REQUEST NO. 11:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201chow computations [Tesla] perform[s] using NVIDIA GPUs are scheduled, ordered,\ncontrolled, queued, synchronized, parallelized, launched, interrupted, resumed, or executed,\nincluding through kernels, CUDA streams, CUDA graphs, events, threads, controllers,\nschedulers, compilers, runtimes, inference engines, run lists, run engines, or custom software.\u201d\n\n\n\n 31\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 33 of 47\n\n\n\n\n Tesla objects to this Request on the grounds that it is facially overbroad, unduly\nburdensome, unnecessary, oppressive, and not proportional to the needs of this case. The\nRequest seeks documents showing how Tesla's GPU computations are \"scheduled, ordered,\ncontrolled, queued, synchronized, parallelized, launched, interrupted, resumed, or executed\"\nthrough an exhaustive and open-ended list of mechanisms, \"kernels, CUDA streams, CUDA\ngraphs, events, threads, controllers, schedulers, compilers, runtimes, inference engines, run lists,\nrun engines, or custom software.\" Tesla objects to the inclusion of \"custom software\" as a catch-\nall that would require disclosure of Tesla's proprietary computing systems in their entirety. Tesla\nobjects to this Request to the extent that this catch-all request would potentially require Tesla\nto produce an unknowable quantum of engineering documentation representing Tesla's valuable\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\nproperty, including custom GPU scheduling systems, proprietary inference engine designs,\ncompute orchestration strategies, compiler optimizations, and proprietary runtime\nenvironments that constitute core competitive advantages. Production of such materials in a\ndispute between Neural AI and NVIDIA would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as scheduling architecture analyses, execution flow documentation, or\norchestration diagrams, that Tesla does not create or maintain in the ordinary course of business.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to\n\n 32\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 34 of 47\n\n\n\n\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Request is therefore facially overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla objects to this Request to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects on the grounds that\ninformation regarding CUDA streams, CUDA graphs, kernel launching, GPU scheduling, and\ncompute execution is extensively documented in NVIDIA's public developer documentation,\nCUDA programming guides, and technical specifications and is more appropriately sought\nfrom the Defendant. Tesla further objects to the extent that this Request is substantially\nduplicative of prior Requests, including Requests Nos. 1, 4, 5, 7, and 8.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\nDOCUMENT REQUEST NO. 12:\n Documents sufficient to show whether and how user inputs, user commands,\nconfiguration changes, parameter changes, model changes, computational-element changes,\ninput changes, interruptions, or display/output changes affect computations You perform\nusing NVIDIA GPUs and/or queue them for GPU computation.\n\n\n\n\n 33\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 35 of 47\n\n\n\n\nRESPONSE TO DOCUMENT REQUEST NO. 12:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions and Instructions, as though fully set forth in this Response.\n Tesla objects to this Request as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\n\u201cwhether and how user inputs, user commands, configuration changes, parameter changes,\nmodel changes, computational-element changes, input changes, interruptions, or\ndisplay/output changes affect computations [Tesla] perform[s] using NVIDIA GPUs and/or\nqueue them for GPU computation.\u201d\n Tesla objects to this Request on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. The Request seeks\ndocuments showing \"whether and how\" an open-ended and effectively limitless list of changes,\n\"user inputs, user commands, configuration changes, parameter changes, model changes,\ncomputational-element changes, input changes, interruptions, or display/output changes,\" affect\nTesla's GPU computations or queue them for processing. This scope is virtually unlimited. Tesla\nobjects to this Request to the extent that this catch-all request would potentially require Tesla\nto produce an unknowable quantum of engineering documentation representing Tesla's valuable\nintellectual property. The Request fails to identify with sufficient particularity what specific\nimplementation or application of these techniques is at issue in the underlying litigation,\nrendering meaningful compliance impossible without speculation.\n Tesla objects to the terms \"computational-element changes,\" \"input changes,\" and\n\"display/output changes\" as vague, ambiguous, and undefined. It is unclear what constitutes a\n\"computational-element change\" or how broadly \"input changes\" is intended to sweep.\n Tesla further objects to this Request to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Request to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary technical\n\n 34\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 36 of 47\n\n\n\n\ninformation, including user interface architectures, real-time computing systems, dynamic\nGPU scheduling implementations, model update pipelines, interactive inference systems, and\nhuman-machine interface designs. Production of such materials in a dispute between Neural\nAI and NVIDIA would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Request seeks to impose a duty on Tesla to create\nmaterials, such as user interaction flow analyses, change-impact documentation, or input-to-\ncomputation mapping diagrams, that Tesla does not create or maintain in the ordinary course\nof business. Tesla objects to this Request to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine.\n Tesla objects to this Request to the extent it uses generic technical terminology to\ndescribe fundamental and ubiquitous computing operations that are not unique to any particular\nproprietary technology, patented method, or party to this litigation. The Request fails to\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Request is therefore facially overbroad, unduly\nburdensome, and disproportionate to the needs of this case, particularly as directed to a non-\nparty.\n Tesla further objects on the grounds that the computing concepts described in this\nRequest, user input handling, parameter configuration, compute queuing, are standard GPU\ncomputing operations documented in NVIDIA's public developer resources and are more\nappropriately sought from the Defendant. Tesla further objects to the extent that this Request\nis substantially duplicative of prior Requests, including Requests Nos. 1, 4, 5, 7, and 11.\n Tesla further objects to this Request as Neural AI has failed to comply with the\nrequirements of Fed. R. Civ. P. 45(d). Tesla further objects to this Request to the extent that it\ncalls for documents not within Tesla\u2019s possession, custody or control or kept in the ordinary\ncourse of business.\n\n\n\n 35\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 37 of 47\n\n\n\n\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Request and the burden it imposes on\nTesla.\n\n\n\n\n 36\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 38 of 47\n\n\n\n\n OBJECTIONS AND RESPONSES TO DEPOSITION TOPICS\nDEPOSITON TOPIC NO. 1:\n The NVIDIA software and libraries You use to perform computations, including but\nnot limited to NVIDIA's Aerial, Clara Parabricks, cuBLAS, cuDNN, cuFFT, cuQuantum,\ncuSOLVER, cuSPARSE, Drive, DriveWorks, Holoscan, Isaac, Isaac Lab, Maxine, Memory\nMap, Merlin, Metropolis, Modulus, Monai, Morpheus, NeMo, PyTorch, RAPIDS, Riva,\nRuntime Driver, TensorFlow, TensorRT, Triton, VSS (Deepstream).\nRESPONSE TO DEPOSITION TOPIC NO. 1:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including\nTesla\u2019s use of a long list of software and libraries. Tesla further objects to this Topic as vague\nand ambiguous as to what \"computations\" are relevant to the underlying litigation.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The Topic identifies almost thirty\nNVIDIA software and libraries, placing undue burden on Tesla, a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's highly confidential and proprietary information\nregarding its internal software stack, computing infrastructure, and technology choices, which\nconstitute competitively sensitive business information. Disclosure of which specific NVIDIA\ntools Tesla does or does not use would reveal Tesla's internal technology strategy and\ncompetitive posture.\n\n\n\n 37\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 39 of 47\n\n\n\n\n Tesla further objects to this Topic on the grounds that information regarding NVIDIA's\nsoftware products and their usage by customers is more appropriately sought from NVIDIA,\nthe Defendant in this action, which possesses licensing records, telemetry data, and customer\nusage information.\n Tesla further objects to this Topic on the grounds that the information sought is more\nreadily obtainable from the parties to this litigation or from other less burdensome sources,\nthus disproportionate to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITION TOPCI NO. 2:\n The NVIDIA sample Source Code You use, in whole or in part, to conduct\ncomputations.\nRESPONSE TO DEPOSITION TOPIC NO. 2:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action. Tesla further\nobjects to this Topic as vague and ambiguous as to what \"computations\" are relevant to the\nunderlying litigation.\n\n 38\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 40 of 47\n\n\n\n\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, and not proportional to the needs of this case. The term \"sample Source Code\" is\nvague and ambiguous, as it is unclear what specific sample code is being referenced, what\nuniverse of NVIDIA sample code is at issue, or how Tesla would be expected to identify\nwhether any portion of its codebase derives from, incorporates, or was inspired by NVIDIA\nsample code. Tesla objects to this Topic to the extent that compliance would require Tesla to\nconduct an exhaustive audit of its entire codebase against an undefined body of NVIDIA\nsample code, a task that is extraordinarily burdensome and disproportionate for a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's proprietary source code and internal software, which\nconstitute trade secrets and core intellectual property. Tesla further objects on the grounds that\nNVIDIA sample source code is publicly available and/or in NVIDIA's possession, custody, or\ncontrol, and information regarding its distribution to customers is more appropriately sought\nfrom the Defendant.\n Tesla further objects to this Topic on the grounds that the information sought is more\nreadily obtainable from the parties to this litigation or from other less burdensome sources,\nthus disproportionate to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n 39\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 41 of 47\n\n\n\n\nDEPOSITION TOPIC NO. 3:\n Your customizations and/or data inputs to NVIDIA software that alter the way in which\nNVIDIA software performs computations and/or a description of the data input to NVIDIA\nsoftware on which computations are run.\nRESPONSE TO DEPOSITON TOPCI NO. 3:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including how\nTesla handles input data related to GPU computations. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic to the extent that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case. Tesla objects to the\nTopic to the extent that it seeks information showing Tesla\u2019s \u201ccustomizations\u201d or \u201calter the way\nin which NVIDIA software performs computations\u201d without providing specificity as to what\nscope of customizations or alterations are relevant to the underlying action. The Topic fails to\nidentify with sufficient particularity what specific implementation or application of these\ntechniques is at issue in the underlying litigation, rendering meaningful compliance impossible\nwithout speculation. As drafted, this Topic would extend to the entirety of Tesla's computing\noperations and is therefore overbroad, unduly burdensome, and disproportionate to the needs of\nthis case, particularly as directed to a non-party.\n\n\n\n\n 40\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 42 of 47\n\n\n\n\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case.\n Tesla objects to this Topic to the extent it seeks information protected by the attorney-\nclient privilege or the work-product doctrine. Tesla further objects to this Request as Neural\nAI has failed to comply with the requirements of Fed. R. Civ. P. 45(d). Tesla further objects\nto the extent this Topic seeks to impose a duty on Tesla to create materials or compile analyses\nthat Tesla does not create or maintain in the ordinary course of business.\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITON TOPIC NO. 4:\n Identification of Your software that uses NVIDIA GPUs to perform computations.\nRESPONSE TO DEPOSITION TOPIC NO. 4:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic t as irrelevant to the extent it seeks information that lacks\nany connection to any specific claims or defenses at issue in the underlying action, including\nTesla or third-party software that uses NVIDA GPUs. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and not proportional to the needs of this case, particularly given that\nTesla is not a party to this litigation. The Topic seeks identification of software used to perform\n\n 41\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 43 of 47\n\n\n\n\ncomputations on NVIDIA GPUs without any specificity, a scope that could potentially\nencompass virtually any software system Tesla operates. The Topic fails to identify with\nsufficient particularity what specific implementation or application of these techniques is at\nissue in the underlying litigation, rendering meaningful compliance impossible without\nspeculation. As drafted, this Topic would extend to the entirety of Tesla's computing operations\nand is therefore overbroad, unduly burdensome, and disproportionate to the needs of this case,\nparticularly as directed to a non-party.