[B! GPGPU] R2Mã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯

R2M id:R2M

GPGPUã«é–¢ã™ã‚‹R2Mã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯ (29)

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

${{author_name}}$
{{author_name}}{{created}}
{{ #comment }}{{ comment }}{{ /comment }}
- {{ label }}

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

DeepSeekãŒGPUã®åˆ¶é™ã‚’çªç ´ã™ã‚‹ã®ã«ä½¿ã‚ã‚ŒãŸPTXã€‚ãã®åˆ¶ç´„ãŒçªç ´ã®åŽŸå‹•åŠ›ã«ãªã£ãŸ - ä¸è¯ITæœ€æ–°äº‹æƒ…
DeepSeekã¯ãƒˆãƒ¬ãƒ¼ãƒ‹ãƒ³ã‚°ã‚³ã‚¹ãƒˆãŒå¾“æ¥ã®1/10ã€‚ãã‚Œã«ã¯å†—é•·è¨ˆç®—ã®ï¼˜å‰²ã‚’ã‚«ãƒƒãƒˆã™ã‚‹ã¨ã„ã†åŠªåŠ›ãŒå¿…è¦ã ã£ãŸã€‚ç±³å›½æ”¿åºœã¯ä¸å›½ã«å¯¾ã—ã¦GPUã®åˆ¶é™ã‚’ã‹ã‘ã€ãã‚Œã‚’DeepSeekãƒãƒ¼ãƒ ã¯çªç ´ã‚’ã—ãŸã¨é‡åä½ãŒå ±ã˜ãŸã€‚ å†—é•·è¨ˆç®—ã‚’ï¼˜å‰²ã‚«ãƒƒãƒˆã—ãŸDeepSeek DeepSeekãŒãªãœ1/10ã®ãƒˆãƒ¬ãƒ¼ãƒ‹ãƒ³ã‚°ã‚³ã‚¹ãƒˆã§GPTã¨è‚©ã‚’ä¸¦ã¹ã‚‹å¤§è¦æ¨¡è¨€èªžãƒ¢ãƒ‡ãƒ«ï¼ˆLLMï¼‰ã‚’é–‹ç™ºã§ããŸã®ã‹ã€‚ãã®ç”ãˆã¯ã²ã¨ã¤ã§ã¯ãªã„ã€‚ãã®æŠ€è¡“è©³ç´°ã‚’é–‹è¨ã—ãŸã€ŒDeepSeek-V3 Technical Reportã€ï¼ˆhttps://arxiv.org/abs/2412.19437ï¼‰ã¯ã€æœ€é©åŒ–æŠ€è¡“ã®ã‚«ã‚¿ãƒã‚°ã®ã‚ˆã†ã«ãªã£ã¦ã„ã‚‹ã€‚é–‹ç™ºãƒãƒ¼ãƒ ã¯ã‚ã‚Šã¨ã‚ã‚‰ã‚†ã‚‹æœ€é©åŒ–ã‚’è©¦ã¿ã¦ã€æœ€çµ‚çš„ã«å†—é•·è¨ˆç®—ã®80ï¼…ã‚’å‰Šæ¸›ã™ã‚‹ã“ã¨ã«æˆåŠŸã—ãŸã¨ä¸»å¼µã—ã¦ã„ã‚‹ã€‚ GPUã‚’ãƒãƒƒã‚¯ã—ãŸDeepSeek ãã®ä¸ã§ã‚‚ãƒã‚¤ãƒ©ã‚¤ãƒˆã¨ã‚‚è¨€ãˆã‚‹ã®ãŒã€CUDAã‹ã‚‰ä¸‹ã‚Šã¦PTXã«ã‚ˆã‚‹
R2M 2025/04/04
GPGPU
ãƒªãƒ³ã‚¯
Pytorch2.0ã§Flash Attentionã‚’ä½¿ã£ã¦ã¿ãŸè©±
ã“ã‚“ã«ã¡ã¯ã€Fusicã®ãƒãƒ³ã§ã™ã€‚æ ªå¼ä¼šç¤¾Fusicã§ã¯æ©Ÿæ¢°å¦ç¿’é–¢é€£ã®PoCã‹ã‚‰é–‹ç™ºãƒ»é‹ç”¨ã¾ã§æ§˜ã€…ãªã”ç›¸è«‡ã«å¯¾å¿œã—ã¦ã¾ã™ã€‚ã‚‚ã—å›°ã£ã¦ã„ã‚‹ã“ã¨ãŒã‚ã‚Šã¾ã—ãŸã‚‰æ°—è»½ã«ãŠå£°ã‹ã‘ã¦ãã ã•ã„ã€‚ ä»Šå›žã¯Flash Attentionã‚’ä½¿ã£ã¦ã¿ãŸã“ã¨ã«ã¤ã„ã¦ã€ç°¡å˜ã«èª¬æ˜Žã—ã‚ˆã†ã¨æ€ã„ã¾ã™ã€‚FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awarenessã§ç´¹ä»‹ã•ã‚ŒãŸã“ã®Attentionæ–¹æ³•ã¯ã€æ—©ãã¦æ£ç¢ºãªAttentionã‚¢ãƒ«ã‚´ãƒªã‚ºãƒ ã‚’å®Ÿç¾ã—ã€ã‚ˆã‚Šé•·ã„Sequenceã§ã®Transf ormerå¦ç¿’ã‚’è¡Œã†ã“ã¨ãŒã§ãã¾ã™ã€‚ ã“ã®è¨˜äº‹ã§ã¯ã€Flash Attentionã®ç†è«–çš„ãªã“ã¨ã‚’è§£æžã™ã‚‹ã“ã¨ã‚ˆã‚Šã¯ã€Pytorch2.0ã§ã®å®Ÿè£…ã‚’è¡Œã†éš›ã€æ³¨æ„ã™ã¹ããªéƒ¨åˆ†ã‚’æ•´ç†ã—ã¾ã™ã®ã§ã€è«–æ–‡ã®å†…å®¹ã«ã¤ã„ã¦ã¯FlashAttention -
R2M 2025/01/03
GPGPU

