New AMD Instinct‚ĄĘ MI200 Series Accelerators Bring Leadership HPC and AI Performance to Power Exascale Systems and More

Novem­ber 08, 2021 12:00pm EST

- With new AMD CDNA‚ĄĘ 2 archi¬≠tec¬≠tu¬≠re, AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors deli¬≠ver ground-brea¬≠king 4.9x advan¬≠ta¬≠ge in HPC per¬≠for¬≠mance1 com¬≠pared to com¬≠pe¬≠ting data cen¬≠ter acce¬≠le¬≠ra¬≠tors, expe¬≠diting sci¬≠ence and discovery -

- MI200 series acce¬≠le¬≠ra¬≠tors are first mul¬≠ti-die GPU, first to sup¬≠port 128GB of HBM2e memo¬≠ry, and deli¬≠ver a sub¬≠stan¬≠ti¬≠al boost for appli¬≠ca¬≠ti¬≠ons cri¬≠ti¬≠cal to the foun¬≠da¬≠ti¬≠on of science -

SANTA CLARA, Calif., Nov. 08, 2021 (GLOBE NEWSWIRE) ‚ÄĒ AMD (NASDAQ: AMD) today announ¬≠ced the new AMD Instinct‚ĄĘ MI200 series acce¬≠le¬≠ra¬≠tors, the first exas¬≠ca¬≠le-class GPU acce¬≠le¬≠ra¬≠tors. AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors includes the world‚Äôs fas¬≠test high per¬≠for¬≠mance com¬≠pu¬≠ting (HPC) and arti¬≠fi¬≠ci¬≠al intel¬≠li¬≠gence (AI) acce¬≠le¬≠ra¬≠tor,1 the AMD Instinct‚ĄĘ MI250X.

Built on AMD CDNA‚ĄĘ 2 archi¬≠tec¬≠tu¬≠re, AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors deli¬≠ver lea¬≠ding appli¬≠ca¬≠ti¬≠on per¬≠for¬≠mance for a broad set of HPC workloads.2 The AMD Instinct MI250X acce¬≠le¬≠ra¬≠tor pro¬≠vi¬≠des up to 4.9X bet¬≠ter per¬≠for¬≠mance than com¬≠pe¬≠ti¬≠ti¬≠ve acce¬≠le¬≠ra¬≠tors for dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64) HPC appli¬≠ca¬≠ti¬≠ons and sur¬≠pas¬≠ses 380 tera¬≠flops of peak theo¬≠re¬≠ti¬≠cal half-pre¬≠cis¬≠i¬≠on (FP16) for AI workloads to enable dis¬≠rup¬≠ti¬≠ve approa¬≠ches in fur¬≠ther acce¬≠le¬≠ra¬≠ting data-dri¬≠ven rese¬≠arch.1

‚ÄúAMD Instinct MI200 acce¬≠le¬≠ra¬≠tors deli¬≠ver lea¬≠der¬≠ship HPC and AI per¬≠for¬≠mance, hel¬≠ping sci¬≠en¬≠tists make gene¬≠ra¬≠tio¬≠nal leaps in rese¬≠arch that can dra¬≠ma¬≠ti¬≠cal¬≠ly shor¬≠ten the time bet¬≠ween initi¬≠al hypo¬≠the¬≠sis and dis¬≠co¬≠very,‚ÄĚ said For¬≠rest Nor¬≠rod, seni¬≠or vice pre¬≠si¬≠dent and gene¬≠ral mana¬≠ger, Data Cen¬≠ter and Embedded Solu¬≠ti¬≠ons Busi¬≠ness Group, AMD. ‚ÄúWith key inno¬≠va¬≠tions in archi¬≠tec¬≠tu¬≠re, pack¬≠a¬≠ging and sys¬≠tem design, the AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors are the most advan¬≠ced data cen¬≠ter GPUs ever, pro¬≠vi¬≠ding excep¬≠tio¬≠nal per¬≠for¬≠mance for super¬≠com¬≠pu¬≠ters and data cen¬≠ters to sol¬≠ve the world‚Äôs most com¬≠plex problems.‚ÄĚ