\n Tesla objects to this Topic to the extent it seeks Tesla confidential, proprietary or trade\nsecret information without making any showing of relevance or that the request is proportional\nto the needs of the case. Tesla objects to this Topic to the extent it seeks the disclosure of Tesla\nconfidential trade secrets and proprietary technical information, including proprietary source\ncode, internal software architectures, and custom-built AI/ML frameworks that constitute core\nintellectual property and are not relevant to this case. Disclosure of such information,\nparticularly on a third-party, would risk competitive harm to Tesla, a non-party.\n Tesla further objects to the extent this Topic seeks to impose a duty on Tesla to create\nmaterials or compile analyses that Tesla does not create or maintain in the ordinary course of\nbusiness. Tesla objects to this Topic to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to this Request as\nNeural AI has failed to comply with the requirements of Fed. R. Civ. P. 45(d).\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\nDEPOSITION TOPIC NO. 5:\n Using Your software, the ways in which output data from a GPU computation(s),\nincluding intermediate results of GPU computations are stored, referenced by a pointer,\ntransferred, copied, streamed, written back, returned, accumulated, reused, or made available\n\n 42\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 44 of 47\n\n\n\n\nincluding asynchronously from NVIDIA GPU memory to CPU memory, host memory, system\nmemory, storage, display, network, or another memory location before, during, or in parallel\nwith computations performed using NVIDIA GPUs\nRESPONSE TO DEPOSITON TOPIC NO. 5:\n\n Tesla restates and incorporates its Preliminary Statement, General Objections,\nObjections to Definitions, and Objections to Instructions, as though fully set forth in this\nResponse.\n Tesla objects to this Topic as irrelevant to the extent it seeks information that lacks any\nconnection to any specific claims or defenses at issue in the underlying action, including how\nTesla handles output data related to GPU computations. Tesla further objects to this Topic as\nvague and ambiguous as to what \"computations\" are relevant to the underlying litigation. Tesla\nfurther objects to this Topic to the extent it seeks information beyond the use of NVIDIA\nproducts.\n Tesla objects to this Topic on the grounds that it is overbroad, unduly burdensome,\nunnecessary, oppressive, and grossly disproportionate to the needs of this case. Tesla objects\nto this Topic to the extent it uses generic technical terminology to describe fundamental and\nubiquitous computing operations that are not unique to any particular proprietary technology,\npatented method, or party to this litigation. The Topic fails to identify with sufficient\nparticularity what specific implementation or application of these techniques is at issue in the\nunderlying litigation, rendering meaningful compliance impossible without speculation. As\ndrafted, this Topic would extend to the entirety of Tesla's computing operations and is\ntherefore overbroad, unduly burdensome, and disproportionate to the needs of this case,\nparticularly as directed to a non-party.\n Tesla further objects to this Topic to the extent it seeks documents containing\nconfidential, proprietary or trade secret information without making any showing of relevance\nor that the request is proportional to the needs of the case. Tesla objects to this Topic to the\nextent it seeks the disclosure of Tesla's confidential trade secrets and proprietary intellectual\n\n 43\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 45 of 47\n\n\n\n\nproperty, including GPU-CPU data transfer strategies. Disclosure of such confidential\nmaterials in a dispute between Neural AI and NVIDIA would risk competitive harm to Tesla,\na non-party.\n Tesla further objects to the extent this Topic seeks to impose a duty on Tesla to create\nmaterials or compile analyses that Tesla does not create or maintain in the ordinary course of\nbusiness. Tesla objects to this Topic to the extent it seeks information protected by the\nattorney-client privilege or the work-product doctrine. Tesla further objects to this Request as\nNeural AI has failed to comply with the requirements of Fed. R. Civ. P. 45(d).\n Subject to and without waiving the foregoing general and specific objections, Tesla is\nwilling to meet and confer regarding the scope of this Topic and the burden it imposes on\nTesla.\n\n\n\n\n 44\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 46 of 47\n\n\n\nDated: July 21, 2026 /s/ Jun Zheng _____________\n\n Jun Zheng\n TX Bar No. 24102681\n zhengjun@tesla.com\n Tesla, Inc.\n 1 Tesla Rd\n Austin, TX 78725\n (512) 417-3528\n\n Gina H. Cremona\n CA Bar No. 305392\n gcremona@tesla.com\n Tesla, Inc.\n 1501 Page Mill Rd.\n Palo Alto, CA 94304\n (650) 647-0015\n\n Ashraf Fawzy\n DC Bar No. 989132\n afawzy@tesla.com\n Tesla, Inc.\n 800 Connecticut Ave. NW\n Washington, DC 20006\n (202) 905-9221\n\n\n\n\n Attorneys for TESLA, INC.\n\n\n\n\n 45\n\f Case 7:26-cv-00318 Document 1-10 Filed 08/17/26 Page 47 of 47\n\n\n\n\n CERTIFICATE OF SERVICE\n\n I hereby certify that a true copy of the above document was served upon Plaintiff Neural\n\nAI\u2019s counsel of record via electronic mail on July 21, 2026.\n\n\n /s/ Jun Zheng\n Jun Zheng\n\n\n\n\n 46\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:18.906153-07:00","document_number":"1","attachment_number":10,"pacer_doc_id":"181037209220","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 9","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294681/","id":490294681,"tags":[],"absolute_url":"/docket/74659430/1/11/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.182389-07:00","date_modified":"2026-08-21T19:37:16.478208-07:00","sha1":"4022b38c5eb39c760ff456e0d98c16808a64b7fb","page_count":10,"file_size":213794,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.11.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.11.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 1 of 10\n\n\n\n\n EXHIBIT\n\n 10\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 2 of 10\n\n\n Friday, August 7, 2026 at 1:58:08 PM Paci\ufb01c Daylight Time\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\nDate: Tuesday, August 4, 2026 at 11:28:50 AM Paci\ufb01c Daylight Time\nFrom: Tanner Laiche\nTo: Rocco Magni, Gina Cremona\nCC: Jun Zheng, Emily Portuguese, Tamar Lusztig, Brian Melton, Max Tribble, Samuel Drezdzon, Richard\n Wojtczak, Rachel Hanna, Ashraf Fawzy\nAttachments: Neural AI, Third-Party Questions.docx, Neural AI, Draft Third-Party Declaration.docx\n\n\nCounsel,\n\nI am following up on Rocco\u2019s email.\n\nDespite the parties\u2019 prior meet-and-confers, document discovery in the underlying action\ncloses on August 11. Unless the parties can promptly reach a resolution, that deadline\nleaves Neural AI no practical alternative but to move to compel by the end of this week or,\nat the latest, August 10, to preserve its rights.\n\nTo reduce burden and potentially avoid motion practice, I am attaching a short set of\nquestions intended to guide your investigation and help identify the responsive\ninformation, and also recirculating the draft declaration we previously shared, and that\nTesla may revise to ensure its accuracy.\n\nIf Tesla commits to provide an executed declaration, Neural AI is willing to consider\naccepting the declaration in lieu of further document production and/or deposition\ntestimony, subject to resolving any material gaps. Otherwise, the discovery deadline will\nforce Neural AI to move to compel by or before August 10 to preserve its rights. Even if a\nmotion becomes necessary, we remain open to resolving the issues promptly and mooting\nor withdrawing the motion through compliance.\n\nWe are available this week to further meet and confer as necessary.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\nFrom: Rocco Magni \nDate: Sunday, August 2, 2026 at 6:57 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\n\n 1 of 9\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 3 of 10\n\n\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nOur discovery deadline is approaching soon. Please let us know when you can confer again\nthis upcoming week. Thanks.\n\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Rocco Magni \nDate: Tuesday, July 28, 2026 at 6:52 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nThanks for speaking today. Attached is a draft of the declaration I referred to on our call.\n\nBest,\n\nRocco\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\n\n 2 of 9\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 4 of 10\n\n\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Gina Cremona \nDate: Thursday, July 23, 2026 at 9:45 AM\nTo: Rocco Magni \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\n\nHi Rocco, we are not available today.\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 22, 2026 12:01 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche\n; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nWould tomorrow work?\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received this message in\n\n 3 of 9\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 5 of 10\n\n\nerror, please notify the sender and delete it immediately.\n\n\n\n On Jul 22, 2026, at 2:54 PM, Gina Cremona wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n We are not available on Friday, but can meet on Tuesday, July 28 between 8-10 am PT.\n\n Thanks,\n Gina\n\n\n From: Rocco Magni \n Sent: Wednesday, July 22, 2026 6:54 AM\n To: Jun Zheng ; Gina Cremona\n \n Cc: Tanner Laiche ; Emily Portuguese\n ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n ; Ashraf Fawzy ; Jun\n Zheng \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n Jun,\n\n Please provide times to meet and confer on Friday 7/24. Thanks.\n\n --\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you received this\n message in error, please notify the sender and delete it immediately.\n\n 4 of 9\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 6 of 10\n\n\nFrom: Jun Zheng \nDate: Tuesday, July 21, 2026 at 7:37 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nEXTERNAL Email\nRocco,\n\nWe understand from the email exchange below that the noticed date for\ntestimony subpoena is now Sept. 14, 2026. In the meantime, attached please \ufb01nd\nTesla's objections and responses to Neural AI's subpoena.\n\nThanks!\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n\nFrom: Gina Cremona \nSent: Monday, July 20, 2026 1:07 PM\nTo: Rocco Magni \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\n\n 5 of 9\n\f Case 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 7 of 10\n\n\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nHi Rocco,\n\nTesla cannot provide an agreed date for deposition until we have reviewed and\nresponded to the subpoenas. However, we can agree to Sept. 14 as the noticed date\nfor testimony subpoena for now, subject to modi\ufb01cation once we have responded to\nthe document subpoena.\n\nRegards,\nGina\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 15, 2026 12:58 PM\nTo: Gina Cremona \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nGina,\n\nWe\u2019ll agree to pull down the July 28 date once we have an agreed replacement date.\nLet us know what date works for you and we\u2019ll withdraw the current notice.\nThanks.\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received\nthis message in error, please notify the sender and delete it immediately.\n\n\n\n\n 6 of 9\n\fCase 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 8 of 10\n\n\n On Jul 15, 2026, at 3:54 PM, Gina Cremona \n wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n Understood. To con\ufb01rm, the deposition date of July 28 is ok calendar, and\n we will work on agreeing to a new date.\n\n Regards,\n Gina\n\n\n From: Rocco Magni \n Sent: Tuesday, July 14, 2026 5:28 PM\n To: Gina Cremona ; Tanner Laiche\n ; Emily Portuguese\n \n Cc: Tamar Lusztig ; Brian\n Melton ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Ms. Cremona,\n\n Discovery has not been extended; only depositions. Written\n discovery will still close on August 11.\n\n We can work with you on a deposition date between now and\n September 14. But we do not have \ufb02exibility on the timing of your\n RFP responses and document production beyond the 1 week\n extension we noted below.\n\n Best,\n\n Rocco\n\n --\n\n\n 7 of 9\n\fCase 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 9 of 10\n\n\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you\n received this message in error, please notify the sender and delete it\n immediately.\n From: Gina Cremona \n Date: Tuesday, July 14, 2026 at 8:19 PM\n To: Tanner Laiche ; Emily\n Portuguese \n Cc: Rocco Magni ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D.\n Tex.) - Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n It has come to our attention that the discovery deadline in this case has\n been extended. Due to summer vacations and people being out of the\n okice, we renew our request for an additional 2-week extension such that\n our deadlines will be Aug. 4 and Aug. 11.\n\n Regards,\n Gina\n\n\n From: Tanner Laiche \n Sent: Monday, July 6, 2026 5:20 PM\n To: Gina Cremona ; Emily Portuguese\n \n Cc: Rocco Magni ; Tamar\n Lusztig ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n\n 8 of 9\n\fCase 7:26-cv-00318 Document 1-11 Filed 08/17/26 Page 10 of 10\n\n\n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Counsel,\n\n Thanks for reaching out. Given the upcoming close of fact\n discovery, Neural AI is not able to agree to a three-week\n extension. That said, we can agree to a one-week extension for\n Tesla\u2019s written objections/responses to the subpoena(s). A copy\n of the Protective Order is attached.\n\n Regards,\n\n Tanner Laiche\n Susman Godfrey LLP\n 206.505.3816 | tlaiche@susmangodfrey.com\n 401 Union Street | Suite 3000 | Seattle, WA 98101\n HOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\n From: Gina Cremona \n Date: Thursday, July 2, 2026 at 7:21 PM\n To: Emily Portuguese \n Subject: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n We are in receipt of your Subpoena to Produce Documents and Subpoena\n for Testimony. Due to the holiday weekend and vacation schedules, we\n request a three-week extension to respond such that our deadlines will be\n Aug. 4 and Aug. 11.\n\n Additionally, please provide us with a copy of the protective order.\n\n Regards,\n Gina\n\n Gina H. Cremona\n Senior Counsel, IP Litigation\n 1501 Page Mill Rd., Palo Alto, CA 94304\n E. gcremona@tesla.com T. 650.647.0015\n\n\n \n\n\n\n\n 9 of 9\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:21.277821-07:00","document_number":"1","attachment_number":11,"pacer_doc_id":"181037209221","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 10","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294682/","id":490294682,"tags":[],"absolute_url":"/docket/74659430/1/12/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.209190-07:00","date_modified":"2026-08-23T05:22:59.254652-07:00","sha1":"5343e4c5121876a6ba58cb7ef512fd4d0f3ac73f","page_count":3,"file_size":89069,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.12.