2023-08-01ãã1ã¶æéã®è¨äºä¸è¦§
ã¯ããã« æ¨æ¥ãGoogle TPU v5e ã«é¢ããããã°ãæ¸ãã¾ããã vengineer.hatenablog.com ãã®ä¸ã§ãv5e ã§ã¯ãMultislice Technology ãªããã®ãåºã¦ãã¾ããã Googleã¯ãããã«ã¤ãã¦ãOverview ãå ¬éãã¦ãã¾ããã Cloud TPU Multislice Overview [Pubâ¦
ã¯ããã« Xã®TLã«æµãã¦ããä¸è¨ã®ãã¤ã¼ããIntel GPU (Ponte Vecchio) ã® die ã£ã½ãã§ãã The base tile of Ponte Vecchio. They put eight Compute Tiles and four times Rambo Cache (the middle lines) on there.Another shot with the Compute Tilesâ¦
8æã®æ ç»éè³ : 24æ¬ 7æã®æ ç»éè³ : 25æ¬ 6æã®æ ç»éè³ : 20æ¬ 5æã®æ ç»éè³ : 29æ¬ 4æã®æ ç»éè³ : 20æ¬ 3æã®æ ç»éè³ : 27æ¬ 2æã®æ ç»éè³ : 17æ¬ 1æã®æ ç»éè³ : 20æ¬ 2007å¹´3æãã2022å¹´12æã¾ã§ : 3348æ¬ + 2023å¹´1æãã8æã¾ã§ : 182æ¬ = 3â¦
ã¯ããã« Google Cloud Next '23 ã«ã¦ãGoogle ã TPU v5e ãçºè¡¨ãã¾ããã www.youtube.com Efficient (Per/$ vs. TPU v4) 2X training 2.5x inference Scalable 10s of Ks of chips (Multislice technology) ã»ãã®ã¡ããã¨ãããªãã£ãã§ãããä¸è¨ã®ãâ¦
ã¯ããã« ãã®ããã°ã§ãä½åº¦ãåãä¸ãã¦ãã Esperanto Technologies ãä¸è¨ã® eetimes ã® Sally -san ã®è¨äºã«ããã¨ãrecommendation ãã HPC/LLMs ã« pivot ããã¨ã www.eetimes.com Esperanto Technologies ã¯ãrecommendation ãã¿ã¼ã²ããã«ãã¦ãâ¦
ã¯ããã« AMD ã® IR (Q2.21 - Q2.23) ã®ã«ãã´ãªå¥ Data Center Client Gaming Embedded ãè¦ãã¦ã¿ã¾ããã å ¨ä½ã®æ¨ç§» ä¸è¨ã Q2.21 - Q2.23 ã®ã«ãã´ãªå¥ããã³4ã«ãã´ãªåè¨ã®ã°ã©ãã§ããQ2.22 ããã¼ã¯ã§ä¸ãã£ã¦ãã¾ãã ä¸ãã£ã¦ããã®ã¯ãã©ããã â¦
ã¯ããã« MediaTekã®å£²ä¸ã2021年度ã®åæãä¸åã£ã¦ããã¨ããããã°ãæ¸ããã®ã 2022.12.13ã vengineer.hatenablog.com ãããã8ãæããã¼ã¿ã®ã¢ãããã¼ãããã¾ããã 2022.12 - 2023.7 ã追å 2023.3ã¯ã2021.3ããã¡ãã£ã¨ä¸ã§ããããã以å¤ã¯ 20â¦
ã¯ããã« Xã®TLã«æµãã¦ããä¸è¨ã®æ稿ããã¤ã¼ããªãã ãã©ãç»åããã¾ã使ããã¦ãã¦ããªããè¨äºãèªãã§ããã¿ããã§ãã Google is working on a new developer option in Android that will swap out your device's Linux kernel with one that uses â¦
ã¯ããã« OpenHW Group ãã RISC-V MUC ã® "CORE-V MCU DevKit" ç¨ã®ãããããã¼ãã¢ã¦ããããã¨ãã¢ãã¦ã³ã¹ãããã¾ããã www.openhwgroup.org CORE-V MCU DevKit CORE-V MUC ã¯ã GlobalFoundriesâ proprietary 22FDX® process technology CV32E40P câ¦
ã¯ããã« ãã®ããã°ã§ã3åãã¨ãããã¦ãã Mipsology (Zebra) ã AMD (ãã¶ããæ§Xilinx)ããè²·ãä¸ããã¾ããã community.amd.com Mipsology AI Inferenceç¨ã¨ãã¦ãXilinxã¨ä»²è¯ããã¦ãã Mipsology ããããã¯ãå Zebra ã AMD ãåãè¾¼ãæãã§ãâ¦
