Recently, Meta (Facebook's parent company) announced the construction of the world's most powerful AI supercomputer, called AI Research SuperCluster (RSC).
The goal is to reach a final operation speed (Linpack) of 220 petaflops (2.2 billion times per second), in fact, the RSC is already in operation, but there will be additional operation nodes in the future.
The picture shows the DGX A100
At present, the RSC configuration is 760 sets of NVIDIA DGX A100 AI system, that is, 1520 AMD 64-core EPYC processors (second generation, codename Rome), 6080 NVIDIA A100 GPUs, and 1240 sets of DGX A100 systems will be added in the future, with a single GPU size of up to 16000 (40GB of single video memory).
In terms of I/O, the network bandwidth is 200Gbps, 175PB FlashArray, 10PB FlashBlade, and 46PB cache.
After investigation, the DGX A100 was released in 2020 and sold for a single system for up to $199,000. In other words, when the RSC is finally completed, it will cost $398 million (about 2.5 billion yuan) to purchase the DGX A100 alone.
