Nvidia: “Grace is twice as efficient as AMD and Intel processors”

Grace is Nvidia’s next new processor for supercomputers and data centers. Unlike competitors AMD and Intel, Nvidia is based on the ARM architecture and not x86-64, which makes the performance comparisons particularly interesting. Now Nvidia itself has published the results of a series of comparative performance tests, with the Grace competing with AMD’s Epyc 9654 and Intel’s Xeon Platinum 8480 Plus.

Each Grace module contains dual circuits that each contain 72 Neoverse V2 cores. They’re linked by NVLink-C2, which provides up to 900GB/s of bandwidth in both directions, and up to 1TB of direct-attached LPDDR5X memory, the same as in Apple’s M processors. Therefore, Nvidia is compared against AMD and Intel systems that also have dual processors — a total of 192 Zen 4 cores with Epyc from AMD and 112 Golden Cove cores with Xeon from Intel.

All tests selected by Nvidia have been developed for supercomputers: Weather WRF, MD CP2K, Climate NEMO, CFD Openfoam and Graph Analytics Gapbs BFS. Results are presented as percentage comparisons with Epyc as an index.

In pure performance, Grace roughly equals Epyc’s double circles, with a 40 percent lead in the Graph Analytics test. The Xeon computer lags 50-70 percent behind the other two in the first four tests and is comparable to Epyc in the fifth test.

However, what Nvidia wants to focus on is performance per watt, which is measured as the total rated computing power of a data center with a total capacity of 5 megawatts. Here Grace is 70-150 percent higher than Epyc and 227-300 percent higher than Xeon. Against Epyc, the difference is greater in the Graph Analytics test and against Xeon in the MD CP2K test.