Meta Platforms gave an enormous thumbs as much as NVIDIA, selecting our applied sciences for what it believes will likely be its strongest analysis system up to now.
The AI Analysis SuperCluster (RSC), introduced right this moment, is already coaching new fashions to advance AI.
As soon as absolutely deployed, Meta’s RSC is predicted to be the biggest buyer set up of NVIDIA DGX A100 techniques.
“We hope RSC will assist us construct solely new AI techniques that may, for instance, energy real-time voice translations to massive teams of individuals, every talking a distinct language, so they may seamlessly collaborate on a analysis challenge or play an AR recreation collectively,” the corporate stated in a weblog.
Coaching AI’s Largest Fashions
When RSC is absolutely constructed out, later this yr, Meta goals to make use of it to coach AI fashions with greater than a trillion parameters. That might advance fields similar to natural-language processing for jobs like figuring out dangerous content material in actual time.
Along with efficiency at scale, Meta cited excessive reliability, safety, privateness and the pliability to deal with “a variety of AI fashions” as its key standards for RSC.
Below the Hood
The brand new AI supercomputer at the moment makes use of 760 NVIDIA DGX A100 techniques as its compute nodes. They pack a complete of 6,080 NVIDIA A100 GPUs linked on an NVIDIA Quantum 200Gb/s InfiniBand community to ship 1,895 petaflops of TF32 efficiency.
Regardless of challenges from COVID-19, RSC took simply 18 months to go from an thought on paper to a working AI supercomputer (proven within the video beneath) thanks partially to the NVIDIA DGX A100 know-how on the basis of Meta RSC.
20x Efficiency Features
It’s the second time Meta has picked NVIDIA applied sciences as the bottom for its analysis infrastructure. In 2017, Meta constructed the primary era of this infrastructure for AI analysis with 22,000 NVIDIA V100 Tensor Core GPUs that handles 35,000 AI coaching jobs a day.
Meta’s early benchmarks confirmed RSC can prepare massive NLP fashions 3x quicker and run pc imaginative and prescient jobs 20x quicker than the prior system.
In a second part later this yr, RSC will increase to 16,000 GPUs that Meta believes will ship a whopping 5 exaflops of blended precision AI efficiency. And Meta goals to increase RSC’s storage system to ship as much as an exabyte of knowledge at 16 terabytes per second.
A Scalable Structure
NVIDIA AI applied sciences can be found to enterprises of any dimension.
NVIDIA DGX, which features a full stack of NVIDIA AI software program, scales simply from a single system to a DGX SuperPOD working on-premises or at a colocation supplier. Prospects may also lease DGX techniques by NVIDIA DGX Foundry.