In die­ser Sek­ti­on wer­den Pres­se­mit­tei­lun­gen der Her­stel­ler 1:1 ver­öf­fent­licht. Pla­net 3DNow! über­nimmt weder für die Rich­tig­keit der Anga­ben noch für die Art der For­mu­lie­rung jed­we­de Ver­ant­wor­tung. Die Anga­ben kön­nen sub­jek­tiv sein und geben nicht not­wen­di­ger­wei­se die Mei­nung der auf Pla­net 3DNow! schrei­ben­den Autoren wieder.

AMD Announces Ambitious Goal to Increase Energy Efficiency of Processors Running AI Training and High Performance Computing Applications 30x by 2025

Verfasst vonpipin

— High-per­for­mance AMD EPYC™ CPUs and AMD Instinct™ acce­le­ra­tors tar­get deli­ve­ring unpre­ce­den­ted advan­ce in ener­gy effi­ci­en­cy for Arti­fi­ci­al Intel­li­gence trai­ning and Super­com­pu­ting applications —

SANTA CLARA, Calif. 09/29/2021

AMD (NASDAQ: AMD) today announ­ced a goal to deli­ver a 30x increase in ener­gy effi­ci­en­cy for AMD EPYC CPUs and AMD Instinct acce­le­ra­tors in Arti­fi­ci­al Intel­li­gence (AI) trai­ning and High Per­for­mance Com­pu­ting (HPC) appli­ca­ti­ons run­ning on acce­le­ra­ted com­pu­te nodes by 2025.1 Accom­pli­shing this ambi­tious goal will requi­re AMD to increase the ener­gy effi­ci­en­cy of a com­pu­te node at a rate that is more than 2.5x fas­ter than the aggre­ga­te indus­try-wide impro­ve­ment made during the last five years.2

Acce­le­ra­ted com­pu­te nodes are the most powerful and advan­ced com­pu­ting sys­tems in the world used for sci­en­ti­fic rese­arch and lar­ge-sca­le super­com­pu­ter simu­la­ti­ons. They pro­vi­de the com­pu­ting capa­bi­li­ty used by sci­en­tists to achie­ve breakth­roughs across many fields inclu­ding mate­ri­al sci­en­ces, cli­ma­te pre­dic­tions, geno­mics, drug dis­co­very and alter­na­ti­ve ener­gy. Acce­le­ra­ted nodes are also inte­gral for trai­ning AI neu­ral net­works that are curr­ent­ly used for acti­vi­ties inclu­ding speech reco­gni­ti­on, lan­guage trans­la­ti­on and expert recom­men­da­ti­on sys­tems, with simi­lar pro­mi­sing uses over the coming deca­de. The 30x goal would save bil­li­ons of kilo­watt hours of elec­tri­ci­ty in 2025, redu­cing the power requi­red for the­se sys­tems to com­ple­te a sin­gle cal­cu­la­ti­on by 97% over five years. 

“Achie­ving gains in pro­ces­sor ener­gy effi­ci­en­cy is a long-term design prio­ri­ty for AMD and we are now set­ting a new goal for modern com­pu­te nodes using our high-per­for­mance CPUs and acce­le­ra­tors when appli­ed to AI trai­ning and high-per­for­mance com­pu­ting deploy­ments,” said Mark Paper­mas­ter, exe­cu­ti­ve vice pre­si­dent and CTO, AMD. “Focu­sed on the­se very important seg­ments and the value pro­po­si­ti­on for lea­ding com­pa­nies to enhan­ce their envi­ron­men­tal ste­ward­ship, AMD’s 30x goal out­paces indus­try ener­gy effi­ci­en­cy per­for­mance in the­se are­as by 150% com­pared to the pre­vious five-year time period.”

“With com­pu­ting beco­ming ubi­qui­tous from edge to core to cloud, AMD has taken a bold posi­ti­on on the ener­gy effi­ci­en­cy of its pro­ces­sors, this time for the acce­le­ra­ted com­pu­te for AI and High Per­for­mance Com­pu­ting appli­ca­ti­ons,” said Addi­son Snell, CEO of Intersect360 Rese­arch. “Future gains are more dif­fi­cult now as the his­to­ri­cal advan­ta­ges that come with Moore’s Law have great­ly dimi­nis­hed. A 30-times impro­ve­ment in ener­gy effi­ci­en­cy in five years will be an impres­si­ve tech­ni­cal achie­ve­ment that will demons­tra­te the strength of AMD tech­no­lo­gy and their empha­sis on envi­ron­men­tal sustainability.” 

