AMD EPYC 9684X Genoa-X Provides Incredible HPC Performance

Written by Michael Larabel in Processors on 19 July 2023 at 09:00 AM EDT. Page 2 of 8. 17 Comments.

The processors tested for this initial Genoa-X comparison include:

- AMD EPYC 7773X 2P Milan-X
- AMD EPYC 9554 2P Genoa
- AMD EPYC 9654 2P Genoa
- AMD EPYC 9754 2P Bergamo
- AMD EPYC 9684X 2P Genoa-X
- Intel Xeon Max 9480 2P Sapphire Rapids HBM2e (DDR5 + HBM2e Caching Mode)
- Intel Xeon Max 9480 2P Sapphire Rapids HBM2e (HBM2e Only Mode / No DDR5)
- Intel Xeon Platinum 8490H 2P Sapphire Rapids

All CPUs were tested on air cooling. The AMD EPYC Genoa(X) and Bergamo processors were tested both in the default performance determinism mode as well as in the power determinism mode and 400W cTDP set from the BIOS.

AMD EPYC 9684X 2P Genoa-X Benchmarks

All benchmarks occurred on Ubuntu 23.04 with the Linux 6.2 kernel while also making use of the GCC 13 compiler. During all CPU tests the "performance" CPU frequency scaling governor was used.

OpenFOAM benchmark with settings of Input: drivaerFastback, Medium Mesh Size, Execution Time. EPYC 9684X 2P - Power 400W was the fastest.

Right off the bat with the OpenFOAM leading open-source computational fluid dynamics software, we see the performance of Genoa-X shine with a noticeable advantage over the 96-core EPYC 9654 Genoa processor. With the small clock speed differences and then the large L3 cache via AMD 3D V-Cache allowed for much faster performance. Genoa-X was also much faster than both the Xeon Platinum 8490H and the new Xeon Max with HBM2e in caching mode. Unfortunately for this test (among others) the Xeon Max 9480 2P failed to run in HBM-only mode due to the less than 2GB of HBM2e (RAM) per core. The Xeon Max HBM-only mode can do wonders when your workload is able to run within the 64GB of HBM2e per socket, but if exceeding that especially for the parallelized workloads can suffer from memory pressure or other problematic behavior due to the memory constraints. The EPYC 9684X delivers a big leap in performance over the prior-generation Milan-X EPYC 7773X flagship.

OpenFOAM benchmark with settings of Input: drivaerFastback, Medium Mesh Size, Execution Time. EPYC 9684X 2P - Power 400W was the fastest.

The EPYC 9684X 2P together was pulling 622 Watts on average with a peak of 655 Watts, which was actually lower than the EPYC 9654 2P results. This was also better than the Xeon Platinum 8490H 2P Sapphire Rapids results.

WRF benchmark with settings of Input: conus 2.5km. EPYC 9684X 2P - Power 400W was the fastest.

The AMD EPYC 9684X processors also performed very well with the WRF weather forecasting software to deliver measurable time savings compared to Genoa or Sapphire Rapids. WRF with the conus 2.6km model is another case where the Xeon Max 9480 in HBM-only mode didn't have enough HBM2e to function and thus the HBM-only metric.

WRF benchmark with settings of Input: conus 2.5km. EPYC 9684X 2P - Power 400W was the fastest.

The EPYC 9684X out-of-the-box was pulling less power still than the Xeon Platinum 8490H 2P except when running in the 400W power determinism mode.

miniFE benchmark with settings of Problem Size: Small. EPYC 9684X 2P was the fastest.
miniFE benchmark with settings of Problem Size: Small. EPYC 9684X 2P was the fastest.
miniFE benchmark with settings of Problem Size: Small. EPYC 9684X 2P was the fastest.

With the miniFE finite element analysis benchmark the Genoa-X pair was demonstrating great performance-per-Watt while Bergamo with its power-efficient Zen 4C cores led.


Related Articles