AMD Unveils World’s First 7nm Datacenter GPUs

‚Äď Powe¬≠ring the Next Era of Arti¬≠fi¬≠ci¬≠al Intel¬≠li¬≠gence, Cloud Com¬≠pu¬≠ting and High Per¬≠for¬≠mance Com¬≠pu¬≠ting (HPC)

‚ÄĒ AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors with super¬≠char¬≠ged com¬≠pu¬≠te per¬≠for¬≠mance, high-speed con¬≠nec¬≠ti¬≠vi¬≠ty, fast memo¬≠ry band¬≠width and updated ROCm open soft¬≠ware plat¬≠form power the most deman¬≠ding deep lear¬≠ning, HPC, cloud and ren¬≠de¬≠ring appli¬≠ca¬≠ti¬≠ons ‚ÄĒ

San Fran¬≠cis¬≠co, Calif. 

AMD (NASDAQ: AMD) today announ¬≠ced the AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors, the world‚Äôs first 7nm dat¬≠a¬≠cen¬≠ter GPUs, desi¬≠gned to deli¬≠ver the com¬≠pu¬≠te per¬≠for¬≠mance requi¬≠red for next-gene¬≠ra¬≠ti¬≠on deep lear¬≠ning, HPC, cloud com¬≠pu¬≠ting and ren¬≠de¬≠ring appli¬≠ca¬≠ti¬≠ons. Rese¬≠ar¬≠chers, sci¬≠en¬≠tists and deve¬≠lo¬≠pers will use AMD Rade¬≠on Instinct‚ĄĘ acce¬≠le¬≠ra¬≠tors to sol¬≠ve tough and inte¬≠res¬≠t¬≠ing chal¬≠lenges, inclu¬≠ding lar¬≠ge-sca¬≠le simu¬≠la¬≠ti¬≠ons, cli¬≠ma¬≠te chan¬≠ge, com¬≠pu¬≠ta¬≠tio¬≠nal bio¬≠lo¬≠gy, dise¬≠a¬≠se pre¬≠ven¬≠ti¬≠on and more.

‚ÄúLega¬≠cy GPU archi¬≠tec¬≠tures limit IT mana¬≠gers from effec¬≠tively addres¬≠sing the con¬≠stant¬≠ly evol¬≠ving demands of pro¬≠ces¬≠sing and ana¬≠ly¬≠zing huge data¬≠sets for modern cloud dat¬≠a¬≠cen¬≠ter workloads,‚ÄĚ said David Wang, seni¬≠or vice pre¬≠si¬≠dent of engi¬≠nee¬≠ring, Rade¬≠on Tech¬≠no¬≠lo¬≠gies Group at AMD. ‚ÄúCom¬≠bi¬≠ning world-class per¬≠for¬≠mance and a fle¬≠xi¬≠ble archi¬≠tec¬≠tu¬≠re with a robust soft¬≠ware plat¬≠form and the industry‚Äôs lea¬≠ding-edge ROCm open soft¬≠ware eco¬≠sys¬≠tem, the new AMD Rade¬≠on Instinct‚ĄĘ acce¬≠le¬≠ra¬≠tors pro¬≠vi¬≠de the cri¬≠ti¬≠cal com¬≠pon¬≠ents nee¬≠ded to sol¬≠ve the most dif¬≠fi¬≠cult cloud com¬≠pu¬≠ting chal¬≠lenges today and into the future.‚ÄĚ

The AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors fea¬≠ture fle¬≠xi¬≠ble mixed-pre¬≠cis¬≠i¬≠on capa¬≠bi¬≠li¬≠ties, powered by high-per¬≠for¬≠mance com¬≠pu¬≠te units that expand the types of workloads the¬≠se acce¬≠le¬≠ra¬≠tors can address, inclu¬≠ding a ran¬≠ge of HPC and deep lear¬≠ning appli¬≠ca¬≠ti¬≠ons. The new AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors were desi¬≠gned to effi¬≠ci¬≠ent¬≠ly pro¬≠cess workloads such as rapidly trai¬≠ning com¬≠plex neu¬≠ral net¬≠works, deli¬≠ve¬≠ring hig¬≠her levels of floa¬≠ting-point per¬≠for¬≠mance, grea¬≠ter effi¬≠ci¬≠en¬≠ci¬≠es and new fea¬≠tures for dat¬≠a¬≠cen¬≠ter and depart¬≠ment¬≠al deploy¬≠ments1.

The AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors pro¬≠vi¬≠de ultra-fast floa¬≠ting-point per¬≠for¬≠man¬≠ceand hyper-fast HBM2 (second-gene¬≠ra¬≠ti¬≠on High-Band¬≠width Memo¬≠ry) with up to 1 TB/s memo¬≠ry band¬≠width speeds. They are also the first GPUs capa¬≠ble of sup¬≠port¬≠ing next-gene¬≠ra¬≠ti¬≠on PCIe¬ģ 4.02 inter¬≠con¬≠nect, which is up to 2X fas¬≠ter than other x86 CPU-to-GPU inter¬≠con¬≠nect tech¬≠no¬≠lo¬≠gies3, and fea¬≠ture AMD Infi¬≠ni¬≠ty Fabric‚ĄĘ Link GPU inter¬≠con¬≠nect tech¬≠no¬≠lo¬≠gy that enables GPU-to-GPU com¬≠mu¬≠ni¬≠ca¬≠ti¬≠ons that are up to 6X fas¬≠ter than PCIe¬ģ Gen 3 inter¬≠con¬≠nect speeds4.

AMD also announ¬≠ced a new ver¬≠si¬≠on of the ROCm open soft¬≠ware plat¬≠form for acce¬≠le¬≠ra¬≠ted com¬≠pu¬≠ting that sup¬≠ports the archi¬≠tec¬≠tu¬≠ral fea¬≠tures of the new acce¬≠le¬≠ra¬≠tors, inclu¬≠ding opti¬≠mi¬≠zed deep lear¬≠ning ope¬≠ra¬≠ti¬≠ons (DLOPS) and the AMD Infi¬≠ni¬≠ty Fabric‚ĄĘ Link GPU inter¬≠con¬≠nect tech¬≠no¬≠lo¬≠gy. Desi¬≠gned for sca¬≠le, ROCm allows cus¬≠to¬≠mers to deploy high-per¬≠for¬≠mance, ener¬≠gy-effi¬≠ci¬≠ent hete¬≠ro¬≠ge¬≠neous com¬≠pu¬≠ting sys¬≠tems in an open environment.

‚ÄúGoog¬≠le belie¬≠ves that open source is good for ever¬≠yo¬≠ne,‚ÄĚ said Rajat Mon¬≠ga, engi¬≠nee¬≠ring direc¬≠tor, Ten¬≠sor¬≠Flow, Goog¬≠le. ‚ÄúWe‚Äô¬≠ve seen how hel¬≠pful it can be to open source machi¬≠ne lear¬≠ning tech¬≠no¬≠lo¬≠gy, and we‚Äôre glad to see AMD embra¬≠cing it. With the ROCm open soft¬≠ware plat¬≠form, Ten¬≠sor¬≠Flow users will bene¬≠fit from GPU acce¬≠le¬≠ra¬≠ti¬≠on and a more robust open source machi¬≠ne lear¬≠ning ecosystem.‚ÄĚ

Key fea¬≠tures of the AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 acce¬≠le¬≠ra¬≠tors include:

  • Opti¬≠mi¬≠zed Deep Lear¬≠ning Ope¬≠ra¬≠ti¬≠ons: Pro¬≠vi¬≠des fle¬≠xi¬≠ble mixed-pre¬≠cis¬≠i¬≠on FP16, FP32 and INT4/INT8 capa¬≠bi¬≠li¬≠ties to meet gro¬≠wing demand for dyna¬≠mic and ever-chan¬≠ging workloads, from trai¬≠ning com¬≠plex neu¬≠ral net¬≠works to run¬≠ning infe¬≠rence against tho¬≠se trai¬≠ned networks.
  • World‚Äôs Fas¬≠test Dou¬≠ble Pre¬≠cis¬≠i¬≠on PCIe¬ģ2 Acce¬≠le¬≠ra¬≠tor5The AMD Rade¬≠on Instinct‚ĄĘ MI60 is the world‚Äôs fas¬≠test dou¬≠ble pre¬≠cis¬≠i¬≠on PCIe 4.0 capa¬≠ble acce¬≠le¬≠ra¬≠tor, deli¬≠ve¬≠ring up to 7.4 TFLOPS peak FP64 per¬≠for¬≠mance5 allo¬≠wing sci¬≠en¬≠tists and rese¬≠ar¬≠chers to more effi¬≠ci¬≠ent¬≠ly pro¬≠cess HPC appli¬≠ca¬≠ti¬≠ons across a ran¬≠ge of indus¬≠tries inclu¬≠ding life sci¬≠en¬≠ces, ener¬≠gy, finan¬≠ce, auto¬≠mo¬≠ti¬≠ve, aero¬≠space, aca¬≠de¬≠mics, govern¬≠ment, defen¬≠se and more. The AMD Rade¬≠on Instinct‚ĄĘ MI50 deli¬≠vers up to 6.7 TFLOPS FP64 peak per¬≠for¬≠mance1, while pro¬≠vi¬≠ding an effi¬≠ci¬≠ent, cost-effec¬≠ti¬≠ve solu¬≠ti¬≠on for a varie¬≠ty of deep lear¬≠ning workloads, as well as enab¬≠ling high reu¬≠se in Vir¬≠tu¬≠al Desk¬≠top Infra¬≠struc¬≠tu¬≠re (VDI), Desk¬≠top-as-a-Ser¬≠vice (DaaS) and cloud environments.
  • Up to 6X Fas¬≠ter Data Trans¬≠fer: Two Infi¬≠ni¬≠ty Fabric‚ĄĘ Links per GPU deli¬≠ver up to 200 GB/s of peer-to-peer band¬≠width ‚Äď up to 6X fas¬≠ter than PCIe 3.0 alo¬≠ne4 ‚Äď and enable the con¬≠nec¬≠tion of up to 4 GPUs in a hive ring con¬≠fi¬≠gu¬≠ra¬≠ti¬≠on (2 hives in 8 GPU servers).
  • Ultra-Fast HBM2 Memo¬≠ry: The AMD Rade¬≠on Instinct‚ĄĘ MI60 pro¬≠vi¬≠des 32GB of HBM2 Error-cor¬≠rec¬≠ting code (ECC) memo¬≠ry6, and the Rade¬≠on Instinct‚ĄĘ MI50 pro¬≠vi¬≠des 16GB of HBM2 ECC memo¬≠ry. Both GPUs pro¬≠vi¬≠de full-chip ECC and Relia¬≠bi¬≠li¬≠ty, Acces¬≠si¬≠bi¬≠li¬≠ty and Ser¬≠vicea¬≠bi¬≠li¬≠ty (RAS)7 tech¬≠no¬≠lo¬≠gies, which are cri¬≠ti¬≠cal to deli¬≠ver more accu¬≠ra¬≠te com¬≠pu¬≠te results for lar¬≠ge-sca¬≠le HPC deployments.
  • Secu¬≠re Vir¬≠tua¬≠li¬≠zed Workload Sup¬≠port: AMD MxG¬≠PU Tech¬≠no¬≠lo¬≠gy, the industry‚Äôs only hard¬≠ware-based GPU vir¬≠tua¬≠liza¬≠ti¬≠on solu¬≠ti¬≠on, which is based on the indus¬≠try-stan¬≠dard SR-IOV (Sin¬≠gle Root I/O Vir¬≠tua¬≠liza¬≠ti¬≠on) tech¬≠no¬≠lo¬≠gy, makes it dif¬≠fi¬≠cult for hackers to attack at the hard¬≠ware level, hel¬≠ping pro¬≠vi¬≠de secu¬≠ri¬≠ty for vir¬≠tua¬≠li¬≠zed cloud deployments.

