The first supercomputer based on Nvidia’s H100 computing GPU has yet to set records in terms of absolute performance, but it has already shown its capabilities in terms of performance per watt.
Henri of Flatiron Labs (opens in new tab) The supercomputer, based on Intel’s Xeon Platinum 8362 (Ice Lake) and accelerated by Nvidia’s H100 computing GPU, made the Top500 and green 500 (opens in new tab) list. Additionally, he ran AMD’s EPYC and Instinct MI250X hardware from the top of the Green500 list, dismissing AMD-powered Frontier test and development systems.
Lenovo built the Henri supercomputer. This is currently his 405th most powerful system on the Top500 list, with an Rmax performance of 2.04 FP64 PFLOPS, which is hardly impressive on its own. It’s worth noting that the machine draws only 31 kW, a world record he shows an energy efficiency of 65.091 GFLOPS/Watt. To put the numbers into context, the Frontier TDS machine reaches 62.684 PFLOPS/W, while Frontier (the world’s fastest supercomputer) scores 52.227 PFLOPS/W, while the Lumi system achieves 58.021 PFLOPS/W.
The Henri machine is a relatively simple supercomputer by today’s standards. I’m using an off-the-shelf air-cooled ThinkSystem SR670 V2 server from Lenovo with 32-core Xeon Platinum 8362 processors from Intel (5,920 total cores) and 80 H100 80GB PCIe cards from Nvidia. About the hopper architecture. Of course, using air cooling for smaller systems can have other implications for performance per watt results. However, Nvidia’s latest computing GPUs generally offer good performance.
Ian Fisk, co-director of the Scientific Computing Core at the Flatiron Institute, said: “It’s a workhorse machine, encouraging researchers to try new things and make discoveries. […] [It offers] Very capable and very efficient, but not particularly quirky. It only took a few people to load the system. This kind of efficiency is now available to more groups than just the largest supercomputing centers. ”
Let’s face it, Nvidia-based supercomputers (often made up of standard servers) have been the performance-per-watt champions of the Green500 list for some time.
AMD EPYC and Instinct MI250X based machines, on the other hand, are no outsiders in terms of performance per watt metrics, especially considering the scale of AMD’s tech-powered Frontier, Lumi, AdAstra, Setonix, and Dardel machines. Additionally, six of the top 10 supercomputers on the Green500 list use AMD CPUs and GPUs, three are accelerated by his Nvidia computing GPUs, and one is Intel’s Xeon Platinum 8260M-based. I am using node.