2024-01-01ãã1å¹´éã®è¨äºä¸è¦§
æ ç»å¥½ãã®æ¯è¨ 4000æ¬ã¾ã§ã100æ¬ãåãã¾ããã ã¨ãããã¨ã¯ãæ¥å¹´ãéæã§ãã¾ãï¼ 2024å¹´ : åè¨ : 296 / 3921 2023å¹´ï¼åè¨: 277 / 3625 2022å¹´ï¼åè¨: 375 / 3348 2021å¹´ï¼åè¨: 431 / 2973 2020å¹´ï¼åè¨: 514 / 2542 2019å¹´ï¼åè¨: 264 / 2030 2018â¦
ã¯ããã« ä¸è¨ã® SystemC Update 2024 ãçºãã¦ã¿ããããªãããè²ã ã¨æ¸ãã¦ãã£ãã®ã§ãè¨é²ã«æ®ãã¾ãã semiwiki.com SystemC Evolution Day 2024 10æã« SystemC Evolution Day 2024 ãéå¬ãããããã§ãã çºè¡¨è³æãå ¬éããã¦ããããã§ãã Welcomâ¦
ã¯ããã« Xã§ã®æ稿ãSiliconéçºã®ã³ã¹ãã«ã¤ãã¦ã è¨é²ã«æ®ãã¾ãã å è¨äº anysilicon.com ä¸å³ã¯ãä¸è¨ã®è¨äºãã説æã®ããã«å¼ç¨ãã¾ãã 7nm : $249M 5nm : $449M 3nm : $581M 2nm : $725M æ¥æ¿ã«å¢ãã¦ããã®ã¯ã·ãªã³ã³ã§ã¯ãªãã ã½ããã¦ã§ã¢ã¨ãâ¦
ã¯ããã« SemiAnalysisãã¯ãªã¹ãã¹ãã¬ã¼ã³ãã¨ãã¦ãå ¬éãã¦ãããä¸è¨ã®è¨äºã semianalysis.com è¨é²ã¨ãã¦æ®ãã¾ãã NVIDIA B300/GB300 B300 ã¯ãTSMC N4P 㧠B100/B200 ã¨åãããã»ã¹ã§æ°è¦éçºã®æ¨¡æ§ãããè¨ç®æ§è½ã¯ 50% ã¢ãããã¡ã¢ãªå®¹éã 50â¦
12æã®æ ç»éè³ : 27 æ¬ 11æã®æ ç»éè³ : 21æ¬ 10æã®æ ç»éè³ : 27æ¬ 9æã®æ ç»éè³ : 27æ¬ 8æã®æ ç»éè³ : 25æ¬ 7æã®æ ç»éè³ : 20æ¬ 6æã®æ ç»éè³ : 25æ¬ 5æã®æ ç»éè³ : 33æ¬ 4æã®æ ç»éè³ : 25æ¬ 3æã®æ ç»éè³ : 17æ¬ 2æã®æ ç»éè³ : 25æ¬ 1æâ¦
ã¯ããã« NVIDIAã® ConnectX-8 ã® Data sheet ããªããã¦ã¿ã ConnectX-8 Data Sheet ConnectX-8 Data Sheet Maximum total bandwidth : 800Gb/s InfiniBand speeds : 800/400/200/100Gb/s Ethernet speeds : 400/200/100/50/25Gb/s Host interface PCIe Genâ¦
ã¯ããã« ä»å¹´ãçµæ§æ¬ãèªã¿ã¾ããããã®ããã°ã«è¨é²ãããã®ããªã¹ãã¢ãããã¾ã 2024å¹´ã«èªãã æ¬ é©å½ã«ã«ãã´ãªåããã¾ãã ãã¸ãã¹æ¸ ãæ æ°ããªãã¦åå¨ããªã è¡åçµæ¸å¦ãæå¼·ã®å¦åã§ãã ããããããã¨ãããããã¹ããã¨ããå ¨é¨ã§ããï¼ â¦