LLM
ãƒªãƒ³ã‚¯
GPU ã¨ FlashAttension ã‚’ã¡ã‚ƒã‚“ã¨ç†è§£ã—ãŸã„
ã¯ã˜ã‚ã« ChatGPT ã‚’ã¯ã˜ã‚ã¦ã¨ã—ã¦ã€å¤šãã® LLM ãŒä¸–ã®ä¸ã«é€ã‚Šå‡ºã•ã‚ŒãŸ 2023 å¹´ã§ã—ãŸã€‚OSSã¨ã—ã¦å…¬é–‹ã•ã‚Œã¦ã„ã‚‹ãƒ¢ãƒ‡ãƒ«ã‚‚å¤šãè©¦ã™ã ã‘ã§ã‚ã‚Œã°éžå¸¸ã« Colab ãªã©ã§è©¦ã›ã¦æ„Ÿå‹•ã—ã¦ã„ã¾ã™ã€‚ ã¨ã¯ã„ãˆã€ã‚„ã¯ã‚Šä¸€åº¦ã¯ LLM ã‚’è‡ªåˆ†ã§å¦ç¿’ã•ã›ã¦ã¿ãŸã„ã€ãŸã åŠ¹çŽ‡çš„ãªå¦ç¿’ã‚’ã•ã›ãªã„ã¨ãŠé‡‘ãŒæº¶ã‘ã‚‹ã ã‘...ã€‚ãã‚“ãªä¸è¦‹ã¤ã‘ãŸè¨˜äº‹ãŒã“ã¡ã‚‰ã§ã™ã€‚ ã•ã¾ã–ã¾ãª tips ãŒè¼‰ã£ã¦ã„ã¾ã™ã€‚ npaka san ãŒã“ã¡ã‚‰ã‚’æ—¥æœ¬èªžã§ã¾ã¨ã‚ã¦ä¸‹ã•ã£ã¦ã„ã¾ã™ã€‚ ã“ã®è¨˜äº‹ã§ã¯ã€ä¸Šã«æŒ™ã’ã‚‰ã‚Œã¦ã„ã‚‹æŠ€è¡“ã®ï¼‘ã¤ã§ã‚ã‚‹ FlashAttension ã«ã¤ã„ã¦ã¿ã¦ã„ãã¾ã™ã€‚ç‰¹ã«ã€ã©ã®ã‚ˆã†ãªæ”¹å–„ãŒè¡Œã‚ã‚Œã¦ã„ã‚‹ã®ã‹ã‚’è¿½ã„ã¾ã™ã€‚ï¼ˆçµæžœçš„ã«ã©ã‚Œãã‚‰ã„é«˜é€Ÿã«ãªã‚‹ã‹ã¯è©³ã—ãè¿°ã¹ãªã„ãŸã‚ä»–ã®è¨˜äº‹ã‚’å‚ç…§ã—ã¦ãã ã•ã„ï¼‰ å®Ÿã¯ FlashAttension(2022) ã®ã•ã‚‰ãªã‚‹æ”¹å–„ã¨ã—ã¦ FlashAttenstion2(20
R2M 2025/01/03
GPGPU

LLM
ãƒªãƒ³ã‚¯
Metaã¯10ä¸‡å°ä»¥ä¸Šã®NVIDIA H100ã‚’ä½¿ç”¨ã—ã¦Llama-4ã‚’ãƒˆãƒ¬ãƒ¼ãƒ‹ãƒ³ã‚°ã—ã¦ã„ã‚‹
Metaã®ãƒžãƒ¼ã‚¯ãƒ»ã‚¶ãƒƒã‚«ãƒ¼ãƒãƒ¼ã‚°CEOãŒ2024å¹´ç¬¬3å››åŠæœŸ(7ï½ž9æœˆ)æ±ºç®—å ±å‘Šã®å ´ã§ã€Œ10ä¸‡å°ä»¥ä¸Šã®NVIDIA H100ã€ã‚ã‚‹ã„ã¯ä»–ç¤¾ãŒå ±å‘Šã—ã¦ã„ã‚‹ã‚‚ã®ã‚ˆã‚Šã‚‚å¤§ãã„è¦æ¨¡ã®ã‚¯ãƒ©ã‚¹ã‚¿ãƒ¼ã§ã€Llama-4ãƒ¢ãƒ‡ãƒ«ã‚’ãƒˆãƒ¬ãƒ¼ãƒ‹ãƒ³ã‚°ã—ã¦ã„ã‚‹ã€ã“ã¨ã‚’æ˜Žã‚‰ã‹ã«ã—ã¾ã—ãŸã€‚ Meta is using more than 100,000 Nvidia H100 AI GPUs to train Llama-4 â€” Mark Zuckerberg says that Llama 4 is being trained on a cluster â€œbigger than anything that Iâ€™ve seenâ€ | Tom's Hardware https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-is-using-
R2M 2024/11/02
ãã®ã†ã¡è‡ªç¤¾é–‹ç™ºGPUã®æµã‚ŒãŒå‡ºã¦ããã†

