Meta says its new AI supercomputer is among the fastest in the world
And it will be the world's fastest when completed in mid-2022, the company claims
Facebook's parent firm Meta has unveiled the AI Research SuperCluster (RSC), a new supercomputer that it says is among the fastest in the world.
The new supercomputing machine was announced by Meta CEO Mark Zuckerberg, who said it will be the world's fastest supercomputer once it is complete later this year.
Meta's RSC has been designed to train a range of machine learning systems across the company's businesses. This includes teaching algorithms to better detect spoken words in audio and specific objects in images, quickly translating across different languages, and moderating content to identify explicit posts and hate speech on Facebook and Instagram.
The RSC will also be used to design tools or run programmes for the Metaverse.
Meta engineers Kevin Lee and Shubho Sengupta wrote in a blog post that the company's artificial intelligence (AI) supercomputer will help develop AI models that can work across hundreds of languages and learn from trillions of examples.
Among other things, the AI supercomputer will also be able to seamlessly analyse images, videos and text simultaneously; design new reality tools; and much more.
"Ultimately, the work done with RSC will pave the way toward building technologies for the next major computing platform - the metaverse, where AI-driven applications and products will play an important role," Meta engineers noted.
"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."
In its current form, the supercomputer consists of 760 Nvidia GGX A 100 modules packing 6,080 GPUs, which are connected through a Quantum InfiniBand network able to transmit data at 200 Gigabits per second. The system features 175 petabytes (PB) of bulk storage and 46 PB of cache storage.
According to Meta, RSC is already providing up to 20 times enhanced performance, compared with the company's older setup. By the end of the second phase, which is expected to happen before the end of 2022, RSC will have a total of 16,000 GPUs, enabling it to deliver 16 terabytes of data per second while boosting its AI training performance by more than 2.5 times. By the end of the second phase, the machine will house a total of 1 exabyte of storage, which is equal to 1,000 PB.
AI supercomputers are somewhat different from traditional supercomputers in the sense that machine learning jobs that they do require less accuracy than the kind of jobs handled by their regular counterparts. This means AI Supercomputers can complete more calculations per second than traditional supercomputers using the same hardware.
In other words, Meta's RSC is not a direct competitor to Fugaku, which is currently the fastest supercomputer in the world. Fugaku was jointly developed by Japanese tech firm Fujitsu and Japanese research institute Riken.
It features chip technology from Arm Ltd and also makes use of Fujitsu's 48-core A64FX system-on-chip. The machine is capable of carrying out over 415.5 quadrillion computations a second.
In 2020, Microsoft also built a massive supercomputer for OpenAI, a company with a mission to build an artificial general intelligence that will benefit humanity as a whole.
Microsoft's supercomputer is hosted in the Azure cloud and is claimed to be among the top five supercomputers in the world. It features 285,000 processor cores and 10,000 GPUs and can offer 400 gigabits per second of network connectivity for each GPU server, according to Microsoft.