ã¯ããã« ã¯ã¦ãªããã°ã®ã©ã³ãã³ã°ãçºè¡¨ããã¾ããã blog.hatenablog.com ãªãã¨ãå¹´éç·åã¯ã¦ãªããã°ã©ã³ãã³ã°ã§ã57ä½ã§ãã 57ä½ã®ããã°ã¨ã¯ã vengineer.hatenablog.com ã§ãã 676ã¦ã¼ã¶ã¼ãããã¯ãã¼ã¯ãä»ãã¦ãããããã§ããã b.hatena.neâ¦
ã¯ããã« ã²ã¨ãèªæ¸ä¼ è亡ã®ä¸çå² æ±ã¤ã³ãä¼ç¤¾ã¨ã¢ã¸ã¢ã®æµ·ã Kindleçã1,375åããã¤ã³ããã㯠562 ç¾å¨ã¯ã963å(ãã¤ã³ã 10) æ±ã¤ã³ãä¼ç¤¾ã£ã¦ããã£ã±ããã£ããã ã¨ããçç±(å¿ãã)ã§ãæ±ã¤ã³ãä¼ç¤¾ã®ãã¨ãç¥ãããã¦ãèªãã ãã§ããã ã§ãâ¦
ã¯ããã« ä¸è¨ã®Xã®æ稿ã¨ãã¦ãNVIDIA RTX 5090 ã®åºæ¿ã®åçãæµãã¦ãã¾ããã®ã§è¨é²ã«æ®ãã¾ãã æ稿ããã³æ稿ã«ãã£ãç»åã説æã®ããã«å¼ç¨ãã¾ãã RTX 5090 PCB?https://t.co/Cky6GZiecF pic.twitter.com/BXUwPhD2Thâ HXL (@9550pro) 2024å¹´12æ2â¦
ã¯ããã« 2023å¹´ãã¯ã©ãããã¼ã«ã³ã¬ã¯ã·ã§ã³ã¯ããã¡ãã 2024 ä»å¹´ãããã¥ã¼ãªãã®ã追å ãã¾ãã ãã¢ã¼ããã¢ã¼ (é岡) ãããã¸ã£ãã³ (ç¦å³¶) 丹å¾çå½ (京é½) åæå±±ãã«ã¯ãªã¼ (åæå±±) æ¹åãã¼ã« (ç¥å¥å·) open air (å µåº«) éããã¡ãã¼ã« (æç¥)â¦
ã¯ããã« æ©ããã®ã®ãã¾ã1å¹´ãçµãããã¨ãã¦ãã¾ãã æ¥é±ããå¹´æ«å¹´å§ä¼ã¿ã¨ãããã¨ã§ãä»æ¥ãä»äºç´ãï¼ã§ãã 12æ30æ¥ãã1æ3æ¥ã¾ã§ãä¼ã¿ã§ããåæ¥ãæãã§ã12æ28æ¥ãã1æ5æ¥ã¾ã§ã®9é£ä¼ã§ãã ãªããã9é£ä¼ã£ã¦ãè¶ ä¹ ãã¶ãã§ãããã¶ãã転â¦
ã¯ããã« Intel ã® GPUã£ã¦ãMicrosoft Windows 11 ã® WSL2 ã®ä¸ã§åããããã èªåã§ã2023å¹´8æã«ç¢ºèªãããã vengineer.hatenablog.com Lunar Lake ã使ããã ä¸è¨ã®è¨äºã«ããã¨ã www.phoronix.com DG1 Alchemist Tiger Lake Alder Lake Raptor Lakeâ¦
ã¯ããã« HotChips 2024ã®è¬æ¼æ¨¡æ§ãYoutubeã«ã¢ããããã¾ããã HC2024-S2: Specialized Processors ã®ä¸ã«ãTenstorrent ã® Blackhole Galaxy ã®èª¬æãããã¾ããã®ã§ãè¨é²ã«æ®ãã¾ãã Tenstorrent ã® Blackhole Galaxy ãããªã® 13:48 é ãã youtu.beâ¦
ã¯ããã« Intel GPUã«é¢ãã¦ã¯ãOpenCLãåãããã«ãªã£ãé ã¯ãããWatchãã¦ãã¾ããã確ããBroadwell ã ã£ãã¨æãã¾ãã ãã®æã¯ãGen 8 ã¨ã Gen 9 ã¨ãè¨ã£ã¦ãã¾ããã Tiger Lake ã«ãªã£ã¦ãXe ã¨ãã®ååã«ãªããLunar Lake ã§ã¯ãXe 2 ã«ãªã£ãâ¦
