NVIDIA at the Supercomputing 2017 conference announced a major upgrade of its new SaturnV AI supercomputer, which when complete, the company claims, will be not just one of the world's top-10 AI supercomputers in terms of raw compute power; but will also the world's most energy-efficient. The SaturnV will be a cluster supercomputer with 660 NVIDIA DGX-1 nodes. Each such node packs eight NVIDIA GV100 GPUs, which takes the machine's total GPU count to a staggering 5,280 (that's GPUs, not CUDA cores). They add up to an FP16 performance that's scraping the ExaFLOP (1,000-petaFLOP or 10^18 FLOP/s) barrier; while its FP64 (double-precision) compute performance nears 40 petaFLOP/s (40,000 TFLOP/s).

SaturnV should beat Summit, a supercomputer being co-developed by NVIDIA and IBM, which in turn should unseat Sunway TaihuLight, that's currently the world's fastest supercomputer. This feat gains prominence as NVIDIA SaturnV and NVIDIA+IBM Summit are both machines built by the American private-sector, which are trying to beat a supercomputing leader backed by the mighty Chinese exchequer. The other claim to fame of SaturnV is its energy-efficiency. Before its upgrade, SaturnV achieved an energy-efficiency of a staggering 15.1 GFLOP/s per Watt, which was already the fourth "greenest." NVIDIA expects the upgraded SaturnV to take the number-one spot.SaturnV powered by "Volta" and NVIDIA-IBM Summit aren't the only "Volta" powered giants. Japan's ABCI supercomputer was recently infused with 4,352 GV100 GPUs, taking its peak FP64 throughput to 37 petaFLOP/s. It's the first exaFLOP-scale supercomputer in the world (with 0.55 exaFLOP peak FP16 throughput).

This dash to the top with incredibly powerful GPUs has more than bragging rights at stake. Each of the top-10 supercomputers in business is taking in time orders with leading AI (artificial intelligence) research teams. China has its own state-backed AI projects that require the time and resources of Sunway TaihuLight and other state-owned supercomputers in its league. Leading supercomputing companies Stateside are selling their own supercomputer time to AI research clients.