Exas¬≠ca¬≠le With AMD
AMD, in col¬≠la¬≠bo¬≠ra¬≠ti¬≠on with the U.S. Depart¬≠ment of Ener¬≠gy, Oak Ridge Natio¬≠nal Labo¬≠ra¬≠to¬≠ry, and HPE, desi¬≠gned the Fron¬≠tier super¬≠com¬≠pu¬≠ter expec¬≠ted to deli¬≠ver more than 1.5 exa¬≠flops of peak com¬≠pu¬≠ting power. Powered by opti¬≠mi¬≠zed 3rd Gen AMD EPYC‚ĄĘ CPUs and AMD Instinct MI250X acce¬≠le¬≠ra¬≠tors, Fron¬≠tier will push the boun¬≠da¬≠ries of sci¬≠en¬≠ti¬≠fic dis¬≠co¬≠very by dra¬≠ma¬≠ti¬≠cal¬≠ly enhan¬≠cing per¬≠for¬≠mance of AI, ana¬≠ly¬≠tics, and simu¬≠la¬≠ti¬≠on at sca¬≠le, hel¬≠ping sci¬≠en¬≠tists to pack in more cal¬≠cu¬≠la¬≠ti¬≠ons, iden¬≠ti¬≠fy new pat¬≠terns in data, and deve¬≠lop inno¬≠va¬≠ti¬≠ve data ana¬≠ly¬≠sis methods to acce¬≠le¬≠ra¬≠te the pace of sci¬≠en¬≠ti¬≠fic discovery.

‚ÄúThe Fron¬≠tier super¬≠com¬≠pu¬≠ter is the cul¬≠mi¬≠na¬≠ti¬≠on of a strong col¬≠la¬≠bo¬≠ra¬≠ti¬≠on bet¬≠ween AMD, HPE and the U.S. Depart¬≠ment of Ener¬≠gy, to pro¬≠vi¬≠de an exas¬≠ca¬≠le-capa¬≠ble sys¬≠tem that pushes the boun¬≠da¬≠ries of sci¬≠en¬≠ti¬≠fic dis¬≠co¬≠very by dra¬≠ma¬≠ti¬≠cal¬≠ly enhan¬≠cing per¬≠for¬≠mance of arti¬≠fi¬≠ci¬≠al intel¬≠li¬≠gence, ana¬≠ly¬≠tics, and simu¬≠la¬≠ti¬≠on at sca¬≠le,‚ÄĚ said Tho¬≠mas Zacha¬≠ria, direc¬≠tor, Oak Ridge Natio¬≠nal Laboratory.

Powe¬≠ring The Future of HPC
The AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors, com¬≠bi¬≠ned with 3rd Gen AMD EPYC CPUs and the ROCm‚ĄĘ 5.0 open soft¬≠ware plat¬≠form, are desi¬≠gned to pro¬≠pel new dis¬≠co¬≠veries for the exas¬≠ca¬≠le era and tack¬≠le our most pres¬≠sing chal¬≠lenges from cli¬≠ma¬≠te chan¬≠ge to vac¬≠ci¬≠ne research.

Key capa­bi­li­ties and fea­tures of the AMD Instinct MI200 series acce­le­ra­tors include:

  • AMD CDNA‚ĄĘ 2 archi¬≠tec¬≠tu¬≠re ‚Äď 2nd Gen Matrix Cores acce¬≠le¬≠ra¬≠ting FP64 and FP32 matrix ope¬≠ra¬≠ti¬≠ons, deli¬≠ve¬≠ring up to 4X the peak theo¬≠re¬≠ti¬≠cal FP64 per¬≠for¬≠mance vs. AMD pre¬≠vious gen GPUs. 1,3,4
  • Lea¬≠der¬≠ship Pack¬≠a¬≠ging Tech¬≠no¬≠lo¬≠gy ‚Äď Indus¬≠try-first mul¬≠ti-die GPU design with 2.5D Ele¬≠va¬≠ted Fanout Bridge (EFB) tech¬≠no¬≠lo¬≠gy deli¬≠vers 1.8X more cores and 2.7X hig¬≠her memo¬≠ry band¬≠width vs. AMD pre¬≠vious gen GPUs, offe¬≠ring the industry‚Äôs best aggre¬≠ga¬≠te peak theo¬≠re¬≠ti¬≠cal memo¬≠ry band¬≠width at 3.2 tera¬≠bytes per second. 4,5,6
  • 3rd Gen AMD Infi¬≠ni¬≠ty Fabric‚ĄĘ tech¬≠no¬≠lo¬≠gy ‚Äď Up to 8 Infi¬≠ni¬≠ty Fabric links con¬≠nect the AMD Instinct MI200 with 3rd Gen EPYC CPUs and other GPUs in the node to enable uni¬≠fied CPU/GPU memo¬≠ry cohe¬≠ren¬≠cy and maxi¬≠mi¬≠ze sys¬≠tem through¬≠put, allo¬≠wing for an easier on-ramp for CPU codes to tap the power of accelerators.