ã¯ããã« Xã®æ稿ã«ãIntel GPU Arc B850 ã® PCIe Board ã«ãM.2 SSD ã2æåºãã£ã¦ããç»åãæµãã¦ãã¾ããã Intel Arc B580 GPU with two M.2 slots smiles for the camera â Maxsun GPU offers storage expansion with two PCIe 4.0 SSDs https://t.co/â¦
ã¯ããã« RISC-Vãã¼ã¹ã®ãã¶ã¼ãã¼ããDC-ROMA RISC-V Mainboard ã¨ãããã®ããããããªã®ã§ãè¨é²ã«æ®ãã¾ãã DC-ROMA RISC-V Mainboard deepcomputing.io ããã«ãä»æ§ãè¼ã£ã¦ãã¾ãã SoC : JH7110 CPU : RISC-V 64-bit Quad-core CPU up to 1.5GHz Gâ¦
ã¯ããã« NVIDIAã®github ã® open-gpu ãé¤ãããã H20 ã§ã¯ãªãã H20 3e ãªããã®ããã£ãã®ã§è¨é²ã«æ®ãã¾ãã NVIDIA H20 ã¨ã¯ï¼ H20 ã«ã¤ãã¦ã¯ãéå»3ååãä¸ãã¦ãã¾ãã vengineer.hatenablog.com vengineer.hatenablog.com vengineer.hatenablog.â¦
ã¯ããã« Tenstorrentã¯ã2種é¡ã®ããããæè¼ãããã¼ãã販売ãã¦ãã¾ãã Grayskull e75 e150 Wormhole n150s n300s ããã¦ãç¾å¨éçºä¸(è©ä¾¡ä¸ã ã¨æã)ãBlackhole æè¼ã®ãã¼ãã®ååãåãã£ãã®ã§ãè¨é²ã«æ®ãã¾ãã Blackhole æè¼ã®åºæ¿ã®ååã¯ï¼â¦
ã¯ããã« ä»åã¯ã Intel ã«è²·åããã Altera AMD ã«è²·åããã Xilinx ã®å£²ä¸ã伸ã³ã¦ããªãã®ã§ããªãã§ãªã®ããå¦æ³ãããã¨æãã¾ãã ããã§ã¯ã Let's å¦æ³ Intel ã«è²·åããã Altera 2015å¹´ã«Intelã«è²·åãããAlteraã2022å¹´1æã«ä¸è¨ã®ããã°ã«ãâ¦
ã¯ããã« AWS Trainium 2ã®éçºãã¼ããã¼ã¯ï¼ã©ãï¼ ä»æ¥ã¯ããã®ä»¶ã«ã¤ãã¦ãå¦æ³ãããã¨æãã¾ãã ããã§ã¯ãLet'så¦æ³ AWS re:invest 2024 ã®ãããªãã ä¸è¨ã¯ AWS re:invent 2024 ã®ãããªããã®å¼ç¨ãTrainium 2ã¯ã2 (die + 2 x HBM) ã1ã¤ã®ããâ¦
ã¯ããã« Broadcomã®ä¸è¨ã®è³æã«ãã£ã Broadcom Enabling AI Infrastructure Investor Meeting ã® Page.92 ã®ä¸è¨ã®ã¹ã©ã¤ã(説æã®ããã«å¼ç¨ãã¾ã) ã®ç´«ã®ä¸¸ã®æå³ãåãã£ãã®ã§ãè¨é²ã«æ®ãã¾ãã 3.5D XDSiP Platform Technology Broadcom ã® 3.5D Xâ¦
