TECH: Meta says its new AI supercomputer will be the world’s fastest by mid-2022.

Meta has com­plet­ed the first phase of a new AI super­com­put­er. Once the AI ​​Research Super­Clus­ter (RSC) is ful­ly built lat­er this year, the com­pa­ny believes it will be the fastest AI super­com­put­er on the plan­et, capa­ble of “oper­at­ing at near­ly 5 exaflops of pre­ci­sion com­put­ing mixed”.

The com­pa­ny says RSC will help researchers devel­op bet­ter AI mod­els that can learn from bil­lions of exam­ples. Among oth­er things, the mod­els will be able to build bet­ter aug­ment­ed real­i­ty tools and “seam­less­ly ana­lyze text, images, and video togeth­er,” accord­ing to Meta. Much of this work is in ser­vice of his vision for the meta­verse, in which he says AI-pow­ered apps and prod­ucts will have a key role.

“We hope RSC will help us build entire­ly new AI sys­tems that can, for exam­ple, pro­vide real-time voice trans­la­tions to large groups of peo­ple, each speak­ing a dif­fer­ent lan­guage, so they can col­lab­o­rate in a mean­ing­ful way. about a research project or play­ing an AR game togeth­er,” tech­ni­cal pro­gram man­ag­er Kevin Lee and soft­ware engi­neer Shub­ho Sen­gup­ta wrote in a blog post.

RSC cur­rent­ly has 760 Nvidia DGX A100 sys­tems with a total of 6,080 GPUs. Meta believes the cur­rent iter­a­tion is already among the fastest AI super­com­put­ers on the plan­et. Based on ear­ly bench­marks, he claims that RSC can run com­put­er vision work­flows up to 20x faster and the NVIDIA col­lec­tive com­mu­ni­ca­tion library more than nine times faster.

Meta says RSC can also train large-scale nat­ur­al lan­guage pro­cess­ing mod­els three times faster. Thus, AI mod­els that deter­mine whether “an action, sound or image is harm­ful or benign” (for exam­ple, to erad­i­cate hate speech) can be trained faster. Accord­ing to the com­pa­ny, this research will help pro­tect peo­ple on cur­rent ser­vices like Face­book and Insta­gram, as well as in the metaverse.

In addi­tion to build­ing the phys­i­cal infra­struc­ture and sys­tems to run RSC, Meta said it needs to ensure secu­ri­ty and pri­va­cy con­trols are in place to pro­tect the actu­al train­ing data it uses. He says that by using real-world data from his pro­duc­tion sys­tems, instead of pub­licly avail­able datasets, he can put his research to more effec­tive use, for exam­ple, by iden­ti­fy­ing harm­ful content.

This year, Meta plans to increase the num­ber of GPUs in RSC to 16,000. It says this will increase AI train­ing per­for­mance by more than 2.5 times. The com­pa­ny, which began work on the project in ear­ly 2020, want­ed RSC to train AI mod­els on datasets up to an exabyte in size (the equiv­a­lent of 36,000 years of high qual­i­ty video).

“We expect that such a step-by-step change in com­put­ing capac­i­ty will not only allow us to cre­ate more accu­rate AI mod­els for our exist­ing ser­vices, but also enable com­plete­ly new user expe­ri­ences, espe­cial­ly in the meta­verse,” Lee and Sen­gup­ta wrote.

Be the first to comment

Leave a Reply

Your email address will not be published.