ïŒNLPïŒ Deep LearningïŒDLïŒã¯ãã³ã³ãã¥ãŒãã£ã³ã°èœåã«å¯ŸããéèŠãé«ãåéã§ãããããGPUã®éžæã«ãã£ãŠããã®åéã§ã®çµéšãåºæ¬çã«æ±ºãŸããŸããããããæ°ããGPUãè³Œå ¥ããéã«èæ ®ãã¹ãéèŠãªããããã£ã¯äœã§ããïŒã¡ã¢ãªãã³ã¢ããã³ãœã«ã³ã¢ïŒãéã®äŸ¡å€ã®èŠ³ç¹ããæè¯ã®éžæãããæ¹æ³ã¯ïŒãã®èšäºã§ã¯ãããããã¹ãŠã®è³ªåããããã誀解ã詳现ã«åæããGPUãçŽæçã«ç解ã§ããããã«ãããšãšãã«ãæ£ããéžæãè¡ãããã®ãã³ããããã€ã玹ä»ããŸãã
ãã®èšäºã¯ãGPUã«ã€ããŠã®ããã€ãã®ç°ãªãã¬ãã«ã®ç解ãæäŸããããã«æžãããŠããŸãã NVIDIAã®æ°ããAmpereã·ãªãŒãºãéžæè¢ããããŸãïŒ
- GPUã®è©³çŽ°ãGPUãæ£ç¢ºã«é«éåããçç±ãããã³NVIDIA RTX 30 Ampereã·ãªãŒãºã®æ°ããGPUã®ç¬èªæ§ã«èå³ããªãå Žåã¯ãèšäºã®åé ãã¹ãããããŠãé床ãš1ãã«ãããã®é床ã®ã°ã©ããããã³æšå¥šäºé ã®ã»ã¯ã·ã§ã³ã«é²ãããšãã§ããŸããããããã®èšäºã®æ žå¿ã§ãããæã䟡å€ã®ããã³ã³ãã³ãã§ãã
- ç¹å®ã®è³ªåã«èå³ãããå Žåã¯ãèšäºã®æåŸã®éšåã§æãé »ç¹ã«åãäžããŸããã
- GPUãšTensorCoreãã©ã®ããã«æ©èœããããæ·±ãç解ããå¿ èŠãããå Žåã¯ããã®èšäºãæåããæåŸãŸã§èªãããšããå§ãããŸããç¹å®ã®äž»é¡ã«é¢ããç¥èã«å¿ããŠã1ã2ç« ãã¹ãããã§ããŸãã
åã»ã¯ã·ã§ã³ã®åã«ã¯ãå šäœãèªããã©ãããå€æããã®ã«åœ¹ç«ã€çãèŠçŽããããŸãã
ã³ã³ãã³ã
GPU?
GPU,
/ L1 /
Ampere
Ampere
Ampere
Ampere / RTX 30
GPU
GPU
GPU
11 ?
11 ?
GPU-
GPU
GPU?
PCIe 4.0?
PCIe 8x/16x?
RTX 3090, 3 PCIe?
4 RTX 3090 4 RTX 3080?
GPU ?
NVLink, ?
. ?
?
?
Intel GPU?
?
AMD GPU + ROCm - NVIDIA GPU + CUDA?
, â GPU?
,
ãã®èšäºã¯æ¬¡ã®ããã«æ§æãããŠããŸãããŸããGPUãé«éåããçç±ã説æããŸããããã»ããµãšGPUã®éãããã³ãœã«ã³ã¢ãã¡ã¢ãªåž¯åå¹ ãGPUã¡ã¢ãªéå±€ãããã³ããããã¹ãŠãGOã¿ã¹ã¯ã®ããã©ãŒãã³ã¹ã«ã©ã®ããã«é¢é£ãããã«ã€ããŠèª¬æããŸãããããã®èª¬æã¯ãå¿ èŠãªGPUãã©ã¡ãŒã¿ãŒãããããç解ããã®ã«åœ¹ç«ã€å ŽåããããŸãã次ã«ãGPUããã©ãŒãã³ã¹ã®çè«çæšå®å€ãšããã€ã¢ã¹ã®ãªãä¿¡é Œæ§ã®é«ãããã©ãŒãã³ã¹ããŒã¿ãååŸããããã®ãããã€ãã®NVIDIAé床ãã¹ããšã®å¯Ÿå¿ã«ã€ããŠèª¬æããŸããè³Œå ¥æã«èæ ®ãã¹ãNVIDIARTX 30Ampereã·ãªãŒãºGPUã®ç¬èªã®æ©èœã«ã€ããŠèª¬æããŸãã次ã«ã1ã2ãããã4ã8ãããã³GPUã¯ã©ã¹ã¿ãŒã®GPUã«é¢ããæšå¥šäºé ã瀺ããŸãã次ã«ãTwitterã§å°ãããããããã質åãžã®åçã®ã»ã¯ã·ã§ã³ããããŸãããŸããäžè¬çãªèª€è§£ãææããã¯ã©ãŠããšãã¹ã¯ããããå·åŽãAMDãšNVIDIAãªã©ã®ããŸããŸãªåé¡ãæµ®ã圫ãã«ããŸãã
GPUã¯ã©ã®ããã«æ©èœããŸããïŒ
GPUãé »ç¹ã«äœ¿çšããå Žåã¯ãGPUãã©ã®ããã«æ©èœããããç解ããŠãããšåœ¹ç«ã¡ãŸãããã®ç¥èã¯ãGPUãé ãå Žåãšéãå Žåãããçç±ãç解ããã®ã«åœ¹ç«ã¡ãŸãããããŠãGPUãå¿ èŠãã©ããããããŠå°æ¥ã©ã®ããŒããŠã§ã¢ãªãã·ã§ã³ãGPUãšç«¶åã§ããããç解ã§ããŸããç¹å®ã®GPUãéžæããããã®æçšãªããã©ãŒãã³ã¹æ å ±ãšåŒæ°ãå¿ èŠãªå Žåã¯ããã®ã»ã¯ã·ã§ã³ãã¹ãããã§ããŸããGPUãã©ã®ããã«æ©èœãããã«ã€ããŠã®æãäžè¬çãªèª¬æã¯ãQuoraã®åçã«ãããŸãã
ããã¯äžè¬çãªèª¬æã§ãããGPUãããã»ããµãããGOã«é©ããŠããçç±ã®åé¡ããã説æããŠããŸãã詳现ã調ã¹ããšãGPUã®éããããããŸãã
åŠçé床ã«åœ±é¿ãäžããæãéèŠãªGPUç¹æ§
ãã®ã»ã¯ã·ã§ã³ã¯ãGOã®åéã§ã®ããã©ãŒãã³ã¹ã«ã€ããŠããçŽæçã«èããã®ã«åœ¹ç«ã¡ãŸãããã®ç解ã¯ãå°æ¥ã®GPUãèªåã§è©äŸ¡ããã®ã«åœ¹ç«ã¡ãŸãã
ãã³ãœã«ã³ã¢
æŠèŠïŒ
- ãã³ãœã«ã«ãŒãã«ã¯ãä¹ç®ãšå ç®ãã«ãŠã³ãããããã«å¿ èŠãªã¯ããã¯ãµã€ã¯ã«æ°ã16ååæžããŸããç§ã®äŸã§ã¯ã32Ã32ãããªãã¯ã¹ã®å Žåã¯128ãã8ã¯ããã¯ãµã€ã¯ã«ã«ãªããŸãã
- ãã³ãœã«ã«ãŒãã«ã¯ãã¡ã¢ãªã¢ã¯ã»ã¹ãµã€ã¯ã«ãç¯çŽããããšã«ãããå ±æã¡ã¢ãªãžã®ç¹°ãè¿ãã¢ã¯ã»ã¹ãžã®äŸåãæžãããŸãã
- ãã³ãœã«ã«ãŒãã«ã¯éåžžã«é«éã§ãããããèšç®ã¯ãã¯ãããã«ããã¯ã§ã¯ãããŸãããå¯äžã®ããã«ããã¯ã¯ãããããžã®ããŒã¿ã®è»¢éã§ãã
ä»æ¥ãå®äŸ¡ãªGPUãéåžžã«å€ããããã»ãšãã©ã®äººããã³ãœã«ã³ã¢ãåããGPUãè³Œå ¥ã§ããŸãããããã£ãŠãç§ã¯åžžã«TensorCoreãåããGPUããå§ãããŸãããããªãã¯ã¹ä¹ç®ã«ç¹åãããããã®èšç®ã¢ãžã¥ãŒã«ã®éèŠæ§ãç解ããã«ã¯ãããããã©ã®ããã«æ©èœããããç解ããããšã圹ç«ã¡ãŸãããã¹ãŠã®è¡åã®ãµã€ãºã32Ã32ã§ããè¡åä¹ç®A * B = Cã®ç°¡åãªäŸã䜿çšããŠããã³ãœã«ã«ãŒãã«ãããå Žåãšãªãå Žåã®ä¹ç®ãã©ã®ããã«èŠãããã瀺ããŸãã
ãããç解ããã«ã¯ãæåã«ããŒã®æŠå¿µãç解ããå¿ èŠããããŸããããã»ããµã¯1 GHzã§åäœããŠããå Žåãããã¯10ãè¡ã91ç§ãããã®ãã£ãã¯æ°ãåã¯ããã¯ã¯èšç®ã®æ©äŒã§ãããã ããã»ãšãã©ã®å Žåãæäœã«ã¯1ã¯ããã¯ãµã€ã¯ã«ããé·ãæéãããããŸãããã€ãã©ã€ã³ãå€æããŸããã1ã€ã®æäœã®å®è¡ãéå§ããã«ã¯ãæåã«ãåã®æäœãå®äºããããã«å¿ èŠãªæ°ã®ã¯ããã¯ãµã€ã¯ã«ãåŸ æ©ããå¿ èŠããããŸããããã¯ãé 延æäœãšãåŒã°ããŸãã
æäœã®éèŠãªæéãŸãã¯é 延ããã£ãã¯åäœã§æ¬¡ã«ç€ºããŸãã
- æ倧48GBã®ã°ããŒãã«ã¡ã¢ãªãžã®ã¢ã¯ã»ã¹ïŒã200ã¯ããã¯ãµã€ã¯ã«ã
- å ±æã¡ã¢ãªã¢ã¯ã»ã¹ïŒã¹ããªãŒãã³ã°ãã«ãããã»ããµãããæ倧164 KBïŒïŒã20ã¯ããã¯ã
- è€åä¹ç®-å ç®ïŒSUSïŒïŒ4å°ç¯ã
- ãã³ãœã«ã«ãŒãã«ã§ã®ãããªãã¯ã¹ä¹ç®ïŒ1ã¯ããã¯ãµã€ã¯ã«ã
ãŸããGPUã®ã¹ã¬ããã®æå°åäœïŒ32ã¹ã¬ããã®ãã±ããïŒã¯ã¯ãŒããšåŒã°ããããšãç¥ã£ãŠããå¿ èŠããããŸããã¯ãŒãã¯éåžžåæããŠæ©èœããŸããã¯ãŒãå ã®ãã¹ãŠã®ã¹ã¬ããã¯äºãã«åŸ æ©ããå¿ èŠããããŸãããã¹ãŠã®GPUã¡ã¢ãªæäœã¯ã¯ãŒãçšã«æé©åãããŠããŸããããšãã°ãã°ããŒãã«ã¡ã¢ãªããã®ããŒãã«ã¯ã32 * 4ãã€ã-32åã®æµ®åå°æ°ç¹æ°ãå¿ èŠã§ããã¯ãŒãå ã®ã¹ã¬ããããšã«1ã€ã®ãã®ãããªæ°ã§ããã¹ããªãŒãã³ã°ãã«ãããã»ããµïŒGPUã®ããã»ããµã³ã¢ã«çžåœïŒã§ã¯ãæ倧32ã®ã¯ãŒã= 1024ã¹ã¬ãããååšããå¯èœæ§ããããŸãããã«ãããã»ããµãªãœãŒã¹ã¯ããã¹ãŠã®ã¢ã¯ãã£ããªã¯ãŒãéã§å ±æãããŸãããããã£ãŠã1ã€ã®ã¯ãŒãã«å€ãã®ã¬ãžã¹ã¿ãå ±æã¡ã¢ãªãããã³ãã³ãœã«ã³ã¢ãªãœãŒã¹ãå«ãŸããããã«ãåäœã«å¿ èŠãªã¯ãŒããå°ãªããªãå ŽåããããŸãã
äž¡æ¹ã®äŸã§ãåãã³ã³ãã¥ãŒãã£ã³ã°ãªãœãŒã¹ããããšä»®å®ããŸãããããã®32Ã32ãããªãã¯ã¹ä¹ç®ã®å°ããªäŸã§ã¯ã8ã€ã®ãã«ãããã»ããµïŒRTX 3090ã®çŽ10ïŒ ïŒãšãã«ãããã»ããµã§8ã€ã®ã¯ãŒãã䜿çšããŸãã
ãã³ãœã«ã«ãŒãã«ã䜿çšããªããããªãã¯ã¹ä¹ç®