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.12.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-12 Filed 08/17/26 Page 1 of 3\n\n\n\n\n EXHIBIT\n\n 11\n\f Case 7:26-cv-00318 Document 1-12 Filed 08/17/26 Page 2 of 3\n\n\n\n\nGeneral Deployment\n \u2022 Do you have architecture diagrams for your hardware and software systems that use\n NVIDIA GPUs?\n \u2022 Do you use PyTorch or TensorRT?\n \u2022 Which of the following, if any, top-layer applications do you use: Modulus, Maxine,\n cuQuantum, Merlin, Aerial, Monai, Triton, Nemo, VSS Blueprint, Riva, Metropolis,\n Holoscan, Clara Parabricks, Rapids, Isaac, Isaac Lab, Drive, DriveWorks, and\n Morpheus?\n \u2022 For each NVIDIA software product identified, in the ordinary course, do you download\n and use it locally, use it via a cloud-based solution offered by NVIDIA, or through a\n third-party cloud provider?\n \u2022 Do you make modifications to the source code for PyTorch, TensorRT, or other\n NVIDIA-provided software when its deployed on NVIDIA GPUs?\n \u2022 In the ordinary course, approximately how frequently do you run computations utilizing\n PyTorch or TensorRT on NVIDIA GPUs (e.g., many times per day, every day, every\n week, or every month)?\n \u2022 Are the relevant systems operated in the United States or used to support U.S.-directed\n operations?\n\nInput Data Path\n \u2022 In the ordinary course, do you use GPUDirect Storage (GDS) to load input data directly\n from storage into GPU memory, bypassing CPU main memory? Or do you use CPU\n main memory?\n \u2022 In the ordinary course, do you use GPUDirect RDMA or any direct NIC-to-GPU memory\n path for receiving live input data?\n \u2022 In the ordinary course, do you use unified memory (cudaMallocManaged), Unified\n Virtual Memory (UVM), or a coherent CPU/GPU memory architecture (e.g., DGX Spark\n UMA, Grace Hopper coherent memory) for the input data path?\n \u2022 In the ordinary course, do you use the PyTorch function torch.cuda.gds.GdsFile or any\n GDS-enabled data loader (e.g., DALI GDS, KvikIO) to load data?\n\nOutput Data Path and Memory Transfers\n \u2022 In the ordinary course, are outputs of GPU computations copied back to CPU/main\n memory? If so, what data is copied (e.g., transcripts, generated tokens, logits,\n embeddings)?\n \u2022 In the ordinary course, do GPU-to-CPU (D2H) or CPU-to-GPU (H2D) data transfers\n occur during or in parallel with GPU computations, or only after each computation\n completes?\n \u2022 In the ordinary course, how are your memory partitions configured for GPU\n computations? Are there separate memory regions for input data, output data, workspace,\n and intermediate results?\n \u2022 In the ordinary course, are output buffers from one GPU computation reused as input\n buffers for a subsequent computation?\n \u2022 Do you use torch.compile, TorchDynamo, Triton, TensorRT-LLM compilation, NVRTC,\n or PTX JIT to generate GPU programs at runtime, or do you use only precompiled\n kernels?\n\n\n\n 1\n\f Case 7:26-cv-00318 Document 1-12 Filed 08/17/26 Page 3 of 3\n\n\n\n\nTensorRT Usage\n \u2022 If you use TensorRT or TensorRT-LLM, how do you populate the input buffers and\n retrieve the outputs? Do you use CPU-side request processing and H2D/D2H copies?\n \u2022 Are TensorRT engines built by your organization, provided pre-built by NVIDIA, or\n obtained from another source?\n\n\n\n\n 2\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:21.520182-07:00","document_number":"1","attachment_number":12,"pacer_doc_id":"181037209222","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 11","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294683/","id":490294683,"tags":[],"absolute_url":"/docket/74659430/1/13/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.222994-07:00","date_modified":"2026-08-21T19:46:41.623298-07:00","sha1":"fc30c45aa644f442841bc2cf9d27e8a54fd5333b","page_count":4,"file_size":84501,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.13.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.13.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-13 Filed 08/17/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 12\n\f Case 7:26-cv-00318 Document 1-13 Filed 08/17/26 Page 2 of 4\n\n\n\n\n UNITED STATES DISTRICT COURT\n WESTERN DISTRICT OF TEXAS\n MIDLAND/ODESSA DIVISION\n\n\n NEURAL AI, LLC\n\n Plaintiff, Case No. 7:24-cv-00221-ADA-DTG\n v. JURY TRIAL DEMANDED\n NVIDIA CORPORATION\n\n Defendant.\n\n\n\n\n I, [DECLARANT NAME], hereby declare as follows:\n\n 1. I am [TITLE] of [COMPANY NAME]. I am over the age of eighteen and\n\ncompetent to make this declaration. I make this declaration based on my personal knowledge,\n\nincluding my general familiarity with [COMPANY NAME]\u2019s technical infrastructure, software\n\ndeployments, use of NVIDIA products, as well as the on the basis of a reasonably diligent\n\ninvestigation.\n\n 2. I understand that this declaration is submitted in connection with a subpoena served\n\non [INSERT DAT] on [COMPANY NAME] in the above-captioned action (the \u201cSubpoena\u201d).\n\n 3. In the ordinary course of business, [COMPANY NAME] uses NVIDIA GPUs (as\n\nthat term is defined in the Subpoena attached as Exhibit A) in its commercial operations.\n\nSpecifically, [COMPANY NAME] deploys GPUs of the [INSERT ARCHITECTURE]\n\narchitecture(s), including but not limited to [INSERT SPECIFIC GPU MODEL(S)] (collectively,\n\nthe \u201cDeployed NVIDIA GPUs\u201d).\n\n 4. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select among cuFFT,\n\f Case 7:26-cv-00318 Document 1-13 Filed 08/17/26 Page 3 of 4\n\n\n\n\ncuSparse, cuSolver, cuBlas, and cuDNN].\n\n 5. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select PyTorch and/or\n\nTensorRT].\n\n 6. In the ordinary course of business, [COMPANY NAME] uses the following\n\nNVIDIA offerings in connection with the Deployed NVIDIA GPUs: [Select among Modulus,\n\nMaxine, cuQuantum, Merlin, Ariel, Monai, Triton, Nemo, Riva, Metropolis, Holoscan, Clara\n\nParabricks, Rapids, Issac, Drive, and Morpheus].\n\n 7. In the ordinary course of business, [COMPANY NAME] uses the NVIDIA\n\nsoftware identified in Paragraphs 4-6 as provided by NVIDIA, without modification to the source\n\ncode.\n\n 8. To the best of [COMPANY NAME]\u2019s knowledge, when it uses the NVIDIA\n\nSoftware on the Deployed NVIDIA GPUs in the ordinary course of business, the hardware and\n\nNVIDIA software function together as designed and intended by NVIDIA.\n\n 9. In the ordinary course, [COMPANY NAME] uses one or more pretrained neural-\n\nnetwork models, model implementations, or model configurations distributed, made available, or\n\nrecommended by NVIDIA. In the ordinary course, [COMPANY NAME] deploys those models,\n\nimplementations, or configurations using PyTorch, TensorRT, or other NVIDIA software without\n\nmodifying their underlying network structure or low-level implementation code.\n\n 10. In the ordinary course of business, the Deployed NVIDIA GPUs are installed in\n\nsystems with separate CPU main memory and GPU memory, and the CPU and GPU in these\n\nsystems are connected via a bus.\n\n 11. In the ordinary course of business and for the operations described in this\n\f Case 7:26-cv-00318 Document 1-13 Filed 08/17/26 Page 4 of 4\n\n\n\n\ndeclaration, [COMPANY NAME] does not use Deployed NVIDIA GPUs with a unified\n\nCPU/GPU memory pool, GPUDirect Storage (\u201cGDS\u201d) to transfer input data directly from storage\n\nto GPU memory, NVIDIA Unified Virtual Memory (\u201cUVM\u201d), or CUDA managed memory to\n\nbypass the standard CPU-memory-to-GPU-memory data transfer path.\n\n 12. In [COMPANY NAME]\u2019s normal operations, when it uses the NVIDIA software\n\ndescribed in Paragraphs 4-6 on the Deployed NVIDIA GPUs, input data is received and processed\n\nby the CPU and stored in CPU main memory before being transferred to the GPU for computation.\n\nTo the best of my knowledge, this is the default and standard method by which data is loaded for\n\nprocessing on NVIDIA GPUs.\n\n 13. In the ordinary course of business and to the best of [COMPANY NAME]\u2019s\n\nknowledge, [COMPANY NAME] does not implement custom code or configurations that alter\n\nthe default data flow path provided by the NVIDIA Software with respect to how input data is\n\nreceived by the CPU, stored in main memory, and transferred to the GPU for computation.\n\n 14. To the best of [COMPANY NAME]\u2019s knowledge, the operations described in this\n\ndeclaration are performed in the United States using systems located in the United States or\n\nsystems that directly support [COMPANY NAME]\u2019s United States operations.\n\n 15. At least once per day, [COMPANY NAME] uses NVIDIA GPUs in conjunction\n\nwith at least one NVIDIA offering identified in each of Paragraphs 4-6.\n\n I declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct. Executed on [DATE], in [CITY, STATE].\n\n_______________________________________\n[DECLARANT NAME]\n[TITLE]\n[COMPANY NAME]\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:21.641219-07:00","document_number":"1","attachment_number":13,"pacer_doc_id":"181037209223","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 12","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294684/","id":490294684,"tags":[],"absolute_url":"/docket/74659430/1/14/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.236447-07:00","date_modified":"2026-08-21T19:35:26.554411-07:00","sha1":"272f61cd4fa53585fce6aee95e2106e120fd7bb9","page_count":4,"file_size":1280957,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.14.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.14.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-14 Filed 08/17/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 13\n\f Case 7:26-cv-00318 Document 1-14 Filed 08/17/26 Page 2 of 4\n Company Blog Artificial Intelligence AI Infrastructure Physical AI Gaming & Creating Industries Subscribe US \uf2bd Sign In\n\n\n\n\nTesla Unveils Top AV Training Supercomputer Powered by NVIDIA A100\nGPUs\n\u2018Incredible\u2019 GPU cluster powers AI development for Autopilot and full self-driving.\nJune 22, 2021 by Danny Shapiro\n\n\n 2 mins 0 Share\n\n\n\n\nTackling one of the largest computing challenges of this lifetime requires larger than life computing.\n\n\nAt CVPR this week, Andrej Karpathy, senior director of AI at Tesla, unveiled the in-house supercomputer the automaker is using to train\ndeep neural networks for Autopilot and self-driving capabilities. The cluster uses 720 nodes of 8x NVIDIA A100 Tensor Core GPUs (5,760\nGPUs total) to achieve an industry-leading 1.8 exaflops of performance.\n\n\n\u201cThis is a really incredible supercomputer,\u201d Karpathy said. \u201cI actually believe that in terms of flops, this is roughly the No. 5 supercomputer\nin the world.\u201d\n\nWith unprecedented levels of compute for the automotive industry at the center of its development cycle, Tesla is making it possible for NVIDIA GTC Berlin\nautonomous vehicle engineers to do their life\u2019s work efficiently and at the cutting edge. Registration Is Now Open\n\nNVIDIA A100 GPUs deliver acceleration at every scale to power the world\u2019s highest-performing data centers. Powered by the NVIDIA October 20-22\n\nAmpere Architecture, the A100 GPU provides up to 20x higher performance over the prior generation and can be partitioned into seven\n Register Now\nGPU instances to dynamically adjust to shifting demands.\n\n\n\n\n Recent News\n Gaming\n\n\n GeForce NOW Shakes Up August With 26\n New Games\n August 6, 2026\n\n\n\n AI\n\n\n Into the Omniverse: How Open World\n Models Push the Frontier of Physical AI\n August 6, 2026\n\n\n\n AI\n\n\n NVIDIA and Partners Build in America, for\n America\n August 5, 2026\n\n\n\n AI Infrastructure\n\n\nThe GPU cluster is part of Tesla\u2019s vertically integrated autonomous driving approach, which uses more than 1 million cars already driving NVIDIA Joins NSF State and Regional AI\non the road to refine and build new features for continuous improvement. Hubs Program to Expand AI Research and\n Education Across the US\n\f Case 7:26-cv-00318 Document 1-14 Filed 08/17/26 Page 3 of 4\n August 4, 2026\nFrom the Car to the\n Company Blog Data Center Subscribe US\n\n\nTesla\u2019s cyclical development begins in the car. A deep neural network running in \u201cshadow mode\u201d quietly perceives and makes predictions\n View All Recent News\nwhile the car is driving without actually controlling the vehicle.\n\nThese predictions are recorded, and any mistakes or misidentifications are logged. Tesla engineers then use these instances to create a\ntraining dataset of difficult and diverse scenarios to refine the DNN.\n\n\nThe result is a collection of roughly 1 million 10-second clips recorded at 36 frames per second, totaling a whopping 1.5 petabytes of\ndata. The DNN is then run through these scenarios in the data center over and over until it operates without a mistake. Finally, it\u2019s sent\nback to the vehicle and begins the process again.\n\n\nKarpathy said training a DNN in this manner and on such a large amount of data requires \u201ca huge amount of compute,\u201d which led Tesla to\nbuild and deploy the current generation supercomputer with high-performance A100 GPUs.\n\n\n\n\nContinuous Iteration\nIn addition to comprehensive training, Tesla\u2019s supercomputer gives autonomous vehicle engineers the performance needed to\nexperiment and iterate in the development process.\n\n\nKarpathy said the current DNN structure the automaker is deploying allows a team of 20 engineers to work on a single network at once,\nisolating different features for parallel development.\n\nThese DNNs can then be run through training datasets at speeds faster than what has been previously possible for rapid iteration.\n\n\n\u201cComputer vision is the bread and butter of what we do and enables Autopilot. For that to work, you need to train a massive neural\nnetwork and experiment a lot,\u201d Karpathy said. \u201cThat\u2019s why we\u2019ve invested a lot into the compute.\u201d\n\nWatch the full CVPR session.\n\n\nCategories: Driving\n\n\n\nTags: NVIDIA DGX Transportation\n\n\n\n\n Comments for this thread are now closed \u00d7\n\n\n0 Comments \ue603\n 1 Login\n\n\n\n \uf109 Share Best Newest Oldest\n\n\n\n\n This discussion has been closed.