Soft­ware for Enab­ling Exas­ca­le Science
AMD ROCm‚ĄĘ is an open soft¬≠ware plat¬≠form allo¬≠wing rese¬≠ar¬≠chers to tap the power of AMD Instinct‚ĄĘ acce¬≠le¬≠ra¬≠tors to dri¬≠ve sci¬≠en¬≠ti¬≠fic dis¬≠co¬≠veries. The ROCm plat¬≠form is built on the foun¬≠da¬≠ti¬≠on of open por¬≠ta¬≠bi¬≠li¬≠ty, sup¬≠port¬≠ing envi¬≠ron¬≠ments across mul¬≠ti¬≠ple acce¬≠le¬≠ra¬≠tor ven¬≠dors and archi¬≠tec¬≠tures. With ROCm 5.0, AMD extends its open plat¬≠form powe¬≠ring top HPC and AI appli¬≠ca¬≠ti¬≠ons with AMD Instinct MI200 series acce¬≠le¬≠ra¬≠tors, incre¬≠asing acces¬≠si¬≠bi¬≠li¬≠ty of ROCm for deve¬≠lo¬≠pers and deli¬≠ve¬≠ring lea¬≠der¬≠ship per¬≠for¬≠mance across key workloads.

Through the AMD Infi¬≠ni¬≠ty Hub, rese¬≠ar¬≠chers, data sci¬≠en¬≠tists and end-users can easi¬≠ly find, down¬≠load and install con¬≠tai¬≠ne¬≠ri¬≠zed HPC apps and ML frame¬≠works that are opti¬≠mi¬≠zed and sup¬≠port¬≠ed on AMD Instinct acce¬≠le¬≠ra¬≠tors and ROCm. The hub curr¬≠ent¬≠ly offers a ran¬≠ge of con¬≠tai¬≠ners sup¬≠port¬≠ing eit¬≠her Rade¬≠on Instinct‚ĄĘ MI50, AMD Instinct‚ĄĘ MI100 or AMD Instinct MI200 acce¬≠le¬≠ra¬≠tors inclu¬≠ding seve¬≠ral appli¬≠ca¬≠ti¬≠ons like Chro¬≠ma, CP2k, LAMMPS, NAMD, OpenMM and more, along with popu¬≠lar ML frame¬≠works Ten¬≠sor¬≠Flow and PyTorch. New con¬≠tai¬≠ners are con¬≠ti¬≠nu¬≠al¬≠ly being added to the hub.

Available Ser­ver Solutions
The AMD Instinct MI250X and AMD Instinct MI250 are available in the open-hard¬≠ware com¬≠pu¬≠te acce¬≠le¬≠ra¬≠tor modu¬≠le or OCP Acce¬≠le¬≠ra¬≠tor Modu¬≠le (OAM) form fac¬≠tor. The AMD Instinct MI210 will be available in a PCIe¬ģ card form fac¬≠tor in OEM servers.

The AMD MI250X acce­le­ra­tor is curr­ent­ly available from HPE in the HPE Cray EX Super­com­pu­ter, and addi­tio­nal AMD Instinct MI200 series acce­le­ra­tors are expec­ted in sys­tems from major OEM and ODM part­ners in enter­pri­se mar­kets in Q1 2022, inclu­ding ASUS, ATOS, Dell Tech­no­lo­gies, Giga­byte, Hew­lett Packard Enter­pri­se (HPE), Leno­vo, Pen­gu­in Com­pu­tin­gand Supermicro.

MI200 Series Specifications

Models Com¬≠pu¬≠te Units Stream Pro¬≠ces¬≠sors FP64 | FP32 Vec¬≠tor (Peak) FP64 | FP32 Matrix (Peak) FP16 | bf16
(Peak)
INT4INT8
(Peak)
HBM2e
ECC
Memory
Memo­ry Bandwidth Form Fac­tor
AMD Instinct MI250x 220 14,080 Up to 47.9 TF Up to 95.7 TF Up to 383.0 TF Up to 383.0 TOPS 128GB 3.2 TB/sec OCP Acce¬≠le¬≠ra¬≠tor Module
AMD Instinct MI250 208 13,312 Up to 45.3 TF Up to 90.5 TF Up to 362.1 TF Up to 362.1 TOPS 128GB 3.2 TB/sec OCP Acce¬≠le¬≠ra¬≠tor Module