Increased ener­gy effi­ci­en­cy for acce­le­ra­ted com­pu­ting appli­ca­ti­ons is part of the company’s new goals in Envi­ron­men­tal, Social, Gover­nan­ce (ESG) span­ning its ope­ra­ti­ons, sup­p­ly chain and pro­ducts. For more than twen­ty-five years, AMD has been trans­par­ent­ly report­ing on its envi­ron­men­tal ste­ward­ship and per­for­mance. For its recent achie­ve­ments in pro­duct ener­gy effi­ci­en­cy, AMD was named to For­tu­ne’s Chan­ge the World list in 2020 that reco­gni­zes out­stan­ding efforts by com­pa­nies to tack­le society’s unmet needs. 

Methodology

In addi­ti­on to com­pu­te node performance/Watt mea­su­re­ments1, to make the goal par­ti­cu­lar­ly rele­vant to world­wi­de ener­gy use, AMD uses seg­ment-spe­ci­fic dat­a­cen­ter power uti­liza­ti­on effec­ti­ve­ness (PUE) with equip­ment uti­liza­ti­on taken into account.3 The ener­gy con­sump­ti­on base­line uses the same indus­try ener­gy per ope­ra­ti­on impro­ve­ment rates as from 2015–2020, extra­po­la­ted to 2025. The mea­su­re of ener­gy per ope­ra­ti­on impro­ve­ment in each seg­ment from 2020–2025 is weigh­ted by the pro­jec­ted world­wi­de volu­mes4 mul­ti­pli­ed by the Typi­cal Ener­gy Con­sump­ti­on (TEC) of each com­pu­ting seg­ment to arri­ve at a meaningful metric of actu­al ener­gy usa­ge impro­ve­ment worldwide.

Dr. Jona­than Koomey, Pre­si­dent, Koomey Ana­ly­tics, said “The ener­gy effi­ci­en­cy goal set by AMD for acce­le­ra­ted com­pu­te nodes used for AI trai­ning and High Per­for­mance Com­pu­ting ful­ly reflects modern workloads, repre­sen­ta­ti­ve ope­ra­ting beha­vi­ors and accu­ra­te bench­mar­king methodology.”

Supporting Resources

About AMD

For more than 50 years AMD has dri­ven inno­va­ti­on in high-per­for­mance com­pu­ting, gra­phics and visua­liza­ti­on tech­no­lo­gies ― the buil­ding blocks for gam­ing, immersi­ve plat­forms and the dat­a­cen­ter. Hundreds of mil­li­ons of con­su­mers, lea­ding For­tu­ne 500 busi­nesses and cut­ting-edge sci­en­ti­fic rese­arch faci­li­ties around the world rely on AMD tech­no­lo­gy dai­ly to impro­ve how they live, work and play. AMD employees around the world are focu­sed on buil­ding gre­at pro­ducts that push the boun­da­ries of what is pos­si­ble. For more infor­ma­ti­on about how AMD is enab­ling today and inspi­ring tomor­row, visit the AMD (NASDAQ: AMD) web­site, blog, Face­book and Twit­ter pages.

AMD, the AMD Arrow logo, EPYC, Instinct and com­bi­na­ti­ons the­reof, are trade­marks of Advan­ced Micro Devices, Inc. Other names are for infor­ma­tio­nal pur­po­ses only and may be trade­marks of their respec­ti­ve owners.
 

 

FOOTNOTES

  1. Includes AMD high per­for­mance CPU and GPU acce­le­ra­tors used for AI trai­ning and High-Per­for­mance Com­pu­ting in a 4‑Accelerator, CPU hos­ted con­fi­gu­ra­ti­on. Goal cal­cu­la­ti­ons are based on per­for­mance scores as mea­su­red by stan­dard per­for­mance metrics (HPC: Lin­pack DGEMM ker­nel FLOPS with 4k matrix size.  AI trai­ning: lower pre­cis­i­on trai­ning-focu­sed floa­ting point math GEMM ker­nels such as FP16 or BF16 FLOPS ope­ra­ting on 4k matri­ces) divi­ded by the rated power con­sump­ti­on of a repre­sen­ta­ti­ve acce­le­ra­ted com­pu­te node inclu­ding the CPU host + memo­ry, and 4 GPU accelerators.
  2. Based on 2015–2020 indus­try trends in ener­gy effi­ci­en­cy gains and data cen­ter ener­gy con­sump­ti­on in 2025. 
  3. The CPU socket and GPU node power con­sump­ti­ons are based on seg­ment-spe­ci­fic uti­liza­ti­on (acti­ve vs. idle) per­cen­ta­ges then mul­ti­pli­ed by PUE to deter­mi­ne actu­al ener­gy use for cal­cu­la­ti­on of the per­for­mance per Watt.
  4. Total 2025 Ser­ver CPUs — 18.8 Mu (IDC — Q1 2021 Tra­cker), Total HPC CPUs – 3.3Mu (Hype­ri­on- Q4 2020 Tra­cker), Total 2025 HPC GPUs 624k (Hype­ri­on HPC Mar­ket Ana­ly­sis, April ’21)