\n\n\n\n\n Subscribe Privacy Do Not Sell My Data\n\n\n\n\nRelated News\n\n\n\n\n AI AI AI AI\n\f Case 7:26-cv-00318 Document 1-14 Filed 08/17/26 Page 4 of 4\nInto the Omniverse: How Open\n Company Blog\n NVIDIA and Partners Build in AI Leaders Propose SAFE Industry Leaders Unite\n Subscribe US\n in Open\nWorld Models Push the America, for America Guidelines for Cybersecurity Secure AI Alliance for AI Safety\nFrontier of Physical AI Aug 5, 2026 Transparency and Security\nAug 6, 2026 Aug 4, 2026 Jul 27, 2026\n\n\n\n\n Follow Us \uf39e \uf16d \uf0e1 \ue61b \uf167\n\n\n United States\n\nPrivacy Policy Your Privacy Choices Terms of Service Accessibility Corporate Policies Product Security Contact\nCopyright \u00a9 2026 NVIDIA Corporation\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:23.393018-07:00","document_number":"1","attachment_number":14,"pacer_doc_id":"181037209224","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 13","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294685/","id":490294685,"tags":[],"absolute_url":"/docket/74659430/1/15/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.249943-07:00","date_modified":"2026-08-21T18:47:01.755326-07:00","sha1":"5979caab8efc51d03d8a733631710ddaea87a166","page_count":4,"file_size":3964442,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.15.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.15.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-15 Filed 08/17/26 Page 1 of 4\n\n\n\n\n EXHIBIT\n\n 14\n\f Case 7:26-cv-00318 Document 1-15 Filed 08/17/26 Page 2 of 4\n US Edition RSS Sign in Search\n\n\n Best Picks CPUs GPUs PC Components News Laptops Desktops Software & AI Coupons Premium Forums\n\n\n\n\nTRENDING Tom's 30th Anniversary AMD Advancing AI RAM Shortage Vera Rubin AI Data Centers RAM Combo Deals TH Premium AMD Instinct MI455X DLSS\n\n\n\n PC Components > GPUs ADVERTISEMENT\n\n\n\n Tesla Brags About In-House\n Supercomputer, Now With 7,360 A100\n GPUs\n News By Mark Tyson Published August 17, 2022\n\n\n With a 28% increase in GPUs, it's now a top-7 supercomputer\n worldwide by GPU count\n\n\n\n\n (Image credit: Tesla)\n\n\n\n\n 9 Follow us Newsletter\n\n\n\n Tesla has boosted its in-house AI supercomputer with thousands of\n additional Nvidia A100 GPUs. The Tesla supercomputer had 5,760 A100\n GPUs about a year ago, and that count has since risen to 7,360 A100 GPUs\n \u2014 that's an additional 1,600 GPUs, or about a 28% increase.\n\n According to Tesla Engineering Manager Tim Zaman, this upgrade makes\n the firm's AI system a top-7 supercomputer worldwide by GPU count.\n\n An Nvidia A100 GPU is a powerful Ampere architecture solution aimed at\n data centers. Yes, it uses the same GPU architecture as GeForce RTX 30\n series GPUs, which are some of the best graphics cards currently\n ADVERTISEMENT\n available. However, there is no close consumer relation to the A100, which\n comes with 80GB of HBM2e memory on board, offers up to 2 TB/s\n bandwidth, and requires up to 400W of power. The architecture of the\n A100 has also been tweaked for accelerating tasks common in AI, data\n analytics, and high-performance computing (HPC) applications.\n\n\n\n Latest Videos From Tom's Hardware\n\n\n\n\n Ad 1 of 2.\n Watch full video here: How to get rid of Google's AI overviews\n\n\n\n The first system Nvidia showed wielding the A100 was the Nvidia DGX\n A100, which packed in eight A100 GPUs linked via six NVSwitch with 4.8\n TBps of bi-directional bandwidth for upJoin\n to 10 PetaOPS\n Tom\u2019s of INT8 today\n Hardware ADVERTISEMENT EXPLORE\n\f Case 7:26-cv-00318 Document 1-15 Filed 08/17/26 Page 3 of 4\nperformance, 5 PFLOPS of FP16, 2.5 TFLOPS of TF32, and 156 TFLOPS of\nFP64 in a single node.\n\nThat was eight A100 GPUs \u2014 Tesla's AI supercomputer now has 7,360 of\nthese. Tesla hasn't publicly benchmarked its AI supercomputer, but the\nsimilarly-equipped GPU-based NERSC Perlmutter, which has 6,144 Nvidia\nA100 GPUs, achieves 70.87 Linpack petaflops. Using this and data from\nother A100 GPU supercomputers as performance reference points, HPC\nWire estimates the Tesla AI supercomputer is capable of achieving about\n100 Linpack petaflops.\n\n YOU MAY LIKE\n\n\nChina bypasses US GPU bans with 1.54-exaflops 'LineShine'\nsupercomputer\n\n\n\nGoogle could build more AI accelerators than Nvidia sells in\n2028, analyst claims\n\n\n\nNvidia's memory costs soar 485%, latest AI systems now cost\n$7.8 million to build\n\n\nTesla doesn\u2019t intend to continue down the Nvidia GPU architecture path\nfor its in-house AI supercomputers long-term. This world\u2019s top-7 machine\nby GPU-count is merely a precursor to the upcoming Dojo supercomputer,\nwhich was first announced by Elon Musk back in 2020. A year ago we got a\nlook at the Tesla D1 Dojo chip, which are designed to supplant Nvidia's\nGPUs for \u201cmaximum performance, throughput and bandwidth at every\ngranularity.\u201d\n ADVERTISEMENT\n\n\n\n\n \uf125\n(Image credit: Tesla)\n\n\nThe Tesla Dojo D1 is a custom ASIC (application-specific integrated circuit)\ndesign, purposed for AI training, and it is one of the first ASICs in this field.\nCurrent D1 test chips are manufactured on TSMC N7 and pack in about 50\nmillion transistors.\n\n\n\n Stay On the Cutting Edge: Get the Tom's\n Hardware Newsletter\n Get Tom's Hardware's best news and in-depth reviews, straight to\n your inbox.\n\n\n Your Email Address SIGN ME UP\n\n\n\n By signing up, you agree to our Terms of services and acknowledge that you\n have read our Privacy Notice. You also agree to receive marketing emails from\n us that may include promotions from our trusted partners and sponsors, which\n you can unsubscribe from at any time.\n\n\n\n\nMore information about the Dojo D1 chip, and the Dojo system, might be\nrevealed at next week's Hot Chips Symposium \u2014 three Tesla\npresentations are schedule for next Tuesday, addressing Dojo D1 chip\narchitecture, Dojo and ML training, and enabling AI through system\nintegration.\n\n\n ADVERTISEMENT\nTOPICS\n\n Nvidia Science GeForce\n\n\n\n Ad 1 of 2.\n \uf123 SEE ALL COMMENTS (9)\n\n\n\n\n Mark Tyson News Editor\n Join Tom\u2019s Hardware today EXPLORE\n\f Case 7:26-cv-00318 Document 1-15 Filed 08/17/26 Page 4 of 4\nREPLY \uf105\n\n\n\ncirdecus\n\n Mandark said:\n\n Nothing like vertical integration. ALWAYS make your OWN stuff to\n control your destiny. Don\u2019t believe the fools that say it\u2019s cheaper to\n outsource it\u2019s not. Outsource companies need to make a profit too.\n And when you outsource you lose the ability to innovate\n\n\n I love to see Tesla vertically integrate. It\u2019s one of the main reasons for\n their success\n\n\n\n\nCompletely agree. It does, however, tend to condense power which\ncould mean less consumer choice, but vertical integration is where it's\nat. I never thought I'd see Amazon buying their own fleet of transport\nvehicles, planes and ships lol.\n\nREPLY \uf105\n\n\n\njtenorj\nNoticed a mistake in your article. You state that Tesla's D1 chip has 50\nmillion transistors when Telsa's slide for the chip clearly shows 50 billion.\nThat's a difference of 3 orders of magnitude. 50 billion is also much\nmore in line with the chip's size in mm squared as well as its 400w power\ndraw on a modern fairly compact process node.\nREPLY \uf105\n\n\n\n\n SHOW MORE COMMENTS \uf105\n\n\n\n\n Tom's Hardware is part of Future US Inc, an international media group and leading digital publisher. Visit our corporate site.\n\n\n\n\n Terms and conditions Contact Future's experts Privacy policy Cookies policy\n\n\n Accessibility Statement Advertise with us About us Coupons\n\n\n Careers Do not sell or share my personal\n information\n\n\n\n \u00a9 Future US, Inc. Full 7th Floor, 130 West 42nd Street, New York, NY 10036.\n\n\n\n\n Ad 1 of 2.\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:25.416705-07:00","document_number":"1","attachment_number":15,"pacer_doc_id":"181037209225","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 14","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294686/","id":490294686,"tags":[],"absolute_url":"/docket/74659430/1/16/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.263488-07:00","date_modified":"2026-08-21T18:43:50.171873-07:00","sha1":"7dd2f0381ddf392df32646f6c7d603093dfbc70a","page_count":8,"file_size":9757818,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.16.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.16.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 1 of 8\n\n\n\n\n EXHIBIT\n\n 15\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 2 of 8\n 00:10:03:51 Get $400 off your Disrupt 2026 ticket: REGISTER NOW.\n\n\n Latest Startups Venture Apple Security AI Apps Disrupt 2026 Events Podcasts Newsletters\n\n\n\n\n TRANSPORTATION\n\n\n\n\n Tesla Dojo: The rise and fall of\n Elon Musk\u2019s AI supercomputer\n\n\n\n Rebecca Bellan 9:18 AM PDT \u00b7 September 2, 2025\n\n\n\n\n IMAGE CREDITS: BRYCE DURBIN/TECHCRUNCH\n\n\n\n\n For years, Elon Musk has spoken of the promise of Dojo, the AI supercomputer\n that was supposed to be the cornerstone of Tesla\u2019s AI ambitions. It was\n important enough to Musk that in July 2024, he said the company\u2019s AI team\n would \u201cdouble down\u201d on Dojo in the lead-up to Tesla\u2019s robotaxi reveal, which\n happened in October.\n\n After six years of hype, Tesla decided last month to shut down Dojo and\n disband the team behind the supercomputer in August 2025. Within weeks of\n projecting that Dojo 2, Tesla\u2019s second supercluster that was meant to be built\n on the company\u2019s in-house D2 chips, would reach scale by 2026, Musk\n reversed course, declaring it \u201can evolutionary dead end.\u201d\n\n Ad\n Manage your stocks and ET\n\n Fidelity Investments\n\n\n Learn more\n\n\n This article originally set out to explain what Dojo was and how it could help\n Tesla achieve full-self driving, autonomous humanoid robots, semiconductor\n autonomy, and more. Now, you can think of it more as an obituary of a project\n that convinced so many analysts and investors that Tesla wasn\u2019t just an\n automaker, it was an AI company.\n\n Dojo was Tesla\u2019s custom-built supercomputer that was designed to train its\n \u201cFull Self-Driving\u201d neural networks.\n\n Beefing up Dojo went hand-in-hand with Tesla\u2019s goal to reach full self-driving\n October 13 \u2013 15 San Francisco\n and bring a robotaxi to market. FSD (Supervised) is Tesla\u2019s advanced driver\n assistance system that\u2019s on hundreds of thousands of Tesla vehicles today Scale faster. Grow your portfolio. Gain\n practical expertise. No matter your goal,\n and can perform some automated driving tasks, but it still requires a human Disrupt can empower you.\n to be attentive behind the wheel. It\u2019s also the basis of similar technology\n Save up to $400 today!\n powering Tesla\u2019s limited robotaxi service that the company launched in Austin\n this June using Model Y SUVs. REGISTER NOW\n Ad : (0:03)\n Even as Dojo\u2019s raison d\u2019\u00eatre started to come to life, Tesla failed to attribute its\n self-driving successes \u2014 controversial as they were \u2014 to the supercomputer.\n In fact, Musk and Tesla had barely mentioned Dojo at all over the past year. In\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 3 of 8\n August 2024, Tesla began promoting Cortex, the company\u2019s \u201cgiant new AI\n training supercluster being built at Tesla HQ in Austin to solve real-world AI,\u201d\n which Musk has said would have \u201cmassive storage for video training of FSD\n Most\n and Optimus.\u201d Popular\n In Tesla\u2019s Q4 2024 shareholder deck, the company shared updates on Cortex, ChatGPT brings unlimited text\n chats to free users\n but nothing on Dojo. It\u2019s not clear whether Tesla\u2019s Dojo shutdown affects\n Cortex.\n Ford\u2019s new electric truck,\n \u2018Fathom,\u2019 starts at $28,350\n\n\n Bending Spoons to buy Airtable\n for $1.28B\n\n\n Influencers draw backlash for\n attending OpenAI\u2019s first luxury\n trip\n\n\n Sequoia\u2019s Shaun Maguire leads\n $1B round for nuclear startup\n Valar Atomics\n\n\n Malaysia is reportedly shutting\n down Balaji Srinivasan\u2019s\n Network School\n\n\n YouTuber Hank Green says his\n The response to Dojo\u2019s disbanding has been mixed. Some see it as another AI usage is \u2018not healthy\u2019\n example of Musk making promises he can\u2019t deliver on that comes at a time of\n falling EV sales and a lackluster robotaxi rollout. Others say the shutdown\n wasn\u2019t a failure, but a strategic pivot from a high-risk, self-reliant hardware to\n a streamlined path that relies on partners for chip development. Ad\n\n\n Dojo\u2019s story reveals what was on the line, where the project fell short, and\n what its shutdown signals for Tesla\u2019s future.\n\n\n\n A recap of Dojo\u2019s shutdown Manage your stocks and ET\n Fidelity Basket Portfolios is a faster and easier way to build a\n\n Tesla disbanded its Dojo team and shut down the project in mid-August 2025. basket of securities and\n\n\n Dojo\u2019s lead, Peter Bannon, left the company as well, following the departure of Fidelity Investments Learn more\n around 20 workers who left to start their own AI chip and infrastructure\n company called DensityAI.\n\n Analysts have pointed out that losing key talent can quickly derail a project,\n especially a highly specialized, internal tech project.\n Advertisement\n The shutdown came a couple of weeks after Tesla signed a $16.5 billion deal\n to get its next-generation AI6 chips from Samsung. The AI6 chip is Tesla\u2019s bet\n on a chip design that can scale from powering FSD and Tesla\u2019s Optimus\n humanoid robots to high-performance AI training in data centers.\n\n \u201cOnce it became clear that all paths converged to AI6, I had to shut down Dojo\n and make some tough personnel choices, as Dojo 2 was now an evolutionary\n dead end,\u201d Musk posted on X, the social media platform he owns. \u201cDojo 3\n arguably lives on in the form of a large number of AI6 [systems-on-a-chip] on\n a single board.\u201d\n\n\n\n Tesla\u2019s Dojo backstory\n\n\n\n\n IMAGE CREDITS:SUZANNE CORDEIRO / AFP / GETTY IMAGES\n\n\n\n Musk has insisted that Tesla isn\u2019t just an automaker, or even a purveyor of\n solar panels and energy storage systems. Instead, he has pitched Tesla as an\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 4 of 8\n AI company, one that has cracked the code to self-driving cars by mimicking\n human perception.\n\n Most other companies building autonomous vehicle technology rely on a\n combination of sensors to perceive the world \u2014 like lidar, radar and cameras\n \u2014 as well as high-definition maps to localize the vehicle. Tesla believes it can\n achieve fully autonomous driving by relying on cameras alone to capture\n visual data and then use advanced neural networks to process that data and\n make quick decisions about how the car should behave.