{"id":59823,"date":"2020-11-16T15:01:31","date_gmt":"2020-11-16T14:01:31","guid":{"rendered":"https:\/\/www.planet3dnow.de\/cms\/?p=59823"},"modified":"2020-11-16T21:52:47","modified_gmt":"2020-11-16T20:52:47","slug":"amd-announces-worlds-fastest-hpc-accelerator-for-scientific-research1","status":"publish","type":"post","link":"https:\/\/www.planet3dnow.de\/cms\/59823-amd-announces-worlds-fastest-hpc-accelerator-for-scientific-research1\/","title":{"rendered":"AMD<\/span> Announces World\u2019s Fastest HPC<\/span> Accelerator for Scientific Research\u00b9"},"content":{"rendered":"

\ua7f7 AMD<\/span> Instinct\u2122 MI100<\/span> acce\u00adle\u00adra\u00adtors revo\u00adlu\u00adtio\u00adni\u00adze high-per\u00adfor\u00admance com\u00adpu\u00adting (HPC<\/span>) and AI<\/span> with indus\u00adtry-lea\u00adding com\u00adpu\u00adte per\u00adfor\u00admance \ua7f7<\/i><\/span><\/span><\/span><\/span><\/p>\n

\ua7f7 First GPU<\/span> acce\u00adle\u00adra\u00adtor with new AMD<\/span> CDNA<\/span> archi\u00adtec\u00adtu\u00adre engi\u00adnee\u00adred for the exas\u00adca\u00adle era \ua7f7 <\/i><\/span><\/span><\/span><\/span><\/p>\n

SANTA<\/span> CLARA<\/span>, Calif. <\/b><\/span><\/span>\u2014 Novem\u00adber 16, 2020 \u2014 <\/b><\/span>AMD<\/span><\/a><\/u><\/span><\/span> (NASDAQ<\/span>: AMD<\/span>) today announ\u00adced the new AMD<\/span> Instinct\u2122 MI100<\/span> acce\u00adle\u00adra\u00adtor \u2013 the world\u2019s fas\u00adtest HPC<\/span> GPU<\/span> and the first x86 ser\u00adver GPU<\/span> to sur\u00adpass the 10 tera\u00adflops (FP64<\/span>) per\u00adfor\u00admance bar\u00adri\u00ader.<\/span>1<\/span><\/sup> Sup\u00adport\u00aded by new acce\u00adle\u00adra\u00adted com\u00adpu\u00adte plat\u00adforms from Dell, Giga\u00adbyte, HPE<\/span>, and Super\u00admi\u00adcro, the MI100<\/span>, com\u00adbi\u00adned with AMD<\/span> EPYC<\/span><\/span>TM<\/span><\/span><\/sup> CPUs and the ROCm\u2122 4.0 open soft\u00adware plat\u00adform, is desi\u00adgned to pro\u00adpel new dis\u00adco\u00adveries ahead of the exas\u00adca\u00adle era. <\/span><\/span><\/span><\/span><\/p>\n

Built on the new AMD<\/span> CDNA<\/span> archi\u00adtec\u00adtu\u00adre, the AMD<\/span> Instinct MI100<\/span> GPU<\/span> enables a new class of acce\u00adle\u00adra\u00adted sys\u00adtems for HPC<\/span> and AI<\/span> when pai\u00adred with 2<\/span>nd<\/span><\/sup> Gen AMD<\/span> EPYC<\/span> pro\u00adces\u00adsors. The MI100<\/span> offers up to 11.5 TFLOPS<\/span> of peak FP64<\/span> per\u00adfor\u00admance for HPC<\/span> and up to 46.1 TFLOPS<\/span> peak FP32<\/span> Matrix per\u00adfor\u00admance for AI<\/span> and machi\u00adne lear\u00adning workloads<\/span>2<\/span><\/sup>. With new AMD<\/span> Matrix Core tech\u00adno\u00adlo\u00adgy, the MI100<\/span> also deli\u00advers a near\u00adly 7x boost in FP16<\/span> theo\u00adre\u00adti\u00adcal peak floa\u00adting point per\u00adfor\u00admance for AI<\/span> trai\u00adning workloads com\u00adpared to AMD<\/span>\u2019s pri\u00ador gene\u00adra\u00adti\u00adon acce\u00adle\u00adra\u00adtors.<\/span>3<\/span><\/sup> <\/span><\/span><\/span><\/p>\n