GPGPU
ãƒªãƒ³ã‚¯
ã€ŒAMD ROCmã€ãŒã€ŒWSL 2ã€ã«ãƒ™ãƒ¼ã‚¿å¯¾å¿œ ï½žNVIDIAã§ã„ã†ã¨ã“ã‚ã®ã€ŒCUDAã€ã«ç›¸å½“ï¼ãƒžãƒ«ãƒGPUã€ã€ŒTensorflowã€ã‚‚ã‚µãƒãƒ¼ãƒˆ
R2M 2024/07/02
GPGPU
ãƒªãƒ³ã‚¯
NVIDIAãŒCUDAã‚’ä»–ã®ãƒãƒ¼ãƒ‰ã‚¦ã‚§ã‚¢ä¸Šã§å®Ÿè¡Œã™ã‚‹ã“ã¨ã‚’ç¦æ¢
NVIDIAãŒã€GPUå‘ã‘ã®ã‚³ãƒ³ãƒ”ãƒ¥ãƒ¼ãƒ†ã‚£ãƒ³ã‚°ãƒ—ãƒ©ãƒƒãƒˆãƒ•ã‚©ãƒ¼ãƒ ã¨ã—ã¦æä¾›ã—ã¦ã„ã‚‹ã€ŒCUDAã€ã®ã‚½ãƒ•ãƒˆã‚¦ã‚§ã‚¢åˆ©ç”¨è¨±è«¾å¥‘ç´„(EULA)ã®ä¸ã§ã€ç¿»è¨³ãƒ¬ã‚¤ãƒ¤ãƒ¼ã‚’é€šã˜ã¦NVIDIAä»¥å¤–ã®ãƒãƒ¼ãƒ‰ã‚¦ã‚§ã‚¢ãƒ—ãƒ©ãƒƒãƒˆãƒ•ã‚©ãƒ¼ãƒ ã§å®Ÿè¡Œã™ã‚‹ã“ã¨ã‚’ç¦æ¢ã—ã¦ã„ã‚‹ã“ã¨ãŒã‚ã‹ã‚Šã¾ã—ãŸã€‚ã‚‚ã¨ã‚‚ã¨ã“ã®æ¡é …ã¯NVIDIAã®ã‚µã‚¤ãƒˆä¸Šã§å…¬é–‹ã•ã‚Œã¦ã„ã‚‹ã‚ªãƒ³ãƒ©ã‚¤ãƒ³ç‰ˆã®EULAã«ã¯å«ã¾ã‚Œã¦ã„ã¾ã—ãŸãŒã€ã‚¤ãƒ³ã‚¹ãƒˆãƒ¼ãƒ«ã—ãŸCUDAã®ãƒ‰ã‚ãƒ¥ãƒ¡ãƒ³ãƒˆã«ã‚‚å«ã¾ã‚Œã‚‹ã‚ˆã†ã«ãªã£ãŸã¨ã®ã“ã¨ã§ã™ã€‚ License Agreement for NVIDIA Software Development Kits â€” EULA https://docs.nvidia.com/cuda/eula/index.html Nvidia bans using translation layers for CUDA software â€” previously the prohi
R2M 2024/03/06
intelã‚‚amdã‚‚zludaã«å¦å®šçš„ã ã£ãŸã®ã¯ã“ã†ã„ã†äº‹ãŒã‚ã‚‹ã‹ã‚‰ã‹

GPU

GPGPU
ãƒªãƒ³ã‚¯
https://twitter.com/norasuzu/status/1763599923088601452
R2M 2024/03/04
GPGPU

windows
ãƒªãƒ³ã‚¯
NVIDIAã®ã€ŒCUDAã€ã¨Intelã®GPUã‚’ã¤ãªãã‚½ãƒ•ãƒˆã‚¦ã‚§ã‚¢ã€ŒZLUDAã€ãŒAMDå‘ã‘ã¨ã—ã¦è»¢èº«å¾©æ´»ã™ã‚‹ã‚‚ä»Šå¾Œã®é–‹ç™ºã¯çµ¶æœ›çš„
NVIDIAã®ã‚³ãƒ³ãƒ”ãƒ¥ãƒ¼ãƒ†ã‚£ãƒ³ã‚°å‘ã‘GPUæ´»ç”¨æŠ€è¡“ã€ŒCUDAã€ã‚’Intel GPUã§å®Ÿè¡Œã§ãã‚‹ã‚ˆã†ã«ã—ãŸã‚½ãƒ•ãƒˆã‚¦ã‚§ã‚¢ã€ŒZLUDAã€ãŒå¾©æ´»ã—ã¾ã—ãŸãŒã€Intelã§ã¯ãªãAMDã®GPUã§å‹•ä½œã™ã‚‹ã‚ˆã†æ”¹å¤‰ãŒåŠ ãˆã‚‰ã‚Œã¦ã„ã¾ã—ãŸã€‚ AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source - Phoronix https://www.phoronix.com/review/radeon-cuda-zluda Software allows CUDA code to run on AMD and Intel GPUs without changes â€” ZLUDA is back but both companies ditched it, nixing future updates | T
R2M 2024/02/16
GPGPU