\n\n The pitch has been that Dojo-trained AI software will eventually be pushed out\n to Tesla customers via over-the-air updates. The scale of FSD also means\n Tesla has been able to rake in millions of miles worth of video footage that it\n uses to train FSD. The idea there is that the more data Tesla can collect, the\n closer the automaker can get to actually achieving full self-driving.\n\n However, some industry experts say there might be a limit to the brute force\n approach of throwing more data at a model and expecting it to get smarter.\n\n \u201cFirst of all, there\u2019s an economic constraint, and soon it will just get too\n expensive to do that,\u201d Anand Raghunathan, Purdue University\u2019s Silicon Valley\n professor of electrical and computer engineering, told TechCrunch. Further,\n he said, \u201cSome people claim that we might actually run out of meaningful\n data to train the models on. More data doesn\u2019t necessarily mean more\n information, so it depends on whether that data has information that is useful\n to create a better model, and if the training process is able to actually distill\n that information into a better model.\u201d\n\n Raghunathan said despite these doubts, the trend of more data appears to be\n here for the short-term at least. And more data means more compute power\n needed to store and process it all to train Tesla\u2019s AI models. That was where\n Dojo, the supercomputer, came in.\n\n\n\n What is a supercomputer?\n Dojo was Tesla\u2019s supercomputer system that was designed to function as a\n training ground for AI, specifically FSD. The name is a nod to the space where\n martial arts are practiced.\n\n A supercomputer is made up of thousands of smaller computers called\n nodes. Each of those nodes has its own CPU (central processing unit) and\n GPU (graphics processing unit). The former handles overall management of\n the node, and the latter does the complex stuff, like splitting tasks into\n multiple parts and working on them simultaneously.\n\n GPUs are essential for machine learning operations like those that power FSD\n training in simulation. They also power large language models, which is why\n the rise of generative AI has made Nvidia the most valuable company on the\n planet.\n\n Even Tesla buys Nvidia GPUs to train its AI (more on that later).\n\n\n\n Why did Tesla need a\n supercomputer?\n Tesla\u2019s vision-only approach was the main reason Tesla needed a\n supercomputer. The neural networks behind FSD are trained on vast amounts\n of driving data to recognize and classify objects around the vehicle and then\n make driving decisions. That means that when FSD is engaged, the neural\n nets have to collect and process visual data continuously at speeds that\n match the depth and velocity recognition capabilities of a human.\n\n In other words, Tesla means to create a digital duplicate of the human visual\n cortex and brain function.\n\n To get there, Tesla needs to store and process all the video data collected\n from its cars around the world and run millions of simulations to train its\n model on the data.\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 5 of 8\n Elon Musk \u00b7 Jul 23, 2024\n @elonmusk \u00b7 Follow\n Replying to @elonmusk and @ajtourville\n And Dojo 1 will have roughly 8k H100-equivalent of training\n online by end of year.\n Not massive, but not trivial either.\n Elon Musk\n @elonmusk \u00b7 Follow\n Dojo pics\n\n\n\n\n 2:24 PM \u00b7 Jul 23, 2024\n 6.3K Reply Copy link\n Read 385 replies\n\n Tesla relied mainly on Nvidia to power its current Dojo training computer, but it\n didn\u2019t want to have all its eggs in one basket \u2014 not least because Nvidia chips\n are expensive. Tesla had hoped to make something better that increased\n bandwidth and decreased latencies. That\u2019s why the automaker\u2019s AI division\n decided to come up with its own custom hardware program that aimed to\n train AI models more efficiently than traditional systems.\n\n At that program\u2019s core was Tesla\u2019s proprietary D1 chips, which the company\n said were optimized for AI workloads.\n\n\n\n Tell me more about these chips\n\n\n\n\n GANESH VENKATARAMANAN, FORMER SENIOR DIRECTOR OF AUTOPILOT HARDWARE, PRESENTING THE\n D1 TRAINING TILE AT TESLA\u2019S 2021 AI DAY.\n\n IMAGE CREDITS:TESLA / SCREENSHOT\n\n\n\n Tesla, like Apple, thinks hardware and software should be designed to work\n together. That\u2019s why Tesla was working to move away from the standard GPU\n hardware and design its own chips to power Dojo.\n\n Tesla unveiled its D1 chip, a silicon square the size of a palm, on AI Day in 2021.\n The D1 chip entered into production around July 2023.\n\n The Taiwan Semiconductor Manufacturing Company (TSMC) manufactured\n the chips using 7 nanometer semiconductor nodes. The D1 has 50 billion\n transistors and a large die size of 645 millimeters squared, according to Tesla.\n This is all to say that the D1 promises to be extremely powerful and efficient\n and to handle complex tasks quickly.\n\n The D1 wasn\u2019t as powerful as Nvidia\u2019s A100 chip, though.\n\n Tesla had been working on a next-gen D2 chip that aimed to solve information\n flow bottlenecks. Instead of connecting the individual chips, the D2 would\n have put the entire Dojo tile onto a single wafer of silicon.\n\n Tesla never confirmed how many D1 chips it ordered or received. The\n company also never provided a timeline for how long it would have taken to\n get Dojo supercomputers running on D1 chips.\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 6 of 8\n What did Dojo mean for Tesla?\n\n\n\n\n VISITORS ARE VIEWING TESLA\u2019S HUMANOID ROBOT OPTIMUS PRIME II AT WAIC IN SHANGHAI,\n CHINA, ON JULY 7, 2024.\n\n IMAGE CREDITS:COSTFOTO / NURPHOTO / GETTY IMAGES\n\n\n\n Tesla\u2019s hope was that by taking control of its own chip production, it might\n one day be able to quickly add large amounts of compute power to AI training\n programs at a low cost.\n\n It also meant not having to rely on Nvidia\u2019s chips in the future, which are\n increasingly expensive and hard to secure. Now, Tesla is going all-in on\n partnerships \u2014 with Nvidia, AMD, and Samsung, which will build its next-gen\n AI6 chip.\n\n During Tesla\u2019s second-quarter 2024 earnings call, Musk said demand for\n Nvidia hardware was \u201cso high that it\u2019s often difficult to get the GPUs.\u201d He said\n he was \u201cquite concerned about actually being able to get steady GPUs when\n we want them, and I think this therefore requires that we put a lot more effort\n on Dojo in order to ensure that we\u2019ve got the training capability that we\n need.\u201d\n\n Dojo was a risky bet, one that Musk hedged several times by saying that Tesla\n might not succeed.\n\n In the long run, Tesla toyed with the idea of creating a new business model\n based on its AI division, with Musk even saying during a Q2 2024 earnings call\n that he saw \u201ca path to being competitive with Nvidia with Dojo.\u201d While D1 was\n more tailored for Tesla computer vision labeling and training \u2014 useful for FSD\n and Optimus training \u2014 it wouldn\u2019t have been useful for much else. Future\n versions would have to be more tailored to general-purpose AI training, Musk\n said.\n\n The problem that Tesla might have come up against is that almost all AI\n software out there has been written to work with GPUs. Using Dojo chips to\n train general-purpose AI models would have required rewriting the software.\n\n That is, unless Tesla rented out its compute, similar to how AWS and Azure\n rent out cloud computing capabilities \u2014 an idea that excited analysts. A\n September 2023 report from Morgan Stanley predicted that Dojo could add\n $500 billion to Tesla\u2019s market value by unlocking new revenue streams in the\n form of robotaxis and software services.\n\n In short, Dojo chips were an insurance policy for the automaker, but one that\n might have paid dividends.\n\n\n\n How far did Tesla Dojo get?\n\fCase 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 7 of 8\n\n\n\n\n NVIDIA CEO JENSEN HUANG AND TESLA CEO ELON MUSK\n\n IMAGE CREDITS:KIM KULISH / CORBIS / GETTY IMAGES\n\n\n\n Musk often provided progress reports, but many of his goals for Dojo were\n never reached.\n\n For instance, Musk suggested in June 2023 that Dojo had been online and\n running useful tasks for a few months.\u201d Around the same time, Tesla said it\n expected Dojo to be one of the top five most powerful supercomputers by\n February 2024 and had planned for total compute to reach 100 exaflops in\n October 2024, which would have required roughly 276,000 D1s, or around\n 320,500 Nvidia A100 GPUs.\n\n Tesla never provided an update or any information that would suggest it ever\n reached these goals.\n\n Tesla and Musk made numerous other pledges for Dojo, including financial\n ones. For instance, Tesla committed in January 2024 to spend $500 million to\n build a Dojo supercomputer at its gigafactory in Buffalo, New York, and has\n already spent $314 million of that, per a 2024 report.\n\n Just after Tesla\u2019s second-quarter 2024 earnings call, Musk posted photos of\n Dojo 1 on X, saying that it would have \u201croughly 8k H100-equivalent of training\n online by end of year. Not massive, but not trivial either.\u201d\n\n Despite all of this activity \u2014 particularly by Musk on X and in earnings calls \u2014\n mention of Dojo abruptly ended August 2024. And talk switched to Cortex.\n\n During the company\u2019s fourth-quarter 2024 earnings call, Tesla said it\n completed the deployment of Cortex, \u201ca ~50k H100 training cluster at\n Gigafactory Texas\u201d and that Cortex helped enable V13 of supervised FSD.\n\n In Q2 2025, Tesla noted it \u201cexpanded AI training compute with an additional\n 16k H200 GPUs at Gigafactory Texas, bringing Cortex to a total of 67k H100\n equivalents.\u201d During that same earnings call, Musk said he expected to have\n a second Dojo cluster operating \u201cat scale\u201d in 2026. He also hinted at potential\n redundancies.\n\n \u201cThinking about Dojo 3 and the AI6 inference chip, it seems like intuitively, we\n want to try to find convergence there, where it\u2019s basically the same chip,\u201d\n Musk said.\n\n A few weeks later, he reversed course and disbanded the Dojo team.\n\n TechCrunch confirmed in late August 2025 that Tesla still plans to commit\n $500 million to a supercomputer in Buffalo \u2014 it just won\u2019t be Dojo.\n\n This story originally published August 3, 2024. The article was updated for a\n final time September 2, 2025, with new information about Tesla\u2019s decision to\n shut down Dojo.\n\n\n Topics: AI Dojo Elon Musk evergreens Tesla Tesla FSD Transportation\n\n\n When you purchase through links in our articles, we may earn a small commission. This doesn\u2019t affect our\n editorial independence.\n\n\n\n\n Advertisement\n\f Case 7:26-cv-00318 Document 1-16 Filed 08/17/26 Page 8 of 8\n\n\n Rebecca Bellan\n Senior Repor ter |\n\n Rebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends\n shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and\u2026\n\n View Bio\n\n\n\n Loading the next article\n\n\n\n\n Advertisement\n\n\n\n\n TechCrunch Terms of Service OpenAI vs Apple\n\n Nous Research\n Staff Privacy Policy\n Space Data Centers\n Contact Us RSS Terms of Use\n Staya Nadella\n Advertise\n SpaceX Starship\n Crunchboard Jobs\n Tech Layoffs\n Site Map ChatGPT\n\n\n\u00a9 2026 TechCrunch Media LLC.\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:28.107323-07:00","document_number":"1","attachment_number":16,"pacer_doc_id":"181037209226","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 15","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294687/","id":490294687,"tags":[],"absolute_url":"/docket/74659430/1/17/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.277236-07:00","date_modified":"2026-08-21T18:35:42.597102-07:00","sha1":"ecc898f4d17388ed405c48f354574a537565c73a","page_count":6,"file_size":6011244,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.17.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.17.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 1 of 6\n\n\n\n\n EXHIBIT\n\n 16\n\f Case 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 2 of 6\n SCIENCE TECHNOLOGY ENVIRONMENT DIY GEAR MERCH NEWSLETTER\n\n\n\n\nTECHNOLOGY VEHICLES SELF DRIVING\n\n\nHow Tesla is using a supercomputer to train its self-driving tech\nTesla's approach to autonomy is controversial: It relies on just cameras to see and understand the roads.\nROB STUMPF / PUBLISHED JUN 26, 2021 5:00 AM EDT / ADD POPULAR SCIENCE\n\n ADVERTISEMENT\n\n\n\n\n Trending\n\n\n\n\n The supercomputer cluster has 5,760 GPUs\u2014processing power it needs to help power its self-driving aspirations. Tesla\n\n\n\n Get the Popular Science daily newsletter \ud83d\udca1 BIRDS\n\n Breakthroughs, discoveries, and DIY tips sent six days a week. Outside eagle spotted on Jackie and\n Shadow\u2019s nest\n Enter your email SIGN UP LAURA BAISAS\n\n By signing up, you confirm you are 16+, will receive newsletters and promotional content and agree to our Terms of Use and\n acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.\n\n\n You can\u2019t buy a fully self-driving car today, but automakers around the globe\n are racing to become the first company to place such a vehicle on dealer\n lots. No two companies are taking the same technological path to achieve\n this plan, either. Some make use of remote sensing methods like Light\n Detection and Ranging (LiDAR), while others rely on radar-based sensors to BEARS\n\n help pick out hard-to-see obstacles in the roadway. And typically, firms Why do mother bears kill cubs? The\n working on autonomous tech will use a combination of LiDAR, radar, and answer is complicated.\n cameras. JENNIFER BYRNE\n\n\n Then there\u2019s Tesla, which believes vision-based image recognition using only\n cameras is the key to affordable and reliable autonomy.\n\fCase 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 3 of 6\n ADVERTISEMENT\n\n\n\n\n Nomatic\n Nomatic\n\n\n\nBut there\u2019s a catch to Tesla\u2019s method: perfecting vision-based autonomy is\ndifficult. It requires the use of a continuously improving system that can\nquickly adapt to new and changing road conditions, and then it must be\ncapable of sharing that information with other vehicles on the roadway. That\nkind of learning takes significantly more processing power than what is\navailable in a single vehicle\u2014it takes a supercomputer.\n[Related: Everything self-driving cars calculate before changing lanes]\nDuring a talk at the International Joint Conference on Computer Vision and\nPattern Recognition earlier this month, Tesla\u2019s senior director of AI, Andrej\nKarpathy, revealed that the automaker has been working on a project to do\nexactly that.