Teslaについて過去ツイートを纏めてみました (Memo from my past tweets)。
Tesla Memo [2/3] : before Tesla AI Day, Aug 19, 2021
Tesla Memo [3/3] : Tesla AI Day, Aug 19, 2021
概要
- "Accelerated Mathematical Engine", Tesla, Patent, Jun 2, 2020 (Jul 24, 2017)
- Full Self-Driving (FSD) Chip, Tesla Autonomy Investor Day, Apr 22 2019
- Full Self-Driving (FSD) Chip, Hot Chips 2019, Aug 20, 2019
- "Instruction Set Architecture for a Vector Computational unit", Tesla Patent Application, Aug 1, 2019 (Jun 1, 2018)
- "Dojo cluster", Andrej Karpathy, PyTorch Developer Conference, Oct 10, 2019
- "Dojo, our training supercomputer", Elon Musk, Feb 3, 2020, and Aug 15, 2020
- "Tesla's next-gen Autopilot and infotainment hardware already in development: report", Aug 18, 2020, 財經要聞
- TSMC InFO_SoW (System-on-Wafer), ECTC 2020, May 2020
- "Dojo uses our own chips & a computer architecture optimized for neural net training", Elon Mask, Sep 21, 2020
- ※ページ内リンク設定が解らないのでスクロールして下さい。
"Accelerated Mathematical Engine", Tesla, Patent, Jun 2, 2020 (Jul 24, 2017)
特許を眺めていたら Teslaの行列演算器の特許申請を発見、この時点はさして気にせず…
"Accelerated mathematical engine", US20190026078A1Peter Joseph Bannon, Kevin Altair Hurd, Emil TALPES
2019年2月27日 https://twitter.com/ogawa_tter/status/1100591325617938432
2019年2月27日 https://twitter.com/ogawa_tter/status/1100591325617938432
2017-07-24: Priority to US201762536399P...2019-01-24: Publication of US20190026078A12020-06-02: Publication of US10671349B22020-06-02: Application granted
2020年6月2日に特許認定。↓は認定された特許から。
"Computational array microprocessor system with variable latency memory access", US20190026237A1
Emil TALPES, Peter Joseph Bannon, Kevin Altair Hurd
https://patents.google.com/patent/US20190026237A1/en2017-07-24: Priority to US201762536399P...2019-01-24: Publication of US20190026237A1”
"Computational array microprocessor system using non-consecutive data formatting", US20190026249A1
Emil TALPES, William McGee, Peter Joseph Bannon
2017-07-24: Priority to US201762536399P...2019-01-24; Publication of US20190026249A1
"Vector computational unit”, US20190026250A1
Debjit Das Sarma, Emil TALPES, Peter Joseph Bannon
2017-07-24; Priority to US201762536399P...2019-01-24: Publication of US20190026250A1
現時点で特許として認定されているのは冒頭の一つだけですが、Priorityは 4件全て 2017-07-24、全てに Peter Joseph Bannonさんと Emil TALPESさんのお名前があります。
Pete Bannonさんは 1984年に DEC入社、DECを買収した Compaqと Intelの Alpha関連知財の譲渡により Intelに移籍
"Compaq and Intel to Accelerate Enterprise Server Roadmaps", Jun 25, 2001
"Compaq to Transfer Technology to Intel and Move Its 64-bit Servers to Intel Itanium processor family, Vows to Redefine Enterprise Server Price/Performance"
"Intel gets more key Alpha alums", Jul 3, 2003
"Pete Bannon, one of the key architects behind the touted Alpha processor, and a number of other Alpha engineers are joining Intel to work on future versions of Itanium."
