"Locality is efficiency, Efficiency is power, Power is performance, Performance is King", Bill Dally ãã«ãã¹ã¬ããã£ã³ã°ã¨ã¯ï¼ CPUã¨GPUã®ãã«ãã¹ã¬ããã£ã³ã°ã®éããããã°ã«ã¾ã¨ãã¦ãããã©ä¾ã«ãã£ã¦èª°ãèå³ãªãããâ arutema47 (@arutema47) 2021å¹´8æ16æ¥ ã¤ã¶ããããèªã¿ããæ¹ãå¤ããã ã£ãã®ã§å®æããã¾ããã ãã«ãã¹ã¬ããã£ã³ã°ã¨ã¯ã¡ã¢ãªé 延ãé è½ãã¹ã«ã¼ããããä¸ãããã¼ãã¦ã§ã¢ã®ãã¯ããã¯ã§ãã ãã CPUã¨GPUã§ä½¿ããæ¹ãããªãç°ãªãããããã®éãã«ã¤ãã¦èãã¦ã¿ãè¨äºã§ãã ï¼SIMDã«ã¤ãã¦ä¸¦åããã°ã©ãã³ã°ã®è¦³ç¹ãã触ããã¹ãã§ããããæéç¡ããã«ãã¹ã¬ããã£ã³ã°ã«æ³¨ç®ããããåçã§ã¯çãã¾ãããï¼ æ¬è¨äºã«ã¤ã㦠æ¬è¨äºã¯CPUã¨G
AMDï¼GPUã¨CPUã§åãããã°ã©ã ãåããHSAãã®ææ°ååãå ¬è¡¨ãJavaã¸ã®å¯¾å¿è¨ç»ãæããã« ã©ã¤ã¿ã¼ï¼å¡©ç°ç´³äº ç±³å½æé2013å¹´8æ25æ¥ãã27æ¥ã¾ã§ï¼ç±³å½ã«ãªãã©ã«ãã¢å·ã®ã¹ã¿ã³ãã©ã¼ã大å¦ã«ã¦ï¼åå°ä½ãããã¤ã¹é¢é£ã®å¦è¡ã¤ãã³ããHot Chipsããéå¬ããã¦ãããAMDã¯ããã§ï¼å社ãæå±ããGPUã³ã³ãã¥ã¼ãã£ã³ã°ãã¬ã¼ã ã¯ã¼ã¯ãHSAãï¼Heterogeneous System Architectureï¼ã®åãçµã¿ã¨ææã«ã¤ãã¦çºè¡¨ããã¨ããã ããã«å ç«ã¡ï¼å ±éé¢ä¿è ã対象ã¨ããé»è©±ä¼è°ã§ï¼ãã®æ¦è¦ã説æãããã®ã§ï¼ä»åã¯ãã®æ¦è¦ãã¬ãã¼ããã¦ã¿ããã CPUã¨GPUãåãããã°ã©ã ã§æ±ããããã«ããHSA㯠OpenCLã®ããããå®è¡ç°å¢ãç®æã HSAã¨ã¯GPUã³ã³ãã¥ã¼ãã£ã³ã°ã容æã«ããããã®ä»çµã¿ã§ããï¼CPUãGPUã¨ãã£ããã¼ãã¦ã§ã¢ã¨ï¼ãã®ä¸ã§
è¦ç¹ã¨ãªãã§ãããé¨åã ãæ¥æ¬èªã§ã¾ã¨ãã¦ãããLee et al. (2010) 1. Introduction ãã®è«æã®è¦ç¹ã¯ä»¥ä¸ã®3ã¤ã éå»ã«å¤æ°åå¨ãããGPUã¯100åãã100åãã®æ§è½ãæã¤ã¨ä¸»å¼µããè«æã®å 容ãåæ¤è¨ãããæ¤è¨ã«ããã¦ãCPUã¨GPU両æ¹ãæé©åããããã®çµæãGPUã¯å¹³å2.5åã®æ§è½ãCPUæ¯ã§æã¤ãã¨ãããã£ããã¤ã¾ããCPUã¨GPUã¯åãå俵ï¼ballparkï¼ã«ãããã®ãªã®ã ã 並ååï¼ä¸¦è¡åãã®æ¹ãè¯ãã®ããªï¼ã®æ段ã«kernelï¼ããã§ã¯ã¢ããªã±ã¼ã·ã§ã³ãããã°ã©ã ã¨ãã£ãæå³ãããï¼ã®å®å¹æ§è½ãã©ã®ããã«ä¾åããããã系統çã«åé¡ãã¦ã¾ã¨ããã CPUã¨GPUã®æ§è½ã®éããåæããã¢ã¼ããã¯ãã£ã®ä¸ã§ãæ§è½ã«å½±é¿ãä¸ãã主è¦ãªé¨åãç¹å®ããã 2. The workload: throughput computing kernels 14
ãã®è¨äºã¯ GPGPU Advent Calender ã®äºæ¥ç®ã®è¨äºã§ãã äºæ¥ç®ã®å 容ãã³ã¬ããï¼ã¨ããã¿ãªããã®ã¤ã£ãã¿ãæ¬ç·¨ã§ãã以ä¸ã®å 容ã¯ãªãã±ã¨ãªãã¾ãã http://pcl.intel-research.net/publications/isca319-lee.pdf æã¯2010å¹´â¦ãã¼ã 使ãã°100åã¨ã1000åã¨ãéããªããã§ããï¼ãªãã§CPUãããªé ãã®ï¼ãªã©ã¨ãä¸çä¸ã§ããã¯ã½ã«è¨ãããIntelã¯ãã¤ãã«ããåãã¦ãããåãã¯å ¨å¡ééã£ã¦ãï¼ä¿ºããã®ééããæ£ãã¦ããï¼ãã¨ããå 容ã®è«æãISCAã«æ稿ããã®ã ã£ã⦠ã¨ããäºæ ã ã£ãã®ãã©ããã¯ç¥ããªãã§ãããå 容ã¨ãã¦ã¯ãâãã®ã°ã©ããå ¨ã¦ãç©èªã£ã¦ãã¦ã ãGPUã ã¨100åéãã¨ãè¨ããã©ãããã¯CPUã®ã³ã¼ããæé©åãã¦ãªãããã§ãCPUãæé©åããã°ãå¹³åãã£ãã®2.5åããéããªããªãããã¨ãã
ãªãªã¼ã¹ãé害æ å ±ãªã©ã®ãµã¼ãã¹ã®ãç¥ãã
ææ°ã®äººæ°ã¨ã³ããªã¼ã®é ä¿¡
å¦çãå®è¡ä¸ã§ã
j次ã®ããã¯ãã¼ã¯
kåã®ããã¯ãã¼ã¯
lãã¨ã§èªã
eã³ã¡ã³ãä¸è¦§ãéã
oãã¼ã¸ãéã
{{#tags}}- {{label}}
{{/tags}}