Following the release last month of the [1]EPYC 9005 series processors, AMD published a BIOS and Workload Tuning Guide of straight-forward settings recommendations for those running new EPYC Turin servers to optimize the performance of different workloads like databases and Java to HPC and AI/ML software. Recently I started running some benchmarks to look at the impact of AMD's recommended BIOS tuning and beginning this comparison by looking at the performance (and power) impact across a range of AI / machine learning workloads on a 5th Gen AMD EPYC server.

The AMD EPYC 9005 BIOS & Workload Tuning Guide is to help those deploying new EPYC Turin servers for specific purposes to cater the platform for their targeted workloads. This guide is not about any hand-optimizing of the code, operating system optimizations, or other tuning but is focused on the BIOS option adjustments depending upon your targeted workload.

Those interested can find the AMD EPYC 9005 BIOS & Workload Tuning Guide at [2]AMD.com . Among the AMD recommendations are around optimal CPU/power settings for different workloads, NUMA and memory settings, Infinity Fabric behavior, and also covering I/O and virtualization settings, among other BIOS tunables.

[3]

This AMD BIOS tuning guide for EPYC Turin is straight-forward and nicely split-up for different workloads. So I figured it would be fun for some interesting benchmarks on Phoronix to compare the difference of the default BIOS against AMD's tuning recommendations. In this article I kicked things off with a variety of artificial intelligence (AI) and machine learning (ML) Linux benchmarks at the default BIOS configuration and with the AI/ML tuning recommendations.

For making it representative of end-user experiences, rather than using the AMD reference server platform I used the recently-assembled [4]Supermicro H13SSL-N Socket SP5 motherboard with EPYC 9005 series support. The 96-core [5]EPYC 9655 processor was used with this 1P Supermicro server platform for conducting this BIOS tuning comparison.

[6]

The Supermicro H13SSL-N with EPYC 9655 was tested with 12 x 64GB Micron DDR5-6000 MTC40F2046S1RC64BDY memory, 3.2TB Micron 7450 MTFDKCB3T2TFS NVMe SSD, and running Ubuntu 24.10 with an up-to-date software stack with GCC 14.2 and manually upgrading to the Linux 6.12 kernel. The acpi-cpufreq performance governor was used throughout all testing. The only changes between runs for this testing was running with the stock/default BIOS settings on this Supermicro EPYC motherboard and then repeating the tests after applying all of the recommended AMD BIOS options for AI/ML workloads.



[1] https://www.phoronix.com/search/EPYC+9005

[2] https://www.amd.com/content/dam/amd/en/documents/epyc-technical-docs/tuning-guides/58467_amd-epyc-9005-tg-bios-and-workload.pdf

[3] https://www.phoronix.com/image-viewer.php?id=amd-epyc-ai-ml-tuning&image=amd_ai_tuning_2_lrg

[4] https://www.phoronix.com/review/supermicro-h13ssln-epyc-turin

[5] https://www.phoronix.com/review/amd-epyc-9655

[6] https://www.phoronix.com/image-viewer.php?id=amd-epyc-ai-ml-tuning&image=amd_ai_tuning_3_lrg