ã‚½ãƒ•ãƒˆã‚¦ã‚§ã‚¢
ãƒªãƒ³ã‚¯
é«˜é€ŸãªPythonäº’æ›è¨€èªžã®ã€ŒMojoã€ã€ãã®ã¾ã¾GPUã§å®Ÿè¡Œå¯èƒ½ã«ã€‚ã€ŒMAX GPUsã€ç™ºè¡¨ã€‚ModCon '23
é«˜é€ŸãªPythonäº’æ›è¨€èªžã®ã€ŒMojoã€ã€ãã®ã¾ã¾GPUã§å®Ÿè¡Œå¯èƒ½ã«ã€‚ã€ŒMAX GPUsã€ç™ºè¡¨ã€‚ModCon '23 é«˜é€ŸãªPythonäº’æ›è¨€èªžã€ŒMojoã€ã®é–‹ç™ºå…ƒã§ã‚ã‚‹Modularç¤¾ã¯ã€åŒç¤¾ã¨ã—ã¦åˆã®ã‚¤ãƒ™ãƒ³ãƒˆã€ŒModCon '23ã€ã‚’12æœˆ4æ—¥ã«ã‚µãƒ³ãƒ•ãƒ©ãƒ³ã‚·ã‚¹ã‚³ã§é–‹å‚¬ã—ã¾ã—ãŸã€‚ åŸºèª¿è¬›æ¼”ã§ã€é«˜é€ŸãªPythonäº’æ›è¨€èªžã®ã€ŒMojoã€ã‚’å«ã‚€AIå‡¦ç†ã‚’é«˜é€Ÿã«å®Ÿè¡Œã™ã‚‹ãƒ—ãƒ©ãƒƒãƒˆãƒ•ã‚©ãƒ¼ãƒ ã€ŒMAXã€ï¼ˆModular Accelerated Xecution (MAX) Platformï¼‰ãŒNVIDIAã®GPUå¯¾å¿œã¨ãªã‚Šã€ã•ã‚‰ã«é«˜é€Ÿãªå‡¦ç†ãŒå¯èƒ½ã¨ãªã‚‹ã€ŒMAX GPUsã€ãŒç™ºè¡¨ã•ã‚Œã¾ã—ãŸã€‚
R2M 2023/12/06
python

GPGPU
ãƒªãƒ³ã‚¯
GPU.js - GPU accelerated Javascript
R2M 2020/08/04
javascript

library

GPGPU
ãƒªãƒ³ã‚¯
CUDAã‚’Radeonã§å‹•ã‹ã™ï¼ˆå°Žå…¥ç·¨ï¼‰ - Fixstars Tech Blog /proc/cpuinfo
ã“ã®ãƒ–ãƒã‚°ã¯ã€æ ªå¼ä¼šç¤¾ãƒ•ã‚£ãƒƒã‚¯ã‚¹ã‚¿ãƒ¼ã‚ºã®ã‚¨ãƒ³ã‚¸ãƒ‹ã‚¢ãŒã€ã‚ã‚‰ã‚†ã‚‹ãƒ†ãƒ¼ãƒžã«ã¤ã„ã¦è‡ªç”±ã«æ›¸ã„ã¦ã„ã‚‹ãƒ–ãƒã‚°ã§ã™ã€‚ ã¿ãªã•ã‚“ã€ä»Šæ—¥ã‚‚å…ƒæ°—ã«GPGPUã—ã¦ã„ã¾ã™ã‹ï¼Ÿ åŽ»å¹´(SC15)ã®è©±ã§ã™ãŒã€ã€ŒRadeonã§CUDAãŒä½¿ãˆã‚‹ã‚ˆã†ã«ã™ã‚‹ã‚ˆï¼ã€ã¨AMDãŒç™ºè¡¨ã—ãŸãƒ‹ãƒ¥ãƒ¼ã‚¹ã‚’è¦šãˆã¦ã„ã‚‹æ–¹ã„ã‚‰ã£ã—ã‚ƒã„ã¾ã™ã§ã—ã‚‡ã†ã‹ã€‚Boltzmann Initiativeã®è©±ã§ã™ã€‚ ã‚ã‚Œã‹ã‚‰ã—ã°ã‚‰ãæ™‚ãŒæµã‚Œã€æ©Ÿã¯ç†Ÿã—ã¾ã—ãŸã€‚ã€ŒCUDAã¯NVIDIAå°‚ç”¨ã€ãã‚“ãªæ™‚æœŸã¯ä»Šã¯æ˜”ã€‚ãã†ã€CUDAã‚’AMDã®GPUã§ã‚ã‚‹Radeonã§å‹•ã‹ã™ã“ã¨ã«æˆåŠŸã—ã¾ã—ãŸï¼ï¼ ã®ã§ã€ã“ã“ã§ã¯ãã®æ–¹æ³•ã¨ã€ãã®æ™‚ã«ä½¿ã£ãŸã‚³ãƒ¼ãƒ‰ã‚’ç´¹ä»‹ã—ãŸã„ã¨æ€ã„ã¾ã™ã€‚ äºˆã‚ãŠæ–ã‚Šã—ã¦ãŠãã¨ã€è¡¨é¡Œã«ã‚‚ã‚ã‚‹é€šã‚Šã€ä»Šå›žã¯å°Žå…¥ç·¨ã¨ã„ã†ã“ã¨ã§ã€Œå‹•ãã“ã¨ã‚’ç¢ºèªã™ã‚‹ã€ã¾ã§ã«ãªã‚Šã¾ã™ã€‚ ãã®ã†ã¡æ™‚é–“ã‚’è¦‹ã¤ã‘ã¦ã€Œæ€§èƒ½æ¸¬å®šç·¨ã€ã‚‚ã‚„ã£ã¦ã¿ãŸã„ã¨æ€ã„ã¾ã™ãŒã€ãœã²ã“ã‚Œã‚’èªã‚“ã æ–¹ã€ãœã²æ€§èƒ½æ¸¬å®šã—ã¦ã¿
R2M 2016/08/02
GPGPU

