NVIDIA Ampere GPUs Come to Google Cloud at Speed of Light

Just weeks after launch, NVIDIA A100 GPUs land on Goog­le Com­pu­te Engine.

The NVIDIA A100 Ten­sor Core GPU has lan­ded on Goog­le Cloud.

Available in alpha on Goog­le Com­pu­te Engi­ne just over a month after its intro­duc­tion, A100 has come to the cloud fas­ter than any NVIDIA GPU in history.

Today’s intro­duc­tion of the Acce­le­ra­tor-Opti­mi­zed VM (A2) ins­tance fami­ly fea­turing A100 makes Goog­le the first cloud ser­vice pro­vi­der to offer the new NVIDIA GPU.

A100, which is built on the new­ly intro­du­ced NVIDIA Ampere archi­tec­tu­re, deli­vers NVIDIA’s grea­test gene­ra­tio­nal leap ever. It boosts trai­ning and infe­rence com­pu­ting per­for­mance by 20x over its pre­de­ces­sors, pro­vi­ding tre­men­dous spee­dups for workloads to power the AI revolution.

Goog­le Cloud cus­to­mers often look to us to pro­vi­de the latest hard­ware and soft­ware ser­vices to help them dri­ve inno­va­ti­on on AI and sci­en­ti­fic com­pu­ting workloads, ” said Manish Saina­ni, direc­tor of Pro­duct Manage­ment at Goog­le Cloud. “With our new A2 VM fami­ly, we are proud to be the first major cloud pro­vi­der to mar­ket NVIDIA A100 GPUs, just as we were with NVIDIA T4 GPUs. We are exci­ted to see what our cus­to­mers will do with the­se new capabilities.”

In cloud data cen­ters, A100 can power a broad ran­ge of com­pu­te-inten­si­ve appli­ca­ti­ons, inclu­ding AI trai­ning and infe­rence, data ana­ly­tics, sci­en­ti­fic com­pu­ting, geno­mics, edge video ana­ly­tics, 5G ser­vices, and more.

Fast-gro­wing, cri­ti­cal indus­tries will be able to acce­le­ra­te their dis­co­veries with the breakth­rough per­for­mance of A100 on Goog­le Com­pu­te Engi­ne. From sca­ling up AI trai­ning and sci­en­ti­fic com­pu­ting, to sca­ling out infe­rence appli­ca­ti­ons, to enab­ling real-time con­ver­sa­tio­nal AI, A100 acce­le­ra­tes com­plex and unpre­dic­ta­ble workloads of all sizes run­ning in the cloud. 

NVIDIA CUDA 11, coming to gene­ral avai­la­bi­li­ty soon, makes acces­si­ble to deve­lo­pers the new capa­bi­li­ties of NVIDIA A100 GPUs, inclu­ding Ten­sor Cores, mixed-pre­cis­i­on modes, mul­ti-ins­tance GPU, advan­ced memo­ry manage­ment and stan­dard C++/Fortran par­al­lel lan­guage constructs.

Breakthrough A100 Performance in the Cloud for Every Size Workload

The new A2 VM ins­tances can deli­ver dif­fe­rent levels of per­for­mance to effi­ci­ent­ly acce­le­ra­te workloads across CUDA-enab­led machi­ne lear­ning trai­ning and infe­rence, data ana­ly­tics, as well as high per­for­mance computing.

For lar­ge, deman­ding workloads, Goog­le Com­pu­te Engi­ne offers cus­to­mers the a2-megag­pu-16g ins­tance, which comes with 16 A100 GPUs, offe­ring a total of 640GB of GPU memo­ry and 1.3TB of sys­tem memo­ry — all con­nec­ted through NVS­witch with up to 9.6TB/s of aggre­ga­te bandwidth.

For tho­se with smal­ler workloads, Goog­le Com­pu­te Engi­ne is also offe­ring A2 VMs in smal­ler con­fi­gu­ra­ti­ons to match spe­ci­fic appli­ca­ti­ons’ needs.

Goog­le Cloud announ­ced that addi­tio­nal NVIDIA A100 sup­port is coming soon to Goog­le Kuber­netes Engi­ne, Cloud AI Plat­form and other Goog­le Cloud ser­vices. For more infor­ma­ti­on, inclu­ding tech­ni­cal details on the new A2 VM fami­ly and how to sign up for access, visit the Goog­le Cloud blog.