Meta, the company behind Facebook, said the supercomputer will be completed as soon as 2022, serving AI research to develop the metaverse.
AI supercomputer image shared by Meta. Photo: Meta |
The system, called AI Research SuperCluster (RSC), was declared by CEO Mark Zuckerberg of Meta as "the most powerful AI supercomputer in the world" when it was released later this year. The product is built for metaverse development in mind.
"Our experience shows that the metaverse will require enormous computing power with billions of billions of calculations per second. RSC enables a new AI model that can learn from trillions of samples, understand hundreds of languages and more. more," wrote Mark Zuckerberg.
In an article on the company's website, Meta's Technical Director Kevin Lee explains their need for increasingly high processing power in speech, language, and vision tasks. For example, to identify harmful content, the system needs to process longer videos, faster sampling rates, speech recognition must work well in situations with background noise, etc. .
In the field of metaverse, RSC is expected to create completely new AI systems. For example, it could provide voice translation for a group of people who speak a different language at the same time, allowing them to work together or play an AR game.
According to Mr. Lee, in the early stages, RSC is equipped with 760 Nvidia DGX A100 systems with 6,080 GPUs for computing tasks. This number will increase to 16,000 GPUs when completed. The GPUs communicate with each other via the Nvidia Quantum transmission system at 200 Gb/s. Triple-tier memory with capacities of 10, 46, and 175 petabytes, respectively.
In the test published by the company, the process of processing computer vision tasks is 20 times faster than the system that Meta is using with 22,000 Nvidia V100 Tensor Core GPUs.
As the most powerful AI supercomputer, Meta's RSC is not necessarily the most powerful supercomputer in the world. According to experts, it is necessary to distinguish between "supercomputer" and "AI supercomputer". RSC cannot be compared with machines used in fields such as cosmology, nuclear physics... because they have different purposes.
The difference between these supercomputers comes from the need for precision, said Bob Sorensen, a supercomputer analyst at Hyperion Research. Operations can occur between very large or very small numbers, including decimals. The precision of the numbers after the comma can be inversely proportional to the calculation speed. AI is a field that requires not too high precision, so Meta's supercomputer will probably have faster computing speeds than supercomputers in the rest of the fields.
In addition, according to Bob Sorensen, Meta's publication is based on theoretical calculations. In fact, most supercomputers won't be able to run with this performance.
Currently, the world's most powerful supercomputer belongs to the Fugaku machine, developed by Fujitsu and Japan's Riken National Research Institute with a computational speed of 442 petaflops, or 442 million billion operations per second. This is almost three times higher than the second-ranked supercomputer, IBM's Summit, with 148 petaflops. In third place is Sierra, the system located at Lawrence Livermore National Laboratory (USA) with a speed of 94.6 petaflops.