{"id":64802,"date":"2022-03-23T09:24:11","date_gmt":"2022-03-23T08:24:11","guid":{"rendered":"https:\/\/www.planet3dnow.de\/cms\/?p=64802"},"modified":"2022-03-23T09:24:11","modified_gmt":"2022-03-23T08:24:11","slug":"amd-instinct-expands-ecosystem-and-delivers-exascale-class-technology-for-hpc-and-ai-applications","status":"publish","type":"post","link":"https:\/\/www.planet3dnow.de\/cms\/64802-amd-instinct-expands-ecosystem-and-delivers-exascale-class-technology-for-hpc-and-ai-applications\/","title":{"rendered":"AMD<\/span> Instinct\u2122 Expands Ecosystem and Delivers Exascale-Class Technology for HPC<\/span> and AI<\/span> Applications"},"content":{"rendered":"

\u2013 Powered by AMD<\/span> CDNA<\/span>\u2122 2 archi\u00adtec\u00adtu\u00adre and AMD<\/span> ROCm\u21225, new AMD<\/span> Instinct MI210<\/span> GPUs acce\u00adle\u00adra\u00adting insights and dis\u00adco\u00advery for main\u00adstream users \u2013<\/em><\/p>\n

SANTA<\/span> CLARA<\/span>, Calif., March 22, 2022 (GLOBE<\/span> NEWSWIRE<\/span>) \u2014 AMD<\/span><\/a> (NASDAQ<\/span>: AMD<\/span>) today announ\u00adced the avai\u00adla\u00adbi\u00adli\u00adty of the AMD<\/span> Instinct\u2122 eco\u00adsys\u00adtem<\/a> with expan\u00added sys\u00adtem sup\u00adport from part\u00adners inclu\u00adding ASUS<\/span>, Dell Tech\u00adno\u00adlo\u00adgies, Giga\u00adbyte, HPE<\/span>, Leno\u00advo and Super\u00admi\u00adcro, the new AMD<\/span> Instinct\u2122 MI210<\/span> acce\u00adle\u00adra\u00adtor<\/a> and the robust capa\u00adbi\u00adli\u00adties of ROCm\u2122 5 soft\u00adware<\/a>. Altog\u00ade\u00adther, the AMD<\/span> Instinct and ROCm eco\u00adsys\u00adtem is offe\u00adring exas\u00adca\u00adle-class tech\u00adno\u00adlo\u00adgy to a broad base of HPC<\/span> and AI<\/span> cus\u00adto\u00admers, addres\u00adsing the gro\u00adwing demand for com\u00adpu\u00adte-acce\u00adle\u00adra\u00adted data cen\u00adter workloads and redu\u00adcing the time to insights and discovery.<\/p>\n

\u201c<\/span>With twice the plat\u00adforms available com\u00adpared to our pre\u00advious gene\u00adra\u00adti\u00adon acce\u00adle\u00adra\u00adtors, gro\u00adwing cus\u00adto\u00admer adop\u00adti\u00adon across HPC<\/span> and AI<\/span> appli\u00adca\u00adti\u00adons, and new sup\u00adport from com\u00admer\u00adcial ISVs in key workloads, we\u2019re con\u00adti\u00adnuing to dri\u00adve adop\u00adti\u00adon of the AMD<\/span> Instinct MI200<\/span> acce\u00adle\u00adra\u00adtors and ROCm 5 soft\u00adware eco\u00adsys\u00adtem,\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>. \u201cNow with the avai\u00adla\u00adbi\u00adli\u00adty of the AMD<\/span> Instinct MI210<\/span> acce\u00adle\u00adra\u00adtor to the MI200<\/span> fami\u00adly, our cus\u00adto\u00admers can choo\u00adse the acce\u00adle\u00adra\u00adtor that works best for their workloads, whe\u00adther they need lea\u00adding-edge acce\u00adle\u00adra\u00adted pro\u00adces\u00adsing for lar\u00adge sca\u00adle HPC<\/span> and AI<\/span> workloads, or if they want access to exas\u00adca\u00adle-class tech\u00adno\u00adlo\u00adgy in a com\u00admer\u00adcial format.\u201d<\/p>\n