CUDA
ãƒªãƒ³ã‚¯
ã€GTC2015ã€‘NVIDIAã€è‡ªå‹•é‹è»¢è»Šé–‹ç™ºã‚ãƒƒãƒˆã€ŒDRIVE PXã€ã§è‡ªå‹•é§è»Šã‚„ADASã®ãƒ‡ãƒ¢
R2M 2016/01/06
ä»®æƒ³ç’°å¢ƒã§ãƒ‡ãƒ¼ã‚¿ã‚’æ§‹ç¯‰ã—ã¦ã€å®Ÿéš›ã®ç’°å¢ƒã§ãƒã‚§ãƒƒã‚¯ã£ã¦ã®ãŒã™ã”ã„ã€‚ã‚„ã‚‹æ°—ã«ãªã‚Œã°ã‚ã‚Šã¨ã‚ã‚‰ã‚†ã‚‹çŠ¶æ³ã‚’ä¸€ä¸‡å°ä¸¦åˆ—ã§å¦ç¿’ã¨ã‹å‡ºæ¥ã‚‹è¨³ã ã—ãã€‚

è‡ªå‹•é‹è»¢

GPGPU
ãƒªãƒ³ã‚¯
AMD GCN isa ã¨ NVIDA SASS ã§ã®ã‚¢ãƒ‰ãƒ¬ã‚¹æ¼”ç®—ã®ãƒãƒ¥ãƒ¼ãƒ‹ãƒ³ã‚° - Fixstars Tech Blog /proc/cpuinfo
ã“ã®ãƒ–ãƒã‚°ã¯ã€æ ªå¼ä¼šç¤¾ãƒ•ã‚£ãƒƒã‚¯ã‚¹ã‚¿ãƒ¼ã‚ºã®ã‚¨ãƒ³ã‚¸ãƒ‹ã‚¢ãŒã€ã‚ã‚‰ã‚†ã‚‹ãƒ†ãƒ¼ãƒžã«ã¤ã„ã¦è‡ªç”±ã«æ›¸ã„ã¦ã„ã‚‹ãƒ–ãƒã‚°ã§ã™ã€‚ ã¡ã‚‡ã£ã¨ã—ãŸãã£ã‹ã‘ãŒã‚ã£ã¦ã€AMDç”¨ã®GPUã¨NVIDIAç”¨ã®GPUä¸¡æ–¹ã§é«˜é€ŸåŒ–ä½œæ¥ã‚’è¡Œã„ã¾ã—ãŸã€‚ ãã®ã¨ãã«å¾—ã‚‰ã‚ŒãŸçŸ¥è¦‹ã‚’æ›¸ã„ã¦ãŠãã¾ã™ã€‚ AMD GCN isa ã¨ NVIDIA SASS GPUã§ãƒ—ãƒã‚°ãƒ©ãƒŸãƒ³ã‚°ã‚’è¡Œã†ã¨ãã¯ã€OpenCLã€CUDA ãªã©ã‚’ä½¿ã†ã¨æ€ã„ã¾ã™ãŒã€ã“ã‚Œã‚‰ã¯ã‚ˆã‚Šä½Žãƒ¬ãƒ™ãƒ«ãªGPUç”¨ã®æ©Ÿæ¢°èªžã«ã‚³ãƒ³ãƒ‘ã‚¤ãƒ«ã•ã‚Œã¾ã™ã€‚ AMD ã®GPUã§ã¯ã€ã“ã‚Œã«â€ISAâ€ã€NVIDIA ã® GPU ã§ã¯ã€ã“ã‚Œã‚’â€SASSâ€ã¨ã„ã†åå‰ãŒä»˜ã‘ã‚‰ã‚Œã¦ã„ã¾ã™ã€‚ISAã¯ä¸€èˆ¬çš„ãªå˜èªžãªã®ã§ã€åˆ¥ã«æ£ã—ã„åå‰ãŒã‚ã‚‹ã‹ã‚‚ã—ã‚Œã¾ã›ã‚“ãŒã€ãƒ“ãƒ«ãƒ‰æ™‚ã®ã‚ªãƒ—ã‚·ãƒ§ãƒ³ãªã©ã§æŒ‡å®šã™ã‚‹å ´åˆã¯ã€â€“isa ç‰ã‚’æŒ‡å®šã—ã¦ã„ã‚‹ã®ã§ã€ã“ã“ã§ã¯ã€GCN isa ã¨ã—ã¾ã™ã€‚ é€šå¸¸ã®GPUãƒ—ãƒã‚°ãƒ©ãƒŸãƒ³ã‚°ã§ã“ã‚Œã‚‰ã‚’è¦‹ã‚‹ã“ã¨ã¯ãªã„ã‹ã‚‚
R2M 2015/11/11
GPGPU
ãƒªãƒ³ã‚¯
AMD Radeon R9 Fury Xã®æ€§èƒ½ã‚’ViennaCLã§è©¦ã™ - Fixstars Tech Blog /proc/cpuinfo