\u201c<\/span>Today AMD<\/span> takes a major step for\u00adward in the jour\u00adney toward exas\u00adca\u00adle com\u00adpu\u00adting as we unveil the AMD<\/span> Instinct MI100<\/span> \u2013 the world\u2019s fas\u00adtest HPC<\/span> GPU<\/span>,\u201d said Brad McCre\u00addie, cor\u00adpo\u00adra\u00adte vice pre\u00adsi\u00addent, Data Cen\u00adter GPU<\/span> and Acce\u00adle\u00adra\u00adted Pro\u00adces\u00adsing, AMD<\/span>. \u201cSqua\u00adre\u00adly tar\u00adge\u00adted toward the workloads that mat\u00adter in sci\u00aden\u00adti\u00adfic com\u00adpu\u00adting, our latest acce\u00adle\u00adra\u00adtor, when com\u00adbi\u00adned with the AMD<\/span> ROCm open soft\u00adware plat\u00adform, is desi\u00adgned to pro\u00advi\u00adde sci\u00aden\u00adtists and rese\u00adar\u00adchers a supe\u00adri\u00ador foun\u00adda\u00adti\u00adon for their work in HPC<\/span>.\u201d <\/span><\/span><\/span><\/span><\/p>\n

Open Soft\u00adware Plat\u00adform for the Exas\u00adca\u00adle Era<\/b><\/span><\/span><\/p>\n

The AMD<\/span> ROCm deve\u00adlo\u00adper soft\u00adware pro\u00advi\u00addes the foun\u00adda\u00adti\u00adon for exas\u00adca\u00adle com\u00adpu\u00adting. As an open source tool\u00adset con\u00adsis\u00adting of com\u00adpi\u00adlers, pro\u00adgramming APIs and libra\u00adri\u00ades, ROCm is used by exas\u00adca\u00adle soft\u00adware deve\u00adlo\u00adpers to crea\u00adte high per\u00adfor\u00admance appli\u00adca\u00adti\u00adons. ROCm 4.0 has been opti\u00admi\u00adzed to deli\u00adver <\/span><\/span>per\u00adfor\u00admance at sca\u00adle for MI100-based sys\u00adtems. ROCm 4.0 has upgraded the com\u00adpi\u00adler to be open source and uni\u00adfied to sup\u00adport both OpenMP\u00ae 5.0 and HIP<\/span>. PyTorch and Ten\u00adsor\u00adflow frame\u00adworks, which have been opti\u00admi\u00adzed with ROCm 4.0, can now achie\u00adve hig\u00adher per\u00adfor\u00admance with MI100<\/span><\/span><\/span>7,8<\/span><\/span><\/sup>. ROCm 4.0 is the latest offe\u00adring for HPC<\/span>, ML<\/span> and AI<\/span> appli\u00adca\u00adti\u00adon deve\u00adlo\u00adpers which allows them to crea\u00adte per\u00adfor\u00admance por\u00adta\u00adble software. <\/span><\/span><\/p>\n

\u201c<\/span>We\u2019ve recei\u00adved ear\u00adly access to the MI100<\/span> acce\u00adle\u00adra\u00adtor, and the preli\u00admi\u00adna\u00adry results are very encou\u00adra\u00adging. We\u2019ve typi\u00adcal\u00adly seen signi\u00adfi\u00adcant per\u00adfor\u00admance boosts, up to 2\u20133x com\u00adpared to other GPUs,\u201d said Bron\u00adson Mes\u00adser, direc\u00adtor of sci\u00adence, Oak Ridge Lea\u00adder\u00adship Com\u00adpu\u00adting Faci\u00adli\u00adty. \u201cWhat\u2019s also important to reco\u00adgni\u00adze is the impact soft\u00adware has on per\u00adfor\u00admance. The fact that the ROCm open soft\u00adware plat\u00adform and HIP<\/span> deve\u00adlo\u00adper tool are open source and work on a varie\u00adty of plat\u00adforms, it is some\u00adthing that we have been abso\u00adlut\u00ade\u00adly almost obses\u00adsed with sin\u00adce we fiel\u00added the very first hybrid CPU<\/span>\/GPU<\/span> system.\u201d<\/span><\/span><\/p>\n

 <\/p>\n

Key capa\u00adbi\u00adli\u00adties and fea\u00adtures of the AMD<\/span> Instinct MI100<\/span> acce\u00adle\u00adra\u00adtor include: <\/span><\/span><\/span><\/span><\/p>\n