ããããã32Ã32ã®ãµã€ãºã®è¡åA * B = Cãä¹ç®ããå¿ èŠãããå Žåãã¢ã¯ã»ã¹é 延ã¯çŽ10åã®1ã§ãããããåžžã«ã¢ã¯ã»ã¹ããŠããã¡ã¢ãªããå ±æã¡ã¢ãªã«ããŒã¿ãããŒãããå¿ èŠããããŸãïŒ 200ããŒãããã³20ããŒïŒãå ±æã¡ã¢ãªå ã®ã¡ã¢ãªã®ãããã¯ã¯ãå€ãã®å Žåãã¡ã¢ãªã¿ã€ã«ããŸãã¯åã«ã¿ã€ã«ãšåŒã°ããŸãã 2ã€ã®32x32ãããŒãã£ã³ã°ãã€ã³ãçªå·ãå ±æã¡ã¢ãªã¿ã€ã«ã«ããŒãããã«ã¯ã2 * 32ã¯ãŒãã䜿çšããŠäžŠè¡ããŠå®è¡ã§ããŸãããããã8ã€ã®ã¯ãŒããæã€8ã€ã®ãã«ãããã»ããµãããããã䞊ååã®ãããã§ãã°ããŒãã«ã¡ã¢ãªããå ±æã¡ã¢ãªãžã®1ã€ã®ã·ãŒã±ã³ã·ã£ã«ããŒããå®è¡ããå¿ èŠããããŸããããã«ã¯200ã¯ããã¯ãµã€ã¯ã«ããããŸãã
è¡åãä¹ç®ããã«ã¯ãå ±æã¡ã¢ãªAãšå ±æã¡ã¢ãªBãã32åã®æ°å€ã®ãã¯ãã«ãããŒãããCMSãå®è¡ããŠãããåºåãã¬ãžã¹ã¿Cã«æ ŒçŽããå¿ èŠããããŸããåãã«ãããã»ããµã8ã€ã®ã¹ã«ã©ãŒç©ïŒ32Ã32ïŒãåŠçããããã«ããã®äœæ¥ãåå²ããŸãã ïŒCã®8ã€ã®åºåããŒã¿ãèšç®ããŸãããªãããããæ£ç¢ºã«8ã€ããã®ãïŒå€ãã¢ã«ãŽãªãºã ã§ã¯-4ïŒãããã¯çŽç²ã«æè¡çãªæ©èœã§ãããããç解ããã«ã¯ãScottGrayã®èšäºãèªãããšããå§ãããŸããããã¯ãå ±æã¡ã¢ãªãžã®ã¢ã¯ã»ã¹ã8åããããããã20ãµã€ã¯ã«ã®ã³ã¹ãããããã8åã®SLSæäœïŒ32䞊åïŒããããã4ãµã€ã¯ã«ã®ã³ã¹ãã«ãªãããšãæå³ããŸããåèšã§ãã³ã¹ãã¯æ¬¡ã®ããã«ãªããŸã
ã200ãã£ãã¯ïŒã°ããŒãã«ã¡ã¢ãªïŒ+ 8 * 20ãã£ãã¯ïŒå ±æã¡ã¢ãªïŒ+ 8 * 4ãã£ãã¯ïŒCMSïŒ= 392ãã£ãã¯æ¬¡ã«
ããã³ãœã«ã³ã¢ã®ãã®ã³ã¹ããèŠãŠã¿ãŸãããã
ãã³ãœã«ã«ãŒãã«ã«ãããããªãã¯ã¹ä¹ç®
ãã³ãœã«ã«ãŒãã«ã䜿çšãããšã1ãµã€ã¯ã«ã§4Ã4ã®è¡åãä¹ç®ã§ããŸãããããè¡ãã«ã¯ãã¡ã¢ãªããã³ãœã«ã³ã¢ã«ã³ããŒããå¿ èŠããããŸããäžèšã®ããã«ãã°ããŒãã«ã¡ã¢ãªïŒ200ãã£ãã¯ïŒããããŒã¿ãèªã¿åããå ±æã¡ã¢ãªã«ä¿åããå¿ èŠããããŸãã32Ã32ã®è¡åãä¹ç®ããã«ã¯ããã³ãœã«ã«ãŒãã«ã§8Ã8 = 64ã®æäœãå®è¡ããå¿ èŠããããŸãã1ã€ã®ãã«ãããã»ããµã«ã¯8ã€ã®ãã³ãœã«ã³ã¢ãå«ãŸããŠããŸãã8ã€ã®ãã«ãããã»ããµã§ã64ã®ãã³ãœã«ã³ã¢ããããŸã-å¿ èŠãªæ°ã ãã§ãïŒå ±æã¡ã¢ãªãããã³ãœã«ã³ã¢ã«1åã®è»¢éïŒ20ã¯ããã¯ãµã€ã¯ã«ïŒã§ããŒã¿ã転éããããã64ã®æäœãã¹ãŠã䞊è¡ããŠïŒ1ã¯ããã¯ãµã€ã¯ã«ïŒå®è¡ã§ããŸããããã¯ããã³ãœã«ã³ã¢ã§ã®ãããªãã¯ã¹ä¹ç®ã®ç·ã³ã¹ãã次ã®ããã«ãªãããšãæå³ããŸãïŒ
200ã¯ããã¯ãµã€ã¯ã«ïŒã°ããŒãã«ã¡ã¢ãªïŒ+ 20ã¯ããã¯ãµã€ã¯ã«ïŒå ±æã¡ã¢ãªïŒ+ 1ã¯ããã¯ãµã€ã¯ã«ïŒãã³ãœã«ã³ã¢ïŒ= 221ã¯ããã¯ãµã€ã¯ã«
ãããã£ãŠããã³ãœã«ã«ãŒãã«ã䜿çšãããšããããªãã¯ã¹ä¹ç®ã®ã³ã¹ãã392ãã221ã¯ããã¯ãµã€ã¯ã«ã«å€§å¹ ã«åæžã§ããŸããç°¡ç¥åããäŸã§ã¯ããã³ãœã«ã«ãŒãã«ã«ãã£ãŠå ±æã¡ã¢ãªã¢ã¯ã»ã¹ãšSNSæäœã®äž¡æ¹ã®ã³ã¹ããåæžãããŸããã
ãã®äŸã¯ããã³ãœã«ã«ãŒãã«ãããå Žåãšãªãå Žåã®äžé£ã®èšç®ã¹ãããã«å€§ãŸãã«åŸããŸãããããã¯éåžžã«åçŽåãããäŸã§ããããšã«æ³šæããŠãã ãããå®éã®å Žåããããªãã¯ã¹ã®ä¹ç®ã«ã¯ã倧ããªã¡ã¢ãªã¿ã€ã«ãšããããã«ç°ãªãäžé£ã®ã¢ã¯ã·ã§ã³ãå«ãŸããŸãã
ãã ãããã®äŸã§ã¯ã次ã®å±æ§ã§ããã¡ã¢ãªåž¯åå¹ ããã³ãœã«ã³ã¢ãåããGPUã«ãšã£ãŠéåžžã«éèŠã§ããçç±ãæããã«ãªã£ãŠããããã«æãããŸãããããªãã¯ã¹ã«ãã³ãœã«ã³ã¢ãä¹ç®ããå Žåãã°ããŒãã«ã¡ã¢ãªã¯æãé«äŸ¡ãªãã®ã§ãããããã°ããŒãã«ã¡ã¢ãªãžã®ã¢ã¯ã»ã¹ã®åŸ ã¡æéãççž®ã§ããã°ãGPUã¯ã¯ããã«é«éã«ãªããŸããããã¯ãã¡ã¢ãªã¯ããã¯é床ãäžããïŒ1ç§ãããã®ã¯ããã¯ãµã€ã¯ã«ãå¢ããããç±ãšé»åã®æ¶è²»éãå¢ããïŒããäžåºŠã«è»¢éã§ããèŠçŽ ã®æ°ïŒãã¹å¹ ïŒãå¢ããããšã«ãã£ãŠå®è¡ã§ããŸãã
ã¡ã¢ãªåž¯åå¹
åã®ã»ã¯ã·ã§ã³ã§ã¯ããã³ãœã«ã«ãŒãã«ã®é床ã«ã€ããŠèª¬æããŸããããããã¯éåžžã«é«éã§ãããããã»ãšãã©ã®æéã¢ã€ãã«ç¶æ ã«ãªããã°ããŒãã«ã¡ã¢ãªããã®ããŒã¿ãå°çããã®ãåŸ ã¡ãŸããããšãã°ãéåžžã«å€§ããªãããªãã¯ã¹ã䜿çšãããBERT Largeãããžã§ã¯ãã®ãã¬ãŒãã³ã°äžïŒãã³ãœã«ã«ãŒãã«ã®å Žåã¯å€§ããã»ã©è¯ãïŒãTFLOPSã§ã®ãã³ãœã«ã«ãŒãã«ã®äœ¿çšçã¯çŽ30ïŒ ã§ãããããã¯ããã³ãœã«ã«ãŒãã«ãã¢ã€ãã«ç¶æ ã§ãã£ãæéã®70ïŒ ãæå³ããŸãã
ããã¯ã2ã€ã®GPUããã³ãœã«ã³ã¢ãšæ¯èŒããå Žåãããããã®æé«ã®ããã©ãŒãã³ã¹ææšã®1ã€ãã¡ã¢ãªåž¯åå¹ ã§ããããšãæå³ããŸããããšãã°ãA100GPUã®åž¯åå¹ ã¯1.555GB / sã§ãããV100ã®åž¯åå¹ ã¯900 GB / sã§ããç°¡åãªèšç®ã«ãããšãA100ã¯V100ããã1555/900 = 1.73åé«éã«ãªããŸãã
å ±æã¡ã¢ãª/ L1ãã£ãã·ã¥/ã¬ãžã¹ã¿
é床å¶éèŠå ã¯ãã³ãœã«ã³ã¢ã®ã¡ã¢ãªãžã®ããŒã¿ã®è»¢éã§ãããããGPUã®ä»ã®ããããã£ã«ç®ãåããå¿ èŠããããŸããããã«ãããããããžã®ããŒã¿ã®è»¢éãé«éåã§ããŸããããã«é¢é£ããã®ã¯ãå ±æã¡ã¢ãªãL1ãã£ãã·ã¥ãããã³ã¬ãžã¹ã¿ã®æ°ã§ããã¡ã¢ãªéå±€ãããŒã¿è»¢éãã©ã®ããã«é«éåããããç解ããã«ã¯ãGPUã§ãããªãã¯ã¹ãã©ã®ããã«ä¹ç®ãããããç解ããããšã圹ç«ã¡ãŸãã
ãããªãã¯ã¹ä¹ç®ã§ã¯ãäœéã®ã°ããŒãã«ã¡ã¢ãªããé«éã®ããŒã«ã«å ±æã¡ã¢ãªããããŠè¶ é«éã¬ãžã¹ã¿ãžãšé²ãã¡ã¢ãªéå±€ã䜿çšããŸãããã ããã¡ã¢ãªãé«éã§ããã»ã©ãã¡ã¢ãªã¯å°ãããªããŸãããããã£ãŠããããªãã¯ã¹ãå°ããªãã®ã«åå²ããŠãããããŒã«ã«å ±æã¡ã¢ãªã§ãããã®å°ããªã¿ã€ã«ãä¹ç®ããå¿ èŠããããŸããããããã°ãããã¯ã¹ããªãŒãã³ã°ãã«ãããã»ããµïŒPMïŒã«ãã°ããè¿ããªããŸããããã¯ããã»ããµã³ã¢ã«çžåœããŸãããã³ãœã«ã³ã¢ã䜿çšãããšããã1ã€ã®ã¹ããããå®è¡ã§ããŸãããã¹ãŠã®ã¿ã€ã«ãååŸãããããã®äžéšããã³ãœã«ã³ã¢ã«ããŒãããŸããå ±æã¡ã¢ãªã¯ãããªãã¯ã¹ã¿ã€ã«ãã°ããŒãã«GPUã¡ã¢ãªããã10ã50åéãåŠçãããã³ãœã«ã³ã¢ã¬ãžã¹ã¿ã¯ãããã°ããŒãã«GPUã¡ã¢ãªããã200åéãåŠçããŸãã
ã¿ã€ã«ã®ãµã€ãºã倧ãããããšãããå€ãã®ã¡ã¢ãªãåå©çšã§ããŸããããã«ã€ããŠã¯ãç§ã®èšäºTPU vsGPUã§è©³ãã説æããŸãããTPUã§ã¯ããã³ãœã«ã³ã¢ããšã«éåžžã«å€§ããªã¿ã€ã«ããããŸããTPUã¯ãã°ããŒãã«ã¡ã¢ãªããã®æ°ãã転éããšã«ãããå€ãã®ã¡ã¢ãªãåå©çšã§ãããããGPUããããããªãã¯ã¹ä¹ç®ã®åŠçããããã«å¹ççã«ãªããŸãã
ã¿ã€ã«ãµã€ãºã¯ãåPMã®ã¡ã¢ãªéã«ãã£ãŠæ±ºãŸããŸããããã¯GPUã®ããã»ããµã³ã¢ã«çžåœããŸããã¢ãŒããã¯ãã£ã«å¿ããŠããããã®ããªã¥ãŒã ã¯æ¬¡ã®ãšããã§ãã
- ãã«ã¿ïŒ96KBå ±æã¡ã¢ãª/ 32KB L1
- ãã¥ãŒãªã³ã°ïŒ64KBå ±æã¡ã¢ãª/ 32KB L1
- ã¢ã³ãã¢ïŒ164KBå ±æã¡ã¢ãª/ 32KB L1