ä»Šå›žã¯ã€FuryXã‚’å«ã‚ã¦ã€ä»Šã€ç¤¾å†…ã§æ‰‹ã«å…¥ã‚‹ãƒ‡ãƒã‚¤ã‚¹ã‚’ã„ãã¤ã‹é›†ã‚ã¦æ¯”è¼ƒã—ã¦ã¿ã¾ã—ãŸã€‚ GPU AMD Radeon R9 Fury X, AMD APP SDK 3.0, Ubuntu 15.04 AMD Radeon 290X, AMD APP SDK 3.0, Ubuntu 14.04 NVIDIA Tesla K20c, CUDA7, Fedora21 NVIDIA GeForce GTX 680, CUDA7, Fedora21 CPU Intel Core-i7 3770K, gcc 4.9.2, Ubuntu 15.04 Intel Core-i7 2600K, gcc 4.8.4, Ubuntu 14.04 å„ãƒ‡ãƒã‚¤ã‚¹ãŒãã‚Œãžã‚Œåˆ¥ã€…ã®è¨ˆç®—æ©Ÿã«ã‚¤ãƒ³ã‚¹ãƒˆãƒ¼ãƒ«ã•ã‚Œã¦ã„ã¦ç’°å¢ƒãŒæƒãˆã‚‰ã‚Œã¦ã„ãªã„ã®ã§åŽ³å¯†ãªæ„å‘³ã§ã®ãƒ‡ãƒã‚¤ã‚¹æ€§èƒ½ã®æ¯”è¼ƒã¯ã§ãã¦ã„ãªã„ã®ã§ã‚ãã¾ã§å‚è€ƒå€¤ã¨æ‰ãˆã¦ãã ã•ã„ã€‚
R2M 2015/11/11
GPGPU
ãƒªãƒ³ã‚¯
introdunction to SIMD programming - primitive: blog
Unite 2015 Tokyo ã®è¬›æ¼”ã§è©³ç´°ã‚’è©±ã›ãªã‹ã£ãŸã®ãŒå¿ƒæ®‹ã‚Šã ã£ãŸã®ã§ã€å¤§é‡ã®ã‚ªãƒ–ã‚¸ã‚§ã‚¯ãƒˆã®æ›´æ–°å‡¦ç†ã«ã¤ã„ã¦ã“ã®å ´ã§æ›¸ã„ã¦ã¿ã¾ã™ã€‚ ä¸»ã« C++ ã§ã€ç°¡å˜ãªãƒ‘ãƒ¼ãƒ†ã‚£ã‚¯ãƒ«ã‚¨ãƒ³ã‚¸ãƒ³ã‚’ä½œã‚Šã€ãã‚Œã‚’ SIMD ã‚’ç”¨ã„ã¦é«˜é€ŸåŒ–ã™ã‚‹æ‰‹é †ã‚’è§£èª¬ã—ã¾ã™ã€‚ è©±ã‚’ç°¡å˜ã«ã™ã‚‹ãŸã‚ã€ä»¥ä¸‹ã®å‰æã‚’è¨ã‘ã¾ã™ã€‚ ãƒ»x86 ç’°å¢ƒã®ã¿è€ƒæ…® ãƒ»ãƒ‘ãƒ¼ãƒ†ã‚£ã‚¯ãƒ«ã¯ä½ç½®ã¨é€Ÿåº¦ã®ã¿ã‚’ä¿æŒ ãƒ»ãƒ‘ãƒ¼ãƒ†ã‚£ã‚¯ãƒ«åŒå£«ã®ç›¸äº’è¡çªã¯ç·å½“ãŸã‚Šã§è¨ˆç®— ç·å½“ãŸã‚Šãªã®ã§è¶…é…ã„ã§ã™ãŒã€å®Ÿè£…ã¯å®¹æ˜“ã§ SIMD ã«ã‚ˆã‚‹æ©æµã‚’å—ã‘ã‚„ã™ãã€é¡Œæã¨ã—ã¦æ‰‹é ƒã§ã™ã€‚ ã“ã®è¨˜äº‹ã®ä¸ã§å¼•ç”¨ã•ã‚Œã¦ã„ã‚‹ã‚½ãƒ¼ã‚¹ã®å…ƒã¯ ã“ã¡ã‚‰ã€ãƒ“ãƒ«ãƒ‰çµæžœ (ä¸Šã®ã‚¹ã‚¯ãƒªãƒ¼ãƒ³ã‚·ãƒ§ãƒƒãƒˆã®ãƒ‡ãƒ¢ãƒ—ãƒã‚°ãƒ©ãƒ ) ã¯ ã“ã¡ã‚‰ ã«ãªã‚Šã¾ã™ã€‚ ç›¸äº’è¡çªã™ã‚‹ãƒ‘ãƒ¼ãƒ†ã‚£ã‚¯ãƒ«ã‚’å®Ÿè£…ã™ã‚‹å ´åˆã€ãŠäº’ã„ã®è·é›¢ã‚’è¨ˆç®—ã—ã€å½“ãŸã£ã¦ã„ãŸã‚‰ã‚ã‚Šè¾¼ã¿å…·åˆã«å¿œã˜ã¦æŠ¼ã—è¿”ã™ã€ã¨ã„ã†ã®ãŒã‚ˆãã‚ã‚‹å®Ÿè£…ã ã¨æ€ã„ã¾ã™ã€‚ã¾ãšã¯ãã‚Œã‚’ã‚¹ãƒˆãƒ¬ãƒ¼ãƒˆã«
R2M 2015/05/27
SIMD