1年余りで P. A. Semi、Appleによる買収を経て 2016年2月から Teslaの Directorに
Debjit Das Sarmaさん、Emil TALPESさん、William McGeeさんらは次エントリーの学習向けエンジン特許の共同発明者ですが (LinkedInによると 3名とも前職 AMD、2016年2、4月に移籍)、Peter Joseph Bannonさんは含まれていません。
Kevin Altair Hurdさんは、2018年1月から Senior Principal Engineer, Intel
Full Self-Driving (FSD) Chip, Tesla Autonomy Investor Day, Apr 22 2019
Teslaが車載コンピュータチップを独自開発し、Tesla Autonomy Investor Dayで発表：
Tesla Autonomy Investor Day, Apr 22 2019
発表された Peter Joseph Bannonさんは先述の特許全ての共同発明者です。
チップ詳細は次の Hot Chips 2019, Aug 20, 2019で。
"Full Self-Driving (FSD) Chip, Hot Chips 2019, Aug 20, 2019
Hot Chips 2019, Aug 20, 2019での発表：
"Compute and redundancy solution for Tesla's Full Self driving computer",
Pete Bannon, Ganesh Venkataramanan, Debjit Das Sarma, Emil Talpes, Bill McGee & Team, Tesla
発表メンバーや内容から冒頭の特許が関連していると推測されますが、詳細は不明。
さらに IEEE Microに掲載：
"Compute Solution for Tesla's Full Self-Driving Computer", Emil Talpes, Debjit Das Sarma, Ganesh Venkataramanan, Peter Bannon, Bill McGee, Benjamin Floering, Ankit Jalote, Christopher Hsiong, Sahil Arora, Atchyuth Gorti, Gagandeep S. Sachdev
IEEE Micro ( Volume: 40, Issue: 2, March-April 1 2020)
"Instruction Set Architecture for a Vector Computational Unit", Tesla Patent Application, Aug 1, 2019 (Jun 1, 2018)
冒頭の特許と関連しそうな発明者が重なる特許申請を発見：
”Systems and methods for converting a matrix input to a vectorized input for a matrix processor", US10747844B2
Peter Joseph Bannon, William A McGee, Emil TALPES
2017-12-12: Application filed by Tesla Inc...2019-06-13: Publication of US20190179870A12020-08-18: Application granted
2020-08-18: Publication of US10747844B2
"Instruction Set Architecture for a Vector Computational Unit", US20190235866A1
Debjit Das Sarma, Emil TALPES, Peter Joseph Bannon
2018-02-01: Priority to US201862625249P...2019-08-01: Publication of US20190235866A1
冒頭の "Vector computational unit”, US20190026250A1, 2019-01-24 (2017-07-24) 改訂版？
"Dojo cluster", Andrej Karpathy, PyTorch Developer Conference, Oct 10, 2019
Andrej Karpathyさん (Tesla) の PyToch Developer Conference, Oct 10, 2019でのビデオを見ていたら、Dojo clusterの紹介が、ただし詳細は不明：
GPU cluster: On premiseDojo cluster
Neural Network training computer and a chipImproving efficiency by in order magnitude
"Dojo" について検索したところ以下の記事が見つかりましたが、やはり詳細は不明：
"Elon Musk hints at Tesla’s secret project ‘Dojo’ making the difference in race to full self-driving", Aug 5, 2019 https://electrek.co/2019/08/05/tesla-project-dojo-full-self-driving-elon-musk/
記事から
Elon Musk, Aug 4, 2019
"Will Dojo to be the difference"
"Dojo, our training supercomputer", Elon Musk, Feb 3, 2020, and Aug 15, 2020
Elon Muskさんの Dojoについてツイート：
Elon Musk, Feb 3, 2020
"Dojo, our training supercomputer, will be able to process vast amounts of video training data & efficiently run hypersparce arrays with a vast number of parameters, ..."
Elon Musk, Aug 15, 2020
"Tesla is developing a NN training computer called Dojo to process truly vast amounts of video data."
"Tesla's next-gen Autopilot and infotainment hardware already in development: report", Aug 18, 2020
2020年8月17日付け台湾の財經要聞に、Teslaが Broadcomと共同で TSMC InFO System-on-Waferを採用した新しい HPCチップ (7nm) を開発、第四四半期に生産開始との噂が掲載、初期生産 2,000チップ：
"台積電7奈米 獲博通HPC訂單"、2020/08/17、財經要聞、中時新聞網
”博通（Broadcom）與電動車大廠特斯拉（Tesla）共同開發的新款高效能運算（HPC）晶片，將以台積電7奈米先進製程投片，並採用台積電整合型扇出（InFO）系統單晶圓（System-on-Wafer，SoW）先進封裝技術，預計第四季開始生產，初期投片約達2,000片規模。”
TSMC InFO_SoW (System-on-Wafer), ECTC 2020, May 2020
TSMCは 2020年5月に開催された Electronic Components and Technology Conference (ECTC) で HPC向けに InFO_SoW (System-on-Wafer) を発表：
※InFO (Integrated Fan-Out)
"InFO_SoW (System-on-Wafer) for High Performance Computing", TSMC, ECTC 2020
TSMC Exhibits Packaging Prowess at Virtual ECTC 2020, Jun 30, 2020 https://www.3dincites.com/2020/06/iftle-454-tsmc-exhibits-packaging-prowess-at-virtual-ectc-2020/
7 x 7 Dieで実装2x Bandwidth density & 97% lower PDN impedance15% Power saving of the InterconnectsTDP: 7000W (Power density: 1.2W/mm^2)
ECTC 2020でのTSMCパッケージングの方向性についてのキーノート講演 (42分余り)：
"Innovative Heterogeneous Integration Technologies
The New Path Going Forward", Douglas Yu, VP of Integrated Interconnect & Packaging under R&D, TSMC, Keynote, ECTC 2020, May 27
ECTC 2020で発表の Info_SoWもほんの少しだけ紹介。
"Dojo uses our own chips & a computer architecture optimized for neural net training", Elon Mask, Sep 21, 2020
Elon Mask, Sep 21, 2020
"Dojo uses our own chips & a computer architecture ...","Custom ultra high speed bus & efficient cooling"
0 件のコメント:
コメントを投稿