Meta has unveiled the AI Analysis SuperCluster (RSC), a brand new supercomputer that’s among the many quickest on this planet. And it’ll solely get quicker – by the top of the 12 months it ought to rank primary, with computing energy on the exascale.
The corporate previously generally known as Fb has had its fingers within the AI pie for a couple of years now, and it’s not onerous to see why. Via Fb, Instagram and Whatsapp et al, the conglomerate generates much more information than any mere mortal might probably course of – and there’s obscene quantities of cash to be made in sifting by means of all of it.
Meta’s RSC can be as much as the duty, utilizing this information and immense computing energy to coach AI algorithms to raised acknowledge objects in photos and spoken phrases in audio, rapidly translate between languages, and establish dangerous content material and misinformation that shouldn’t be on social media. And naturally, it’ll in all probability find yourself powering the metaverse, which the corporate is making an attempt to persuade the world it wants.
In its present type, the supercomputer consists of 6,080 Nvidia A100 GPUs, linked by means of a Quantum InfiniBand community that may transmit information at 200 Gigabits per second. Storage-wise, the system packs an astounding 175 petabytes (PB) of bulk storage, 46 PB of cache storage and an extra 10 PB of NFS storage. For reference, one petabyte is the same as 1,000,000 gigabytes.
In benchmarking exams, RSC was in a position to practice pure language processing fashions 3 times quicker than Meta’s earlier AI supercomputer, and was 20 occasions quicker at pc imaginative and prescient duties.
Importantly, Meta says RSC has plenty of safety and privateness measures in place. For one, it’s not related to the web at giant, with information solely being fed in from from the corporate’s personal information facilities. That information is encrypted right through the system, solely being decrypted when it’s about for use to coach fashions.
And it’s simply getting began. By mid-2022, RSC can have been upgraded with 10,000 extra GPUs, a quicker community that may ship 16 terabytes of information per second, and home a grand whole of 1 exabyte of storage – that’s 1,000 PB.
At that time, Meta says that RSC can have a peak efficiency of an insane 5 exaflops. The prime supercomputer on this planet in the mean time – Japan’s Fugaku – peaks at “solely” 2 exaflops, which means Meta would have a mammoth lead. Nevertheless, it’s price noting that not solely is RSC not absolutely operational but, but it surely hasn’t been examined on the identical benchmarking processes as Fugaku, so the numbers may not be instantly comparable.
Both manner, RSC will sit comfortably among the many prime 10 supercomputers on this planet when it debuts later this 12 months.
Post a Comment