By akademiotoelektronik, 13/05/2022

AI Research SuperCluster: World's fastest supercomputer by mid-2022

ShareTweetShareShareEmailComments

Construction of the AI ​​Research SuperCluster (RSC) will be completed sometime this year, according to Meta Group's Artificial Intelligence Research Division. It will be the fastest supercomputer in the world when it is finished in mid-2022.

Meta's AI Research SuperCluster (RSC)

On January 24, Meta announced the progress of its supercomputer. According to the firm, this should be completed by mid-2022. It will be the most powerful supercomputer dedicated to artificial intelligence in the world.

The AI ​​Research SuperCluster would already be effective, but we will have to wait for the summer to see it at full power. Indeed, Meta exploits only part of the capabilities of the machine. The latter should have a mixed precision calculation capacity of 5 exaflops.

The AI ​​RSC embeds 6,080 GPUs for 760 Nvidia DGX A100 systems interconnected by a 200 Gb/s InfiniBand network. It also has 231 petabytes of Pure Storage FlashArray storage, and 46 petabytes of cache. On the first benchmarks, the supercomputer can run computer vision workflows 20 times faster, compared to the previous installation.

AI Research SuperCluster :  le superordinateur le plus rapide du monde vers la moitié de 2022

Indeed, Meta currently uses an infrastructure composed of 20,000 Nvidia V100 GPUs gathered in a cluster running 35,000 training tasks per day. The start of work for the AI ​​Research SuperCluster began in 2020 to take advantage of new technologies.

A supercomputer to develop the Metaverse

The new supercomputer will be used for the development of the Metaverse of the Meta group. For this, the AI ​​Research SuperCluster will help with better performing AI model searches. This includes a learning capacity on the order of trillions of examples.

This AI will then improve the construction of tools for augmented reality, or the analysis of text, images or videos in a transparent way. This will help the firm to prevent all kinds of abuse or illegal content on the Metaverse, or on its platforms (Facebook, Instagram and WhatsApp).

All applications and products derived from the developed AI will contribute to the Metaverse. For this, Meta plans a total number of 16,000 Nvidia GPUs for the RSC this year. This will multiply the learning capabilities of the AI ​​by, at least 2.5 times. RSC should eventually be able to train AI models on up to an exabyte of data.

With such computing power, the AI ​​RSC should surpass the current number one: the "Fugaku" and its 442,010 teraflops in high precision. However, other supercomputers are also under construction. These are the Department of Energy's Aurora at Argonne National Laboratory and El Capitan which will oversee the US nuclear stockpile, with 2 exaflops each.

Leave a review postRelated Items:ShareTweetShareShareEmailLeave a comment

Leave a Reply

Cancel reply

Your email address will not be published. Mandatory fields are marked with *

Tags: