NVIDIA GOES ALL-IN ON AI
While scientific research using FP64 remains important, Nvidia is betting big on the increasing importance of AI workloads. Its current Selene supercomputer already ranks among the top ten fastest for traditional workloads, but in AI performance Nvidia says it can do
2.8 exaflops. If Selene sounds potent, just wait for the upcoming Eos supercomputer, built on Nvidia’s latest Hopper H100 processor.
Eos will consist of 18 DGX H100 SuperPODs, each with 32 DGX H100 servers. The servers will be powered by dual AMD EPYC processors, but Nvidia doesn’t even make a point of mentioning what CPUs are used as they are child’s play compared to the GPU hardware. There are eight H100 GPUs per DGX H100 or 4,608 GPUs in total.
Eos looks pretty tame in terms of its traditional supercomputer Linpack potential. Each H100 has a theoretical performance of 60 teraflops FP64, giving a maximum of 276 petaflops. That’s enough for second place on the current top 500, but the AI potential is much greater.
In FP16/BF16 workloads, Eos provides up to 9.2 exaflops of compute. Nvidia is taking this a step further by supporting FP8 number formats, which doubles again the theoretical performance to 18.4 exaflops.
Other companies are pursuing AI as well, but Nvidia’s CEO Jensen Huang is particularly bullish on where AI is going. Where traditional weather modeling and forecasts have been done using FP64, Nvidia has trained an AI model to predict weather patterns and it is claimed that it can achieve comparable accuracy in a fraction of the time, thanks to the power of deep learning.