Sup­port­ing Resources

About AMD
For more than 50 years AMD has dri¬≠ven inno¬≠va¬≠ti¬≠on in high-per¬≠for¬≠mance com¬≠pu¬≠ting, gra¬≠phics and visua¬≠liza¬≠ti¬≠on tech¬≠no¬≠lo¬≠gies ‚Äē the buil¬≠ding blocks for gam¬≠ing, immersi¬≠ve plat¬≠forms and the data cen¬≠ter. Hundreds of mil¬≠li¬≠ons of con¬≠su¬≠mers, lea¬≠ding For¬≠tu¬≠ne 500 busi¬≠nesses and cut¬≠ting-edge sci¬≠en¬≠ti¬≠fic rese¬≠arch faci¬≠li¬≠ties around the world rely on AMD tech¬≠no¬≠lo¬≠gy dai¬≠ly to impro¬≠ve how they live, work and play. AMD employees around the world are focu¬≠sed on buil¬≠ding gre¬≠at pro¬≠ducts that push the boun¬≠da¬≠ries of what is pos¬≠si¬≠ble. For more infor¬≠ma¬≠ti¬≠on about how AMD is enab¬≠ling today and inspi¬≠ring tomor¬≠row, visit the AMD (NASDAQ: AMDweb¬≠siteFace¬≠bookLin¬≠ke¬≠dIn and Twit¬≠ter pages.