Updated ROCm Open Software Platform

AMD today also announ¬≠ced a new ver¬≠si¬≠on of its ROCm open soft¬≠ware plat¬≠form desi¬≠gned to speed deve¬≠lo¬≠p¬≠ment of high-per¬≠for¬≠mance, ener¬≠gy-effi¬≠ci¬≠ent hete¬≠ro¬≠ge¬≠neous com¬≠pu¬≠ting sys¬≠tems. In addi¬≠ti¬≠on to sup¬≠port for the new Rade¬≠on Instinct‚ĄĘ acce¬≠le¬≠ra¬≠tors, ROCm soft¬≠ware ver¬≠si¬≠on 2.0 pro¬≠vi¬≠des updated math libra¬≠ri¬≠es for the new DLOPS; sup¬≠port for 64-bit Linux ope¬≠ra¬≠ting sys¬≠tems inclu¬≠ding Cent¬≠OS, RHEL and Ubun¬≠tu; opti¬≠miza¬≠ti¬≠ons of exis¬≠ting com¬≠pon¬≠ents; and sup¬≠port for the latest ver¬≠si¬≠ons of the most popu¬≠lar deep lear¬≠ning frame¬≠works, inclu¬≠ding Ten¬≠sor¬≠Flow 1.11, PyTorch (Caffe2) and others. Learn more about ROCm 2.0 soft¬≠ware here.


The AMD Rade¬≠on Instinct‚ĄĘ MI60 acce¬≠le¬≠ra¬≠tor is expec¬≠ted to ship to dat¬≠a¬≠cen¬≠ter cus¬≠to¬≠mers by the end of 2018. The AMD Rade¬≠on Instinct‚ĄĘ MI50 acce¬≠le¬≠ra¬≠tor is expec¬≠ted to begin ship¬≠ping to data cen¬≠ter cus¬≠to¬≠mers by the end of Q1 2019. The ROCm 2.0 open soft¬≠ware plat¬≠form is expec¬≠ted to be available by the end of 2018.

Supporting Resources

  • Visit the AMD Next Hori¬≠zon event web¬≠page to get the event materials
  • Learn more about AMD Rade¬≠on Instinct‚ĄĘ MI60 and MI50 accelerators
  • Learn more about AMD 7nm tech¬≠no¬≠lo¬≠gy here
  • Learn more about the ROCm 2.0 open soft¬≠ware plat¬≠form here
  • Learn more about ROCm & MIO¬≠pen Docker Hub here
  • Beco¬≠me a fan of AMD on Face¬≠book
  • Fol¬≠low AMD Rade¬≠on Instinct on Twit¬≠ter

About AMD

For more than 45 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 dat¬≠a¬≠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¬≠siteblogFace¬≠book and Twit¬≠ter pages.