Into the Metaverse with the new AI Research SuperCluster

Social media conglomerate, Meta Platforms, requires better AI to realise their metaverse plans. In summer 2022, the company plans to launch a new AI Research SuperCluster (RSC), expected to be one of the fastest machines of its type. The company has stated that the new supercluster needed to advance AI and pave the way into interconnected virtual spaces will require “powerful new computers capable of quintillions of operations per second”.

RSC will be used to train content moderation algorithms used to detect hate speech on Facebook and Instagram, train large models in natural language processing (NLP) and computer vision for research and develop augmented features for the company’s future AR hardware. Among most interesting applications to be developed with RSC is the real-time voice translation to large groups of people, each speaking a different language which will be used as a multi-simultaneous translation for project collaborators from different countries, for example, or for gamers.

The 1st generation cluster has 22,000 NVIDIA V100 Tensor Core GPUs capable to carry out 35,000 trainings a day and is now believed to be too limiting for Meta’s researchers. The 2nd generation cluster, expected to become operational in July will contain 16,000 NVIDIA A100 GPUs and is expected to cope with exabyte-size data sets. Early benchmarks suggest that the system will run computer vision workflows 20 times faster, NVIDIA Collective Communication Library (NCCL) nine times faster and trains large scale NLP models three times faster than the previous generation machine.

Meta’s main competitors, Microsoft and Nvidia, not to be left behind, also announced their own AI supercomputers. However, it is good to bear in mind that the value of modern supercomputers is not determined by their theoretical peak performance but by what they enable the company to achieve, in the application of all this insane computation. For Meta, the true utility of this new infrastructure is in improving their moderation systems to recover their public image at the time when the trust in the company is exceptionally low.