ã¯ããã« ã²ã¨ãèªæ¸ä¼ ç©¢ããèå°å·¡ç¤¼ã«ã¤ãã¦ãèç Kindleæ¬ã1,401åããã¤ã³ãããã¯(361)ã26 %ãã¤ã³ãããã¯ãã¦ããã®ã§ããããªãä¸æ°ã«èªã¿ã¾ããã ææ³ è¿ç¿å°æ¹ã®ããå ´æã«ã¤ã㦠ã¨ä¼¼ãæãã®ã¹ãã¼ãªã¼å±éã§ãããèªã¿ãããã£ãã§ãããªâ¦
ã¯ããã« Google TPU Trillium (v6e) ã GA ã«ãªã£ãããã§ãã cloud.google.com v6e ã¯ãTraining ç¨ï¼ ä¸è¨ã®ããã°ã«ã¯ãä¸è¨ã®ãããªã°ã©ããããã¾ããã Figure 3. Source data: MLPerf⢠4.1 Training Closed results for Trillium (Preview) and v5pâ¦
ã¯ããã« ã²ã¨ãèªæ¸ä¼ æ»ãã ãé£ã¹ã ã¹ãã£ã¼ã´ã³ã»ãã³ã°ãã¸ã§ã¼ã»ãã«ã®æ¸ãä¸ããä½åãªã©å訳10ç¯ãå«ããå ¨16ç¯ã®ææã®ãã©ã¤ãã¸ããããã ä¸çªããã£ãã®ã¯ã ãåºæãã¹ãã£ã¼ã´ã³ã»ãã³ã°ï¼ç½ç³æ訳 -ã解æ¾ãã¸ã§ã¼ã»ãã«ï¼ç½ç³æ訳 â å訳 -â¦
ã¯ããã« ä¸è¨ã®ãã¹ããç¨SoCã¨ã¹ããã¡ã¼ã«ã¼ã®é¢ä¿ãã®ããã°ã«ã¦ãQ1.2024 ã® Smartphone ã®å£²ä¸ 3ä½ã§ãã Xiaomiã vengineer.hatenablog.com ãã£ããå¹´éã1.6åå°ã®Smartphoneã売ã£ã¦ãã¾ãã ãã®ãããã®å°æ°ã売ããã®ãªããç¬èªã«Smartphoneç¨â¦
ã¯ããã« Tenstorrentã® tt-xla Googleã®ä¸è¨ã®ãPJRT: Simplifying ML Hardware and Framework Integrationãã«å¯¾å¿ãããã¨ãã¦ããããã§ãã opensource.googleblog.com ãããã§ããã¨ã Tensorflow JAX PyTorch ãåãããã«ãªãããã§ããã OpenXLA â¦
ã¯ããã« AI Chip ã®ä¸ã«ã¯ãPCIe ããããéæ¥ç¶ç¨éã¨ãã¦ä½¿ã£ã¦ãããã®ãããã¾ãã Graphcore : IPU (PCIe Gen4ç¸å½) SambaNova Systems : SN10/SN20/SN30ãSN40ã¯æè¨ãã¦ããªãããã¶ãããã ã¨å¦æ³ AWS : Inferentia 2/Trainium ã¨ã AWS Trainium 2â¦
ã¯ããã« AWSãTrainium 2ã®ä¸è¬å ¬éãçºè¡¨ãã¾ããã æ§è½ã¯ã BF16 : 650 T FLOPs B8 : 1300 T TOPs ã¨ãNVIDIA H100ããã¯ä½ããã®ã®ãTrn2 UltraClusters ã«ã¦ã3Dæ§æã®Clusterãæ§ç¯ã§ããããã«ãªã£ã¦ãã¾ãã ãããªãAWS ã§ãããä¸è¨ã®ããã«ããâ¦
ã¯ããã« AWS Trainium 㨠Trainium 2 ã®åºæ¿ã re:invent 2024 ã®ãããªã«ãã£ãã®ã§è¨é²ã«æ®ãã¾ãã Trainumã®åºæ¿ã¯ãä¸è¨ã®ããã°ã«ãåçãå¼ç¨ãã¦ãã¾ãã vengineer.hatenablog.com AWS Trainium 㨠Trainium 2 ã®åºæ¿ã®æ¯è¼ ä¸è¨ã®ãããªã® AWS reâ¦