Ampereã«ã¯ã¯ããã«å€ãã®å ±æã¡ã¢ãªãããããšãããããŸããããã«ããããã倧ããªã¿ã€ã«ã䜿çšã§ããããã«ãªããã°ããŒãã«ã¡ã¢ãªã¢ã¯ã»ã¹ã®æ°ãæžããŸãããããã£ãŠãAmpereã¯GPUã¡ã¢ãªåž¯åå¹ ãããå¹ççã«äœ¿çšããŸããããã«ãããããã©ãŒãã³ã¹ã2ã5ïŒ åäžããŸãããã®å¢å ã¯ã巚倧ãªãããªãã¯ã¹ã§ç¹ã«é¡èã§ãã
ã¢ã³ãã¢ãã³ãœã«ã«ãŒãã«ã«ã¯å¥ã®å©ç¹ããããŸããè€æ°ã®ã¹ã¬ããã«å ±éãã倧éã®ããŒã¿ããããŸããããã«ãããã¬ãžã¹ã¿åŒã³åºãã®æ°ãæžããŸããã¬ãžã¹ã¿ã®ãµã€ãºã¯ãPMããã64 kããŸãã¯ã¹ã¬ããããã255ã«å¶éãããŠããŸãã Voltaãšæ¯èŒããŠãAmpere Tensor Coreã¯3åã®1ã®ã¬ãžã¹ã¿ã䜿çšãããããå ±æã¡ã¢ãªå ã®ã¿ã€ã«ããšã«ããå€ãã®ã¢ã¯ãã£ããªTensorCoreããããŸããã€ãŸããåãæ°ã®ã¬ãžã¹ã¿ã§3åã®ãã³ãœã«ã³ã¢ãããŒãã§ããŸãããã ãã垯åå¹ ã¯äŸç¶ãšããŠããã«ããã¯ã§ãããããå®éã®TFLOPSã®å¢å ã¯ãçè«å€ãšæ¯èŒããŠãããããã§ããæ°ãããã³ãœã«ã«ãŒãã«ã«ãããããã©ãŒãã³ã¹ãçŽ1ã3ïŒ åäžããŸããã
å šäœãšããŠãAmpereã¢ãŒããã¯ãã£ã¯ãã°ããŒãã«ã¡ã¢ãªããå ±æã¡ã¢ãªã¿ã€ã«ããã³ãœã«ã³ã¢ã¬ãžã¹ã¿ãŸã§ãæ¹åãããéå±€ãéããŠã¡ã¢ãªåž¯åå¹ ãããå¹ççã«äœ¿çšããããã«æé©åãããŠããããšãããããŸãã
GOã«ãããAmpereã®æå¹æ§ã®è©äŸ¡
æŠèŠïŒ
- Ampere GPUã®ã¡ã¢ãªåž¯åå¹ ãšæ¹åãããã¡ã¢ãªéå±€ã«åºã¥ãçè«äžã®æšå®ã§ã¯ã1.78ã1.87åã®å éãäºæž¬ãããŸãã
- NVIDIAã¯ãTeslaA100ããã³V100GPUã®é床枬å®ã«é¢ããããŒã¿ããªãªãŒã¹ããŸããã圌ãã¯ããããŒã±ãã£ã³ã°çã§ãããåãã®ãªãã¢ãã«ã¯ãããã«åºã¥ããŠæ§ç¯ããããšãã§ããŸãã
- åãã®ãªãã¢ãã«ã¯ãV100ãšæ¯èŒããŠãTesla A100ã¯èªç¶èšèªåŠçã§1.7åãã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã§1.45åé«éã§ããããšã瀺åããŠããŸãã
ãã®ã»ã¯ã·ã§ã³ã¯ãAmpereGPUã®ããã©ãŒãã³ã¹ã¹ã³ã¢ãååŸããæ¹æ³ã®æè¡çãªè©³çŽ°ã詳ãã調ã¹ããæ¹ã察象ãšããŠããŸããèå³ããªãå Žåã¯ãã¹ãããããŠãåé¡ãããŸããã
çè«äžã®é床ã®æšå®
äžèšã®è°è«ãèãããšããã³ãœã«ã³ã¢ãåãã2ã€ã®GPUã¢ãŒããã¯ãã£ã®éãã¯ãäž»ã«ã¡ã¢ãªåž¯åå¹ ã«ããã¯ãã§ããè¿œå ã®å©ç¹ã¯ãå ±æã¡ã¢ãªãšL1ãã£ãã·ã¥ã®å¢å ãããã³ã¬ãžã¹ã¿ã®å¹ççãªäœ¿çšããåŸãããŸãã
Tesla A100 GPUã®åž¯åå¹ ã¯ãTeslaV100ãšæ¯èŒããŠ1555/900 = 1.73åã«å¢å ããŠããŸãããŸããç·ã¡ã¢ãªã倧ããããã«é床ã2ã5ïŒ åäžãããã³ãœã«ã³ã¢ãæ¹åãããããã«1ã3ïŒ åäžãããšäºæ³ããã®ã劥åœã§ããå éã¯1.78ãã1.87åã§ãªããã°ãªããªãããšãããããŸãã
Ampere
AmpereãTuringãVoltaãªã©ã®ã¢ãŒããã¯ãã£ã®GPUã¹ã³ã¢ã1ã€ãããšããŸãããããã®çµæãåãã¢ãŒããã¯ãã£ãŸãã¯ã·ãªãŒãºã®ä»ã®GPUã«å€æ¿ããã®ã¯ç°¡åã§ãã幞ããNVIDIAã¯ãã³ã³ãã¥ãŒã¿ãŒã®ããžã§ã³ãšèªç¶ãªèšèªã®ç解ã«é¢é£ããããŸããŸãªã¿ã¹ã¯ã«ã€ããŠãA100ãšV100ãæ¯èŒãããã³ãããŒã¯ããã§ã«å®æœããŠããŸããæ®å¿µãªãããNVIDIAã¯ããããã®æ°å€ãçŽæ¥æ¯èŒã§ããªãããã«å¯èœãªéãã®ããšãè¡ã£ãŠããŸãããã¹ãã§ã¯ãç°ãªãããŒã¿ãã±ãããµã€ãºãšç°ãªãæ°ã®GPUã䜿çšãããããA100ã¯åãŠãŸããã§ããããããã£ãŠãããæå³ã§ãåŸãããããã©ãŒãã³ã¹ææšã¯ãäžéšã¯æ£çŽã§ãäžéšã¯å®£äŒã§ããäžè¬ã«ãA100ã®ã¡ã¢ãªãå€ããããããŒã¿ãã±ãããµã€ãºã®å¢å ã¯æ£åœåããããšäž»åŒµã§ããŸãããGPUã¢ãŒããã¯ãã£ãæ¯èŒããã«ã¯ãåãããŒã¿ãã±ãããµã€ãºã®ã¿ã¹ã¯ã§åãã®ãªãããã©ãŒãã³ã¹ããŒã¿ãæ¯èŒããå¿ èŠããããŸãã
åãã®ãªãèŠç©ãããååŸããã«ã¯ã2ã€ã®æ¹æ³ã§V100ãšA100ã®æž¬å®å€ãã¹ã±ãŒãªã³ã°ã§ããŸããããŒã¿ãã±ãããµã€ãºã®éããèæ ®ããããGPUã®æ°ã®éãïŒ1ãš8ïŒãèæ ®ããŸãã幞éãªããšã«ãNVIDIAãæäŸããããŒã¿ã§ãäž¡æ¹ã®ã±ãŒã¹ã§åæ§ã®èŠç©ãããèŠã€ããããšãã§ããŸãã
ãã±ãããµã€ãºã2åã«ãããšãã¹ã«ãŒãããã1ç§ãããã®ç»åæ°ã§13.6ïŒ å¢å ããŸãïŒç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãCNNã®å ŽåïŒã RTX Titanã®Transformerã¢ãŒããã¯ãã£ã䜿çšããŠåãã¿ã¹ã¯ã®é床ã枬å®ãããšãããé©ãã¹ãããšã«ãåãçµæïŒ13.5ïŒ ïŒãåŸãããŸãããããã¯ä¿¡é Œã§ããèŠç©ããã®ââããã§ãã
ãããã¯ãŒã¯ã®äžŠååãå¢ãããGPUã®æ°ãå¢ãããšããããã¯ãŒã¯ã«é¢é£ãããªãŒããŒãããã®ããã«ããã©ãŒãã³ã¹ãäœäžããŸãããã ããA100 8x GPUã¯ãV100 8x GPUïŒNVLink 2.0ïŒãšæ¯èŒããŠãããã¯ãŒãã³ã°ïŒNVLink 3.0ïŒã§ã®ããã©ãŒãã³ã¹ãåªããŠããŸãããããæ··ä¹±ãæãèŠå ã§ãã NVIDIAããã®ããŒã¿ãèŠããšãSNSãåŠçããããã«ã8çªç®ã®A100ãåããã·ã¹ãã ã®ãªãŒããŒãããã8çªç®ã®V10000ãåããã·ã¹ãã ããã5ïŒ å°ãªãããšãããããŸããã€ãŸãã1çªç®ã®A10000ãã8çªç®ã®A10000ãžã®é·ç§»ã§7.0åã®å éãåŸãããå Žåã1çªç®ã®V10000ãã8çªç®ã®V10000ãžã®é·ç§»ã§ã¯6.67åã®å éããåŸãããŸãããå€å§åšã®å Žåããã®æ°å€ã¯7ïŒ ã§ãã
ãã®æ å ±ã䜿çšããŠãNVIDIAããæäŸãããããŒã¿ãããç¹å®ã®GOã¢ãŒããã¯ãã£ã®å éãçŽæ¥èŠç©ããããšãã§ããŸããTesla A100ã«ã¯ãTeslaV100ã«æ¯ã¹ãŠæ¬¡ã®é床äžã®å©ç¹ããããŸãã
- SE-ResNeXt101ïŒ1.43åã
- Masked-R-CNNïŒ1.47åã
- ãã©ã³ã¹ãã©ãŒããŒïŒ12å±€ãæ©æ¢°å€æãWMT14 en-deïŒïŒ1.70åã
ãããã£ãŠãã³ã³ãã¥ãŒã¿ããžã§ã³ã®å Žåãæ°å€ã¯çè«äžã®æšå®å€ãäžåã£ãŠååŸãããŸããããã¯ããã³ãœã«æž¬å®å€ãå°ããããšãimg2colãFFTãªã©ã®è¡åä¹ç®ãæºåããããã«å¿ èŠãªæäœã®ãªãŒããŒãããããŸãã¯GPUã飜åãããããšãã§ããªãæäœïŒçµæã®ã¬ã€ã€ãŒãæ¯èŒçå°ããããšãå€ãïŒãåå ã§ããå¯èœæ§ããããŸãããŸããç¹å®ã®ã¢ãŒããã¯ãã£ïŒã°ã«ãŒãåãããç³ã¿èŸŒã¿ïŒã®ã¢ãŒãã£ãã¡ã¯ãã§ããå¯èœæ§ããããŸãã
å€å§åšã®é床ã®å®éçãªè©äŸ¡ã¯ãçè«çãªè©äŸ¡ã«éåžžã«è¿ããã®ã§ããããããã倧ããªè¡åãæäœããããã®ã¢ã«ãŽãªãºã ãéåžžã«åçŽã ããã§ããGPUã®è²»çšå¯Ÿå¹æãèšç®ããããã«ãå®éã®èŠç©ããã䜿çšããŸãã
èŠç©ããã®ââäžæ£ç¢ºãã®å¯èœæ§
äžèšã¯A100ãšV100ã®æ¯èŒè©äŸ¡ã§ãããããŸã§ãNVIDIAã¯ãã²ãŒã ãRTX GPUã®ããã©ãŒãã³ã¹ãå¯ãã«äœäžãããŠããŸããããã³ãœã«ã³ã¢ã®äœ¿çšçãäœäžããå·åŽçšã®ã²ãŒã ãã¡ã³ãè¿œå ãããGPUéã®ããŒã¿è»¢éãçŠæ¢ãããŠããŸãããRT30ã·ãªãŒãºãAmpereA100ã«å¯ŸããŠæªç¥ã®é害ãåŒãèµ·ãããå¯èœæ§ããããŸãã
Ampere / RTX30ã®å Žåã«ä»ã«èæ ®ãã¹ãããš
æŠèŠïŒ
- Ampereã䜿çšãããšãã¹ããŒã¹ãããªãã¯ã¹ã«åºã¥ããŠãããã¯ãŒã¯ããã¬ãŒãã³ã°ã§ããŸããããã«ããããã¬ãŒãã³ã°ããã»ã¹ãæ倧2åé«éåãããŸãã
- ã¹ããŒã¹ãããã¯ãŒã¯ãã¬ãŒãã³ã°ã¯ãŸã ãã£ãã«äœ¿çšãããŸãããããã®ãããã§ãAmpereã¯ããã«æ代é ãã«ãªãããšã¯ãããŸããã
