Meta Platforms Inc reportedly revealed that its research team has built a new artificial intelligence supercomputer, which is named "AI Research Super Cluster" (RSC), which contains 16,000 NVIDIA A100 GPUs with a computing power of 5 EFLOPS (mixed precision). Meta believes that by mid-2022, when the assembly is complete, it will be the world's fastest AI supercomputer.
Meta says its RSC will help companies build better AI models. These models can be learned from trillions of instances, working across hundreds of languages, putting text, images, and videos together to analyze to determine if content is harmful. This research will not only help people stay safe when using Meta's services today, but will also play the same role when Meta builds metaversms in the future.
It is understood that RSC is the result of nearly two years of work, led by Meta's artificial intelligence and infrastructure team, and partners include several well-known companies in the industry, such as NVIDIA, Penguin Computing Inc and Pure Storage Inc. The first phase of the RSC is now up and running, and it consists of 760 Nvidia DGX A100 systems, containing a total of 6,080 GPUs.
Meta pointed out that in the standard machine vision research tasks, the performance of RSC has been improved by 20 times, and the second phase is expected to be completed in 2022, and when it is fully completed, RSC will have a total of 16,000 GPUs, which can use more than one trillion parameters on data as large as 1 exabyte to train artificial intelligence, which will become the supercomputer with the largest number of A100.
In addition, Meta's research team said they also hope that RSC can help build a new AI system, in order to meet the growing bandwidth and capacity demands of AI training, Meta has developed a storage service, the Artificial Intelligence Research Store (AIRStore), which can provide 16TB/s of storage bandwidth and exabytes of storage capacity.
It is reported that in order to meet privacy and security requirements, the entire link of data from the storage system to the GPU is end-to-end encrypted and will not be decrypted until training. And before it is imported into RSC, data must go through a privacy review process to ensure that it is properly anonymized. And RSC is also isolated from the Internet, there is no direct inbound or outbound connection, and traffic can only come from Meta's production data center.