CAUTIONARY STATEMENT
This press release con¬≠ta¬≠ins for¬≠ward-loo¬≠king state¬≠ments con¬≠cer¬≠ning Advan¬≠ced Micro Devices, Inc. (AMD) such as the fea¬≠tures, func¬≠tion¬≠a¬≠li¬≠ty, per¬≠for¬≠mance, avai¬≠la¬≠bi¬≠li¬≠ty, timing and expec¬≠ted bene¬≠fits of AMD pro¬≠ducts inclu¬≠ding the AMD Instinct‚ĄĘ MI200 series acce¬≠le¬≠ra¬≠tors, which are made pur¬≠su¬≠ant to the Safe Har¬≠bor pro¬≠vi¬≠si¬≠ons of the Pri¬≠va¬≠te Secu¬≠ri¬≠ties Liti¬≠ga¬≠ti¬≠on Reform Act of 1995. For¬≠ward-loo¬≠king state¬≠ments are com¬≠mon¬≠ly iden¬≠ti¬≠fied by words such as ‚Äúwould,‚ÄĚ ‚Äúmay,‚ÄĚ ‚Äúexpects,‚ÄĚ ‚Äúbelie¬≠ves,‚ÄĚ ‚Äúplans,‚ÄĚ ‚Äúintends,‚ÄĚ ‚Äúpro¬≠jects‚ÄĚ and other terms with simi¬≠lar mea¬≠ning. Inves¬≠tors are cau¬≠tio¬≠ned that the for¬≠ward-loo¬≠king state¬≠ments in this press release are based on cur¬≠rent beliefs, assump¬≠ti¬≠ons and expec¬≠ta¬≠ti¬≠ons, speak only as of the date of this press release and invol¬≠ve risks and uncer¬≠tain¬≠ties that could cau¬≠se actu¬≠al results to dif¬≠fer mate¬≠ri¬≠al¬≠ly from cur¬≠rent expec¬≠ta¬≠ti¬≠ons. Such state¬≠ments are sub¬≠ject to cer¬≠tain known and unknown risks and uncer¬≠tain¬≠ties, many of which are dif¬≠fi¬≠cult to pre¬≠dict and gene¬≠ral¬≠ly bey¬≠ond AMD‚Äôs con¬≠trol, that could cau¬≠se actu¬≠al results and other future events to dif¬≠fer mate¬≠ri¬≠al¬≠ly from tho¬≠se expres¬≠sed in, or impli¬≠ed or pro¬≠jec¬≠ted by, the for¬≠ward-loo¬≠king infor¬≠ma¬≠ti¬≠on and state¬≠ments. Mate¬≠ri¬≠al fac¬≠tors that could cau¬≠se actu¬≠al results to dif¬≠fer mate¬≠ri¬≠al¬≠ly from cur¬≠rent expec¬≠ta¬≠ti¬≠ons include, wit¬≠hout limi¬≠ta¬≠ti¬≠on, the fol¬≠lo¬≠wing: Intel Corporation‚Äôs domi¬≠nan¬≠ce of the micro¬≠pro¬≠ces¬≠sor mar¬≠ket and its aggres¬≠si¬≠ve busi¬≠ness prac¬≠ti¬≠ces; glo¬≠bal eco¬≠no¬≠mic uncer¬≠tain¬≠ty; loss of a signi¬≠fi¬≠cant cus¬≠to¬≠mer; impact of the COVID-19 pan¬≠de¬≠mic on AMD‚Äôs busi¬≠ness, finan¬≠cial con¬≠di¬≠ti¬≠on and results of ope¬≠ra¬≠ti¬≠ons; com¬≠pe¬≠ti¬≠ti¬≠ve mar¬≠kets in which AMD‚Äôs pro¬≠ducts are sold; mar¬≠ket con¬≠di¬≠ti¬≠ons of the indus¬≠tries in which AMD pro¬≠ducts are sold; cycli¬≠cal natu¬≠re of the semi¬≠con¬≠duc¬≠tor indus¬≠try; quar¬≠ter¬≠ly and sea¬≠so¬≠nal sales pat¬≠terns; AMD‚Äôs abili¬≠ty to ade¬≠qua¬≠te¬≠ly pro¬≠tect its tech¬≠no¬≠lo¬≠gy or other intellec¬≠tu¬≠al pro¬≠per¬≠ty; unfa¬≠vorable cur¬≠ren¬≠cy exch¬≠an¬≠ge rate fluc¬≠tua¬≠tions; abili¬≠ty of third par¬≠ty manu¬≠fac¬≠tu¬≠r¬≠ers to manu¬≠fac¬≠tu¬≠re AMD‚Äôs pro¬≠ducts on a time¬≠ly basis in suf¬≠fi¬≠ci¬≠ent quan¬≠ti¬≠ties and using com¬≠pe¬≠ti¬≠ti¬≠ve tech¬≠no¬≠lo¬≠gies; avai¬≠la¬≠bi¬≠li¬≠ty of essen¬≠ti¬≠al equip¬≠ment, mate¬≠ri¬≠als, sub¬≠stra¬≠tes or manu¬≠fac¬≠tu¬≠ring pro¬≠ces¬≠ses; abili¬≠ty to achie¬≠ve expec¬≠ted manu¬≠fac¬≠tu¬≠ring yields for AMD‚Äôs pro¬≠ducts; AMD‚Äôs abili¬≠ty to intro¬≠du¬≠ce pro¬≠ducts on a time¬≠ly basis with expec¬≠ted fea¬≠tures and per¬≠for¬≠mance levels; AMD‚Äôs abili¬≠ty to gene¬≠ra¬≠te reve¬≠nue from its semi-cus¬≠tom SoC pro¬≠ducts; poten¬≠ti¬≠al secu¬≠ri¬≠ty vul¬≠nerabi¬≠li¬≠ties; poten¬≠ti¬≠al secu¬≠ri¬≠ty