Meta is constructing the world's largest AI-specific supercomputer
Fb’s proprietor desires extraordinary computing energy to develop AI fashions to recognise speech, translate languages and energy 3D worlds
Meta’s RSC supercomputer Meta
Fb’s father or mother firm, Meta, is constructing the world’s strongest AI-specific supercomputer to develop higher speech-recognition instruments, routinely translate between totally different languages and assist construct its 3D digital metaverse.
Though removed from full, the AI Analysis SuperCluster (RSC) is up and working and has already overtaken Meta’s earlier quickest supercomputer. That machine was designed in 2017 and ran on 22,000 highly effective graphics processing models (GPUs) which, regardless of being designed for taking part in video games, are extremely efficient instruments to coach synthetic intelligence fashions with.
RSC presently has solely 6080 GPUs, however they're extra highly effective than these within the older machine and it's already 3 times sooner at coaching giant AI fashions than its predecessor. Its present efficiency is on a par with the Perlmutter supercomputer on the Nationwide Vitality Analysis Scientific Computing Middle in California, which is presently positioned at quantity 5 within the TOP500 world supercomputer rankings.
When RSC is full, it is going to encompass 16,000 GPUs and be virtually 3 times extra highly effective than it's now. Meta says that at this level, it will likely be the quickest AI-optimised supercomputer on this planet, acting at practically 5 exaflops.
Supercomputers may be designed to excel at sure duties. Meta’s machine is specialised to coach and run giant AI fashions. There will likely be extra highly effective computer systems on this planet when it's full, however only some, and none that shares its actual structure or supposed use.
The chopping fringe of AI analysis has relied on scale in recent times, and ever extra highly effective machines to coach fashions with. One of many largest neural networks, the Megatron-Turing Pure Language Technology mannequin, has 530 billion parameters, that are roughly equal to the connections between mind cells. Meta says its machine will finally run fashions with trillions of parameters.
James Knight on the College of Sussex, UK, says the proposed laptop is “big” in scale, however could not overcome a number of the challenges in AI analysis. “A system this massive is unquestionably going to allow them to construct bigger fashions,” he says. “Nonetheless, I don’t suppose that merely growing the scale of language fashions will tackle the well-documented issues of present fashions repeating sexist and racist language or failing primary checks of logical reasoning.”
Post a Comment