- Ampereã«ã¯ãäœç²ŸåºŠã®äœ¿çšãã¯ããã«å®¹æã«ããæ°ããäœç²ŸåºŠããŒã¿ã¿ã€ãããããŸãããå¿ ããã以åã®GPUãããé床ãåäžãããšã¯éããŸããã
- æ°ãããã¡ã³ã®èšèšã¯ãGPUéã«ç©ºãã¹ããŒã¹ãããå Žåã«é©ããŠããŸãããäºãã«è¿ãã«ç«ã£ãŠããGPUãå¹æçã«å·åŽããããã©ããã¯æããã§ã¯ãããŸããã
- RTX 3090ã®3ã¹ãããèšèšã¯ã4ã€ã®GPUãã«ãã«ãšã£ãŠèª²é¡ãšãªããŸããèãããã解決çã¯ã2ã¹ããããªãã·ã§ã³ãŸãã¯PCIeãšã¯ã¹ãã³ãã䜿çšããããšã§ãã
- 4ã€ã®RTX3090ã¯ãåžå Žã«åºåã£ãŠããæšæºã®PSUãæäŸã§ãããããå€ãã®é»åãå¿ èŠãšããŸãã
æ°ããNVIDIAAmpere RTX 30ã«ã¯ãNVIDIA Turing RTX 20ã«æ¯ã¹ãŠè¿œå ã®å©ç¹ããããŸããã€ãŸããã¹ããŒã¹ãã¬ãŒãã³ã°ãšãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ããããŒã¿åŠçã®æ¹åã§ããæ°ããããŒã¿ã¿ã€ããªã©ã®æ®ãã®ããããã£ã¯ãåçŽãªå©äŸ¿æ§ã®åäžãšèŠãªãããšãã§ããŸããè¿œå ã®ããã°ã©ãã³ã°ãå¿ èŠãšããã«ãTuringã·ãªãŒãºãšåãæ¹æ³ã§åŠçãé«éåããŸãã
ã¹ããŒã¹ã©ãŒãã³ã°
Ampereã䜿çšãããšãã¹ããŒã¹ãããªãã¯ã¹ãé«éãã€èªåçã«ä¹ç®ã§ããŸããããã¯æ¬¡ã®ããã«æ©èœããŸãããããªãã¯ã¹ãååŸããŠ4ã€ã®èŠçŽ ã«åå²ãããšãã¹ããŒã¹ãããªãã¯ã¹ããµããŒããããã³ãœã«ã«ãŒãã«ã«ãããããã4ã€ã®èŠçŽ ã®ãã¡2ã€ããŒãã«ããããšãã§ããŸããããã«ããããããªãã¯ã¹ä¹ç®äžã®åž¯åå¹ èŠä»¶ãååã«ãªãããã2åã®é床åäžãå®çŸããŸãã
ç§ã®ç 究ã§ã¯ããŸã°ããªåŠç¿ãããã¯ãŒã¯ã䜿çšããŠããŸãããç¹ã«ãããããã¯ãŒã¯ã«å¿ èŠãªFLOPSãæžãããŸãããGPUã¯ã¹ããŒã¹è¡åããã°ããä¹ç®ã§ããªããããé床ãäžããªãããšããäºå®ã«ã€ããŠããã®äœæ¥ã¯æ¹å€ãããŸããããŸã-ã¹ããŒã¹ãããªãã¯ã¹ä¹ç®ã®ãµããŒãã¯ããã³ãœã«ã«ãŒãã«ãç§ã®ã¢ã«ãŽãªãºã ããŸãã¯ä»ã®ã¢ã«ãŽãªãºã ïŒãªã³ã¯ãlinkãlinkãlinkïŒã¯ãã¹ããŒã¹ãããªãã¯ã¹ãæäœããããããã¬ãŒãã³ã°äžã«å®éã«2åã®é床ã§åäœã§ããããã«ãªããŸããã
ãã®ããããã£ã¯çŸåšå®éšçãªãã®ãšèŠãªãããŠãããã¹ããŒã¹ãããã¯ãŒã¯ãã¬ãŒãã³ã°ã¯æ®éçã«é©çšãããŠããŸããããGPUããã®ãã¯ãããžãŒããµããŒãããŠããå Žåã¯ãã¹ããŒã¹ãã¬ãŒãã³ã°ã®å°æ¥ã«åããããšãã§ããŸãã
äœç²ŸåºŠã®èšç®
æ°ããããŒã¿ã¿ã€ããç§ã®ä»äºã§å¿ å®åºŠã®äœãéäŒæã®å®å®æ§ãã©ã®ããã«æ¹åã§ããã ããã§ã«ç€ºããŸããããããŸã§ã®ãšããã16ãããã®æµ®åå°æ°ç¹æ°ã䜿çšããå®å®ããéäŒæã®åé¡ã¯ãéåžžã®ããŒã¿ã¿ã€ããã¹ãã³[-65,504ã65,504]ã®ã¿ããµããŒãããããšã§ããåŸé ããã®ã®ã£ãããè¶ ãããšãççºããŠNaNå€ãçæãããŸãããããé²ãããã«ãéåžžãéäŒæããåã«å€ã«å°ããªæ°ãæããŠå€ãã¹ã±ãŒãªã³ã°ããåŸé ã®ççºãåé¿ããŸãã
Brain Float 16ïŒBF16ïŒåœ¢åŒã¯ææ°ã«å€ãã®ãããã䜿çšãããããå¯èœãªå€ã®ç¯å²ã¯FP32ãšåãã§ãïŒ[-3 * 10 ^ 38ã3 * 10 ^ 38]ã BF16ã®ç²ŸåºŠã¯äœããªããŸããéèŠãªæ¡æ°ã¯å°ãªããªããŸããããããã¯ãŒã¯ããã¬ãŒãã³ã°ãããšãã®åŸé ã®ç²ŸåºŠã¯ããã»ã©éèŠã§ã¯ãããŸããããããã£ãŠãBF16ã䜿çšãããšãã¹ã±ãŒãªã³ã°ãè¡ã£ãããåŸé ã®ççºãå¿é ãããããå¿ èŠããªããªããŸãããã®åœ¢åŒã§ã¯ã粟床ããããã«äœäžããŸããããã¬ãŒãã³ã°ã®å®å®æ§ãåäžããã¯ãã§ãã
ãããæå³ããããšïŒBF16ã®ç²ŸåºŠã¯FP16ã®ç²ŸåºŠãããäžè²«ããŠããå¯èœæ§ããããŸãããé床ã¯åãã§ããTF32ã®ç²ŸåºŠã«ãããã»ãŒFP32ã®ãããªå®å®æ§ãšãã»ãŒFP16ã®ãããªå éãåŸãããŸããããã«ããããã®ããŒã¿ã¿ã€ãã䜿çšãããšãã³ãŒããäœãå€æŽããã«ãFP32ãTF32ã«ãFP16ãBF16ã«å€æŽã§ããŸãã
äžè¬ã«ããããã®æ°ããããŒã¿ã¿ã€ãã¯ãå€ãããŒã¿ã¿ã€ããšå°ãã®ããã°ã©ãã³ã°ïŒæ£ããã¹ã±ãŒãªã³ã°ãåæåãæ£èŠåãApexã䜿çšïŒã䜿çšããŠãã¹ãŠã®å©ç¹ãåŸãããšãã§ãããšããæå³ã§ãæ æ°ãšèŠãªãããšãã§ããŸãããããã£ãŠããããã®ããŒã¿ã¿ã€ãã¯å éãæäŸããŸãããããã¬ãŒãã³ã°ã§å¿ å®åºŠã®äœããã®ã䜿çšããããããŸãã
æ°ãããã¡ã³ã®èšèšãšç±æŸæ£ã®åé¡
RTX 30ã·ãªãŒãºã®æ°ãããã¡ã³èšèšã«ã¯ããšã¢ãããŒãã¡ã³ãšãšã¢ãã«ãã¡ã³ããããŸãããã¶ã€ã³èªäœã¯ç¬åµçã§ãGPUéã«ç©ºãã¹ããŒã¹ãããå Žåã¯éåžžã«å¹ççã«æ©èœããŸãããã ããGPUãçžäºã«åŒ·å¶ãããå Žåã«GPUãã©ã®ããã«åäœãããã¯æ確ã§ã¯ãããŸããããããŒãã¡ã³ã¯ä»ã®GPUãã空æ°ãå¹ãé£ã°ãããšãã§ããŸããããã®åœ¢ç¶ã以åãšã¯ç°ãªãããããããã©ã®ããã«æ©èœããããå€æããããšã¯ã§ããŸããã 4ã€ã®ã¹ããããããå Žæã«1ã€ãŸãã¯2ã€ã®GPUãé 眮ããããšãèšç»ããŠããå Žåã¯ãåé¡ã¯ãªãã¯ãã§ãããã ãã3ã4åã®RTX 30 GPUã䞊ã¹ãŠäœ¿çšããå Žåã¯ãæåã«æž©åºŠæ¡ä»¶ã«é¢ããã¬ããŒããåŸ ã£ãŠããããã¡ã³ãPCIeãšãã¹ãã³ããŒããŸãã¯ãã®ä»ã®ãœãªã¥ãŒã·ã§ã³ãããã«å¿ èŠãã©ãããå€æããŸããã
ãããã«ãããæ°Žå·ã¯ããŒãã·ã³ã¯ã®åé¡ã解決ããã®ã«åœ¹ç«ã¡ãŸããå€ãã®ã¡ãŒã«ãŒãRTX3080 / RTX 3090ã«ãŒãã«ãã®ãããªãœãªã¥ãŒã·ã§ã³ãæäŸããŠããããã4æã§ãæãããªããŸããããã ãã4ã€ã®GPUãæèŒããã³ã³ãã¥ãŒã¿ãŒãæ§ç¯ããå Žåã¯ãã»ãšãã©ã®å Žåéåžžã«é£ãããããæ¢è£œã®GPUãœãªã¥ãŒã·ã§ã³ãè³Œå ¥ããªãã§ãã ãããã©ãžãšãŒã¿ãŒãé åžããŸãã
å·åŽã®åé¡ã«å¯Ÿããå¥ã®è§£æ±ºçã¯ãPCIeãšãã¹ãã³ããŒãè³Œå ¥ããã±ãŒã¹å ã«ã«ãŒããé åžããããšã§ããããã¯éåžžã«å¹æçã§ã-ç§ãšããã³ãã³å€§åŠã®ä»ã®å€§åŠé¢çã¯ãã®ãªãã·ã§ã³ã䜿çšããŠå€§æåãåããŠããŸããèŠãç®ã¯ããŸãè¯ããããŸããããGPUã¯ç±ããªããŸããïŒãŸãããã®ãªãã·ã§ã³ã¯ãGPUãå容ããã®ã«ååãªã¹ããŒã¹ããªãå Žåã«åœ¹ç«ã¡ãŸããã±ãŒã¹ã«äœè£ãããå Žåã¯ãããšãã°ã3ã€ã®ã¹ããããåããæšæºã®RTX 3090ãè³Œå ¥ãããšã¯ã¹ãã³ããŒã䜿çšããŠã±ãŒã¹å šäœã«é åžã§ããŸãããããã£ãŠã4ã€ã®RTX3090ã®ã¹ããŒã¹ãšå·åŽã®åé¡ãåæã«è§£æ±ºããããšãå¯èœ
ã§ãã1ïŒPCIeãšãã¹ãã³ããŒãåãã4 GPU
3ã¹ãããã«ãŒããšé»æºã®åé¡
RTX 3090ã¯3ã€ã®ã¹ããããå æãããããNVIDIAã®ããã©ã«ãã®ãã¡ã³ã§ãããã4ã€äœ¿çšããããšã¯ã§ããŸããã 350Wã®TDPãå¿ èŠãªãããããã¯é©ãã¹ãããšã§ã¯ãããŸããã RTX 3080ã¯ãããã«å£ã£ãŠããã320Wã®TDPãå¿ èŠã§ããã4ã€ã®RTX3080ã§ã·ã¹ãã ãå·åŽããããšã¯éåžžã«å°é£ã§ãã
350W = 1400Wã®4æã®ã«ãŒãã§ã·ã¹ãã ã«é»åãäŸçµŠããããšãå°é£ã§ãã 1600 Wã®é»æºïŒPSUïŒããããŸãããããã»ããµãšãã¶ãŒããŒãã«ã¯200Wã§ã¯äžååãªå ŽåããããŸããæ倧é»åæ¶è²»ã¯å šè² è·ã§ã®ã¿çºçããHEã®éãããã»ããµã¯éåžžè»œè² è·ã§ãããããã£ãŠã1600WPSUã¯4ã€ã®RTX3080ã«é©ããŠããå¯èœæ§ããããŸããã4ã€ã®RTX 3090ã«ã¯ã1700W以äžã®PSUãæ¢ãããšããå§ãããŸããçŸåšããã®ãããªPSUã¯åžå Žã«åºåã£ãŠããŸããããµãŒããŒPSUãŸãã¯ã¯ãªãããã€ããŒçšã®ç¹å¥ãªãããã¯ãæ©èœããå ŽåããããŸããããããã«ã¯ç°åžžãªãã©ãŒã ãã¡ã¯ã¿ãŒãããå ŽåããããŸãã
