AMD EPYC 9684X Genoa-X Provides Incredible HPC Performance

Of the benchmarks run focused on the HPC/AI workloads, the EPYC 9684X processor easily stole the show. Even the EPYC 9684X 2P in its default (performance determinism mode) outperformed the EPYC 9654 2P running in the 400W cTDP and power determinism mode. Thanks to AVX-512, DDR5 memory. 12 channel memory, and other Zen 4 upgrades, Genoa-X is a tremendous upgrade over Milan-X. For those normally waiting 2+ generations at least between HPC server upgrades, Milan-X to Genoa-X is already significant enough to warrant an upgrade for relevant workloads. This geometric mean is also limited to the tests where all processors could run -- including the Xeon Max 9480 2P in HBM-only mode. See more of the raw AMD EPYC 9684X benchmarking data here.

In some workloads the Xeon Max 9480 2P configuration of Sapphire Rapids HBM2e had a nice showing when running in the HBM2e-only mode with terrific uplift over the HBM2e+DDR5 caching mode. However, the main bottleneck there is ensuring your workloads can fit within 64GB HBM2e per socket / less than 2GB per core with the Xeon Max 9480. Thus for a number of the benchmarks used I wasn't able to successfully run them with the combined 112 cores fighting for a total of 128GB of HBM2e. Xeon Max shows much potential but until it can compete better in core count and allowing more HBM2e-per-core, Genoa-X is ultimately better positioned and can provide up to 96 cores per socket. Or for highly parallelized workloads like CPU-based rendering and other workloads shown in the EPYC 9754 benchmarking, there is now Bergamo with up to 128 cores / 256 threads per socket. The competitiveness of Xeon Max 9480 though was fascinating considering the 56 core limitation of this initial Xeon Max generation. The ace for Intel Xeon Max is workloads fitting within the HBM2e size constraints and supporting AMX. For AI workloads able to make use of Advanced Matrix Extensions, this new capability with Sapphire Rapids has shown to help level the playing field even with the core count disadvantage. However, software support around AMX is still rather limited outside of some key software libraries nor is its presence relevant to all areas of high performance computing.

Generationally I continue to be in awe at the uplift from Milan-X to Genoa-X thanks in large part to Zen 4 with AVX-512 paired with AMD 3D V-Cache. Genoa has been -- and continues to be -- an excellent server processor from my months of benchmarking it while now I am starstruck by Genoa-X with so quickly advancing the HPC performance even further. Like with Milan-X, Genoa-X also leverages the existing hardware and software ecosystem laid out by Genoa over the months prior. With running the latest system BIOS/firmware you are good to go with dropping Genoa-X CPUs into existing AMD SP5 server motherboards. The Linux support also is in good shape on the latest upstreams as well as the major enterprise Linux distributions.

For those looking to maximize their HPC performance right now, the AMD EPYC 9684X is easily one of the best processors I've looked at in the past 19 years of Phoronix. The larger L3 cache paired with AVX-512, 12 channel DDR5 memory, and other Zen 4 benefits can really deliver phenomenal results in a wide range of high performance computing workloads.

Thanks to AMD for providing the EPYC 9684X processors for launch-day testing on Phoronix. Over the days ahead will be some additional follow-up Genoa-X benchmarking looking more closely at the 3D V-Cache impact and other areas along with more Bergamo testing too.