GPU-accelerated supercomputers reshape TOP500 world rankings of fastest computers.

For the first time in history, 56 percent of the flops added to the TOP500 list came from GPUs instead of CPUs, with this shift exemplified by three new systems: Summit, which dethroned the 93-petaflop Sunway TaihuLight as the number one system with a Linpack score of 122.3 petaflops; Sierra, now ranked third fastest at 71.6 Linpack petaflops; and Japan's AI Bridging Cloud Infrastructure (ABCI), ranked fifth at 19.9 Linpack petaflops.

Summit is powered by IBM servers equipped with two Power9 CPUs and six V100 GPUs each, with its 27,686 GPUs delivering 95 percent of the system's 187.7 petaflops peak performance. Sierra employs a similar architecture with four V100 GPUs per dual-socket Power9 node, with 17,280 GPUs representing the lion's share of the system's performance. ABCI pairs two Intel Xeon Gold CPUs with four V100 GPUs per server, with 4,352 V100s delivering the vast majority of its 19.9 Linpack petaflops.

The 56 percent figure likely understates GPU adoption in HPC markets. According to Ian Buck, vice president of NVIDIA's Accelerated Computing business unit, more than half the Tesla GPUs sold into the HPC/AI/data analytics space are purchased by customers who never submit systems for TOP500 consideration, either unconcerned with the rankings or unwilling to advertise their hardware-buying habits to competitors. The three new systems represent a significant milestone for AI and HPC convergence, as Tensor Cores in the V100 GPUs with specialized 16-bit matrix math capability endow them with unprecedented deep learning potential. Summit alone boasts over three peak exaflops of deep learning performance, Sierra approximately two peak exaflops, and ABCI around half an exaflop—together representing more deep learning capability than the other 497 systems on the TOP500 list combined. As TOP500 author Jack Dongarra, Professor at University of Tennessee and Oak Ridge National Lab, observed, "This year's TOP500 list represents a clear shift towards systems that support both HPC and AI computing."

While companies like Intel, Google, Fujitsu, Wave Computing, and Graphcore are developing specialized deep learning accelerators, NVIDIA's integrated AI-HPC design for its Tesla GPU line is proving advantageous as machine learning increasingly accelerates traditional HPC applications across weather forecasting, financial analytics, genomics, and oil & gas exploration. Although Buck acknowledges this interplay between traditional HPC modeling and machine learning is still in its earliest stages, he maintains "it's only going to get more intertwined," arguing that the benefits of supporting 64-bit HPC, machine learning, and visualization on the same chip far outweigh advantages of single-purpose accelerators. With 554 codes ported to V100s—including all of the top 15 HPC applications—and V100s capable of seven double-precision teraflops for conventional supercomputing, mixed-workload machines are now becoming routine, from MareNostrum at Barcelona Supercomputing Centre adding three racks of Power9/V100 nodes to Nimbus at Pawsey Supercomputing Centre adding 12 V100 GPUs.