\nTesla\u2019s new supercomputer hasn\u2019t been named, at least not publicly. The\ncluster itself consists of 720 individual computers called nodes. Each node\nhas eight Nvidia A100 80GB Graphics Processing Units (GPUs) capable of\nperforming high-intensity floating point calculations with nearly 500 times as\nmuch power compared to a standard desktop processor.\nIn total, the cluster has 5,760 GPUs, or enough hardware to achieve an\ninsane 1.8 exaflops of processing power. Karpathy believes this makes\nTesla\u2019s supercomputer the fifth most powerful computing environment in the\nentire world, at least on paper.\nModern Teslas utilize an advanced driver assistance system called Autopilot.\nThis suite of features allows the vehicle to make use of eight exterior-facing\ncameras to gather data about the vehicle\u2019s surroundings and, when engaged\nand where applicable, performs lateral (steering) and longitudinal\n(acceleration and braking) controls under driver supervision. While this\nshouldn\u2019t be confused with Waymo\u2019s advanced self-driving, it is an interim\nstep that uses partial automation to bridge the gap between manual driving\nand fully autonomous control.\n ADVERTISEMENT\n\n\n\n\n[Related: How Waymo is teaching self-driving cars to deal with the chaos\nof parking lots]\nAutopilot uses information gathered from all Tesla vehicles on the road to\nimprove its driving decisions. As a Tesla steers along the street, its exterior\ncameras are constantly gathering data on the outside environment.\nComputers within the car study this data and make predictions of how to\nbehave in any given scenario without actually sending controls to the vehicle\nitself.\nThis information is shared on a machine learning architecture called a neural\nnetwork. The predictions are then recorded and sent back to Tesla to\ndetermine if the decision was correct or if any data was misidentified. If it\nwas, then the data then continually runs through the supercomputer\ntweaking its behavior until it processes without a mistake, effectively training\nTesla\u2019s ever-improving Autopilot model.\n[Related: Intel\u2019s new chip puts a teraflop in your desktop. Here\u2019s what that\nmeans]\nThis method not only consumes a large amount of processing power, but it\nalso requires significant storage in order to stockpile the one million 10-\nsecond clips used to make up the proprietary Tesla dataset training for\n\fCase 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 4 of 6\nAutopilot. These clips alone require 1.5 petabytes of storage, whereas the\nsystem itself is capable of hoarding approximately 10 petabytes of data on\nultra-fast NVMe flash storage.\nRelatedly, Tesla CEO Elon Musk has previously teased \u201cProject Dojo,\u201d a\nsupercomputer built on proprietary Tesla silicon specifically architectured for\nneural net model training. Musk noted that establishing high speed\ncommunication between components and efficient cooling was an ongoing\nchallenge in late 2020, though the project was ongoing.\n ADVERTISEMENT\n\n\n\n\nBecause Karpathy\u2019s cluster uses Nvidia-based GPUs, it doesn\u2019t appear to be\naffiliated with Project Dojo. However, it still plays an important role in Tesla\u2019s\nultimate goal of being the first automaker capable of a fully self-driving\nvehicle on public roads.\nTesla\u2019s rather ambitious goal has been met with quite a bit of skepticism by\nindustry leaders and naysayers of vision-only vehicle autonomy, especially\nsince the automaker rejected the use of ultra-precision LiDAR as part of its\nautonomy suite.\n[Related: This supercomputer will perform 1,000,000,000,000,000,000\noperations per second]\nNo Tesla vehicle on the road today makes use of LiDAR. In fact, Elon Musk\ncalled LiDAR a \u201ccrutch\u201d in 2018, denouncing the technology in favor of\nTesla\u2019s own vision-based system before doing away with supplemental\nradars earlier this year. That decision alone cost Tesla safety endorsements\nfrom the National Highway Traffic Safety Administration.\nMeanwhile, Volvo has chosen to implement LiDAR as a standard feature on\nthe upcoming successor to its XC90 SUV.\nAs for Tesla, its current-generation supercomputer will help to train its\nAutopilot model, and its upcoming Project Dojo likely even moreso. But only\ntime will tell if its vision-based technology will prevail over competitors,\nmeaning that it\u2019s a gambit that could make or break its position as a leader in\nthe autonomy segment.\n\n\n\n 2026 Popular Science Home of the Future Awards\n\n 39 products that will improve your everyday life.\n\n SEE IT\n\n\n\n\n ROB STUMPF\n Contributor, Tech\n Rob has been covering emerging car tech for PopSci since 2021, and the automotive beat for its\n motoring-focused sibling publication, The Drive, since early 2017. He brings both a tech and\n automotive background to his work about what the future of mobility holds.\n\f Case 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 5 of 6\n\n\n\n\n More in Self Driving\n\n\n\n\nSELF DRIVING SELF DRIVING ELECTRIC VEHICLES SELF DRIVING\nTeslas keep hitting emergency vehicles, and The new Mission Master XT is a massive military We\u2019re getting closer to a world filled with self- What we know so far about the fatal Tesla crash\nnow the government is investigating self-driving pack mule parking cars in Paris\nROB VERGER KELSEY D. ATHERTON ROB STUMPF COLLEEN HAGERTY\n\n\n\n\nSELF DRIVING ELECTRIC VEHICLES SELF DRIVING SELF DRIVING\nThe 6 terms you need to know to understand The government is investigating why Tesla Tesla caps off a tumultuous year with a massive \u2018Autopilot\u2019 systems in cars are finally getting\nself-driving cars drivers can play solitaire at the wheel recall graded on safety\nDAN CARNEY COLLEEN HAGERTY CHARLOTTE HU ROB STUMPF\n\n\n\n SEE MORE\n\n\n\n\n More in Vehicles\n\n\n\n\nELECTRIC VEHICLES ELECTRIC VEHICLES ELECTRIC VEHICLES ELECTRIC VEHICLES\nVolvo\u2019s sleek electric car concept has big Ikea Tesla\u2019s new adapter will let other car companies GM is recalling all its Bolts, but there\u2019s no need to These are the most useless car tech features\nvibes use its Fast Charging stations panic about EV safety ROB STUMPF\nROB STUMPF ROB STUMPF ROB VERGER\n\f Case 7:26-cv-00318 Document 1-17 Filed 08/17/26 Page 6 of 6\n\n\n\n\nELECTRIC VEHICLES ELECTRIC VEHICLES SELF DRIVING ELECTRIC VEHICLES\nThe Tesla Model S \u2018Plaid\u2019 will go from 0-60 with Tesla\u2019s new battery tech promises a road to a Self-driving cars still have major perception This company plans to build a self-driving car\nrecord-breaking acceleration cheap self-driving electric car problems with a brain that runs on light\nROB VERGER STAN HORACZEK YULONG CAO & Z. MORLEY MAO/THE CONVERSATION ROB VERGER\n\n\n\n SEE MORE\n\n\n\n\n More in Technology\n\n\n\n\nAI DRONES TECHNOLOGY TECHNOLOGY\nTesla wants to make humanoid robots. Here\u2019s Elvis (the helicopter) is cheating death by Director Of FBI Addresses Congress About San Your McDonald\u2019s Happy Meal Box Is Now A Virtual\ntheir competition. becoming a drone Bernardino iPhone Reality Headset\nCHARLOTTE HU KELSEY D. ATHERTON XAVIER HARDING CARL FRANZEN\n\n\n\n\nCANCER TECHNOLOGY WEAPONS TECHNOLOGY\nHere\u2019s How Virtual Reality Could Help Doctors Feed Your 3D Printer Recycled Plastic Burn After Shooting: Army Employees Patent This Is The Defining Photo Of Virtual Reality (So\nTreat Cancer XAVIER HARDING Self-Destructing Bullets Far)\nALEXANDRA OSSOLA KELSEY D. ATHERTON CARL FRANZEN\n\n\n\n SEE MORE\n\n\n\n\n SCIENCE\n\n TECHNOLOGY\n\n ENVIRONMENT\nGet the Popular Science daily newsletter\n\ud83d\udca1 EDITORIAL STANDARDS HOW WE TEST AND REVIEW PRODUCTS ABOUT US CONTACT US\n DIY\n\nBreakthroughs, discoveries, and DIY tips sent six GEAR\n TV CHANNEL PRIVACY POLICY TERMS OF USE AFFILIATE DISCLOSURE SITEMAP\ndays a week.\n MERCH\n Enter your email\n SHOP\n\n SIGN UP DISCLAIMER(S)\n ARTICLES MAY CONTAIN AFFILIATE LINKS WHICH ENABLE US TO SHARE IN THE REVENUE OF ANY PURCHASES MADE. REGISTRATION\nBy signing up, you confirm you are ON OR USE OF THIS SITE CONSTITUTES ACCEPTANCE OF OUR TERMS OF SERVICE.\n16+, will receive newsletters and\npromotional content and agree to our \u00a9 2026 ZIFF DAVIS. ALL RIGHTS RESERVED.\nTerms of Use and acknowledge the\n WITHDRAW CONSENT / MANAGE PRIVACY PREFERENCES\ndata practices in our Privacy Policy.\n DO NOT SELL MY PERSONAL INFORMATION\nYou may unsubscribe at any time.\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:28.719236-07:00","document_number":"1","attachment_number":17,"pacer_doc_id":"181037209227","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 16","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294688/","id":490294688,"tags":[],"absolute_url":"/docket/74659430/1/18/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.292459-07:00","date_modified":"2026-08-21T19:36:20.480707-07:00","sha1":"4338549fe7d6f78c72ef7741ace9af39d6c4d9c6","page_count":2,"file_size":13695016,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.18.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.18.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-18 Filed 08/17/26 Page 1 of 2\n\n\n\n\n EXHIBIT\n\n 17\n\f Case 7:26-cv-00318 Document 1-18 Filed 08/17/26 Page 2 of 2\nCareers Explore Jobs Manufacturing AI Terafab Vehicle Software Internships About Us Profile US\n\n\n\n\n Build your Career at Tesla\n\n\n\n\n 14 Results Clear Filters (5)\n Field Service Technician, Utility\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n Search by role or keyword\n\n\n Job Category Field Service Technician, Crew Based, Commercial\n Learn More\n Energy - Solar & Storage\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n\n\n Job Type\n Licensed Journeyman Electrician, Commercial Self Perform\n Learn More\n - Select - Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n\n Region\n Construction Manager, Energy Projects\n Learn More\n North America\n Energy - Solar & Storage \u30fb Part-Time Austin, Texas\n\n\n Location\n Excavator Operator, Commercial Self Perform\n United States of America Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n State\n\n Texas Carpenter & Concrete Finisher, Commercial Self Perform\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\n City Search by ZIP Code\n\n Austin Technical Support Engineer, Commercial Charging\n Learn More\n Energy - Solar & Storage \u30fb Full-Time Austin, Texas\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:38.346728-07:00","document_number":"1","attachment_number":18,"pacer_doc_id":"181037209228","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 17","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294689/","id":490294689,"tags":[],"absolute_url":"/docket/74659430/1/19/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.307491-07:00","date_modified":"2026-08-21T18:44:21.434689-07:00","sha1":"bbfd9fdb4424a1ea41cb39a725c87792b2d604b4","page_count":6,"file_size":7347993,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.19.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.19.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 1 of 6\n\n\n\n\n EXHIBIT\n\n 18\n\f Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 2 of 6\n\n\n\n\nBUSINESS\n\n\nTesla now employs 20,000 in Austin, could triple that as\nit ramps up Cybertruck production\nBy Eric Killelea , Staff writer\nUpdated Sep 25, 2023 3:54 p.m.\n\n\n\n\nEmployees at Tesla\u2019s factory in Austin pose in mid-July with a Cybertruck the company said was the first to be built at\nits Gigafactory Texas in Austin. A company official says the plant now employs about 20,000 people \u2014 and that number is\nexpected to triple.\nTesla/Courtesy\n\n\n\n\n Listen Now:\n Tesla now employs 20,000 in Austin, could triple that as it ramps up\n 1x\n Cybertruck production\n About 5 Minutes\n\n\n\n\n Tesla Inc. now employs 20,000 people at its plant near Austin and has the potential to\n triple that as it\u2019s quickly become one of the largest employers in Central Texas, the\n Austin Regional Manufacturers Association said.\n\n\n\n\n Want more Express-News?\n\n Add Preferred Source\n\n\n\n\n MORE NEWS\n Watch More\n San Antonio doctor\n loses 4 properties to\n foreclosure with bids\n totaling $26M\n\f Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 3 of 6\nThe electric vehicle maker has added more than 7,700 employees since the start of the International\n homebuyers are\nyear, when it had just shy of 12,300, according to the Texas-based company\u2019s annual flocking to Texas.\ncompliance report to Travis County economic development officials. Tesla\u2019s growth in Where are they moving\n from?\nTexas has been rapid; it reported about 3,500 employees there in 2021, the year it\nbegan production. Home buyers need less\n income to buy starter\n homes, according to\n ADVERTISEMENT Redfin report\n Article continues below this ad\n\n Governor's aim to break\n up CPS rattles S.A.\n business community\n\n\n Alamo Heights\n neighbors push back on\n planned 5-story\n apartment complex\n\n\n\n\nCurrent employment was disclosed last week by the plant\u2019s director of manufacturing\nduring a meeting of the manufacturers association.\n\n\nWith this year\u2019s new hires, Elon Musk-led Tesla is now Austin\u2019s second-largest private\nemployer, just behind San Antonio\u2019s H-E-B, which earlier this year reported 22,955\nemployees in Austin, according to the Austin Business Journal. Dell Technologies\nemployed 13,000 workers in the metro last year, according to filings with the U.S.\nSecurities and Exchange Commission.\n\n\nRELATED: Heading toward Cybertruck, Tesla triples employment and builds 4,000\nModel Ys a week in Texas\n SALE: ONLY 25\u00a2 Sign in\n\n\n\n\n Want more Express-News?\n Make us a Preferred Source on Google to see more of us when you search.\n\n\n Add Preferred Source\n\n\n\n\nTesla now is stepping up its hiring efforts to push employment to 60,000 at its sprawling\nAustin plant as it continues producing its Model Y mid-sized SUV and ramps up\nproduction of its long-delayed Cybertruck.\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\nThe disclosure reportedly was made by Jason Shawhan, Tesla\u2019s director of\nmanufacturing, when he spoke at the manufacturers association\u2019s 2023 State of\nManufacturing Conference and Expo last week.\n\n\nNeither Tesla nor the Austin Regional Manufacturers Association responded to requests\nfor comment Monday.\n\n\n845 openings Watch More\n\n\n\n\nAs of Monday morning, Tesla had about 845 job openings listed in Texas on its website.\nMost of those were based in Austin.