深局åŠç¿ã«ãããGPUå¹ç
次ã®ãã¹ãã«ã¯ãTeslaA100ãšTeslaV100ã®æ¯èŒã ãã§ãªãããã®ããŒã¿ã«é©åããã¢ãã«ãäœæããTitan VãTitan RTXãRTX 2080 TiãRTX 2080ããã¹ããã4ã€ã®ç°ãªããã¹ãïŒãªã³ã¯ããªã³ã¯ããªã³ã¯ããªã³ã¯ïŒãäœæããŸããã
ãŸãããã¹ãããŒã¿ãã€ã³ããè£éããããšã«ãããRTX 2070ãRTX 2060ãQuadroRTXãªã©ã®ãããã¬ã³ãžã«ãŒãã®ãã³ãããŒã¯çµæãã¹ã±ãŒãªã³ã°ããŸãããéåžžãGPUã¢ãŒããã¯ãã£ã§ã¯ããã®ãããªããŒã¿ã¯ããããªãã¯ã¹ã®ä¹ç®ãšã¡ã¢ãªåž¯åå¹ ã«é¢ããŠç·åœ¢ã«ã¹ã±ãŒãªã³ã°ãããŸãã
FP32çªå·ã䜿çšãããã¬ãŒãã³ã°ã䜿çšããçç±ããªããããæ··å粟床ã®FP16ãã¬ãŒãã³ã°ãã¹ãããã®ã¿ããŒã¿ãåéããŸããã
å³ïŒå³2ïŒRTX 2080Tiã«ãã£ãŠæ£èŠåãããããã©ãŒãã³ã¹RTX2080 Ti
ãšæ¯èŒããŠãRTX 3090ã¯ãç³ã¿èŸŒã¿ãããã¯ãŒã¯ã§ã¯1.57åããã©ã³ã¹ã§ã¯1.5åé«éã§åäœããã³ã¹ãã¯15ïŒ é«ããªããŸããAmpere RTX 30ã¯ãTuring RTX20ã·ãªãŒãºä»¥éå€§å¹ ãªæ¹åã瀺ããŠããããšãããããŸããã
ã³ã¹ããããã®GPUãã£ãŒãã©ãŒãã³ã°ã¬ãŒã
ã©ã®GPUããéã«èŠåãæé«ã®äŸ¡å€ã§ããããïŒããã¯ãã¹ãŠãã·ã¹ãã ã®ç·ã³ã¹ãã«äŸåããŸããé«äŸ¡ãªå Žåã¯ãããé«äŸ¡ãªGPUã«æè³ããã®ãçã«ããªã£ãŠããŸãã
以äžã¯ãPCIe 3.0äžã®3ã€ã®ã¢ã»ã³ããªã«é¢ããããŒã¿ã§ããããã¯ã2ã€ãŸãã¯4ã€ã®GPUãåããã·ã¹ãã ã®ã³ã¹ãã®ããŒã¹ã©ã€ã³ãšããŠäœ¿çšããŸãããã®åºæ¬ã³ã¹ããååŸããŠãGPUã³ã¹ããè¿œå ããŸããç§ã¯åŸè ãAmazonãšeBayããã®ãªãã¡ãŒéã®å¹³åäŸ¡æ ŒãšããŠèšç®ããŸããæ°ããAmpereã®å Žåãç§ã¯1ã€ã®äŸ¡æ Œã®ã¿ã䜿çšããŸããäžèšã®ããã©ãŒãã³ã¹ããŒã¿ãšåãããŠãããã¯1ãã«ãããã®ããã©ãŒãã³ã¹å€ã瀺ããŸããGPUã8ã€ããã·ã¹ãã ã®å ŽåãRTXãµãŒããŒã®æ¥çæšæºãšããŠSupermicroãã¢ããŒã³ãæ¡çšããŠããŸãã衚瀺ãããŠããã°ã©ãã«ã¯ãã¡ã¢ãªèŠä»¶ã¯å«ãŸããŠããŸãããæåã«å¿ èŠãªã¡ã¢ãªã«ã€ããŠèãã次ã«ã°ã©ãã§æé©ãªãªãã·ã§ã³ãæ¢ãå¿ èŠããããŸããã¡ã¢ãªã®ãã³ãã®äŸïŒ
- äºåã«ãã¬ãŒãã³ã°ããããã©ã³ã¹ãã©ãŒããŒã䜿çšããããå°ããªãã©ã³ã¹ãã©ãŒããŒãæåãããã¬ãŒãã³ã°ãã> = 11GBã
- ç 究ãŸãã¯çç£ã«ããã倧èŠæš¡ãªå€å§åšãŸãã¯ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ïŒ> = 24GBã
- ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ãããã¿ã€ãã³ã°ïŒãã©ã³ã¹ãã©ãŒããŒãŸãã¯ã³ã³ããªã¥ãŒã·ã§ã³ãããã¯ãŒã¯ïŒ> = 10GBã
- Kaggleã³ã³ãã¹ããžã®åå > = 8GBã
- ã³ã³ãã¥ãŒã¿ãŒããžã§ã³> = 10GBã
å³ïŒ
å³3ïŒRTX3080ã«å¯Ÿããæ£èŠåããããã«ã®ããã©ãŒãã³ã¹ãå³4ïŒRTX3080ã«å¯Ÿããæ£èŠåããããã«ã®ããã©ãŒãã³ã¹
ã5ïŒRTX3080ã«å¯Ÿããæ£èŠåããããã«ã®ããã©ãŒãã³ã¹ã
GPUã®æšå¥šäºé
ç¹°ãè¿ãã«ãªããŸãããGPUãéžæãããšãã¯ããŸããã¿ã¹ã¯ã«ååãªã¡ã¢ãªãããããšã確èªããŠãã ãããGPUãéžæããæé ã¯æ¬¡ã®ãšããã§ãã
- , GPU: Kaggle, , , , - .
- , .
- GPU, .
- GPU - ? , RTX 3090, ? GPU? , GPU?
äžéšã®æé ã§ã¯ãå¿ èŠãªãã®ã«ã€ããŠèããåãããšãããŠããä»ã®äººã䜿çšããŠããã¡ã¢ãªã®éã«ã€ããŠå°ã調æ»ããå¿ èŠããããŸããã¢ããã€ã¹ã¯ã§ããŸããããã®åéã®ãã¹ãŠã®è³ªåã«å®å šã«çããããšã¯ã§ããŸããã
11GBãè¶ ããã¹ãã¬ãŒãžãå¿ èŠã«ãªãã®ã¯ãã€ã§ããïŒ
å€å§åšã䜿çšããå Žåã¯å°ãªããšã11GBãå¿ èŠã§ããããã®åéã§ç 究ãè¡ãå Žåã¯å°ãªããšã24GBãå¿ èŠã§ããããšã¯ãã§ã«è¿°ã¹ãŸããã以åã®äºåãã¬ãŒãã³ã°æžã¿ã¢ãã«ã®ã»ãšãã©ã¯ãéåžžã«é«ãã¡ã¢ãªèŠä»¶ããããå°ãªããšã11GBã®ã¡ã¢ãªãåããRTX2080Ti以äžã®GPUã§ãã¬ãŒãã³ã°ãããŠããŸãããããã£ãŠãã¡ã¢ãªã11 GBæªæºã®å Žåãäžéšã®ã¢ãã«ã®èµ·åãå°é£ã«ãªãããäžå¯èœã«ãªãå¯èœæ§ããããŸãã
倧éã®ã¡ã¢ãªãå¿ èŠãšããä»ã®é åã¯ãå»çç»åãé«åºŠãªã³ã³ãã¥ãŒã¿ããžã§ã³ã¢ãã«ãããã³ãã¹ãŠå€§ããªç»åã§ãã
å šäœãšããŠãç 究ãç£æ¥çšã¢ããªã±ãŒã·ã§ã³ãKaggleã®ç«¶äºãªã©ã競äºããã®ãããšãã§ããã¢ãã«ã®éçºãæ€èšããŠããå Žåã¯ãã¡ã¢ãªãè¿œå ããããšã§ç«¶äºåãé«ããããšãã§ããŸãã
11 GBæªæºã®ã¡ã¢ãªã§ãã€åé¡ã解決ã§ããŸããïŒ
RTX3070ããã³RTX3080ã«ãŒãã¯åŒ·åã§ãããã¡ã¢ãªãäžè¶³ããŠããŸãããã ããå€ãã®ã¿ã¹ã¯ã§ã¯ããã®éã®ã¡ã¢ãªã¯å¿ èŠãªãå ŽåããããŸãã
RTX 3070ã¯ãGOãã¬ãŒãã³ã°ã«æé©ã§ããã»ãšãã©ã®ã¢ãŒããã¯ãã£ã®åºæ¬çãªãããã¯ãŒãã³ã°ã¹ãã«ã¯ããããã¯ãŒã¯ãçž®å°ããããããå°ããªã€ã¡ãŒãžã䜿çšããããšã§ç¿åŸã§ããŸããGOãåŠã¶å¿ èŠãããå Žåã¯ãRTX 3070ãéžæããŸããäœè£ãããã°ãRTX3070ãéžæããŸãã
RTX 3080ã¯ãä»æ¥æãè²»çšå¹æã®é«ãã«ãŒãã§ããããããããã¿ã€ãã³ã°ã«æé©ã§ãããããã¿ã€ãã³ã°ã«ã¯å€§éã®ã¡ã¢ãªãå¿ èŠã§ãããã¡ã¢ãªã¯å®äŸ¡ã§ãããããã¿ã€ãã³ã°ãšã¯ãç 究ãKaggleã³ã³ãã¹ããã¹ã¿ãŒãã¢ããã®ã¢ã€ãã¢ã®è©Šè¡ãç 究ã³ãŒãã®å®éšãªã©ãããããåéã§ã®ãããã¿ã€ãã³ã°ãæå³ããŸããããããã¹ãŠã®ã¢ããªã±ãŒã·ã§ã³ã«ã¯ãRTX3080ãæé©ã§ãã
ããšãã°ãç 究æãã¹ã¿ãŒãã¢ãããçµå¶ããŠããå Žåãç·äºç®ã®66ã80ïŒ ãRTX 3080ãã·ã³ã«è²»ãããä¿¡é Œæ§ã®é«ãæ°Žå·ãåããRTX 3090ãã·ã³ã«20ã33ïŒ ãè²»ãããŸãã RTX 3080ã¯ããè²»çšå¹æãé«ããSlurmããã¢ã¯ã»ã¹ã§ããŸãã..ããããã¿ã€ãã³ã°ã¯ã¢ãžã£ã€ã«ã¢ãŒãã§å®è¡ããå¿ èŠããããããããå°ããªã¢ãã«ãšããŒã¿ã»ããã§å®è¡ããå¿ èŠããããŸãããããŠãRTX3080ã¯ãã®ããã«æé©ã§ããåŠç/ååãåªãããããã¿ã€ãã¢ãã«ãäœæãããããããRTX 3090ã«å±éããŠããã倧ããªã¢ãã«ã«ã¹ã±ãŒã«ã¢ããã§ããŸãã
äžè¬çãªæšå¥šäºé
å šäœãšããŠãRTX 30ã·ãªãŒãºã¢ãã«ã¯éåžžã«åŒ·åã§ããã絶察ã«ãå§ãããŸããåè¿°ã®ã¡ã¢ãªèŠä»¶ãããã³é»åãšå·åŽã®èŠä»¶ãèæ ®ããŠãã ããã GPUéã«ç©ºãã¹ããããããå Žåã¯ãå·åŽã«åé¡ã¯ãããŸããããã以å€ã®å Žåã¯ãRTX 30ã«ãŒãã«æ°Žå·ãPCIeãšãã¹ãã³ããŒããŸãã¯å¹ççãªãã¡ã³ä»ãã«ãŒããæäŸããŸãã
å šäœãšããŠãRTX3090ãè³Œå ¥ã§ãã人ã«ã¯ãå§ãããŸããããã¯ä»ããªãã«åãã ãã§ãªãã次ã®3-7幎ééåžžã«å¹æçã§ããç¶ããã§ããããä»åŸ3幎éã§HBMã¡ã¢ãªãå€§å¹ ã«å®ããªãå¯èœæ§ã¯äœãããã次ã®GPUã¯RTX 3090ããã25ïŒ ã ãåªããŠããŸãã5ã7幎åŸã«ã¯ãããããå®äŸ¡ãªHBMã¡ã¢ãªã衚瀺ãããŸãããã®åŸãå¿ ãããªãŒããæŽæ°ããå¿ èŠããããŸãã ..ã
