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 [1/3] : Tesla Autonomy Investor Day, Apr 2019 - Sep 2020
Tesla Memo [3/3] : Tesla AI Day, Aug 19, 2021
概要
- "Scalable matrix node engine with configurable data formats", Tesla Patent Application, Nov 5, 2020 (May 5 2019)
- Tesla In-House Supcrecomputer, Andrej Karpathy, Tesla, Jun 20, 2021
- "Looking at holding Tesla AI Day", Elon Mask, Jun 22, 2021
- Tesla AI Day, Aug 19, 2021, Professor Dennis Hong, UCLA
- Dennis Hong教授の投稿画像から推測の関連特許
- "Voltage regulator module with cooling structure", Patent, Jan 2021 (Mar 2018)
- "Mechanical architecture for a multi-chip module", Patent Application, Mar 2020 (Sep 2018)
- "Packaged device having embedded array of components", Patent Application, Jul 2020 (Jan 2019)
- System on Wafer (SoW), TSMC, Patents
- "The FSD improvement will come as a quantum leap, ", Elon Mask, Aug 15, 2020
- "Generating ground truth for machine learning from time series elements", May 2021 (Feb 2019): "AI Processor"
- ※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。
推論用エンジンとは異なる特許申請を発見
※発明者は推論用行列演算器の特許に関係されている方々：
"Scalable matrix node engine with configurable data formats", US20200348909A1
Debjit Das Sarma, William McGee, Emil TALPES
2019-05-03: Priority to US16/403,083...2020-11-05: Publication of US20200348909A1
"Data path for scalable matrix node engine with mixed data formats", US20200349216A1
Debjit Das Sarma
2019-05-03: Application filed by Tesla Inc...2020-11-05: Publication of US20200349216A1
"The matrix processor instruction specifies a floating-point operand formatted using a first floating-point representation format. The matrix computational unit accumulates an intermediate result value calculated using the floating-point operand. The intermediate result value is in a second floating-point representation format."
"Compute and redundancy solution for Tesla's Full Self driving computer", Hot Chips 2019 https://old.hotchips.org/hc31/HC31_2.3_Tesla_Hotchips_ppt_Final_0817.pdf
"Instruction Set Architecture for a Vector Computational Unit", US20190235866A1, (Feb 1, 2018), Debjit Das Sarma, Emil TALPES, Peter Joseph Bannon
"Scalable matrix node engine with configurable data formats", US20200348909A1 (May 3, 2019), Debjit Das Sarma, William McGee, Emil TALPES
参考：Google TPUv1 (推論専用) / TPUv2 (学習に対応)
"Google's Training Chips Revealed: TPUv2 and TPUv3", Hot Chip 2020 https://hc32.hotchips.org/assets/program/conference/day2/HotChips2020_ML_Training_Google_Norrie_Patil.v01.pdf
※他にも重要なキーワードや仕様が含まれていますが、それらは後で気づきました。
Tesla In-House Suprecomputer, Andrej Karpathy, Tesla, Workshop on Autonomous Driving, CVPR 2021, Jun 20, 2021
CVPR 2021併設ワークショップ：Workshop on Autonomous Drivingで、Andrej Karpathyさんが既存 In-House Supercomputerを紹介
Dojoには触れず。
"Looking at holding Tesla AI Day", Elon Mask, Jun 22, 2021
すると
Elon Mask, Jun 22, 2021
”Looking at holding Tesla AI Day in about a month or so. Will go over progress with Tesla AI software & hardware, both training & inference. Purpose is recruiting.”