\n\n\n ADVERTISEMENT\n\f Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 4 of 6\n Article continues below this ad\n\n\n\n\nAmong them are Austin metro area-based positions for manufacturing engineers,\nconstruction managers, supply chain parts advisors, environmental specialists, project\ndesigners for the company\u2019s solar and EV charging stations, attorneys \u2014 even\nan athletic trainer to focus on \u201csetting the right conditions of work \u2014 from\nmanufacturing and maintenance to sales and service.\u201d\n\n\nRELATED: Tesla continues growth, will lease 1 million square feet in Hays County\nindustrial park\n\n\nTesla relocated to Texas from California in 2021 and began limited production of the\nModel Y at its sprawling factory later that year. By the end of 2022, the company said it\nhad invested $5.81 billion into the sprawling facility.\n\n\nDevelopment has continued in 2023. In January, Tesla said it was planning to spend\n$775 million on five construction projects covering 1.5 million square feet of its land off\nTexas 45 and U.S. 130 near Austin-Bergstrom International Airport. It also plans to open\na new cathode production facility at the site this year. Cathodes are a key component\nfor the batteries that power its electric vehicles.\n\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\nMusk also told investors the company had installed Cybertruck production equipment at\nthe Texas plant in anticipation of kicking off production this summer and ramping up\nmanufacturing efforts by 2024.\n\n\nAs Tesla moves toward production of the long-delayed pickup, the company on Monday\nhad 65 Austin job openings specifically related to manufacture of the stainless-steel\nelectric truck.\n\n\nElsewhere in Texas\nThe company also is building a $375 million lithium refinery in Robstown, near Corpus\nChristi. Musk has said it would begin producing enough of the material for batteries in\nabout 1 million EVs a year by 2025.\n\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n Watch More\n\f Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 5 of 6\n As of Monday, Tesla had 39 job openings in Robstown.\n\n\n RELATED: Greg Abbott and Elon Musk celebrate construction of Tesla\u2019s lithium refinery\n in South Texas\n\n\n Tesla\u2019s hiring push comes as tech giants with offices in the Austin area have been laying\n off workers. Google, Microsoft, Amazon, Facebook parent Meta and Dell all have\n announced cuts. It also reflects Musk\u2019s expanding business footprint across the state.\n\n\n Hawthorne, Calif.-based SpaceX, where Musk is CEO, had about 1,700 employees at\n its Starbase facility in Boca Chica in April, according to former Brownsville Mayor Juan\n \u201cTrey\u201d Mendez III.\n\n\n ADVERTISEMENT\n Article continues below this ad\n\n\n\n\n As of Monday, SpaceX was advertising 103 openings for its Starship program in Boca\n Chica and eight at its rocket testing site in McGregor, near Waco.\n\n\n Other Musk-owned businesses with Texas operations \u2014 tunneling firm The Boring Co.\n and brain implant company Neuralink \u2014 also were advertising dozens of job openings\n Monday.\n\n\n Sep 25, 2023 | Updated Sep 25, 2023 3:54 p.m.\n\n\n Eric Killelea\n\n\n\n\nAround the Web Powered by\n\n\n\n\nAfter 60, Leg Strength Comes From The Lawn Problem Most Neurologists Beg Seniors With\nOne Simple Daily Move Homeowners Notice but Can't Fix Neuropathy: Stop Doing This Now\nBy ApexLabs By Glosrity By Health Weekly\n\n\n\n\nSurgeons: This Simple Trick Will Honey: The Greatest Enemy of Women Are Obsessed With These\nEnd Knee Pain & Arthritis Quickly Memory Loss (See How to Use It) Beautiful Floral Caps\n(Try It) By Health Weekly\n By Peoasis\n\nBy Health Weekly\n\n\n\n\n Watch More\n\n\n\n\nA 78-Year-Old Master Craftsman Sciatica is Not From a Slipped 1 Simple Tip to Cut Your Electric\nMade This Hummingbird House. Disc. Meet The Real Enemy of Bill (Try Tonight)\nThen This Happened Sciatica (Stop This) By MadeInGenius\nBy Ribili By SmoothSpine\n\f Case 7:26-cv-00318 Document 1-19 Filed 08/17/26 Page 6 of 6\n\n\n\n\nDoctor Begs Seniors: Do This to Neuropathy is Not From Low Endocrinologist: If You Have\nStop Losing Muscle Vitamin B (Meet The Real Enemy) Diabetes, Read This Before It's\nBy ApexLabs By Health Weekly Removed!\n By Health Weekly\n\n\n\n\nMOST POPULAR\n SAN ANTONIO REAL ESTATE\n Alamo Heights neighbors push back on planned 5-\n story apartment complex\n Homeowners near the site of the proposed development want\n the city to enforce its heights restrictions; the developer says\n the project isn\u2019t feasible without keeping all five floors.\n\n\n\n\nNEWS FORECASTS\n\nBankruptcy court records offer first look at Camp Mystic\u2019s finances When is South Texas\u2019 next rain chance? Here\u2019s the\n latest update\nThe Texas Hill Country retreat\u2019s gross revenue dropped from $11 million in 2024 to $5.7 million in\n2025, the year that a flash flood killed 27 guests and the camp\u2019s owner.\n\n\nBUSINESS NEWS\nInternational\nNEWS homebuyers are Annelise\n BUSINESS Camp, 2-year-old at\nflocking to Texas. Where are center of Texas brain-death\n\u2018Your body belongs to me\u2019: Governor's aim to break up\nthey moving from? battle, dies\nFeds release incarcerated CPS rattles S.A. business\npastor's explicit texts community\n\n\n Return To Top\n\n About Contact Services Account\n\n Our Company Customer Service Archives Subscribe\n\n Careers Frequently Asked Questions Newspaper Archive Newsletters\n\n Our Use of AI Newsroom Contacts Advertising e-Edition\n\n Standards and Practices Copyright and Reprints Photo Store\n\n\n\n\n \u00a9 2026 Hearst Newspapers, LLC Terms of Use Privacy Notice DAA Industry Opt Out Your Privacy Choices (Opt Out of Sale/Targeted Ads)\n\n\n\n\n Watch More\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:32.987582-07:00","document_number":"1","attachment_number":19,"pacer_doc_id":"181037209229","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 18","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294690/","id":490294690,"tags":[],"absolute_url":"/docket/74659430/1/20/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.322216-07:00","date_modified":"2026-08-21T19:46:47.778421-07:00","sha1":"7cf27f8da32ac4577981a1f72026b6a773a16d32","page_count":3,"file_size":11994330,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.20.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.20.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-20 Filed 08/17/26 Page 1 of 3\n\n\n\n\n EXHIBIT\n\n 19\n\f Case 7:26-cv-00318 Document 1-20 Filed 08/17/26 Page 2 of 3\nCareers Explore Jobs Manufacturing AI Terafab Vehicle Software Internships About Us Profile US\n\n\n\n\n Gigafactory\n Austin, Texas\n\f Case 7:26-cv-00318 Document 1-20 Filed 08/17/26 Page 3 of 3\n\n\n Join Us at Our Global Headquarters\n Covering 2,500 acres along the Colorado River with over 10 million square feet of factory\n floor, Gigafactory Texas is a U.S. manufacturing hub for Model Y and the home of Cybertruck.\n\n It doesn\u2019t matter where you come from, where you went to school or what industry you\u2019re in\u2014we\u2019re\n hiring individuals of all levels. If you\u2019ve done exceptional work, join us in solving the next generation of\n engineering, manufacturing and operational challenges.\n\n\n\n View Jobs\n\n\n\n\nManufacturing\nProducing vehicles quickly, efficiently and safely is one of the biggest challenges our Manufacturing team\nfaces. Implement manufacturing processes that evolve to meet rising global demand.\n\n\n\nSupervisor Technician Production Associate\nApply to lead a motivated team of individuals who are Apply to improve the devices, equipment and systems Apply to work across all areas of the vehicle\nresponsible for achieving our ambitious quality and that are critical to the efficiency of our production manufacturing process\u2014on-the-job training provided,\nproduction goals. lines. no prior experience required.\n\n\n\n\nEngineering From powerful battery systems to in-car technology, our Engineering team is developing new processes,\n equipment and tooling to help us shape the future of sustainable transportation. Create the next Cybertruck\n or continue to improve our current vehicle lineup.\n\n Learn more about Tesla battery cell development and apply for open positions to join our Cell team.\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:31.864558-07:00","document_number":"1","attachment_number":20,"pacer_doc_id":"181037209230","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 19","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294691/","id":490294691,"tags":[],"absolute_url":"/docket/74659430/1/21/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.336977-07:00","date_modified":"2026-08-21T19:20:50.490580-07:00","sha1":"45b593686e2cb10248c417349d7a2471c1681e05","page_count":3,"file_size":650572,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.21.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.21.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document1-21 Filed 08/17/26 Pagelof3\n\nEXHIBIT\n\n20\n\fCase 7:26-cv-00318 Document1-21 Filed 08/17/26 Page2of3\n\nUNITED STATES DISTRICT COURT\nWESTERN DISTRICT OF TEXAS\nMIDLAND/ODESSA DIVISION\n\nNEURAL AI, LLC\n\nPlaintiff, Case No. 7:24-cv-00221-ADA-DTG\n\u2122 JURY TRIAL DEMANDED\nNVIDIA CORPORATION\n\nDefendant.\n\nI, Alon Daks, hereby declare as follows:\n\n1. I am a Senior Staff Software Engineer at Tesla Inc. (\u201cTesla\u201d). Iam over the age of\neighteen and competent to make this declaration. I make this declaration based on my personal\nknowledge, including my general familiarity with Tesla\u2019s technical infrastructure, software\ndeployments, use of NVIDIA products, as well as the on the basis of a reasonably diligent\ninvestigation.\n\n2. I understand that this declaration is submitted in connection with a subpoena Neural\nAl served on Tesla on June 25, 2026 in the above-captioned action (the \u201cSubpoena\u201d.\n\nq. In the ordinary course of business, Tesla deploys the following NVIDIA GPUs in\nthe United States: A100, H100, H200, GB300, V100, and A16.\n\n4. In the ordinary course of business, Tesla uses the following software and/or\nlibraries identified in the Subpoena (specifically, Request No. 2 for Production of Documents):\ncuBLAS, cuDNN, cuFFT, cuSOLVER, Isaac Lab, PyTorch, and Triton. I understand that\ncuBLAS, cuDNN, cuFFT, cuSOLVER, and Isaac Lab are provided by NVIDIA, but PyTorch is\n\nmanaged by the Linux Foundation and Triton is managed by https://triton-lang.org.\n\fCase 7:26-cv-00318 Document1-21 Filed 08/17/26 Page 3of3\n\n5. In the ordinary course of business, Tesla uses at least cuBLAS, cuDNN, cuFFT,\ncuSOLVER, and Triton as provided. I understand that Isaac Lab, PyTorch, and Triton are open\nsource software and/or libraries, but cuBLAS, cuDNN, cuFFT, and cuSOLVER are not open\n\nsource.\n\nI declare under penalty of perjury under the laws of the United States of America that the\n\nforegoing is true and correct.\n\nExecuted on August 10, 2026.\n\nAlon Daks\n\nAlon Daks (Aug 10, 2026 18:35:30 PDT)\n\nAlon Daks\n","ocr_status":1,"date_upload":"2026-08-17T14:54:40.418747-07:00","document_number":"1","attachment_number":21,"pacer_doc_id":"181037209231","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 20","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294692/","id":490294692,"tags":[],"absolute_url":"/docket/74659430/1/22/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.357906-07:00","date_modified":"2026-08-21T19:20:58.629941-07:00","sha1":"581aba1682d8d60db6284a78898f34a4374373d9","page_count":14,"file_size":242583,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.22.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.22.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":"Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 1 of 14\n\n\n\n\n EXHIBIT\n\n 21\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 2 of 14\n\n\n Monday, August 17, 2026 at 6:52:05 AM Paci\ufb01c Daylight Time\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\nDate: Tuesday, August 11, 2026 at 4:18:05 PM Paci\ufb01c Daylight Time\nFrom: Tanner Laiche\nTo: Jun Zheng, Rocco Magni, Gina Cremona\nCC: Emily Portuguese, Tamar Lusztig, Brian Melton, Max Tribble, Samuel Drezdzon, Richard Wojtczak, Rachel\n Hanna, Ashraf Fawzy\nAttachments: Outlook-6C509A71.png, Outlook-6C509A71.png\n\n\nCounsel,\n\nWe have reviewed the declaration. Neural AI does not consider this matter concluded. The\ndeclaration is materially insu`icient in several respects:\n\n a. The declaration addresses only GPU and partial software identi\ufb01cation. It omits\n entirely the paragraphs addressing unmodi\ufb01ed use of NVIDIA software (draft \u00b6 7),\n hardware/software functioning as designed by NVIDIA (draft \u00b6 8), use of NVIDIA-\n distributed pretrained models (draft \u00b6 9), CPU/GPU memory architecture (draft \u00b6\u00b6\n 10\u201311), standard data \ufb02ow from CPU memory to GPU memory including th use of\n GPUDirect (draft \u00b6\u00b6 12\u201313), U.S. operations (draft \u00b6 14), and frequency of use (draft \u00b6\n 15). These are key issues within the scope of the subpoena that you unilaterally\n refused to address.\n j. The draft declaration organized software into three categories (draft \u00b6\u00b6 4, 5, 6)\n re\ufb02ecting distinct layers of the accused stack. Tesla collapsed these into a single\n undi`erentiated list and did not address application-layer platforms.\n k. The declaration states that PyTorch and Triton are not NVIDIA-provided without\n con\ufb01rming whether Tesla uses them on NVIDIA GPUs. NVIDIA distributes and\n optimizes PyTorch for use within its GPU-accelerated ecosystem, and Neural AI\u2019s\n infringement contentions identify PyTorch as part of the accused software\n stack. Additionally, \"Triton\" refers to NVIDIA Triton Inference Server\u2014not the Triton\n compiler language\u2014which the declaration does not address.\n\nNeural AI's o`er to accept a declaration was expressly \"subject to resolving any material\ngaps.\" The gaps identi\ufb01ed above are material.\n\nGiven these de\ufb01ciencies, Neural AI intends to pursue (1) document production responsive\nto the outstanding requests, and (2) deposition testimony from a Tesla Rule 30(b)(6)\ndesignee. The testimony subpoena noticed for September 14, 2026 remains in e`ect.\n\nWe are available to meet and confer this week to resolve these issues short of motion\npractice. Please provide available times.\n\nRegards,\n\n 1 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 3 of 14\n\n\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\nThis e-mail contains privileged and con\ufb01dential information, which may be subject to the attorney-client\nprivilege and/or attorney work product protection. If you received this message in error, please notify the\nsender and delete it immediately.