Meta is Building a Massive New Supercomputer
(Photo: Meta)Meta, formerly known as Facebook, is working on what it says will be the fastest AI supercomputer in the world with the goal of advancing into a new generation of AI. Known as the AI Research SuperCluster (RSC), it’s already among the fastest AI supercomputers in operation.
Meta’s goal with the supercomputer is to use AI to power real-time interactions, such as the impressive feat of helping “large groups of people, each speaking a different language… seamlessly collaborate on a research project or play an AR game together.” Speech recognition, computer vision, and neuro-linguistic programming are among the tech giant’s top priorities. But given that this is Meta-slash-Facebook, the metaverse can’t not make an appearance on the RSC’s press release; the company admits that it will use its powerful new supercomputer to build AI-driven applications and products that ultimately support the virtual world.
This means the supercomputer has to be extremely reliable, especially because Meta imagines it will run experiments involving thousands of GPUs for weeks at a time. The user interface has to be decent, too, if the company wants to facilitate the groundbreaking research and engineering it says it’s aiming for. Such ease of use and dependability (plus the obvious computing power) relies on a pretty notable build, particularly for a project that started off as fully remote.
Like with any other AI supercomputer, Meta is assembling RSC by combining multiple GPUs into compute nodes, then connecting them with a high-performance network fabric that allows for ultra-efficient communication. At its current build phase, RSC’s compute nodes are made up of 760 of Nvidia’s DGX A100 systems, which interact through an Nvidia Quantum 1600 Gb/s InfiniBand two-level Clos fabric. (The end product will connect 16,000 GPUs.) Its storage tier boasts 175 petabytes of Pure Storage FlashArray, 46 petabytes of cache storage from Penguin Computing Altus systems, and 10 petabytes of Pure Storage FlashBlade.
Meta estimates it will be finished building out the supercomputer around the middle of this year. The project’s location is under wraps for now; from a cyber standpoint, Meta is enforcing end-to-end encryption across its entire data path and requiring that all information goes through a privacy review process to confirm it has been correctly anonymized prior to use.