The supercomputer arms race is on, and NVIDIA (NVDA -2.30%) is leaving the competition in the dust. Of the world's 500 most powerful supercomputers, NVIDIA graphics processing units (GPUs) or networking equipment (by way of its Mellanox acquisition) can be found in 333 of them.
The title of most powerful is currently claimed by Japan's Fugaku, built with Fujitsu (FJTSY -1.04%) CPUs, but it is set to be surpassed by other systems in 2021. Eventually, however, all of them will be dethroned by Italy's NVIDIA GPU-powered Leonardo once it is built.
Pulling away from the pack
In an announcement today, NVIDIA says Leonardo will house over 14,000 of its data center GPUs, all stitched together using Mellanox networking hardware. In terms of capabilities for a supercomputer (that is, a really powerful data center), Leonardo will reportedly be able to handle a max of 10 exaflops of performance for high-performance computing tasks like AI. For reference, Fugaku delivers about 2 exaflops in comparable performance. One exaflop is equal to 1,000 petaflops and represents one quintillion floating-point operations per second. Suffice to say this will be a powerful computing system.
Other supercomputers were homing in on the one-exaflop benchmark before NVIDIA's announcement. The U.S. Department of Energy's (DoE) Aurora, built with Intel's (INTC -2.00%) hardware, will max out at one exaflop. And the DoE's Frontier, using AMD (AMD -2.44%) chips, will be able to handle 1.5 exaflops.
Both are set to begin operating next year. But Leonardo will be a monumental leap as data centers become a basic computing unit of the future. Italy's CINECA -- a tech consortium serving universities and other researchers -- is building Leonardo, along with three other NVIDIA-based supercomputers in the Czech Republic, Luxembourg, and Slovenia (with four more systems planned in Bulgaria, Finland, Portugal and Spain).
If the claimed computing power can be delivered, Leonardo will illustrate the power of NVIDIA's GPU chips, which are quickly becoming a standard in high-performance systems needed for AI applications. Because they can compute massive amounts of data in shorter periods of time, the GPUs -- long a staple of high-end video game graphics -- are well equipped to train and then deploy large AI algorithms and software. Leonardo will be used for medical research, space exploration, and weather forecasting.