\u201c<\/span>The Lumi super\u00adcom\u00adpu\u00adter powered by AMD<\/span> EPYC<\/span> pro\u00adces\u00adsors and AMD<\/span> Instinct MI200<\/span> acce\u00adle\u00adra\u00adtors will pro\u00advi\u00adde a gene\u00adra\u00adtio\u00adnal leap in per\u00adfor\u00admance for lar\u00adge-sca\u00adle simu\u00adla\u00adti\u00adons and mode\u00adling as well as AI<\/span> and deep lear\u00adning workloads to sol\u00adve some of the big\u00adgest ques\u00adti\u00adons in rese\u00adarch,\u201d said Pek\u00adka Man\u00adni\u00adnen, Direc\u00adtor of the LUMI<\/span> Lea\u00adder\u00adship and Com\u00adpu\u00adting Faci\u00adli\u00adty, CSC<\/span>. \u201cWe\u2019ve uti\u00adli\u00adzed AMD<\/span> Instinct MI210<\/span> acce\u00adle\u00adra\u00adtors to get hands on expe\u00adri\u00adence with the Instinct MI200<\/span> fami\u00adly, pre\u00adpa\u00adring our sci\u00aden\u00adtists to tack\u00adle the many chal\u00adlen\u00adging and com\u00adplex pro\u00adjects they will run once Lumi is ful\u00adly deployed.\u201d<\/p>\n

Powe\u00adring The Future of HPC<\/span> and AI<\/span><\/strong>
\nThe AMD<\/span> Instinct MI200<\/span> series acce\u00adle\u00adra\u00adtors are desi\u00adgned to power dis\u00adco\u00adveries in exas\u00adca\u00adle sys\u00adtems, enab\u00adling rese\u00adar\u00adchers, sci\u00aden\u00adtists and engi\u00adneers to tack\u00adle our most pres\u00adsing chal\u00adlenges, from cli\u00adma\u00adte chan\u00adge to vac\u00adci\u00adne rese\u00adarch. The AMD<\/span> Instinct MI210<\/span> acce\u00adle\u00adra\u00adtors<\/a> spe\u00adci\u00adfi\u00adcal\u00adly enable exas\u00adca\u00adle-class tech\u00adno\u00adlo\u00adgies for cus\u00adto\u00admers who need fan\u00adta\u00adstic HPC<\/span> and AI<\/span> per\u00adfor\u00admance in a PCIe\u00ae for\u00admat. Powered by the AMD<\/span> CDNA<\/span>\u2122 2 archi\u00adtec\u00adtu\u00adre, AMD<\/span> Instinct MI210<\/span> acce\u00adle\u00adra\u00adtors extend AMD<\/span> per\u00adfor\u00admance lea\u00adder\u00adship in dou\u00adble pre\u00adcis\u00adi\u00adon (FP64<\/span>) com\u00adpu\u00adte on PCIe form fac\u00adtor cards1<\/sup>. They also deli\u00adver a robust solu\u00adti\u00adon for acce\u00adle\u00adra\u00adted deep lear\u00adning trai\u00adning offe\u00adring a broad ran\u00adge of mixed-pre\u00adcis\u00adi\u00adon capa\u00adbi\u00adli\u00adties based on the AMD<\/span> Matrix Core Technology.<\/p>\n

Dri\u00adving the ROCm Adoption<\/strong>
\nAn open soft\u00adware plat\u00adform that allows rese\u00adar\u00adchers, sci\u00aden\u00adtists and engi\u00adneers to tap the power of AMD<\/span> Instinct acce\u00adle\u00adra\u00adtors to dri\u00adve sci\u00aden\u00adti\u00adfic dis\u00adco\u00adveries, the AMD<\/span> ROCm plat\u00adform is built on the foun\u00adda\u00adti\u00adon of num\u00ade\u00adrous appli\u00adca\u00adti\u00adons and libra\u00adri\u00ades powe\u00adring top HPC<\/span> and AI<\/span> applications.<\/p>\n