ã¯ããã« AWS Trainium / Inferentia2 ã® NeuronCore-v2 ã«å ¥ã£ããGPSIMD Engineã å³ãè¦ãã¨ã3åãããå ¥ã£ã¦ããã®ã§ããããããä½ã§ããããï¼ ä¸å³ã¯AWSã®ãµã¤ãããæã£ã¦ãã¾ããã説æã®ããã«å¼ç¨ãã¾ãã AWS Neuron Documention ã®ä¸ãæ¢ã£ã¦â¦
ã¯ããã« Xã«Supercomputer Auroraã«é¢ããæ å ±ãæµãã¦ãã¾ããã ICYMI: Preparing for #Exascale on #Aurora @argonne@IXPUG1 Webinar by Scott Parker @argonne_lcf Aug 10 2023o Aurora Overviewo Preparing Applicationso Early performance resultsSlâ¦
ã¯ããã« AWSã®Trainium 㨠Inferentia2ã«ã¯ãNeutronLink-v2ã¨ãããããéãæ¥ç¶ããæ©è½ãå ¥ã£ã¦ããã¨ãããã¨ã¯ããã®ããã°ã§ãç´¹ä»ãã¾ããã ãã® NeuronLink-v2 ã«ã¯ãCollective Communication ãå ¥ã£ã¦ããã¨ãããã®ä»åã®ã話ã Neuron Collectâ¦
ã¯ããã« Groq ã The Language Processing Unit ãªããã®ãçºè¡¨ãã¾ããã futurumgroup.com Groqã®ãã¬ã¹ãªãªã¼ã¹ã¯ããã¡ã www.prnewswire.com The Language Processing Unit (LPU) ä¸è¨ã®è¨äºã«ããã¨ã LLM Llama-2 70B ãï¼ç§ããã100ãã¼ã¯ã³ä»¥ä¸ã®â¦
ã¯ããã« Xã®TLã«ãOccamyãä¸ãã£ã¦ããã¨ããæ稿ãæµãã¦ãã¾ãããä¸è¨ã®æ稿ã説æã®ããã«å¼ç¨ãã¾ãã Live from PULP, you are witnessing the unboxing of Occamy by @LucaBeniniZhFe. Our 432-core Multi-TFLOPs RISC-V-Based 2.5D Chiplet Systemâ¦
ã¯ããã« Xã®TLã«æµãã¦ãããTier IVã® Kato -san ã®æ稿 å½æã®ä»²éãã¡ã¨ãã¼ã ãåçµæããæ°ããéçºãããããããã¡ãã§ãã試ä½ã§ããããããã5å¹´éããã¦éç£ãã¦ããã¾ããhttps://t.co/gXvrEh0n5R https://t.co/fM3Wx8cr50 pic.twitter.com/wDKâ¦
ã¯ããã« æ¨å¹´ã®ï¼æã« Tier IV ã®ããããä¸ãã£ã¦ãããã¨ã¯ãã®ããã°ã§ãåãä¸ãã¾ããã vengineer.hatenablog.com ãã®ãããã使ã£ãã·ã¹ãã ã®éç¨ãå§ã¾ã£ãããã§ãã medium.com ãããã®åçª ä¸è¨ã®è¨äºãããããã®åçªã¯ãAX91101 ã£ã½ãã§ãâ¦
ã¯ããã« NVIDIAã®GPUã®ä»¶ãH100ã売ãåããA800ã売ãã¦ãããã¨ãããã¨ããã®ããã°ã§åãä¸ãã¾ããã ã§ã¯ä½æ ï¼NVIDIAã®GPUã ãã売ããã®ãï¼ ããã¯ãNVIDIAãé·å¹´ä½ãä¸ãã¦ããã¨ã³ã·ã¹ãã ãããããã§ããç¹ã«ã2006å¹´11æçºè¡¨ãããã2007å¹´7â¦
ã¯ããã« NVIDIAã®RTX 40 (Ada Lovelace) ã®æ¬¡ã®GPUã§ãã RTX 50ã¯ãCB20Xã«ãªãããã§ãã videocardz.com GB202, GB203, GB205, GB206, GB207 ä¸è¨ã®è¨äºã®å 容ã§ã¯ã GB202, GB203, GB205, GB206, GB207 ã¨ããã®ã§ãã³ã³ã·ã¥ã¼ãã¼åãã®GPUã«ãªã£ã¦ãâ¦
ã¯ããã« å é±ãAWS Trainiumã®å®ç©ãã¨ãããã¨ã§ããã°ãæ¸ãã¾ããããæ¨è«ãããã® Inferentia2 ã Trainium ã¨ã»ã¼åããããªæ§æã§ããã Trainium Inferentia2 ä¸è¨ã®ãµã¤ãããããããã®çµµã説æã®ããã«å¼ç¨ãã¾ããéãã¯ãNeuronLink-v2ã®æ¬æ°ã â¦