inci¬≠dents inclu¬≠ding IT outa¬≠ges, data loss, data brea¬≠ches and cyber-attacks; uncer¬≠tain¬≠ties invol¬≠ving the orde¬≠ring and ship¬≠ment of AMD‚Äôs pro¬≠ducts; AMD‚Äôs reli¬≠ance on third-par¬≠ty intellec¬≠tu¬≠al pro¬≠per¬≠ty to design and intro¬≠du¬≠ce new pro¬≠ducts in a time¬≠ly man¬≠ner; AMD‚Äôs reli¬≠ance on third-par¬≠ty com¬≠pa¬≠nies for design, manu¬≠fac¬≠tu¬≠re and sup¬≠p¬≠ly of mother¬≠boards, soft¬≠ware and other com¬≠pu¬≠ter plat¬≠form com¬≠pon¬≠ents; AMD‚Äôs reli¬≠ance on Micro¬≠soft and other soft¬≠ware ven¬≠dors‚Äô sup¬≠port to design and deve¬≠lop soft¬≠ware to run on AMD‚Äôs pro¬≠ducts; AMD‚Äôs reli¬≠ance on third-par¬≠ty dis¬≠tri¬≠bu¬≠tors and add-in-board part¬≠ners; impact of modi¬≠fi¬≠ca¬≠ti¬≠on or inter¬≠rup¬≠ti¬≠on of AMD‚Äôs inter¬≠nal busi¬≠ness pro¬≠ces¬≠ses and infor¬≠ma¬≠ti¬≠on sys¬≠tems; com¬≠pa¬≠ti¬≠bi¬≠li¬≠ty of AMD‚Äôs pro¬≠ducts with some or all indus¬≠try-stan¬≠dard soft¬≠ware and hard¬≠ware; cos¬≠ts rela¬≠ted to defec¬≠ti¬≠ve pro¬≠ducts; effi¬≠ci¬≠en¬≠cy of AMD‚Äôs sup¬≠p¬≠ly chain; AMD‚Äôs abili¬≠ty to rely on third par¬≠ty sup¬≠p¬≠ly-chain logi¬≠stics func¬≠tions; AMD‚Äôs abili¬≠ty to effec¬≠tively con¬≠trol sales of its pro¬≠ducts on the gray mar¬≠ket; impact of govern¬≠ment actions and regu¬≠la¬≠ti¬≠ons such as export admi¬≠nis¬≠tra¬≠ti¬≠on regu¬≠la¬≠ti¬≠ons, tariffs and trade pro¬≠tec¬≠tion mea¬≠su¬≠res; AMD‚Äôs abili¬≠ty to rea¬≠li¬≠ze its defer¬≠red tax assets; poten¬≠ti¬≠al tax lia¬≠bi¬≠li¬≠ties; cur¬≠rent and future claims and liti¬≠ga¬≠ti¬≠on; impact of envi¬≠ron¬≠men¬≠tal laws, con¬≠flict mine¬≠rals-rela¬≠ted pro¬≠vi¬≠si¬≠ons and other laws or regu¬≠la¬≠ti¬≠ons; impact of acqui¬≠si¬≠ti¬≠ons, joint ven¬≠tures and/or invest¬≠ments on AMD‚Äôs busi¬≠ness, inclu¬≠ding the announ¬≠ced acqui¬≠si¬≠ti¬≠on of Xilinx, and abili¬≠ty to inte¬≠gra¬≠te acqui¬≠red busi¬≠nesses; AMD‚Äôs abili¬≠ty to com¬≠ple¬≠te the Xilinx mer¬≠ger; impact of the announce¬≠ment and pen¬≠den¬≠cy of the Xilinx mer¬≠ger on AMD‚Äôs busi¬≠ness; impact of any impair¬≠ment of the com¬≠bi¬≠ned company‚Äôs assets on the com¬≠bi¬≠ned company‚Äôs finan¬≠cial posi¬≠ti¬≠on and results of ope¬≠ra¬≠ti¬≠on; rest¬≠ric¬≠tions impo¬≠sed by agree¬≠ments gover¬≠ning AMD‚Äôs notes and the revol¬≠ving cre¬≠dit faci¬≠li¬≠ty; AMD‚Äôs indeb¬≠ted¬≠ness; AMD‚Äôs abili¬≠ty to gene¬≠ra¬≠te suf¬≠fi¬≠ci¬≠ent cash to meet its working capi¬≠tal requi¬≠re¬≠ments or gene¬≠ra¬≠te suf¬≠fi¬≠ci¬≠ent reve¬≠nue and ope¬≠ra¬≠ting cash flow to make all of its plan¬≠ned R&D or stra¬≠te¬≠gic invest¬≠ments; poli¬≠ti¬≠cal, legal, eco¬≠no¬≠mic risks and natu¬≠ral dis¬≠as¬≠ters; future impairm¬≠ents of good¬≠will and tech¬≠no¬≠lo¬≠gy licen¬≠se purcha¬≠ses; AMD‚Äôs abili¬≠ty to attract and retain qua¬≠li¬≠fied per¬≠son¬≠nel; AMD‚Äôs stock pri¬≠ce vola¬≠ti¬≠li¬≠ty; and world¬≠wi¬≠de poli¬≠ti¬≠cal con¬≠di¬≠ti¬≠ons. Inves¬≠tors are urged to review in detail the risks and uncer¬≠tain¬≠ties in AMD‚Äôs Secu¬≠ri¬≠ties and Exch¬≠an¬≠ge Com¬≠mis¬≠si¬≠on filings, inclu¬≠ding but not limi¬≠t¬≠ed to AMD‚Äôs most recent reports on Forms 10‚ÄĎK and 10‚ÄĎQ.