With ROCm 5, AMD<\/span> extends its soft\u00adware plat\u00adform by adding new hard\u00adware sup\u00adport for the AMD<\/span> Instinct MI200<\/span> series acce\u00adle\u00adra\u00adtors and the AMD<\/span> Rade\u00adon\u2122 PRO<\/span> W6800<\/span> pro\u00adfes\u00adsio\u00adnal gra\u00adphics card, plus Red Hat\u00ae Enter\u00adpri\u00adse Linux\u00ae 8.5 sup\u00adport, incre\u00adasing acces\u00adsi\u00adbi\u00adli\u00adty of ROCm for deve\u00adlo\u00adpers and enab\u00adling out\u00adstan\u00adding per\u00adfor\u00admance across key workloads.<\/p>\n

Addi\u00adtio\u00adnal\u00adly, through the AMD<\/span> Infi\u00adni\u00adty Hub<\/a>, the cen\u00adtral loca\u00adti\u00adon for open-source appli\u00adca\u00adti\u00adons that are por\u00adted and opti\u00admi\u00adzed on AMD<\/span> GPUs, end-users can easi\u00adly find, down\u00adload and install con\u00adtai\u00adne\u00adri\u00adzed HPC<\/span> apps and ML<\/span> frame\u00adworks. AMD<\/span> Infi\u00adni\u00adty Hub appli\u00adca\u00adti\u00adon con\u00adtai\u00adners are desi\u00adgned to redu\u00adce the tra\u00addi\u00adtio\u00adnal\u00adly dif\u00adfi\u00adcult issue of obtai\u00adning and instal\u00adling soft\u00adware releases while allo\u00adwing users to learn based on shared expe\u00adri\u00aden\u00adces and pro\u00adblem-sol\u00adving opportunities.<\/p>\n

Expan\u00adding Part\u00adner and Cus\u00adto\u00admer Ecosystem<\/strong>
\nAs more pur\u00adpo\u00adse-built appli\u00adca\u00adti\u00adons are opti\u00admi\u00adzed to work with ROCm and AMD<\/span> Instinct acce\u00adle\u00adra\u00adtors, AMD<\/span> con\u00adti\u00adnues to grow its soft\u00adware eco\u00adsys\u00adtem with the addi\u00adti\u00adon of com\u00admer\u00adcial ISVs, inclu\u00adding Ansys\u00ae, Cas\u00adca\u00adde Tech\u00adno\u00adlo\u00adgies, and Tem\u00adpo\u00adQuest. The\u00adse ISVs pro\u00advi\u00adde appli\u00adca\u00adti\u00adons for acce\u00adle\u00adra\u00adted workloads inclu\u00adding Com\u00adpu\u00adta\u00adtio\u00adnal Flu\u00adid Dyna\u00admics (CFD<\/span>), wea\u00adther fore\u00adcas\u00adting, Com\u00adpu\u00adter Aided Engi\u00adnee\u00adring (CAE<\/span>) and more. The\u00adse updates are on top of exis\u00adting appli\u00adca\u00adti\u00adon sup\u00adport in ROCm which includes HPC<\/span>, AI<\/span> and Machi\u00adne Lear\u00adning appli\u00adca\u00adti\u00adons, AMBER<\/span>, Chro\u00adma, CP2K<\/span>, GRID<\/span>, GRO\u00adMACs, LAAMPS<\/span>, MILC<\/span>, Mini-HAAC, NAMD<\/span>, NAMD<\/span> 3.0, ONNX-RT<\/span>, OpenMM, PyTorch, RELION<\/span>, SPECFEM3D<\/span> Car\u00adte\u00adsi\u00adan, SPECFEM3D<\/span> Glo\u00adbe, and TensorFlow.<\/p>\n

AMD<\/span> is also enab\u00adling part\u00adners like ASUS<\/span>, Dell Tech\u00adno\u00adlo\u00adgies, Giga\u00adbyte, HPE<\/span>, Leno\u00advo, Super\u00admi\u00adcro, and Sys\u00adtem Inte\u00adgra\u00adtors inclu\u00adding Col\u00adfax, Exxact, KOI<\/span> Com\u00adpu\u00adters, Nor-Tech, Pen\u00adgu\u00adin and Sym\u00adme\u00adtric to offer dif\u00adfe\u00adren\u00adtia\u00adted solu\u00adti\u00adons to address next gene\u00adra\u00adti\u00adon com\u00adpu\u00adting chal\u00adlenges. Super\u00adcom\u00adpu\u00adting cus\u00adto\u00admers are alre\u00ada\u00addy taking advan\u00adta\u00adge of the bene\u00adfits offe\u00adred via the\u00adse new cus\u00adto\u00admer wins inclu\u00adding the Fron\u00adtier instal\u00adla\u00adti\u00adon at Oak Ridge Natio\u00adnal Labo\u00adra\u00adto\u00adry<\/a>, KTH<\/span>\/Dardel<\/a>, CSC<\/span>\/LUMI<\/span><\/a> and Cines\/Adastra<\/a>.<\/p>\n

