Meta has joined the list of tech companies with their supercomputers. What’s more, the company claims that its supercomputer will be the world’s fastest by mid-2022.
The AI Research Cluster or RSC will be used to train several systems across Meta businesses, including but not limited to content moderation algorithms to detect hate speech on its platforms. The RSC will also be used to design experiences for the Metaverse.
The increasing number of these AI supercomputers specifically designed to train machine learning systems shows just how vital AI research is to companies like Microsoft, Nvidia and now Meta.
In the News: Valve introduces Dynamic Cloud Saves for Steam Cloud
AI supercomputer powered platforms?
In an announcement by Meta AI, the company’s in-house AI research department, the company shared some numbers comparing the AI RSC to their existing V100 based clusters.
The computer is claimed to be 20 times faster than their current V100-based production clusters, nine times faster than their v100-based research clusters when it comes to Nvidia NCCL collectives and three times faster for large-scale projects NLP workflows.
These AI supercomputers combine multiple GPUs into compute nodes connected by a high-performing network for communication. Currently, the RSC contains 760 Nvidia DGX A100 systems as compute nodes, totalling 6080 GPUs. The supercomputer also gets 175 petabytes of Pure Storage FlashArray has its storage tier along with 46 petabytes of cache storage in Penguin Computing Altus systems and 10 petabytes of Pure Storage FlashBlade.
The end goal for the RSC is to have 16000 GPUs acting as endpoints eventually. The caching and storage system can serve 16TB/s of training data with plans to scale it up to 1 exabyte.
Finally, in terms of security, the RSC is kept completely isolated from the internet with no direct inbound or outbound connections. Additionally, traffic only flows from Meta’s production data centres. The entire data path from storage systems to the actual GPUs is end to end encrypted, and the data goes through a privacy review process before being imported to RSC to ensure proper data anonymisation.
The infrastructure is up and running but has only completed the first development phase. Meta claims that by the second phase ends; the AI RSC will compute at about five exaflops. All expansion plans will be implemented through 2022.
While they’re busy building the world’s fastest AI supercomputer, as revealed by a Wall Street Journal investigation, a “software glitch” in Facebook’s free version ended up charging its users an estimated $7.8million by July 2021.
Facebook has partnered with telecom providers in developing countries to allow users to access the platform without data charges. However, due to this glitch, they ended up getting charged anyway. The company is seemingly aware of the issue and is investigating.
In the News: FIFA 22 TOTY [Team of the Year] announced
