Enlarge image
»AI Research SuperCluster« (RSC): The supercomputer is already running, but it is set to become even bigger
Photo: Meta
When complete, Meta's new supercomputer will help enable real-time translation in large groups, detect malicious content on Facebook and Instagram, develop new augmented reality tools and, of course, technology "for the next big computing platform" - the Metaverse.
The expansion of the "AI Research SuperCluster", or RSC for short, should be completed by mid-2022.
The supercomputer is already running, but by then it should be even bigger and faster than it is now.
Meta assumes that it is already "among the fastest AI supercomputers in the world and will be the fastest" when it is finished.
However, RSC will probably not lead the annually updated TOP500 list of the fastest supercomputers.
Expert Steve Conway explains the difference in the supercomputer blog »Inside HPC«: »General-purpose supercomputers are designed to handle problems with up to 64-bit precision, while 16-bit precision is usually sufficient for AI supercomputers In other words, computers like the new RSC are jerks that are extremely fast in the area for which they were designed – artificial intelligence – but would fail in others.
So you shouldn't ask Facebook's supercomputer for a weather forecast or try to use it to predict what's going on in a fusion reactor.
No information on location and costs
Meta's currently fastest computer system is ranked 139th on the Supercomputer Top 500 list.
In the end, the RSC could actually lead the category of specialized AI supercomputers and significantly improve the company's position in the overall ranking, according to Conway.
However, the addition of "AI" as "artificial intelligence" before the term supercomputer makes a direct comparison with other supercomputers difficult.
More importantly for Meta, RSC will significantly speed up the training of AI models.
What takes nine weeks today should be possible in three weeks in the future.
Among other things, 16,000 Nvidia graphics processors of the type A100 and flash memory systems with a total capacity of 175 petabytes from Pure Storage should ensure this.
Meta does not reveal where the RSC is located and how much it cost to build, even on request.
"Protocol" estimates the price for the Nvidia components alone at 150 million dollars.
The secrecy is apparently part of the security concept.
Meta writes: The training data for the system is Meta's user data, which is anonymized and then sent end-to-end encrypted from Meta's data centers to the RSC and only decrypted immediately before the AI training.
Because this only happens in one place and, so to speak, in the main memory, the data is safe “even in the unlikely event of a break-in to the system”.
The RSC is also not connected to the general Internet.
pbe/mak
