Meta has announced that it’s building the «world’s fastest» AI supercomputer as part of its plans to build a virtual metaverse. 

The AI Research SuperCluster (RSC), which Facebook founder Mark Zuckerberg claims is already the fifth-fastest AI supercomputer in the world, uses Nvidia’s DGX A100 system and is already training new models to advance artificial intelligence capabilities. According to early Meta benchmarks, RSC can train large natural-language processing (NLP) models three times faster and run computer vision jobs 20 times faster than its previous systems. 

The project is expected to be fully constructed midway through the year and will add various AI-based features to Meta’s platforms, including its proposed metaverse. According to a blog post from Zuckerberg, the company aims to use the RSC to train AI models to advance fields such as NLP and image enhancement.

More specifically, Meta hopes to generate models that can work across multiple languages, analyse text, images and video simultaneously, and also develop augmented reality tools.

Key use cases will include the identification of harmful content and the ability to translate multiple languages of audio, in real-time from large groups of people, which appears to be a Metaverse-specific use case.  

The RSC currently uses 760 Nvidia DGX A100 systems as its computer nodes and these pack a total of 6,080 A100 GPUs onto a Nvidia Quantum network said to deliver 1,896 petaflops of performance. Speed was also shown in the installation, with the RCS taking just 18 months to go from idea to fully working supercomputer, according to Nvidia. 

The second phase of the project, which begins later in the year, will expand the RSC to 16,000 GPUs, which Meta suggests will deliver 5 exaflops of «mixed precision AI performance». This will be buoyed by further plans to expand RSC’s storage systems so that it can deliver up to an exabyte of data at 16 terabytes per second. 

This is the second time Meta has partnered with Nvidia for research infrastructure, with a project conducted in 2017, also for the advancement of AI.