Tesla AI Day, Aug 19, 2021, Professor Dennis Hong, UCLA
Elon Mask氏の 6月22日のツイートだと Tesla AI Dayは 7月末開催でしたが続報が無かったところ、Dennis Hong教授 (Mechanical & Aerospace Engineering, UCLA)
http://www.romela.org/ が画像付きツイートを投稿
Dennis Hong, Aug 3, 2021
Tesla AI day, August 19, 2021
以下この画像をもとに推測を開始。
Dennis Hong教授の投稿画像からの推測
上から 2層目画像から TSMC Info_SoW採用と推測。こちらの記事でも
"Tesla AI Day Supercomputer Chip Teaser | Is This The First Deployment Of TSMC InFO_SoW?", Aug 5, 2021 https://semianalysis.substack.com/p/tesla-ai-day-supercomputer-chip-teaser
Cerebrasの Wafer Scale Engineとの方向性の大きな違いは、後述するように 2020年8月に Elon Mask氏が車載用学習チップに言及していたので、学習チップは小規模化も可能なモジュール構成の可能性があり、Wafer Scaleサイズでの 2次元モジュール構成が実現出来る TSMC Info_SoWは最適では・・・
※特許申請から、Cerebrasも 2次元モジュール構成を検討している様子 (続編に掲載)。
冒頭の特許申請："Scalable matrix node engine with configurable data formats", Nov 5, 2020 (May 3, 2019) をあらためて確認したところ、気になる記述が
"In some embodiments, the node engines are arranged in a mesh-like network. Each node engine includes a control unit, a memory, registers, multiple matrix processors, and a post-processing unit such as a vector computational unit. "
Tesla実装系特許申請を調べたところ：
"Voltage regulator module with cooling structure", US10887982B2
Mohamed Nasr, Aydin Nabovati, Jin Zhao
2018-03-22: Priority to US201862646835P...2019-09-26: Publication of US20190297723A12021-01-05: Publication of US10887982B22021-01-05: Application granted
TSMC Info_SoWはチップに対して垂直方向からの給電なのでこの VRMは合致し、かつモジュール単位なので自由度も高く公開された画像の 3層目から下にも近そう。
※特許画像と実装が一致するとは限りませんね。
さらに、上記 VRM採用の Info_SoW的な実装の特許出願
"Mechanical architecture for a multi-chip module", WO2020061330A1
Robert Yinan CAO, Mitchell Heschke, Mengzhi Pang, Shishuang Sun, Vijaykumar Krithivasan
2018-09-19: Priority to US201862733559P...2020-03-26: Publication of WO2020061330A1
※英語版、ツイートは勘違いしてのフランス語版を紹介。
"Packaged device having embedded array of components", US20200221568A1
Jin Zhao, Vijaykumar Krithivasan, Mengzhi Pang, Steven Wayne Butler, Ganesh VENKATARAMANAN, Yang Sun
2019-01-07: Priority to US201962789021P...2020-07-09: Publication of US20200221568A1
TSMCは InFO_SoW (System-on-Wafer) を ECTC 2020で発表：
同社の System on Wafer (SoW) に関連するだろうな特許を確認
"Semiconductor device methods of manufacture", US10978373B2
Shih-Chang Ku, Hung-Chi Li, Tsung-Shu Lin, Tsung-Yu Chen, Wensen Hung
2018-06-19: Priority to US201862687112P...2019-12-19: Publication of US20190385929A12021-04-13: Application granted2021-04-13: Publication of US10978373B2
"Integrated Circuit Package and Method", US20200211922A1
Shu-Rong Chun, Kuo Lung Pan, Pei-Hsuan Lee, Chien Ling Hwang, Yu-Chia Lai, Tin-Hao KuoHao-Yi Tsai, Chen-Hua Yu
2018-12-26: Priority to US201862784941P...2020-07-02: Publication of US20200211922A1
System on Wafer (SoW) テスト環境の特許申請：
"System and method for semiconductor device testing", US20210096173A1
Hao Chen, Mill-Jer Wang
019-09-27: Priority to US201962906962P...2021-04-01: Publication of US20210096173A1
※非常に特殊な工法なので、顧客の採用には専用のテスト環境の提供が必須でしょう。。
"The FSD improvement will come as a quantum leap", Elon Mask, Aug 15, 2020
前述した車載用学習チップについての Elon Mask氏のツイート：
Elon Mask, Aug 15, 2020
"The FSD improvement will come as a quantum leap, because it’s a fundamental architectural rewrite, not an incremental tweak. "
Tesla (TSLA) Q2 2020 Earnings Call Transcript, Jun 30, 2020
Elon Mask:
"things like isolated pictures and doing image recognition on pictures that are harshly correlated in time but not very well and transitioning to kind of a 4D, where it's like ""So what we've been doing, thus far, has really just been like 2D -- mostly 2D, and like I said, well correlated in time. So just hard to convey just how much better a fully 4D system would work "
Elon Mask, Aug 15, 2020
"This is what I meant by "4D"",
リンク先
"Tesla Published Patent 'Generating Ground Truth For Machine Learning From Time Series Elements", Aug 6, 2020 https://www.tesmanian.com/blogs/tesmanian-blog/tesla-published-a-patent-generating-ground-truth-for-machine-learning-from-time-series-elements
リンク先で参照されている特許、2021年5月4日に認定。
"Generating ground truth for machine learning from time series elements", US10997461B2
Ashok Kumar Elluswamy, Matthew Bauch, Christopher Payne, Andrej Karpathy, Joseph Polin
2019-02-01: Application filed by Tesla Inc...2020-08-06: Publication of US20200250473A12021-05-04: Application granted2021-05-04; Publication of US10997461B2
車載 "AI processor 107"
関連特許申請
"Predicting three-dimensional features for autonomous driving", US20200249685A1
Ashok Kumar Elluswamy, Matthew Bauch, Christopher Payne, Andrej Karpathy, Dhaval Shroff, Arvind Ramanandan, James Robert Howard Hakewill
2019-02-01; Application filed by Tesla Inc...2020-08-06: Publication of US20200249685A1
0 件のコメント:
コメントを投稿