Meta is developing a record-breaking supercomputer to power the metaverse

Following Meta’s (previously Fb) October announcement that it’s pushing to stake its claim on the metaverse, the corporate right this moment introduced that it has developed the AI Analysis SuperCluster (RSC), which it claims is among the many quickest AI supercomputers operating right this moment. As soon as it’s absolutely constructed, Meta says will probably be the quickest working supercomputer — the corporate is aiming to finish it by the center of this yr.

CEO Mark Zuckerberg famous that the experiences the corporate is constructing for the metaverse require monumental compute energy — reaching into quintillions of operations per second. The RSC will allow new AI fashions to study from trillions of examples, perceive lots of of languages, and extra.

Knowledge storage firm Pure Storage and chip-maker Nvidia are a part of the supercluster that Fb has constructed. Significantly, Nvidia has been a key participant supporting the metaverse, with its omniverse product billed as “metaverse for engineers.” 

After full deployment, Meta’s RSC would be the largest buyer set up of Nvidia DGX A100 programs, mentioned Nvidia in its press release right this moment. 

Rob Lee, CTO at Pure Storage, instructed VentureBeat by way of e mail that the RSC is critical to different corporations outdoors Meta as a result of the applied sciences (corresponding to AI and AR/VR) powering the metaverse are extra broadly relevant and in-demand in industries throughout the board.

Based on Lee, technical choice makers are all the time trying to study from bleeding-edge practitioners, and the RSC offers nice validation of the core parts which might be powering the world’s largest AI supercomputer. 

“Meta’s world-class workforce noticed the worth of pairing the efficiency, density and ease of Pure Storage merchandise to energy Nvidia GPUs created for this groundbreaking work pushing the boundaries of efficiency and scale,” mentioned Lee. He added that enterprises of all sizes will be capable of profit from Meta’s work, experience, and learnings in advancing how they pursue their knowledge, analytics, and AI methods.

Scale is turning into a giant deal

In a weblog launched right this moment, Meta claims that AI supercomputing is required at scale. Based on Meta, realizing the advantages of self-supervised studying and transformer-based fashions requires numerous domains — whether or not imaginative and prescient, speech, language, or for vital functions like figuring out dangerous content material.

AI at Meta’s scale would require massively highly effective computing options able to immediately analyzing ever-increasing quantities of knowledge. Meta’s RSC is a breakthrough in supercomputing that may result in new applied sciences and buyer experiences enabled by AI, mentioned Lee.

“Scale is necessary right here in a number of methods,” mentioned Lee. He famous that firstly, Meta processes an amazing quantity of knowledge on a continuous foundation, and so there’s a certain quantity of scale in knowledge processing efficiency and capability that requires.

“Secondly, AI initiatives rely upon massive volumes of knowledge — with extra diversified and full knowledge units offering higher outcomes. Thirdly, all of this infrastructure must be managed on the finish of the day, and so house and energy effectivity and ease of administration at scale is vital as effectively. Every of those parts is equally necessary, whether or not in a extra conventional enterprise undertaking or working at Meta’s scale,” Lee mentioned.

Tackling the safety and privateness points that include supercomputing

Over the previous few years, Meta has obtained a number of backlashes on its privateness and knowledge insurance policies, with the Federal Commerce Fee (FTC) saying it was investigating substantial concerns on Facebook’s privacy practices in 2018. Meta desires to sort out safety and privateness points from the get-go, stating that the corporate safeguards knowledge in RSC by designing RSC from the bottom up with privateness and safety in thoughts.

Meta claims this can allow its researchers to soundly practice fashions utilizing encrypted user-generated knowledge that’s not decrypted till proper earlier than coaching.

“For instance, RSC is remoted from the bigger web, with no direct inbound or outbound connections, and visitors can stream solely from Meta’s manufacturing knowledge facilities. To satisfy our privateness and safety necessities, the complete knowledge path from our storage programs to the GPUs is end-to-end encrypted and has the required instruments and processes to confirm that these necessities are met always.” mentioned the corporate weblog.

Meta explains that knowledge should undergo a privateness assessment course of to verify it has been accurately anonymized earlier than it’s then imported into the RSC. The corporate additionally claims that the info can also be encrypted earlier than it may be used to coach AI fashions, and decryption keys are deleted repeatedly to make sure previous knowledge is not accessible.

To construct this supercomputer, Nvidia offered the compute layer — together with the Nvidia DGX A100 programs as its compute nodes. The GPUs talk by way of an Nvidia Quantum 200 Gbps InfiniBand two-level Clos cloth. Lee famous that contributions from Penguin Computing {hardware} and software program are “the glue” that unite Penguin, Nvidia, and Pure Storage. Collectively, these three companions have been essential to offering Meta with an enormous supercomputing resolution.