è€æ°ã®RTX3090ããã·ã¹ãã ãæ§ç¯ããŠããå Žåã¯ãååãªå·åŽãšé»åãäŸçµŠããŠãã ããã
競äºäžã®åªäœæ§ã«å¯Ÿããå³ããèŠä»¶ããªãéããRTX 3080ããå§ãããŸããããã¯ããè²»çšå¹æã®é«ããœãªã¥ãŒã·ã§ã³ã§ãããã»ãšãã©ã®ãããã¯ãŒã¯ã«è¿ éãªãã¬ãŒãã³ã°ãæäŸããŸããå¿ èŠãªã¡ã¢ãªããªãã¯ãå®è¡ããäœåãªã³ãŒããèšè¿°ããŠãããŸããªãå Žåã¯ã24GBãããã¯ãŒã¯ã10GBGPUã«è©°ã蟌ãããã®ããªãã¯ããããããããŸãã
RTX 3070ã¯ãGOãã¬ãŒãã³ã°ããããã¿ã€ãã³ã°ã«ãæé©ãªã«ãŒãã§ãããRTX 3080ããã200ãã«å®ãã§ããRTX3080ãè³Œå ¥ã§ããªãå Žåã¯ãRTX3070ãéžæããŠãã ããã
äºç®ãå³ãããRTX 3070ãé«ãããå Žåã¯ãeBayã§äžå€ã®RTX2070ãçŽ260ãã«ã§èŠã€ããããšãã§ããŸããRTX 3060ãçºå£²ããããã©ããã¯ãŸã æ確ã§ã¯ãããŸããããäºç®ãå³ããå Žåã¯åŸ ã€äŸ¡å€ããããããããŸãããRTX2060ããã³GTX1060ãšäžèŽããäŸ¡æ Œã®å ŽåãçŽ250ãã«ãã300ãã«ã§ãããè¯å¥œã«æ©èœããã¯ãã§ãã
GPUã¯ã©ã¹ã¿ãŒã«é¢ããæšå¥šäºé
GPUã¯ã©ã¹ã¿ãŒã®ã¬ã€ã¢ãŠãã¯ããã®çšéã«å€§ããäŸåããŸãã1024 GPU以äžã®ã·ã¹ãã ã®å Žåãäž»ãªãã®ã¯ãããã¯ãŒã¯ã®ååšã§ãããäžåºŠã«32å以äžã®GPUã䜿çšããå Žåã匷åãªãããã¯ãŒã¯ã®æ§ç¯ã«æè³ããæå³ã¯ãããŸããã
äžè¬ã«ãRTXã«ãŒãã¯ãCUDAå¥çŽã«åºã¥ãããŒã¿ã»ã³ã¿ãŒã§ã¯äœ¿çšã§ããŸããããã ãã倧åŠã¯ãã®èŠåã®äŸå€ãšãªãããšããããããŸãããã®ãããªèš±å¯ãååŸãããå Žåã¯ãNVIDIAã®æ åœè ã«é£çµ¡ããããšããå§ãããŸãã RTXã«ãŒãã䜿çšã§ããå Žåã¯ãSupermicroã®æšæºã®8 GPU RTX3080ãŸãã¯RTX3090ã·ã¹ãã ããå§ãããŸãïŒå·åŽç¶æ ãç¶æã§ããå ŽåïŒã 8ã€ã®A10000ããŒãã®å°ããªã»ããã«ãããç¹ã«8ã€ã®RTX 3090ãåããå·åŽãµãŒããŒãäžå¯èœãªå Žåã«ããããã¿ã€ãã³ã°åŸã«ã¢ãã«ãå¹ççã«äœ¿çšã§ããŸãããã®å ŽåãA10000ã¯éåžžã«è²»çšå¯Ÿå¹æãé«ããããã«å€ããªãããšã¯ãªããããRTX 6000 / RTX8000ãããA10000ããå§ãããŸãã
GPUã¯ã©ã¹ã¿ãŒïŒ256 GPU以äžïŒã§éåžžã«å€§èŠæš¡ãªãããã¯ãŒã¯ããã¬ãŒãã³ã°ããå¿ èŠãããå Žåã¯ãA10000ãåããNVIDIA DGXSuperPODã·ã¹ãã ããå§ãããŸãã 256ã®GPUããããããã¯ãŒãã³ã°ãäžå¯æ¬ ã«ãªããŸãã 256 GPUãè¶ ããŠæ¡åŒµããå Žåã¯ãæšæºãœãªã¥ãŒã·ã§ã³ãæ©èœããªããªãé«åºŠã«æé©åãããã·ã¹ãã ãå¿ èŠã«ãªããŸãã
ç¹ã«1,024GPUã¹ã±ãŒã«ä»¥äžã§ã¯ãåžå Žã§ç«¶äºåã®ãããœãªã¥ãŒã·ã§ã³ã¯Google TPUPodãšNVIDIADGXSuperPodã ãã§ãããã®èŠæš¡ã§ã¯ãå°çšã®ãããã¯ãŒã¯ã€ã³ãã©ã¹ãã©ã¯ãã£ãNVIDIA DGX SuperPodãããèŠæ ãããããããGoogle TPU Podã奜ã¿ãŸãããã ããååãšããŠã2ã€ã®ã·ã¹ãã ã¯ããªãè¿ãã§ããã¢ããªã±ãŒã·ã§ã³ãšããŒããŠã§ã¢ã§ã¯ãGPUã·ã¹ãã ã¯TPUãããæè»æ§ããããŸãããTPUã·ã¹ãã ã¯ãã倧ããªã¢ãã«ããµããŒãããããé©åã«æ¡åŒµã§ããŸãããããã£ãŠãã©ã¡ãã®ã·ã¹ãã ã«ãé·æãšçæããããŸãã
ã©ã®GPUãè³Œå ¥ããªãã»ããããã
å·åŽã®åé¡ã«å¯ŸåŠããPCIeãšã¯ã¹ãã³ããŒããªãéããäžåºŠã«è€æ°ã®RTX FoundersEditionãŸãã¯RTXTitansãè³Œå ¥ããããšã¯ãå§ãããŸããããããã¯ãã ãŠã©ãŒã ã¢ããããã°ã©ãã«ç€ºãããŠãããã®ãšæ¯èŒããŠé床ãåçã«äœäžããŸãã 4ã€ã®RTX2080 Ti Founders Editionã¯ã90°CãŸã§æ¥éã«å ç±ãããã¯ããã¯é床ãäœäžããéåžžå·åŽãããŠããRTX2070ãããäœéã§åäœããŸãã
Tesla V100ãŸãã¯A100ã¯ãäŒæ¥ã®ããŒã¿ã»ã³ã¿ãŒã§ã®äœ¿çšãçŠæ¢ãããŠããããã極端ãªå Žåã«ã®ã¿è³Œå ¥ããããšããå§ãããŸãããŸãã¯ã巚倧ãªGPUã¯ã©ã¹ã¿ãŒã§éåžžã«å€§èŠæš¡ãªãããã¯ãŒã¯ããã¬ãŒãã³ã°ããå¿ èŠãããå Žåã¯ããããè³Œå ¥ããŠãã ãã-ãããã®äŸ¡æ Œ/ããã©ãŒãã³ã¹æ¯ã¯çæ³çã§ã¯ãããŸããã
ãã£ãšè¯ããã®ãè²·ãäœè£ããããªããGTX16ã·ãªãŒãºã«ãŒããè²·ããªãã§ãã ããããã³ãœã«ã³ã¢ããªããããGOã§ã®ããã©ãŒãã³ã¹ãäœäžããŸãã代ããã«ãäžå€ã®RTX 2070 / RTX 2060 / RTX 2060Superã䜿çšããŸããäºç®ãéåžžã«éãããŠããå Žåã¯åããããšãã§ããŸãã
æ°ããGPUãè³Œå ¥ããªãã»ããããã®ã¯ãã€ã§ããã
ãã§ã«RTX2080 Ti以äžãææããŠããå ŽåãRTX3090ã«ã¢ããã°ã¬ãŒãããŠãã»ãšãã©æå³ããããŸããã GPUã¯ãã§ã«åªããŠãããååŸããé»åãšå·åŽã®åé¡ãšæ¯èŒããŠãé床ã®ã¡ãªããã¯ãããããã§ããããã ãã®äŸ¡å€ã¯ãããŸããã
4ã€ã®RTX2080Tiãã4ã€ã®RTX3090ã«ã¢ããã°ã¬ãŒããããå¯äžã®çç±ã¯ãèšç®èœåã«å€§ããäŸåããéåžžã«å€§ããªå€å§åšããã®ä»ã®ãããã¯ãŒã¯ãç 究ããŠããå Žåã§ãããã ããã¡ã¢ãªã«åé¡ãããå Žåã¯ãæåã«ããŸããŸãªããªãã¯ãæ€èšããŠã倧ããªã¢ãã«ãæ¢åã®ã¡ã¢ãªã«è©°ã蟌ãå¿ èŠããããŸãã
1ã€ä»¥äžã®RTX2070ãææããŠããå Žåãã¢ããã°ã¬ãŒãããåã«ç§ãããªãã§ãããã©ãããããèããŸãããããã¯ããªãè¯ãGPUã§ããä»ã®å€ãã®GPUã®å Žåãšåæ§ã«ã8GBã§ã¯äžååãªå Žåã¯ãeBayã§è²©å£²ããŠRTX3090ãè³Œå ¥ããã®ãçã«ããªã£ãŠãããããããŸãããååãªã¡ã¢ãªããªãå Žåã¯ãæŽæ°ãè¡ãããŠããŸãã
質åã誀解ãžã®åç
æŠèŠïŒ
- PCIeã¬ãŒã³ãšPCIe4.0ã¯ããã¥ã¢ã«GPUã·ã¹ãã ã«ã¯é¢ä¿ãããŸãããGPUã4ã€ããã·ã¹ãã ã®å Žåãå®éã«ã¯ããã§ã¯ãããŸããã
- RTX3090ããã³RTX3080ã®å·åŽã¯å°é£ã§ãããŠã©ãŒã¿ãŒã¯ãŒã©ãŒãŸãã¯PCIeãšãã¹ãã³ããŒã䜿çšããŠãã ããã
- NVLinkã¯ãGPUã¯ã©ã¹ã¿ãŒã«ã®ã¿å¿ èŠã§ãã
- åãã³ã³ãã¥ãŒã¿ãŒã§ç°ãªãGPUã䜿çšã§ããŸããïŒããšãã°ãGTX 1080 + RTX 2080 + RTX 3090ïŒãå¹ççãªäžŠååã¯æ©èœããŸããã
- 3å°ä»¥äžã®ãã·ã³ã䞊è¡ããŠå®è¡ããã«ã¯ãInfinibandãš50Gbpsãããã¯ãŒã¯ãå¿ èŠã§ãã
- AMDããã»ããµã¯Intelããã»ããµãããå®äŸ¡ã§ãããIntelããã»ããµã«ã¯ã»ãšãã©å©ç¹ããããŸããã
- ãšã³ãžãã¢ã®è±éçãªåªåã«ãããããããAMD GPU + ROCmã¯ãä»åŸ1ã2幎ã§ã³ãã¥ããã£ãšåçã®ãã³ãœã«ã³ã¢ãäžè¶³ãããããNVIDIAãšç«¶åããããšã¯ã»ãšãã©ã§ããŸããã
- ã¯ã©ãŠãGPUã¯ã䜿çšæéã1幎æªæºã®å Žåã«åœ¹ç«ã¡ãŸãããã®åŸããã¹ã¯ãããçãå®ããªããŸãã
PCIe 4.0ãå¿ èŠã§ããïŒ
éåžžã¯ããã§ã¯ãããŸãããPCIe4.0ã¯GPUã¯ã©ã¹ã¿ãŒã«æé©ã§ãã8GPUãã·ã³ã䜿çšããŠããå Žåã«äŸ¿å©ã§ãããã以å€ã®å Žåãã»ãšãã©å©ç¹ããããŸããã䞊ååãæ¹åãããããŒã¿ãå°ãéã転éãããŸããããããããŒã¿è»¢éã¯ããã«ããã¯ã§ã¯ãããŸãããã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã§ã¯ãããã«ããã¯ã¯ããŒã¿ã¹ãã¬ãŒãžã§ããå¯èœæ§ããããŸãããGPUããGPUãžã®PCIeããŒã¿è»¢éã§ã¯ãããŸããããããã£ãŠãã»ãšãã©ã®äººãPCIe4.0ã䜿çšããçç±ã¯ãããŸãããããã«ããã4ã€ã®GPUã®äžŠååã1ã7ïŒ åäžããå¯èœæ§ããããŸãã
