Facebook Parent Meta Built the World's Fastest AI Supercomputer

Meta (formerly Facebook) is the latest company to build an AI supercomputer. Meta's AI Research SuperCluster has been designed to train machine learning systems, and the company says it will be the world's fastest supercomputer once it is complete later in 2022. Meta's CEO, Mark Zuckerberg, announced the new supercomputer who said it would be finished later this year. "Meta has developed what we believe is the world's fastest AI supercomputer. We're calling it RSC or AI Research SuperCluster, and it'll be complete later this year," Zuckerberg said in a statement.

The supercomputer reflects the heavy investments that big tech companies are making in AI research. Other tech companies like Microsoft and Nvidia have also announced their own "AI supercomputers." These are slightly different from traditional supercomputers, reports suggest. The Meta AI Research SuperCluster will be used to train various systems in the company's businesses, including content moderation algorithms to detect hate speech and explicit content on Facebook and Instagram and even AR programs available on Meta AR devices in the future.

Unsurprisingly, the RSC will be used to design things or run programs for the Metaverse. In a blog post, Meta engineers Kevin Lee and Shubho Sengupta say the RSC will help build the company's AI models that can learn from "trillions of examples and work in hundreds of different languages. It will also seamlessly analyze text, images, and video together, develop new reality tools, and much more."

Meta began work on AI Research SuperCluster about a year and a half ago, with the company's engineers designing the machine's components for cooling, power, networking, and cabling entirely from scratch.

The supercomputer consists of 760 Nvidia GGX A 100 systems containing 6080 GPUs. Meta says that it is already providing up to 20x improved performance. It will have a total of 16,000 GPUs at the end of the second phase of its creation. This is expected to happen before the end of 2022.

Now AI supercomputers are different from regular computers in that machine learning requires less precision than the kind of tasks traditional supercomputers handle. Therefore, 'AI supercomputers' can perform more calculations per second than their regular counterparts using the same hardware. Basically, this means that the Meta AI Research SuperCluster is not a direct competitor to supercomputers like the Summit 500 or Fugaku, which is currently the fastest supercomputer in the world. While the performance of a traditional supercomputer is measured in 64-bit floating-point operations per second (FLOPs), an AI supercomputer is often measured in 32-bit or 16-bit FLOPs.