Enab\u00adling Access for Cus\u00adto\u00admers and Partners<\/strong>
\nThe AMD<\/span> Acce\u00adle\u00adra\u00adtor Cloud<\/a> offers cus\u00adto\u00admers an envi\u00adron\u00adment to remo\u00adte\u00adly access and eva\u00adlua\u00adte AMD<\/span> Instinct acce\u00adle\u00adra\u00adtors and AMD<\/span> ROCm soft\u00adware. Whe\u00adther it\u2019s port\u00ading lega\u00adcy code, bench\u00admar\u00adking an appli\u00adca\u00adti\u00adon or test\u00ading mul\u00adti-GPU or mul\u00adti-node sca\u00adling, the AMD<\/span> Acce\u00adle\u00adra\u00adtor Cloud gives pro\u00ads\u00adpec\u00adti\u00adve cus\u00adto\u00admers and part\u00adners quick and easy access to the latest GPUs and soft\u00adware. The AMD<\/span> Acce\u00adle\u00adra\u00adtor Cloud is also used to power various events such as hacka\u00adthons and ROCm trai\u00adning ses\u00adsi\u00adons offe\u00adred to both exis\u00adting and pro\u00ads\u00adpec\u00adti\u00adve cus\u00adto\u00admers, allo\u00adwing deve\u00adlo\u00adpers to hone their skills and learn how to get the most out of AMD<\/span> Instinct accelerators.<\/p>\n

MI200<\/span> Series Specifications<\/strong><\/p>\n\n\n\n\n\n\n
Model<\/td>\nCom\u00adpu\u00adte Units<\/td>\nStream Pro\u00adces\u00adsors<\/td>\nFP64<\/span> | FP32<\/span> Vec\u00adtor (Peak)<\/td>\nFP64<\/span> | FP32<\/span> Matrix (Peak)<\/td>\nFP16<\/span> | bf16
\n(Peak)<\/td>\n		INT8<\/span>
\n(Peak)<\/td>\n		HBM2e
\nECC<\/span>
\nMemory<\/td>\n		Memo\u00adry Bandwidth<\/td>\nForm Fac\u00adtor<\/td>\n<\/tr>\n
AMD<\/span> Instinct MI210<\/span><\/td>\n104<\/td>\n6,656<\/td>\nUp to 22.6 TF<\/span><\/td>\nUp to 45.3 TF<\/span><\/td>\nUp to 181.0 TF<\/span><\/td>\nUp to 181.0 TOPS<\/span><\/td>\n64GB<\/span><\/td>\nUp to 1.6 TB<\/span>\/sec<\/td>\nPCIe\u00ae<\/td>\n<\/tr>\n
AMD<\/span> Instinct MI250<\/span><\/td>\n208<\/td>\n13,312<\/td>\nUp to 45.3 TF<\/span><\/td>\nUp to 90.5 TF<\/span><\/td>\nUp to 362.1 TF<\/span><\/td>\nUp to 362.1 TOPS<\/span><\/td>\n128GB<\/span><\/td>\n3.2 TB<\/span>\/sec<\/td>\nOCP<\/span> Acce\u00adle\u00adra\u00adtor Modu\u00adle (OAM<\/span>)<\/td>\n<\/tr>\n
AMD<\/span> Instinct MI250x<\/td>\n220<\/td>\n14,080<\/td>\nUp to 47.9 TF<\/span><\/td>\nUp to 95.7 TF<\/span><\/td>\nUp To 383.0 TF<\/span><\/td>\nUp to 383.0 TOPS<\/span><\/td>\n128GB<\/span><\/td>\n3.2 TB<\/span>\/sec<\/td>\nOCP<\/span> Acce\u00adle\u00adra\u00adtor Modu\u00adle (OAM<\/span>)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Sup\u00adport\u00ading Resources<\/strong><\/p>\n