©2021 Advan­ced Micro Devices, Inc. All rights reser­ved. AMD, the AMD Arrow logo, AMD CDNA, EPYC, AMD Instinct, Infi­ni­ty Fabric, Rade­on Instinct, ROCm and com­bi­na­ti­ons the­reof are trade­marks of Advan­ced Micro Devices, Inc. PyTorch is a trade­mark or regis­tered trade­mark of PyTorch. Ten­sor­Flow, the Ten­sor­Flow logo and any rela­ted marks are trade­marks of Goog­le Inc. Other pro­duct names used in this publi­ca­ti­on are for iden­ti­fi­ca­ti­on pur­po­ses only and may be trade­marks of their respec­ti­ve companies.

Addi¬≠tio¬≠nal bench¬≠mark data is available on AMD.com

  1. World‚Äôs fas¬≠test data cen¬≠ter GPU is the AMD Instinct‚ĄĘ MI250X. Cal¬≠cu¬≠la¬≠ti¬≠ons con¬≠duc¬≠ted by AMD Per¬≠for¬≠mance Labs as of Sep 15, 2021, for the AMD Instinct‚ĄĘ MI250X (128GB HBM2e OAM modu¬≠le) acce¬≠le¬≠ra¬≠tor at 1,700 MHz peak boost engi¬≠ne clock resul¬≠ted in 95.7 TFLOPS peak theo¬≠re¬≠ti¬≠cal dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64 Matrix), 47.9 TFLOPS peak theo¬≠re¬≠ti¬≠cal dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64), 95.7 TFLOPS peak theo¬≠re¬≠ti¬≠cal sin¬≠gle pre¬≠cis¬≠i¬≠on matrix (FP32 Matrix), 47.9 TFLOPS peak theo¬≠re¬≠ti¬≠cal sin¬≠gle pre¬≠cis¬≠i¬≠on (FP32), 383.0 TFLOPS peak theo¬≠re¬≠ti¬≠cal half pre¬≠cis¬≠i¬≠on (FP16), and 383.0 TFLOPS peak theo¬≠re¬≠ti¬≠cal Bfloat16 for¬≠mat pre¬≠cis¬≠i¬≠on (BF16) floa¬≠ting-point per¬≠for¬≠mance. Cal¬≠cu¬≠la¬≠ti¬≠ons con¬≠duc¬≠ted by AMD Per¬≠for¬≠mance Labs as of Sep 18, 2020 for the AMD Instinct‚ĄĘ MI100 (32GB HBM2 PCIe¬ģ card) acce¬≠le¬≠ra¬≠tor at 1,502 MHz peak boost engi¬≠ne clock resul¬≠ted in 11.54 TFLOPS peak theo¬≠re¬≠ti¬≠cal dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64), 46.1 TFLOPS peak theo¬≠re¬≠ti¬≠cal sin¬≠gle pre¬≠cis¬≠i¬≠on matrix (FP32), 23.1 TFLOPS peak theo¬≠re¬≠ti¬≠cal sin¬≠gle pre¬≠cis¬≠i¬≠on (FP32), 184.6 TFLOPS peak theo¬≠re¬≠ti¬≠cal half pre¬≠cis¬≠i¬≠on (FP16) floa¬≠ting-point per¬≠for¬≠mance. Published results on the NVi¬≠dia Ampere A100 (80GB) GPU acce¬≠le¬≠ra¬≠tor, boost engi¬≠ne clock of 1410 MHz, resul¬≠ted in 19.5 TFLOPS peak dou¬≠ble pre¬≠cis¬≠i¬≠on ten¬≠sor cores (FP64 Ten¬≠sor Core), 9.7 TFLOPS peak dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64). 19.5 TFLOPS peak sin¬≠gle pre¬≠cis¬≠i¬≠on (FP32), 78 TFLOPS peak half pre¬≠cis¬≠i¬≠on (FP16), 312 TFLOPS peak half pre¬≠cis¬≠i¬≠on (FP16 Ten¬≠sor Flow), 39 TFLOPS peak Bfloat 16 (BF16), 312 TFLOPS peak Bfloat16 for¬≠mat pre¬≠cis¬≠i¬≠on (BF16 Ten¬≠sor Flow), theo¬≠re¬≠ti¬≠cal floa¬≠ting-point per¬≠for¬≠mance. The TF32 data for¬≠mat is not IEEE com¬≠pli¬≠ant and not included in this com¬≠pa¬≠ri¬≠son. https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-ampere-architecture-whitepaper.pdf, page 15, Table 1. MI200-01
  2. AMD Instinct MI250X acce­le­ra­tor appli­ca­ti­on and bench­mark per­for­mance can be found at https://www.amd.com/en/graphics/server-accelerators-benchmarks.