PCIe 8x / 16xã¬ãŒã³ãå¿ èŠã§ããïŒ
PCIe 4.0ãšåæ§ã«ãéåžžã¯ããã§ã¯ãããŸãããPCIeã¬ãŒã³ã¯ã䞊ååãšé«éããŒã¿è»¢éã«å¿ èŠã§ããããããããã«ããã¯ã«ãªãããšã¯ã»ãšãã©ãããŸãããGPUã2ã€ããå Žåã¯ã4è¡ã§ååã§ãã4ã€ã®GPUã®å ŽåãGPUããšã«8ã€ã®ã©ã€ã³ã䜿çšããããšããå§ãããŸããã4ã€ã®ã©ã€ã³ãããå Žåãããã©ãŒãã³ã¹ã¯5ã10ïŒ ããäœäžããŸããã
ããããã3ã€ã®PCIeã¹ãããã䜿çšããå Žåã4ã€ã®RTX 3090ãã©ã®ããã«é©åãããŸããïŒ
1ã€ã®ã¹ãããã«å¯ŸããŠ2ã€ã®ãªãã·ã§ã³ã®ãããããè³Œå ¥ããããPCIeãšã¯ã¹ãã³ãã䜿çšããŠããããé åžã§ããŸããã¹ããŒã¹ã«å ããŠãããã«å·åŽãšé©åãªé»æºã«ã€ããŠèããå¿ èŠããããŸããã©ããããæãç°¡åãªè§£æ±ºçã¯ãå°çšã®æ°Žå·åŽã«ãŒããåãã4 x RTX 3090EVGAãã€ããã«ãããŒãè³Œå ¥ããããšã§ããEVGAã¯é·å¹Žã«ããã£ãŠé 補ã®æ°Žå·ããŒãžã§ã³ã®ã«ãŒãã補é ããŠãããGPUã®å質ãä¿¡é Œã§ããŸãããããããã£ãšå®ããªãã·ã§ã³ããããŸãã
PCIeãšãã¹ãã³ããŒã¯ã¹ããŒã¹ãšå·åŽã®åé¡ã解決ã§ããŸãããã±ãŒã¹ã«ã¯ãã¹ãŠã®ã«ãŒããåçŽã§ããååãªã¹ããŒã¹ãå¿ èŠã§ãããããŠããšã¯ã¹ãã³ããŒãååã«é·ãããšã確èªããŠãã ããïŒ
4 RTX3090ãŸãã¯4RTX 3080ãå·åŽããæ¹æ³ã¯ïŒ
åã®ã»ã¯ã·ã§ã³ãåç §ããŠãã ããã
è€æ°ã®ç°ãªãGPUã¿ã€ãã䜿çšã§ããŸããïŒ
ã¯ãããã ããäœæ¥ãå¹æçã«äžŠååããããšã¯ã§ããŸããã3 RTX 3070 + 1 RTX 3090ãå®è¡ããŠããã·ã¹ãã ãæ³åã§ããŸããäžæ¹ãã¢ãã«ããããã«è©°ã蟌ããšã4ã€ã®RTX3070éã®äžŠååãéåžžã«è¿ éã«æ©èœããŸãããããŠããããå¿ èŠã«ãªããã1ã€ã®çç±ã¯ãå€ãGPUã䜿çšããŠããããšã§ããåäœããŸãããæéã®GPUãåæãã€ã³ãã§æãé ãGPUãåŸ æ©ããããïŒéåžžã¯åŸé æŽæ°æïŒã䞊ååã¯å¹æçã§ã¯ãããŸããã
NVLinkãšã¯äœã§ããïŒå¿ èŠã§ããïŒ
éåžžãNVLinkã¯å¿ èŠãããŸãããè€æ°ã®GPUéã®é«ééä¿¡ã§ãã128以äžã®GPUã®ã¯ã©ã¹ã¿ãŒãããå Žåã«å¿ èŠã§ãããã®ä»ã®å Žåãæšæºã®PCIeããŒã¿è»¢éã«åãå©ç¹ã¯ã»ãšãã©ãããŸããã
ç§ã¯ããªãã®æãå®ãæšèŠã®ããã«ãããéãæã£ãŠããŸãããäœããã¹ããïŒ
ééããªãäžå€GPUãè³Œå ¥ããã䜿çšæžã¿ã®RTX2070ïŒ$ 400ïŒããã³RTX 2060ïŒ$ 300ïŒã§åé¡ãããŸãããããããè²·ãäœè£ããªãå Žåã次åã®éžæè¢ã¯äžå€ã®GTX 1070ïŒ$ 220ïŒãŸãã¯GTX 1070 TiïŒ$ 230ïŒã§ãããããããé«ãããå Žåã¯ãäžå€ã®GTX 980 TiïŒ6GB $ 150ïŒãŸãã¯GTX 1650 SuperïŒ$ 190ïŒãèŠã€ããŠãã ããããããé«é¡ãªå Žåã¯ãã¯ã©ãŠããµãŒãã¹ã䜿çšããããšããå§ãããŸãããããã¯éåžžãGPUã«æéãŸãã¯é»åå¶éãæäŸãããã®åŸã¯æéãæ¯æãå¿ èŠããããŸããç¬èªã®GPUãè³Œå ¥ã§ããããã«ãªããŸã§ããµãŒãã¹ã亀æããŸãã
2å°ã®ãã·ã³éã§ãããžã§ã¯ãã䞊ååããã«ã¯äœãå¿ èŠã§ããïŒ
2å°ã®ãã·ã³éã§äžŠååããŠäœæ¥ãé«éåããã«ã¯ã50Gbps以äžã®ãããã¯ãŒã¯ã«ãŒããå¿ èŠã§ããå°ãªããšãEDRInfinibandãã€ãŸãå°ãªããšã50Gbpsã®é床ã®ãããã¯ãŒã¯ã«ãŒããã€ã³ã¹ããŒã«ããããšããå§ãããŸããeBayã®ã±ãŒãã«ä»ãã®2æã®EDRã«ãŒãã¯ã500ãã«ãæ»ããŸãã
å Žåã«ãã£ãŠã¯ã10 Gbpsã€ãŒãµãããã§ããŸãããããšããããŸãããããã¯éåžžãç¹å®ã®ã¿ã€ãã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ïŒç¹å®ã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ïŒãŸãã¯ç¹å®ã®ã¢ã«ãŽãªãºã ïŒMicrosoft DeepSpeedïŒã§ã®ã¿æ©èœããŸãã
ã¹ããŒã¹ãããªãã¯ã¹ä¹ç®ã¢ã«ãŽãªãºã ã¯ãã¹ããŒã¹ãããªãã¯ã¹ã«é©ããŠããŸããïŒ
ã©ãããããã§ã¯ãããŸããããããªãã¯ã¹ã¯4ã€ã®èŠçŽ ããšã«2ã€ã®ãŒããæã€å¿ èŠããããããã¹ããŒã¹ãããªãã¯ã¹ã¯é©åã«æ§é åãããŠããå¿ èŠããããŸãã4ã€ã®å€ã2ã€ã®å€ã®å§çž®è¡šçŸãšããŠåŠçããããšã§ã¢ã«ãŽãªãºã ããããã«åŸ®èª¿æŽããããšã¯ããããå¯èœã§ãããããã¯ãAmpereã«ããã¹ããŒã¹ãããªãã¯ã¹ã®æ£ç¢ºãªä¹ç®ãå©çšã§ããªãããšãæå³ããŸãã
è€æ°ã®GPUãå®è¡ããã«ã¯Intelããã»ããµãå¿ èŠã§ããïŒ
Kaggleã³ã³ãã¹ãïŒããã»ããµã«ç·åœ¢ä»£æ°èšç®ãããŒããããŠããïŒã§ããã»ããµã«éè² è·ããããŠããå Žåãé€ããŠãIntelããã»ããµã®äœ¿çšã¯ãå§ãããŸããããããŠããã®ãããªç«¶äºã§ããAMDããã»ããµã¯çŽ æŽãããã§ããAMDããã»ããµã¯ãGOã«ãšã£ãŠå¹³åããŠå®äŸ¡ã§åªããŠããŸãã4-GPUãã«ãã®å ŽåãThreadripperãç§ã®æ±ºå®çãªéžæã§ããç§ãã¡ã®å€§åŠã§ã¯ããã®ãããªããã»ããµãããŒã¹ã«ããæ°åã®ã·ã¹ãã ãåéããŠããããããã¯ãã¹ãŠåé¡ãªãå®å šã«æ©èœããŸããGPUã8ã€ããã·ã¹ãã ã®å Žåã補é å ãçµéšããããã»ããµã䜿çšããŸãã8ã«ãŒãã·ã¹ãã ã®ããã»ããµãšPCIeã®ä¿¡é Œæ§ã¯ãé床ãã³ã¹ãå¹çãããéèŠã§ãã
ã±ãŒã¹ã®åœ¢ç¶ã¯å·åŽã«éèŠã§ããïŒ
çªå·ãéåžžãGPUéã«ããããªã®ã£ãããããå Žåã§ããGPUã¯å®å šã«å·åŽãããŸããããŠãžã³ã°ãç°ãªããš1ã3°Cã®å·®ãçããã«ãŒãã®ééãç°ãªããš10ã30°Cã®å·®ãçããå¯èœæ§ããããŸããäžè¬çã«ãã«ãŒãéã«ééãããã°ãå·åŽã«åé¡ã¯ãããŸãããã®ã£ããããªãå Žåã¯ãé©åãªãã¡ã³ïŒãããŒãã¡ã³ïŒãŸãã¯å¥ã®ãœãªã¥ãŒã·ã§ã³ïŒæ°Žå·ãPCIeãšãã¹ãã³ããŒïŒãå¿ èŠã§ãããããã«ãããã±ãŒã¹ã®çš®é¡ãšãã®ãã¡ã³ã¯é¢ä¿ãããŸããã
AMD GPU + ROCmã¯NVIDIAGPU + CUDAããã£ããããŸããïŒ
ä»åŸæ°å¹Žã§ã¯ãããŸããã 3ã€ã®åé¡ããããŸãïŒãã³ãœã«ã«ãŒãã«ããœãããŠã§ã¢ãã³ãã¥ããã£ã
AMDã®GPUã¯ãªã¹ã¿ã«èªäœã¯åªããŠããŸããFP16ã§ã®åªããããã©ãŒãã³ã¹ãåªããã¡ã¢ãªåž¯åå¹ ã§ãããã ãããã³ãœã«ã³ã¢ãŸãã¯ããã«çžåœãããã®ããªããããNVIDIAã®GPUãšæ¯èŒããŠããã©ãŒãã³ã¹ãäœäžããŸãããŸããããŒããŠã§ã¢ã«ãã³ãœã«ã³ã¢ãå®è£ ããªããšãAMDGPUã競äºåãæã€ããšã¯ãããŸãããåã«ãããšã2020幎ã«ã¯ãã³ãœã«ã³ã¢ã«é¡äŒŒããããŒã¿ã»ã³ã¿ãŒçšã®ã«ãŒããèšç»ãããŠããŸãããæ£ç¢ºãªããŒã¿ã¯ãŸã ãããŸããããµãŒããŒçšã®TensorCoreãšåçã®ã«ãŒããããªãå Žåã¯ãAMD GPUãè³Œå ¥ã§ãã人ãã»ãšãã©ããªãããšãæå³ããNVIDIAã«ç«¶äºåãäžããŸãã
AMDãå°æ¥ãã³ãœã«ã³ã¢ã®ãããªãã®ãåããããŒããŠã§ã¢ãå°å ¥ãããšããŸãããããã®åŸãå€ãã®äººãããèšããŸãããããããAMD GPUã§åäœããããã°ã©ã ã¯ãããŸããïŒã©ãããã°äœ¿ããŸããïŒãããã¯äž»ã«èª€è§£ã§ãã ROCmãå®è¡ããAMDãœãããŠã§ã¢ã¯ãã§ã«ååã«éçºãããŠãããPyTorchã§ã®ãµããŒãã¯é©åã«ç·šæãããŠããŸãããŸããAMD GPU + PyTorchã®åäœã«é¢ããã¬ããŒãã¯ããŸãèŠãããšããããŸãããããã¹ãŠã®ãœãããŠã§ã¢æ©èœãããã«çµ±åãããŠããŸããã©ããããä»»æã®ãããã¯ãŒã¯ãéžæããŠãAMDGPUã§å®è¡ã§ããŸãããããã£ãŠãAMDã¯ãã®åéã§ãã§ã«ååã«éçºãããŠããããã®åé¡ã¯å®è³ªçã«è§£æ±ºãããŠããŸãã
