Performance Evaluation of a Supercomputer Based on AMD Rome and Intel Cascade Lake Processors

We present a performance evaluation comparing a cluster of nodes based on the AMD EPYC 7742, a 64-core processor code-named Rome, to one based on the Intel Xeon Gold 6248, a 20-core processor code-named Cascade Lake. The Rome-based nodes have higher memory bandwidth, larger caches, more memory controllers, more cores, and support PCIe Gen4 with more lanes than the Cascade Lake nodes. We critically evaluate these new features with four industry-standard benchmarks (STREAM, HPL, HPCG, and HPCC) and 11 production scientific and engineering applications representing many disciplines: aerospace, aero engineering, climate, material designing, computational chemistry, condensed matter, life sciences, and oceanography. The performance gain of Rome over Cascade Lake on single-node runs of 9 applications ranges from 1.37-3.08x, with a geometric mean of 2.23; comparing 2016-rank runs, a Cascade Lake core has 1.08-1.39x the performance of a Rome core. The Cascade nodes result in a more balanced system, but the Rome nodes provide better per-node application performance.