Meta unveils AI supercomputer with 16,000 Nvidia A100 GPUs
Meta has shown a new supercomputer, which should become the fastest ‘AI supercomputer’ in the world later this year. The system currently has 760 clusters with a total of 6080 Nvidia A100 GPUs, but this will be expanded to 16,000 this year.
Facebook parent company Meta calls the supercomputer AI Research SuperCluster, abbreviated RSC, and shares details in a blog post. Currently, the system has 700 Nvidia DGX A100 systems functioning as nodes. Those systems each have eight A100 GPUs. These cards communicate with each other via an Nvidia Quantum InfiniBand fabric with a bandwidth of 200Gbit/s. RSC also includes 175 petabytes of Pure Storage FlashArray storage, 46PB of cache, and 10PB of Pure Storage FlashBlade storage.
Source: Meta
According to the company, at the current stage, RSC is up to 20 times faster than its current system for machine vision tasks. The system will be expanded further in the course of this year. According to Meta, the supercomputer will then have 16,000 GPUs. The Facebook parent company also claims to have designed a cache and storage system that can process 16TB/s of data. In time, this system must be able to handle data sets totaling 1 exabyte. The system should be ready by mid-2022.
Meta wants to use the AI supercomputer to train systems and AI models that can be used to recognize images, language and spoken text, which can be used to detect harmful content on its platforms. According to the company, RSC will also be used in the future to build the ‘metaverse’.
Meta writes that RSC is isolated from the internet and that all connections run through Meta’s own data centers. Content used by the AI supercomputer to train models will be anonymized for use, according to the company. According to Meta, end-to-end encryption is used between the storage system and the GPUs. The data is only decrypted in-memory, according to the company, and decryption keys for older data are “regularly deleted,” according to the company.
Source: Meta