Using the 5th Gen EPYC BIOS tuning guide published by AMD, I recently looked at the impact of AI and machine learning optimized performance by adjusting some simple BIOS knobs as well as the Java throughput, latency and power efficiency for the EPYC 9005 class processors. In this article is following the AMD BIOS tuning guide to see what performance difference there is for high performance computing (HPC) workloads following the BIOS tuning recommendations compared to the defaults with an AMD EPYC 9575F server.
The AMD EPYC 9005 BIOS and Workload Tuning Guide is publicly available and outlines various BIOS tuning recommendations based upon the class of workloads for your intended AMD EPYC server deployment. For HPC workloads AMD recommends disabling SMT, engaging the “high performance” power profile, running in performance determinism mode, running the respective CPU at its maximum configurable TDP (cTDP) value, and running with four NUMA nodes per socket (4 NPS). The rest of the BIOS settings should be sufficient at their defaults.
In the case of this round of HPC performance tuning benchmarks for AMD EPYC 9005 series, I was using the 64-core AMD EPYC 9575F processor. This 64-core / 128-thread processor can boost up to 5.0GHz and has a 400 Watt default TDP. The configurable TDP rating on the EPYC 9575F is 320 to 400 Watts, thus running at OPN max is already the same 400 Watts… For some AMD EPYC 9005 SKUs where the default TDP isn’t already at the top-end of the cTDP, there will be more headroom for performance in following the AMD BIOS tuning guide – so just keep that in mind depending upon your EPYC Turin processor.
For this testing the AMD EPYC 9575F was running with the Supermicro H13SSL-N retail EPYC 9004/9005 server motherboard. The SilverStone XE360-SP5 AIO liquid cooling was keeping this frequency-optimized Zen 5 server processor running efficiently across all of the benchmarking.
From this 64-core AMD EPYC Zen 5 Supermicro server with Ubuntu Linux, the “default” run was with all the BIOS defaults and then the “HPC Tuning Recommendations” run was when making use of all the recommended AMD BIOS tunables for HPC workloads. In addition to looking at the raw performance, the overall 1P server wall power consumption was monitored along with the CPU power consumption of the EPYC 9575F. It’s quite a straight-forward comparison as with the other tuning benchmarks so let’s get straight to the data.
