Looks like Meta is swinging for the cheap seats.
The social media superpower Meta (formerly Facebook) has announced that it has built an "AI supercomputer" — an unconscionably fast computer designed to train and enhance machine-learning systems, according to a Monday post from Meta CEO Mark Zuckerberg.
"Meta has developed what we believe is the world's fastest AI supercomputer," said Zuckerberg in his post. "We're calling it RSC for AI Research SuperCluster and it'll be complete later this year."
Meta's new supercomputer: the AI Research SuperCluster
It's pretty hard to deny that AI research has risen to a central role for mega-firms like Meta. Its rivals, including Nvidia and Microsoft, have already come forward with ambitious designs to deploy their own "AI supercomputer' systems, which are a variant of the conventional notion of supercomputers. RSC will have a broad range of applications throughout Meta's global reach of capital gains. This will include augmented reality features that will eventually show up in the company's AR hardware, and even content moderation algorithms that detect high-liability language (like "hate speech") on Instagram and Facebook.
Needless to say, Meta will use RSC to develop novel services on the metaverse — the firm's privately-owned virtual supplement for the rapidly vanishing public commons, where office- and advertisement-friendly spaces will pervade nearly all aspects of user experience. And, at least according to Meta, this is all great. "RSC will help Meta's AI researchers build new and better AI models that can learn from trillions of examples; work across hundreds of different languages; seamlessly analyze text, images, and video together; develop new augmented reality tools; and much more," wrote Shubho Sengupta and Kevin Lee, two Meta engineers, in a blog post from the company.
"We hope RSC will help us build entirely new AI systems that can, for example, power real-time voice translations to large groups of people, each speaking a different language, so they can seamlessly collaborate on a research project or play an AR game together," added the Meta engineers. And this project began 1.5 years ago, when Meta engineers began work on the various systems that support the supercomputing AI. This meant designing power, cabling, networking, and cooling systems, from nothing. Well, almost nothing. As of January 2022, Meta had a market cap of nearly $860 billion.
Meta wants to outclass Google, Apple, and Microsoft
But phase one of RSC is already completed, and features 760 Nvidia DGX A100 systems, comprised of 6,080 connected GPUs. And this setup has already improved performance over conventional machine vision research operations by 20-fold, according to the firm. But the biggest claim is the boldest of all: before 2022 is finished, RSC's phase two will be complete. And when it is, it'll have roughly 16,000 total GPUs, enabling it to train AI systems "with more than a trillion parameters on data sets as large as an exabyte." Microsoft's AI supercomputer, built by the research lab OpenAI, uses 10,000 GPUs, according to a report from The Verge.
In case you missed it, this might make Meta's supercomputer AI the fastest in the game.
This might sound like just a bunch of numbers, but the real utility of supercomputers is in their applications, not theoretical speed or peak performance. Meta's applications will start by creating new moderation systems, in a bid to improve the company's PR with consumers. If this happens, Meta could become an intimidating force to rivals like Apple, Google, and Microsoft.
But that's a big "if".