\n\n\nFrom: Jun Zheng \nDate: Monday, August 10, 2026 at 11:57 PM\nTo: Tanner Laiche ; Rocco Magni\n; Gina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\nCounsel,\n\nAs discussed during the meet-and-confers on July 28 and Aug. 7, and in view of Tesla\u2019s objections\nserved on July 21, Neural AI stated it would accept a declaration in lieu of document production and\ndeposition testimony. Accordingly, Tesla conducted a reasonable investigation into which NVIDIA\nGPUs and software identi\ufb01ed in Neural AI\u2019s subpoena Tesla uses and whether Tesla uses that\nsoftware o` the shelf. The attached declaration provides that information. Tesla considers this\nmatter concluded.\n\nRegards,\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\nFrom: Tanner Laiche \nSent: Thursday, August 6, 2026 12:14 PM\nTo: Jun Zheng ; Rocco Magni ;\nGina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n\n 2 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 4 of 14\n\n\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nCounsel,\n\nI am available tomorrow at 9:30 am CT. I will circulate a calendar invite.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\nFrom: Jun Zheng \nDate: Wednesday, August 5, 2026 at 1:31 PM\nTo: Tanner Laiche ; Rocco Magni\n; Gina Cremona \nCc: Emily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\nTanner,\n\nWe are available to meet and confer this Friday between 9-10 AM CT. Please let us\nknow if that works for your team.\n\nThanks,\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n 3 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 5 of 14\n\n\n\n\nFrom: Tanner Laiche \nSent: Tuesday, August 4, 2026 1:28 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Jun Zheng ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nCounsel,\n\nI am following up on Rocco\u2019s email.\n\nDespite the parties\u2019 prior meet-and-confers, document discovery in the underlying action\ncloses on August 11. Unless the parties can promptly reach a resolution, that deadline\nleaves Neural AI no practical alternative but to move to compel by the end of this week or,\nat the latest, August 10, to preserve its rights.\n\nTo reduce burden and potentially avoid motion practice, I am attaching a short set of\nquestions intended to guide your investigation and help identify the responsive\ninformation, and also recirculating the draft declaration we previously shared, and that\nTesla may revise to ensure its accuracy.\n\nIf Tesla commits to provide an executed declaration, Neural AI is willing to consider\naccepting the declaration in lieu of further document production and/or deposition\ntestimony, subject to resolving any material gaps. Otherwise, the discovery deadline will\nforce Neural AI to move to compel by or before August 10 to preserve its rights. Even if a\nmotion becomes necessary, we remain open to resolving the issues promptly and mooting\nor withdrawing the motion through compliance.\n\nWe are available this week to further meet and confer as necessary.\n\nRegards,\n\nTanner Laiche\nSusman Godfrey LLP\n206.505.3816 | tlaiche@susmangodfrey.com\n401 Union Street | Suite 3000 | Seattle, WA 98101\nHOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n 4 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 6 of 14\n\n\n\nFrom: Rocco Magni \nDate: Sunday, August 2, 2026 at 6:57 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nOur discovery deadline is approaching soon. Please let us know when you can confer again\nthis upcoming week. Thanks.\n\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Rocco Magni \nDate: Tuesday, July 28, 2026 at 6:52 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nGina and Ashraf,\n\nThanks for speaking today. Attached is a draft of the declaration I referred to on our call.\n\nBest,\n\n\n 5 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 7 of 14\n\n\nRocco\n\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\nFrom: Gina Cremona \nDate: Thursday, July 23, 2026 at 9:45 AM\nTo: Rocco Magni \nCc: Jun Zheng ; Tanner Laiche ;\nEmily Portuguese ; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak ;\nRachel Hanna ; Ashraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to Tesla\n\nEXTERNAL Email\n\nHi Rocco, we are not available today.\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 22, 2026 12:01 PM\nTo: Gina Cremona \nCc: Jun Zheng ; Tanner Laiche\n; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton ; Max\nTribble ; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna ;\nAshraf Fawzy \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) - Subpoena to\nTesla\n\nWould tomorrow work?\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\n\n 6 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 8 of 14\n\n\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\nThis e-mail may contain privileged and confidential information. If you received this message in\nerror, please notify the sender and delete it immediately.\n\n\n\n On Jul 22, 2026, at 2:54 PM, Gina Cremona wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n We are not available on Friday, but can meet on Tuesday, July 28 between 8-10 am PT.\n\n Thanks,\n Gina\n\n\n From: Rocco Magni \n Sent: Wednesday, July 22, 2026 6:54 AM\n To: Jun Zheng ; Gina Cremona\n \n Cc: Tanner Laiche ; Emily Portuguese\n ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n ; Ashraf Fawzy ; Jun\n Zheng \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n Jun,\n\n Please provide times to meet and confer on Friday 7/24. Thanks.\n\n --\n Rocco F. Magni\n\n 7 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 9 of 14\n\n\nPartner | Susman Godfrey LLP\nO\ufb03ce: 713.653.7861\nCell: 512.514.3519\nFirm Bio\nThis e-mail may contain privileged and confidential information. If you received this\nmessage in error, please notify the sender and delete it immediately.\nFrom: Jun Zheng \nDate: Tuesday, July 21, 2026 at 7:37 PM\nTo: Rocco Magni ; Gina Cremona\n\nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nEXTERNAL Email\nRocco,\n\nWe understand from the email exchange below that the noticed date for\ntestimony subpoena is now Sept. 14, 2026. In the meantime, attached please \ufb01nd\nTesla's objections and responses to Neural AI's subpoena.\n\nThanks!\n\nJun Zheng\nSr. Counsel, IP Litigation\n1 Tesla Road, Austin, TX 78725\nE. zhengjun@tesla.com\n\n\n\n\n\nFrom: Gina Cremona \nSent: Monday, July 20, 2026 1:07 PM\nTo: Rocco Magni \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n\n 8 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 10 of 14\n\n\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n; Ashraf Fawzy ; Jun\nZheng \nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nHi Rocco,\n\nTesla cannot provide an agreed date for deposition until we have reviewed and\nresponded to the subpoenas. However, we can agree to Sept. 14 as the noticed date\nfor testimony subpoena for now, subject to modi\ufb01cation once we have responded to\nthe document subpoena.\n\nRegards,\nGina\n\n\n\nFrom: Rocco Magni \nSent: Wednesday, July 15, 2026 12:58 PM\nTo: Gina Cremona \nCc: Tanner Laiche ; Emily Portuguese\n; Tamar Lusztig\n; Brian Melton\n; Max Tribble\n; Samuel Drezdzon\n; Richard Wojtczak\n; Rachel Hanna\n\nSubject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\nSubpoena to Tesla\n\nGina,\n\nWe\u2019ll agree to pull down the July 28 date once we have an agreed replacement date.\nLet us know what date works for you and we\u2019ll withdraw the current notice.\nThanks.\n--\nRocco F. Magni\nPartner | Susman Godfrey LLP\nOffice: 713.653.7861\nCell: 512.514.3519\n\nFirm Bio\n\n 9 of 13\n\f Case 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 11 of 14\n\n\n\nThis e-mail may contain privileged and confidential information. If you received\nthis message in error, please notify the sender and delete it immediately.\n\n\n\n On Jul 15, 2026, at 3:54 PM, Gina Cremona \n wrote:\n\n\n\n\n EXTERNAL Email\n\n Hi Rocco,\n\n Understood. To con\ufb01rm, the deposition date of July 28 is o` calendar, and\n we will work on agreeing to a new date.\n\n Regards,\n Gina\n\n\n From: Rocco Magni \n Sent: Tuesday, July 14, 2026 5:28 PM\n To: Gina Cremona ; Tanner Laiche\n ; Emily Portuguese\n \n Cc: Tamar Lusztig ; Brian\n Melton ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Ms. Cremona,\n\n Discovery has not been extended; only depositions. Written\n discovery will still close on August 11.\n\n We can work with you on a deposition date between now and\n September 14. But we do not have \ufb02exibility on the timing of your\n RFP responses and document production beyond the 1 week\n extension we noted below.\n\n\n 10 of 13\n\fCase 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 12 of 14\n\n\n Best,\n\n Rocco\n\n --\n Rocco F. Magni\n Partner | Susman Godfrey LLP\n O\ufb03ce: 713.653.7861\n Cell: 512.514.3519\n Firm Bio\n This e-mail may contain privileged and confidential information. If you\n received this message in error, please notify the sender and delete it\n immediately.\n From: Gina Cremona \n Date: Tuesday, July 14, 2026 at 8:19 PM\n To: Tanner Laiche ; Emily\n Portuguese \n Cc: Rocco Magni ; Tamar Lusztig\n ; Brian Melton\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D.\n Tex.) - Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n It has come to our attention that the discovery deadline in this case has\n been extended. Due to summer vacations and people being out of the\n o`ice, we renew our request for an additional 2-week extension such that\n our deadlines will be Aug. 4 and Aug. 11.\n\n Regards,\n Gina\n\n\n From: Tanner Laiche \n Sent: Monday, July 6, 2026 5:20 PM\n To: Gina Cremona ; Emily Portuguese\n \n Cc: Rocco Magni ; Tamar\n Lusztig ; Brian Melton\n\n 11 of 13\n\fCase 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 13 of 14\n\n\n ; Max Tribble\n ; Samuel Drezdzon\n ; Richard Wojtczak\n ; Rachel Hanna\n \n Subject: Re: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221\n (W.D. Tex.) - Subpoena to Tesla\n\n Counsel,\n\n Thanks for reaching out. Given the upcoming close of fact\n discovery, Neural AI is not able to agree to a three-week\n extension. That said, we can agree to a one-week extension for\n Tesla\u2019s written objections/responses to the subpoena(s). A copy\n of the Protective Order is attached.\n\n Regards,\n\n Tanner Laiche\n Susman Godfrey LLP\n 206.505.3816 | tlaiche@susmangodfrey.com\n 401 Union Street | Suite 3000 | Seattle, WA 98101\n HOUSTON \u2022 LOS ANGELES \u2022 SEATTLE \u2022 NEW YORK\n\n\n From: Gina Cremona \n Date: Thursday, July 2, 2026 at 7:21 PM\n To: Emily Portuguese \n Subject: Neural AI, LLC v. Nvidia Corp., 7:24-cv-00221 (W.D. Tex.) -\n Subpoena to Tesla\n\n EXTERNAL Email\n Counsel,\n\n We are in receipt of your Subpoena to Produce Documents and Subpoena\n for Testimony. Due to the holiday weekend and vacation schedules, we\n request a three-week extension to respond such that our deadlines will be\n Aug. 4 and Aug. 11.\n\n Additionally, please provide us with a copy of the protective order.\n\n Regards,\n Gina\n\n Gina H. Cremona\n Senior Counsel, IP Litigation\n 1501 Page Mill Rd., Palo Alto, CA 94304\n E. gcremona@tesla.com T. 650.647.0015\n\n\n 12 of 13\n\fCase 7:26-cv-00318 Document 1-22 Filed 08/17/26 Page 14 of 14\n\n\n\n \n\n\n\n\n 13 of 13\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:31.196311-07:00","document_number":"1","attachment_number":22,"pacer_doc_id":"181037209232","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Exhibit 21","acms_document_guid":""},{"resource_uri":"https://www.courtlistener.com/api/rest/v4/recap-documents/490294693/","id":490294693,"tags":[],"absolute_url":"/docket/74659430/1/23/neural-ai-llc-v-tesla-inc/","date_created":"2026-08-17T14:19:04.373173-07:00","date_modified":"2026-08-21T19:20:44.999988-07:00","sha1":"044b719bc6af90a19f4a26e8a0bcebc68983fc8c","page_count":2,"file_size":219982,"filepath_local":"recap/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.23.pdf","filepath_ia":"https://archive.org/download/gov.uscourts.txwd.1172927763/gov.uscourts.txwd.1172927763.1.23.pdf","ia_upload_failure_count":null,"thumbnail":null,"thumbnail_status":0,"plain_text":" Case 7:26-cv-00318 Document 1-23 Filed 08/17/26 Page 1 of 2\n\n\n\n\n IN THE UNITED STATES DISTRICT COURT\n FOR THE WESTERN DISTRICT OF TEXAS\n\n NEURAL AI, LLC,\n Misc. Case No.\n Petitioner,\n Principal case pending in Western District of\n v. Texas, Civil Action No. 7:24-cv-00221-LS-\n DTG\n TESLA, INC.,\n\n Respondent.\n\n\n\n [PROPOSED] ORDER ON PETITIONER'S MOTION TO COMPEL\n COMPLIANCE WITH SUBPOENA SERVED ON THIRD-PARTY TESLA, INC.\n\n Before the Court is Petitioner Neural AI, LLC\u2019s Motion to Compel Compliance with\n\nSubpoena Served on Third-Party Tesla, Inc. Having considered the Motion, the responses and\n\nobjections thereto, and the applicable law, the Court finds that the Motion should be and hereby is\n\nGRANTED.\n\n IT IS THEREFORE ORDERED that Third-Party Tesla, Inc. shall produce nonprivileged\n\ndocuments responsive to Requests for Production Nos. 1\u201312 on a rolling basis, with production to\n\nbegin within seven (7) days of the date of this Order and be completed within twenty-one (21)\n\ndays of this Order.\n\n IT IS FURTHER ORDERED that Third-Party Tesla, Inc. shall designate and produce a\n\nknowledgeable witness for deposition testimony on Deposition Topics 1\u20135 within thirty (30) days\n\nof the date of this Order.\n\n\n\n\n 1\n\f Case 7:26-cv-00318 Document 1-23 Filed 08/17/26 Page 2 of 2\n\n\n\n\nSO ORDERED: this _____ day of _____________________, 2026.\n\n\n\n\n United States District Judge\n\n\n\n\n 2\n\f","ocr_status":1,"date_upload":"2026-08-17T14:54:32.310812-07:00","document_number":"1","attachment_number":23,"pacer_doc_id":"181037209234","is_available":true,"is_free_on_pacer":null,"is_sealed":null,"document_type":2,"description":"Proposed Order","acms_document_guid":""}],"date_created":"2026-08-17T14:05:48.230485-07:00","date_modified":"2026-08-17T14:18:44.152782-07:00","date_filed":"2026-08-17","time_filed":"15:23:31","entry_number":1,"recap_sequence_number":"2026-08-17.001","pacer_sequence_number":3,"description":"MOTION to Compel Compliance with Subpoena Served on Third-Party Tesla, Inc. by Neural AI, LLC. (Attachments: # 1 Affidavit of Tanner Laiche, # 2 Exhibit 1, # 3 Exhibit 2, # 4 Exhibit 3, # 5 Exhibit 4, # 6 Exhibit 5, # 7 Exhibit 6, # 8 Exhibit 7, # 9 Exhibit 8, # 10 Exhibit 9, # 11 Exhibit 10, # 12 Exhibit 11, # 13 Exhibit 12, # 14 Exhibit 13, # 15 Exhibit 14, # 16 Exhibit 15, # 17 Exhibit 16, # 18 Exhibit 17, # 19 Exhibit 18, # 20 Exhibit 19, # 21 Exhibit 20, # 22 Exhibit 21, # 23 Proposed Order)(Magni, Rocco) (Entered: 08/17/2026)","tags":[]}],"entries_total":"https://www.courtlistener.com/api/rest/v4/docket-entries/?count=on&docket=74659430&page_size=40"}