ãããããœãããŠã§ã¢ã®åé¡ãšãã³ãœã«ã³ã¢ã®æ¬ åŠã解決ããåŸãAMDã¯ãã1ã€ãã³ãã¥ããã£ã®æ¬ åŠã«çŽé¢ããŠããŸãã NVIDIA GPUã§åé¡ãçºçããå Žåã¯ãGoogleã§è§£æ±ºçãæ€çŽ¢ããŠèŠã€ããããšãã§ããŸããããã«ãããNVIDIAGPUãžã®ä¿¡é Œãé«ãŸããŸãã NVIDIA GPUã®äœ¿çšã容æã«ããã€ã³ãã©ã¹ãã©ã¯ãã£ãç»å ŽããŠããŸãïŒGOãæ©èœããããã®ãããããã©ãããã©ãŒã ãããããç§åŠçã¿ã¹ã¯ããµããŒããããŠããŸãïŒã NVIDIA GPUïŒããšãã°ãapexïŒãã¯ããã«ç°¡åã«äœ¿çšã§ããããã«ããããã¯ãããªãã¯ããããããããŸãã NVIDIA GPUã®å°é家ãšããã°ã©ããŒã¯ãã¹ãŠã®èã¿ã®äžã«ããŸãããAMDGPUã®å°é家ã¯ã¯ããã«å°ãªããšæããŸãã
ã³ãã¥ããã£ã®èŠ³ç¹ããã¯ãAMDã®ç¶æ³ã¯Julia察Pythonã®ç¶æ³ãšäŒŒãŠããŸãããžã¥ãªã¢ã«ã¯å€ãã®å¯èœæ§ããããå€ãã®äººããã®ããã°ã©ãã³ã°èšèªãç§åŠçç 究ã«ããé©ããŠããããšãæ£ããææããã§ãããããã ããJuliaã¯Pythonãšæ¯èŒããŠã»ãšãã©äœ¿çšãããŸãããPythonã³ãã¥ããã£ãéåžžã«å€§ãããšããã ãã§ããNumpyãSciPyãPandasãªã©ã®åŒ·åãªããã±ãŒãžã®åšãã«ã¯ããããã®äººãéãŸã£ãŠããŸãããã®ç¶æ³ã¯ãNVIDIAãšAMDã®ç¶æ³ã«äŒŒãŠããŸãã
ãããã£ãŠãAMDãNVIDIAã«è¿œãã€ãã®ã¯ããã³ãœã«ã³ã¢ã«çžåœãããã®ãšROCmãäžå¿ã«æ§ç¯ããã匷åºãªã³ãã¥ããã£ãå°å ¥ããããŸã§ã§ããAMDã¯ãåžžã«ç¹å®ã®ãµãã°ã«ãŒãïŒæå·é貚ãã€ãã³ã°ãããŒã¿ã»ã³ã¿ãŒïŒã§åžå Žã·ã§ã¢ãç²åŸããŸããããããNVIDIAã¯ããããããã«2幎éç¬å ãç¶æããã§ãããã
ã¯ã©ãŠããµãŒãã¹ã䜿çšããæ¹ãããã®ã¯ãã€ã§ãããå°çšã®GPUã³ã³ãã¥ãŒã¿ãŒã¯ãã€ã§ããã
ç°¡åãªã«ãŒã«ïŒ1幎以äžGOãå®è¡ããäºå®ã®å Žåã¯ãGPUãæèŒããã³ã³ãã¥ãŒã¿ãŒãè³Œå ¥ããæ¹ãå®äŸ¡ã§ãããã以å€ã®å Žåã¯ãã¯ã©ãŠãããã°ã©ãã³ã°ã®è±å¯ãªçµéšããããGPUã®æ°ãèªç±ã«ã¹ã±ãŒãªã³ã°ã§ããããã«ãããå Žåãé€ããŠãã¯ã©ãŠããµãŒãã¹ã䜿çšããããšããå§ãããŸãã
ã¯ã©ãŠãGPUãèªåã®ã³ã³ãã¥ãŒã¿ãŒãããé«äŸ¡ã«ãªãæ£ç¢ºãªè»¢æç¹ã¯ã䜿çšãããµãŒãã¹ã«å€§ããäŸåããŸããèªåã§èšç®ããããšããå§ãããŸãã以äžã¯ã1ã€ã®V100ãåããAWS V100ãµãŒããŒã®èšç®äŸã§ãããããã©ãŒãã³ã¹ãè¿ã1ã€ã®RTX3090ãåãããã¹ã¯ãããã³ã³ãã¥ãŒã¿ãŒã®ã³ã¹ããšæ¯èŒããŠããŸãã RTX 3090 PCã®äŸ¡æ Œã¯2200ãã«ïŒ2-GPUãã¢ããŒã³+ RTX 3090ïŒã§ããç±³åœã«ããå Žåã¯ãé»æ°çšã«kWhããã0.12ãã«ãè¿œå ããŸããããããAWSã®ãµãŒããŒããã1æéããã2.14ãã«ãšæ¯èŒããŠãã ããã
幎é15ïŒ ã®ãªãµã€ã¯ã«ã§ãã³ã³ãã¥ãŒã¿ãŒã¯
ïŒ350 WïŒGPUïŒ+ 100 WïŒCPUïŒïŒ* 0.15ïŒãªãµã€ã¯ã«ïŒ* 24æé* 365æ¥= 591 kWh /幎ã䜿çšããŸãã
幎é591kWhã¯ãããã«71ãã«ãæäŸããŸãã
ã³ã³ãã¥ãŒã¿ãŒãšã¯ã©ãŠãã®äŸ¡æ Œã15ïŒ ã®äœ¿çšçã§æ¯èŒãããšã転æç¹ã¯300æ¥ç®ãããã«ãªããŸãïŒ2,311ãã«å¯Ÿ2,270ãã«ïŒïŒ
2.14ãã«/æé* 0.15ïŒãªãµã€ã¯ã«ïŒ* 24æé* 300æ¥= 2,311ãã«
èšç®ãããšã GOã¢ãã«ã®å¯¿åœã300æ¥ãè¶ ããå Žåã¯ãAWSã䜿çšãããããã³ã³ãã¥ãŒã¿ãŒãè³Œå ¥ããããšããå§ãããŸãã
ã³ã³ãã¥ãŒã¿ãšã¯ã©ãŠãã®ã©ã¡ãã䜿çšãããã決å®ããããã«ãä»»æã®ã¯ã©ãŠããµãŒãã¹ã«å¯ŸããŠåæ§ã®èšç®ãè¡ãããšãã§ããŸãã
èšç®èœåã®å©çšã«é¢ããäžè¬çãªæ°å€ã¯æ¬¡ã®ãšããã§ãã
- PhDã³ã³ãã¥ãŒã¿ãŒïŒ<15ïŒ ;
- PhD Slurmã®GPUã¯ã©ã¹ã¿ãŒïŒ> 35ïŒ
- Slurmã®äŒæ¥ç 究ã¯ã©ã¹ã¿ãŒïŒ> 60ïŒ ã
äžè¬ã«ãå®çšçãªãœãªã¥ãŒã·ã§ã³ãéçºãããããæå 端ã®ã¢ã€ãã¢ãèããããšãéèŠãªåéã§ã¯ããªãµã€ã¯ã«çãäœããªããŸããäžéšã®é åã§ã¯äœ¿çšçãäœãïŒè§£éå¯èœæ§ã®èª¿æ»ïŒãä»ã®é åã§ã¯ã¯ããã«é«ããªã£ãŠããŸãïŒæ©æ¢°å€æãèšèªã¢ããªã³ã°ïŒãäžè¬çã«ãèªå®¶çšè»ã®ãªãµã€ã¯ã«ã¯åžžã«é倧è©äŸ¡ãããŠããŸããéåžžãã»ãšãã©ã®ããŒãœãã«ã·ã¹ãã ã¯5ã10ïŒ ãªãµã€ã¯ã«ãããŸãããããã£ãŠãç 究ããŒã ãäŒæ¥ã¯ãåå¥ã®ãã¹ã¯ãããã§ã¯ãªããSlurmã§GPUã¯ã©ã¹ã¿ãŒãç·šæããããšã匷ããå§ãããŸãã
æ æ°ãããŠèªããªã人ã®ããã®ãã³ã
å šäœçã«æé«ã®GPUïŒRTX 3080ããã³RTX3090ã
é¿ããã¹ãGPUïŒç 究è ãšããŠïŒïŒTeslaã«ãŒããQuadroãFounders EditionãTitan RTXãTitan VãTitanXP ã
åªããããã©ãŒãã³ã¹/äŸ¡æ Œæ¯ããããé«äŸ¡ïŒRTX3080ã
åªããããã©ãŒãã³ã¹/äŸ¡æ Œæ¯ãããå®ãïŒRTX 3070ãRTX2060ã¹ãŒããŒã
ãéãå°ãªãïŒäžå€ã«ãŒããè²·ããéå±€ïŒRTX 2070ïŒ$ 400ïŒãRTX 2060ïŒ$ 300ïŒãGTX 1070ïŒ$ 220ïŒãGTX 1070 TiïŒ$ 230ïŒãGTX 1650 SuperïŒ$ 190ïŒãGTX 980 TiïŒ6GB $ 150ïŒã
ç§ã«ã¯ã»ãšãã©ãéããããŸãããå€ãã®ã¹ã¿ãŒãã¢ãããã¯ã©ãŠããµãŒãã¹ã宣äŒããŠããŸããã¯ã©ãŠãã§ç¡æã®ã¯ã¬ãžããã䜿çšããGPUãè³Œå ¥ã§ããããã«ãªããŸã§ãµãŒã¯ã«ã§å€æŽããŸãã
Kaggleã³ã³ããã£ã·ã§ã³ã«åå ããŸãïŒRTX3070ã
ã³ã³ãã¥ãŒã¿ãŒããžã§ã³ãäºååŠç¿ããŸãã¯æ©æ¢°å€æã§ç«¶äºã«åãšããšããŠããŸãïŒ4åã®RTX3090ããã ããååãªå·åŽãšååãªé»åãåããã¢ã»ã³ããªãããããšãå°é家ã確èªãããŸã§åŸ ã¡ãŸãã
ç§ã¯èªç¶ãªèšèªåŠçãåŠãã§ããŸããæ©æ¢°ç¿»èš³ãèšèªã¢ããªã³ã°ããŸãã¯äºååŠç¿ã«èå³ããªãå Žåã¯ãRTX3080ã§ååã§ãã
ç§ã¯GOãå§ããŠãæ¬åœã«ããã«å€¢äžã«ãªããŸãããRTX3070ããå§ããŸãã6ã9ãæã§é£œããªãå Žåã¯ã4ã€ã®RTX 3080ã販売ããã³è³Œå ¥ããŸãã次ã«éžæãããã®ïŒèµ·åãKaggleã調æ»ãé©çšGOïŒã«å¿ããŠãæ°å¹Ž3ã€ã§ãGPUã販売ããããè¯ããã®ïŒæ¬¡äžä»£RTX GPUïŒãè³Œå ¥ããŸãã
GOãè©Šãããã®ã§ãããçå£ãªæå³ã¯ãããŸãããRTX2060Superã¯åªããéžæè¢ã§ãããPSUã®äº€æãå¿ èŠã«ãªãå ŽåããããŸãããã¶ãŒããŒãã«PCIex16ã¹ãããããããPSUãçŽ300ã¯ãããçæããå ŽåãGTX 1050 Tiã¯ä»ã®æè³ãå¿ èŠãšããªããããåªãããªãã·ã§ã³ã«ãªããŸãã
128 GPUæªæºã®äžŠåã·ãã¥ã¬ãŒã·ã§ã³çšã®GPUã¯ã©ã¹ã¿ãŒïŒã¯ã©ã¹ã¿ãŒçšã«RTXãè³Œå ¥ã§ããå ŽåïŒ66ïŒ 8x RTX 3080ããã³33ïŒ 8x RTX 3090ïŒã¢ã»ã³ããªãååã«å·åŽã§ããå Žåã®ã¿ïŒãå·åŽãäžååãªå Žåã¯ã33ïŒ RTX 6000GPUãŸãã¯8xTeslaA100ãè³Œå ¥ããŠãã ããã RTX GPUãè³Œå ¥ã§ããªãå Žåã¯ã8ã€ã®SupermicroA100ããŒããŸãã¯8ã€ã®RTX6000ããŒããéžæããŸã
ã128ãè¶ ããGPUã䜿çšãã䞊åã·ãã¥ã¬ãŒã·ã§ã³çšã®GPUã¯ã©ã¹ã¿ãŒïŒ8ãã¹ã©A100ã®è»ã«ã€ããŠèããŠã¿ãŠãã ããã512ãè¶ ããGPUãå¿ èŠãªå Žåã¯ãDGX A100SuperPODã·ã¹ãã ãæ€èšããŠãã ããã