  3. Cal¬≠cu¬≠la¬≠ti¬≠ons con¬≠duc¬≠ted by AMD Per¬≠for¬≠mance Labs as of Sep 15, 2021, for the AMD Instinct‚ĄĘ MI250X acce¬≠le¬≠ra¬≠tor (128GB HBM2e OAM modu¬≠le) at 1,700 MHz peak boost engi¬≠ne clock resul¬≠ted in 95.7 TFLOPS peak dou¬≠ble pre¬≠cis¬≠i¬≠on matrix (FP64 Matrix) theo¬≠re¬≠ti¬≠cal, floa¬≠ting-point per¬≠for¬≠mance. Published results on the NVi¬≠dia Ampere A100 (80GB) GPU acce¬≠le¬≠ra¬≠tor resul¬≠ted in 19.5 TFLOPS peak dou¬≠ble pre¬≠cis¬≠i¬≠on (FP64 Ten¬≠sor Core) theo¬≠re¬≠ti¬≠cal, floa¬≠ting-point per¬≠for¬≠mance. Results found at:https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-ampere-architecture-whitepaper.pdf, page 15, Table 1.MI200-02
  4. Cal¬≠cu¬≠la¬≠ti¬≠ons con¬≠duc¬≠ted by AMD Per¬≠for¬≠mance Labs as of Sep 21, 2021, for the AMD Instinct‚ĄĘ MI250X and MI250 (128GB HBM2e) OAM acce¬≠le¬≠ra¬≠tors desi¬≠gned with AMD CDNA‚ĄĘ 2 6nm Fin¬≠Fet pro¬≠cess tech¬≠no¬≠lo¬≠gy at 1,600 MHz peak memo¬≠ry clock resul¬≠ted in 128GB HBM2e memo¬≠ry capa¬≠ci¬≠ty and 3.2768 TFLOPS peak theo¬≠re¬≠ti¬≠cal memo¬≠ry band¬≠width per¬≠for¬≠mance. MI250/MI250X memo¬≠ry bus inter¬≠face is 4,096 bits times 2 die and memo¬≠ry data rate is 3.20 Gbps for total memo¬≠ry band¬≠width of 3.2768 TB/s ((3.20 Gbps*(4,096 bits*2))/8).The hig¬≠hest published results on the NVi¬≠dia Ampere A100 (80GB) SXM GPU acce¬≠le¬≠ra¬≠tor resul¬≠ted in 80GB HBM2e memo¬≠ry capa¬≠ci¬≠ty and 2.039 TB/s GPU memo¬≠ry band¬≠width performance.https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/nvidia-a100-datasheet-us-nvidia-1758950-r4-web.pdf MI200-07
  5. The AMD Instinct‚ĄĘ MI250X acce¬≠le¬≠ra¬≠tor has 220 com¬≠pu¬≠te units (CUs) and 14,080 stream cores. The AMD Instinct‚ĄĘ MI100 acce¬≠le¬≠ra¬≠tor has 120 com¬≠pu¬≠te units (CUs) and 7,680 stream cores. MI200-027
  6. Cal¬≠cu¬≠la¬≠ti¬≠ons con¬≠duc¬≠ted by AMD Per¬≠for¬≠mance Labs as of Sep 21, 2021, for the AMD Instinct‚ĄĘ MI250X and MI250 (128GB HBM2e) OAM acce¬≠le¬≠ra¬≠tors desi¬≠gned with AMD CDNA‚ĄĘ 2 6nm Fin¬≠Fet pro¬≠cess tech¬≠no¬≠lo¬≠gy at 1,600 MHz peak memo¬≠ry clock resul¬≠ted in 3.2768 TFLOPS peak theo¬≠re¬≠ti¬≠cal memo¬≠ry band¬≠width per¬≠for¬≠mance. MI250/MI250X memo¬≠ry bus inter¬≠face is 4,096 bits times 2 die and memo¬≠ry data rate is 3.20 Gbps for total memo¬≠ry band¬≠width of 3.2768 TB/s ((3.20 Gbps*(4,096 bits*2))/8). Cal¬≠cu¬≠la¬≠ti¬≠ons by AMD Per¬≠for¬≠mance Labs as of OCT 5th, 2020 for the AMD Instinct‚ĄĘ MI100 acce¬≠le¬≠ra¬≠tor desi¬≠gned with AMD CDNA 7nm Fin¬≠FET pro¬≠cess tech¬≠no¬≠lo¬≠gy at 1,200 MHz peak memo¬≠ry clock resul¬≠ted in 1.2288 TFLOPS peak theo¬≠re¬≠ti¬≠cal memo¬≠ry band¬≠width per¬≠for¬≠mance. MI100 memo¬≠ry bus inter¬≠face is 4,096 bits and memo¬≠ry data rate is 2.40 Gbps for total memo¬≠ry band¬≠width of 1.2288 TB/s ((2.40 Gbps*4,096 bits)/8) MI200-33