ãƒ—ãƒã‚°ãƒ©ãƒŸãƒ³ã‚°

gpgpu
ãƒªãƒ³ã‚¯
OpenCLã‚„ã‚‹å‰ã«SIMDä½¿ã„åˆ‡ã‚Œã£ã¦ã„ã†å¹»æƒ³ - aokomoriuta's blog
ãŠè©«ã³ ã‚¢ãƒ©ã‚¤ãƒ³å¿˜ã‚Œã¦ã¾ã—ãŸã”ã‚ã‚“ãªã•ã„ã€ã§ã‚‚æ™‚é–“ã‚ã‚“ã¾ã‚Šå¤‰ã‚ã£ã¦ãªã„ã‹ã‚‰è¨±ã—ã¦ãƒ»ãƒ»ãƒ» https://t.co/JNtq2U2kMqâ€” é’åå®ˆæŒ (@aokomoriuta) April 29, 2015 ã§ã¯æœ¬ç·¨ã©ã†ãžâ†“ æœ¬ç·¨ è‹¥å¹²è©±é¡Œã«ä¹—ã‚Šé…ã‚ŒãŸæ„Ÿã‚ã‚Šã¾ã™ãŒã€‚ d.hatena.ne.jp ã‘ã©ã€SSEã‚‚çŸ¥ã‚‰ããƒ¼ã€SIMDã‚‚çŸ¥ã‚‰ããƒ¼ã€ãªã‚“ã‹ä¿ºãŒæ›¸ã„ãŸã‚¢ãƒ«ã‚´ãƒªã‚ºãƒ é…ã„ã‘ã©ã¨ã‚Šã‚ãˆãšOpenCLã¨ã‹ã§é«˜é€ŸåŒ–ã—ã‚ˆã£ã‹ãªãƒ¼ã¨ã‹ãã€ç”˜ã£ãŸã‚Œã¦ã‚“ã˜ã‚ƒããˆã‚ˆã€‚CPUã‚’ã‚‚ã£ã¨ä½¿ã„ãã‚Œã‚ˆã€‚ãŠå‰ã®ã‚¢ã‚ªã„ã‚³ãƒ¼ãƒ‰ã®ã›ã„ã§CPUãŒæ³£ã„ã¦ã‚‹ã‚ˆã€‚ã£ã¦ã„ã†è©±ã§ã™ã‚ˆã€‚ GPGPUãªã‚“ã¦ã®ã¯SIMDã‚’ä½¿ã„åˆ‡ã£ãŸå¾Œã®è©±ã§ã—ã‚‡ã€‚ GPGPUã™ã‚‹ã®ã«GPUã®ãƒ‘ãƒ¯ãƒ¼ã¨ãƒ¡ãƒ¢ãƒªãŒè¶³ã‚Šã¾ã›ã‚“ã¨ã‹è¨€ã†å‰ã«ã¾ãšSIMDã‹ã‚‰ã ã‚ã€‚ ã¨ã‹è¨€ã‚ã‚Œã¦ãŸã®ã§ã€æ¤œè¨¼ã™ã‚‹ã“ã¨ã«ã—ãŸï¼ˆã‚„ã£ã¤ã‘ï¼‰ã€‚ ç’°å¢ƒ OS: Windows 7 Profession
R2M 2015/05/20
SIMD

GPGPU
ãƒªãƒ³ã‚¯
GPGPUãŒPostgreSQLã‚’åŠ é€Ÿã™ã‚‹ ï¼ˆPostgreSQLã‚«ãƒ³ãƒ•ã‚¡ãƒ¬ãƒ³ã‚¹2014ï¼‰
https://www.postgresql.jp/events/jpug-pgcon2014 2014å¹´ æ—¥æœ¬PostgreSQLãƒ¦ãƒ¼ã‚¶ä¼šä¸»å‚¬ã® PostgreSQLã‚«ãƒ³ãƒ•ã‚¡ãƒ¬ãƒ³ã‚¹ã§ã®è¬›æ¼”å†…å®¹ã§ã™ã€‚
R2M 2014/12/05
APUã§ã©ã†ãªã‚‹ã®ã‹æ°—ã«ãªã‚‹ã€‚

database

GPGPU
ãƒªãƒ³ã‚¯
(JP) GPGPUãŒPostgreSQLã‚’åŠ é€Ÿã™ã‚‹
http://www.insight-tec.com/dbts-tokyo-2014.html -------- CPU/GPUã¨ã„ã†ç‰¹æ€§ã®ç•°ãªã‚‹ãƒ—ãƒã‚»ãƒƒã‚µã®å¾—æ„åˆ†é‡Žã‚’çµ„ã¿åˆã‚ã›ã¦è²»ç”¨ãƒ»é›»åŠ›å¯¾æ€§èƒ½ã®å‘ä¸Šã‚’ç‹™ã†ãƒ˜ãƒ†ãƒã‚¸ãƒ‹ã‚¢ã‚¹è¨ˆç®—æŠ€è¡“ã¯ã€è¿‘å¹´ã®åŠå°Žä½“æŠ€è¡“ã®ãƒˆãƒ¬ãƒ³ãƒ‰ã®ä¸€ã¤ã§ã™ã€‚ æœ¬ã‚»ãƒƒã‚·ãƒ§ãƒ³ã§ã¯ã€SQLãƒ¯ãƒ¼ã‚¯ãƒãƒ¼ãƒ‰ã®ã†ã¡CPUè² è·ã®é«˜ã„ãƒã‚¸ãƒƒã‚¯ã‚’GPUã«ã‚ªãƒ•ãƒãƒ¼ãƒ‰ã™ã‚‹äº‹ã§å®‰ä¾¡ã«æ¤œç´¢å‡¦ç†ã®é«˜é€ŸåŒ–ã‚’å®Ÿç¾ã™ã‚‹PostgreSQLç”¨ã®æ‹¡å¼µãƒ¢ã‚¸ãƒ¥ãƒ¼ãƒ« PG-Strom ã¨ã€GPGPUã‚’ã¯ã˜ã‚ã¨ã™ã‚‹å‘¨è¾ºæŠ€è¡“ã«ã¤ã„ã¦ã”ç´¹ä»‹ã—ã¾ã™ã€‚
R2M 2014/11/14
database

