Meta is Building New Supercomputer With 16,000 Nvidia A100 GPUs

Meta (previously known as Facebook) is building a new AI supercomputer, called the AI Research SuperCluster, at a currently undisclosed location. According to Meta, it'll be the world's fastest supercomputer to date when complete with a performance metric of over 220 Linpack petaflops. RSC is operational right now but is still being built out.

Specs for this new supercomputer are quite impressive, as the RSC currently features 760 of Nvidia's cutting-edge DGX A100 AI compute units and includes eight Nvidia A100 GPUs in each unit. This makes for a total of 6,080 GPUs for the entire supercomputer. But, once complete, RSC will pack a whopping 16,000 GPUs with an additional 1,240 DGX nodes to its name. The second installation phase will be completed sometime in July.

For networking, the supercomputer will come equipped with Nvidia's Quantum InfiniBand networking system, which outputs up to 200Gb/s of bandwidth. For storage, the supercomputer comes with 175PB of Pure Storage FlashArray, 10PB of Pure Storage FlashBlade, and 46PB of cache storage.

The RSC will be completely isolated from the internet, with all connections from the internet passing through Meta's own data servers first. This is because the supercomputer will be handing internal production data from Meta, and not just public data, making securing RSC a big concern.

Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.