ã¯ããã« AWSãAIå¦ç¿ç¨ãããã§ãã Trainium ãçºè¡¨ããã®ããAWS re:Invent 2020 ã®ã㨠2022å¹´10æ12æ¥ã®æè¡ããã°ã ç¬èªè¨è¨ããã AWS Trainium æè¼ Amazon EC2 Trn1 ã¤ã³ã¹ã¿ã³ã¹ã§ ML ãã¬ã¼ãã³ã°ãé«éå®è¡ï¼åºç¤ç·¨ï¼ ã«ããã¨ããããã«ã¯ HBMâ¦
ã¯ããã« Xã«æµãã¦ãããä¸è¨ã®æ稿ã®è¨äºã #China's largest web and cloud providers are lining up to buy as many @Nvidia #GPUs as they can Alibaba, Baidu, ByteDance, and Tencent ordered 100,000 #A800 GPUs worth ~$1B and another $4B worth oâ¦
ã¯ããã« ã¹ãã£ã¼ãã³ã»ãã³ã°ã® 11/22/63 ãèªã¿çµãã¾ããã 7/27 - 8/12 ã§ãã8/11 - 8/12 ã§ä¸æ°ã« æå¾ã® 1/3 ããããèªã¿ã¾ããã ã¹ãã£ã¼ãã³ã»ãã³ã°11/22/63ã²ã¨ãèªæ¸ä¼çµäºããªãè¯ãã£ã次㯠The Stand(å ¨5å·»)2000é ãããã㪠https://t.coâ¦
ã¯ããã« æ¨å¹´å¾åã®ChatGPTã«å§ã¾ã£ãçæAIãã¼ã ã¨ãããããã«ã®å½±é¿ã§NVIDIAã®GPUã売ãåãã¨ããã話ã www.barrons.com H100ã¯ãQ1.2024/Q2.2024 ã¾ã§è²·ããªã ä¸è¨ã®è¨äºã«ããã¨ãH100ã¯ãQ1.2024/Q2.2024 ã¾ã§è²·ããªããã¨ãããã¨ã§ãã vengineâ¦
ã¯ããã« Bluespec SystemVerilog ã®ä¸ã«ãbluetcl ãªã tcl ã¤ã³ã¿ããªã¿ãããã¾ããä»åã¯ããã® bluetcl ã使ã£ã¦ã¿ããã¨æãã¾ãã bluetcl ã使ããã¨ã§ã¤ã³ã¿ã©ã¯ãã£ãã«ã·ãã¥ã¬ã¼ã·ã§ã³ãé²ãããã¨ãã§ããããã§ãã ä¾é¡ ä¾é¡ (examples/smokeâ¦
ã¯ããã« PCI Express ãç¾å¨ã¯ 5.0 ã§ãããæ°å¹´ããã¨ã6.0 ã«ãªãã¾ãã 7.0 ã® Specification ã®çå®ããã¦ãã¾ãã www.businesswire.com 6.0ã§ã¯ã64Gbps (32Gbps-PAM4) ã§ãã7.0ã§ã¯ 128Gbps (64Gbps-PAM4) ã§ãã 5.0 ããµãã¼ããããIntel Sapphiâ¦
ã¯ããã« NVIDIAã® Grace ç¨ã® Linux Kernel ãgithub ã«ããã¾ãã 64K PAGE SIZE ã¡ãã£ã¨æ£çãã¦ãããã64K PAGE SIZE ããµãã¼ãããããã«ãdefconfig ãä¿®æ£ããã¦ãã¾ããã CONFIG_ARCH_VISCONTI=y CONFIG_ARCH_XGENE=y CONFIG_ARCH_ZYNQMP=y CONFâ¦
ã¯ããã« ãã¤ãã®ããã«ãgithub ãæ£çãã¦ããããNVIDIAã® open-gpu-kernel-modules ã®ä¸ã« EGM ãªããã®ããããã¨ãç¥ãã¾ããã EGM ã¨ã¯ï¼ ããã«ã説æãããã¾ããã #define ADDR_EGM 7 // Extended GPU Memory (EGM) EGM == Extended GPU Memoryâ¦
ã¯ããã« å æ¥ãä¸è¨ã®ãããªãã¨ãXã«æ稿ãã¾ããã(Twitterã«ãã¤ã¼ããã¾ãã) å¤ãã®ã²ã¨ã¯æ¯æ¥æéãç¡ãã¨åãã¦ããã¨æãã¾ãã©ãããã°æéãä½ãåºãã1. æ¯æ¥24æé1é±éãä½ã«æéã使ã£ã¦ããããè¨é²ãã2. è¨é²ãããã®ãåé¡ãã絶対ã«å¿ è¦â¦
ã¯ããã« NVIDIAãGrace Hopper Superchipã GH200 ã¨ãã¦çºè¡¨ãããã¨ã¯ãã®ããã°ã§ããç¥ãããã¦ãã¾ãã vengineer.hatenablog.com ãã¤ãã®ããã«ãgithub ãæ£çãã¦ããããGH180ãªããã®ãè¦ã¤ãã¾ããã ä¸è¨ã®ããã°ã§ã¯ãGH180 == TH500 (Grace)â¦