GPGPU
ãƒªãƒ³ã‚¯
ãƒ•ãƒ«ãƒ‘ãƒ¯ãƒ¼ã§åŠ é€Ÿã™ã‚‹GPUã®å’†å“®ã‚’èžã‘!æœ€å…ˆç«¯GPUã§PostgreSQLã‚’20å€é«˜é€ŸåŒ–!?:é›»è„³ãƒ’ãƒƒãƒãƒã‚¤ã‚¯ã‚¬ã‚¤ãƒ‰: é›»è„³ç©ºé–“ã‚«ã‚¦ãƒœãƒ¼ã‚¤ã‚ºZZ(é›»è„³ç©ºé–“ã‚«ã‚¦ãƒœãƒ¼ã‚¤ã‚º) - ãƒ‹ã‚³ãƒ‹ã‚³ãƒãƒ£ãƒ³ãƒãƒ«ï¼šç”Ÿæ´»
GPUã£ã¦ã‚ã‚‹ã ã‚? ã‚°ãƒ©ãƒ•ã‚£ãƒƒã‚¯ã‚¹ãƒ»ãƒ—ãƒã‚»ãƒƒã‚·ãƒ³ã‚°ãƒ»ãƒ¦ãƒ‹ãƒƒãƒˆã€‚ ã‚³ãƒ³ãƒ”ãƒ¥ãƒ¼ã‚¿ã®ä¸ã§æœ€ã‚‚é«˜é€Ÿãªéƒ¨å“ã¯ãªã‚“ã¨ã„ã£ã¦ã‚‚GPUã ã€‚ ãŸã¨ãˆã°nViditaã®æœ€æ–°SoCã§ã‚ã‚‹Tegra K1ã€‚ CPUéƒ¨ã¯å˜ç²¾åº¦ã§73.6ã‚®ã‚¬FLOPS(ãƒ•ãƒãƒƒãƒ—ã‚¹)ã€ã ãŒGPUéƒ¨ã¯364.8ã‚®ã‚¬FLOPSã ã€‚ ã“ã‚ŒãŒ30ä¸‡å††ã™ã‚‹æœ€æ–°ã®GPUãƒœãƒ¼ãƒ‰ã€GeForce GTX TITAN Zã«ãªã‚‹ã¨ã€ãªã‚“ã¨ã³ã£ãã‚Šã€8ãƒ†ãƒ©FLOPSã ã€‚ ã“ã‚Œã¯Intelã®æœ€é«˜ç´šIA-64ãƒãƒƒãƒ—ã€Xeon ES-2687Wã®198.4ã‚®ã‚¬FLOPSã‚’å¤§ããä¸Šå›žã‚‹ã€‚ Core i7ã¯92ã‚®ã‚¬FLOPSã«éŽãŽãªã„ã€‚ ã¡ãªã¿ã«ã‹ã¤ã¦æ•°å„„å††ã—ãŸã‚¹ãƒ¼ãƒ‘ãƒ¼ã‚³ãƒ³ãƒ”ãƒ¥ãƒ¼ã‚¿ã€Cray-2ã¯ã‚ãšã‹1.9ã‚®ã‚¬FLOPSã«éŽãŽãšã€ä»Šã®ã‚³ãƒ³ãƒ”ãƒ¥ãƒ¼ã‚¿ãŒã©ã‚Œã ã‘é€Ÿã„ã‹çªºã„çŸ¥ã‚Œã‚‹ã€‚JAMSTECã®åœ°çƒã‚·ãƒŸãƒ¥ãƒ¬ãƒ¼ã‚¿ã¯131ãƒ†ãƒ©FLOPSã ãŒã€è¿½ã„ã¤ã‹ã‚Œã‚‹ã®ã¯æ™‚é–“ã®å•é¡Œã ã‚
R2M 2014/06/16
GPGPU

database
ãƒªãƒ³ã‚¯
æµä½“ã‚’ã‚·ãƒŸãƒ¥ãƒ¬ãƒ¼ãƒˆã—ã¦çµµã®å…·ã®åŽšã¿ã‚’å†ç¾ã™ã‚‹ãƒšã‚¤ãƒ³ãƒˆã‚½ãƒ•ãƒˆã€ŒVerve Painterã€NOT SUPPORTED
R2M 2014/03/17
Windows

ãƒ•ãƒªãƒ¼ã‚½ãƒ•ãƒˆ

Art

GPGPU
ãƒªãƒ³ã‚¯
1 2 æ¬¡ã®ãƒšãƒ¼ã‚¸