Meta Works with NVIDIA to Build Massive AI Research Supercomputer

meta-works-with-nvidia-to-build-massive-ai-research-supercomputer

Meta Platforms gave a massive thumbs up to NVIDIA, picking our technologies for what it believes will be its most strong research procedure to day.

The AI Exploration SuperCluster (RSC), announced nowadays, is presently instruction new styles to progress AI.

After totally deployed, Meta’s RSC is anticipated to be the major buyer installation of NVIDIA DGX A100 methods.

“We hope RSC will support us make completely new AI methods that can, for example, power real-time voice translations to substantial groups of persons, each talking a distinctive language, so they could seamlessly collaborate on a investigation project or enjoy an AR recreation with each other,” the company claimed in a web site.

Instruction AI’s Major Styles

When RSC is entirely designed out, afterwards this 12 months, Meta aims to use it to coach AI types with much more than a trillion parameters. That could progress fields this sort of as normal-language processing for positions like determining hazardous written content in actual time.

In addition to effectiveness at scale, Meta cited excessive dependability, protection, privateness and the versatility to cope with “a large variety of AI models” as its important criteria for RSC.


Meta RSC system
Meta’s AI Analysis SuperCluster functions hundreds of NVIDIA DGX methods connected on an NVIDIA Quantum InfiniBand community to accelerate the perform of its AI investigate groups.

Beneath the Hood

The new AI supercomputer now uses 760 NVIDIA DGX A100 programs as its compute nodes. They pack a complete of six,080 NVIDIA A100 GPUs connected on an NVIDIA Quantum 200Gb/s InfiniBand community to deliver one,895 petaflops of TF32 general performance.

Even with problems from COVID-19, RSC took just 18 months to go from an strategy on paper to a performing AI supercomputer (revealed in the video down below) many thanks in element to the NVIDIA DGX A100 technological innovation at the basis of Meta RSC.

20x Efficiency Gains

It is the 2nd time Meta has picked NVIDIA systems as the foundation for its analysis infrastructure. In 2017, Meta built the to start with generation of this infrastructure for AI exploration with 22,000 NVIDIA V100 Tensor Core GPUs that handles 35,000 AI teaching work a working day.

Meta’s early benchmarks showed RSC can coach large NLP products 3x a lot quicker and run pc eyesight work opportunities 20x speedier than the prior technique.

In a 2nd period later this year, RSC will grow to 16,000 GPUs that Meta thinks will provide a whopping five exaflops of mixed precision AI performance. And Meta aims to expand RSC’s storage technique to supply up to an exabyte of information at 16 terabytes for every 2nd.

A Scalable Architecture

NVIDIA AI systems are offered to enterprises of any measurement.

NVIDIA DGX, which consists of a whole stack of NVIDIA AI software, scales quickly from a solitary procedure to a DGX SuperPOD jogging on-premises or at a colocation provider. Customers can also rent DGX programs by NVIDIA DGX Foundry.

Leave a comment

Your email address will not be published.


*