Meta (formerly known as Facebook) is constructing the AI Research SuperCluster, a new AI supercomputer at an undisclosed site. With a performance measure of over 220 Linpack petaflops, it will be the world’s fastest supercomputer to date, according to Meta. RSC is currently active, however, it is still being developed out.

The RSC’s specifications are rather amazing since it presently incorporates 760 of Nvidia’s cutting-edge DGX A100 AI compute units, each with eight Nvidia A100 GPUs. This brings the overall number of GPUs on the supercomputer to 6,080. RSC, on the other hand, will have a whopping 16,000 GPUs and 1,240 DGX nodes when it’s finished. The second phase of installation will be completed in July.

The supercomputer will have Nvidia’s Quantum InfiniBand networking solution for networking, which can deliver up to 200Gb/s of bandwidth. The supercomputer has 175 PB of Pure Storage FlashArray storage, 10 PB of Pure Storage FlashBlade storage, and 46 PB of cache storage.

The RSC will be fully disconnected from the internet, with all internet connections going through Meta’s data servers first. Because the supercomputer will be handling internal Meta production data rather than merely public data, protecting RSC will be a major priority.