Java Throughput/Latency & Power Efficiency Tuning For AMD EPYC Turin Review

Last month I looked at the impact of AMD’s BIOS tuning guide on AI / machine learning workloads for new 5th Gen EPYC “Turin” processors. In today’s article I am looking at the performance and power efficiency impact of AMD EPYC 9005 series processors with AMD’s BIOS tuning recommendations for Java workloads on Linux.

The AMD EPYC 9005 BIOS and Workload Tuning Guide has a section dedicated to Java workloads. This includes various recommendations whether you are more concerned about latency or throughput for your Java workloads. There is also a recommended section for power efficiency settings if most concerned about delivering the greatest power efficiency on AMD EPYC Turin servers. These sections cover various power management, NUMA NPS settings, and APBDIS settings from within the BIOS. With the straight-forward recommendations, I carried out some performance/power tests following the AMD recommendations.

For this testing — like the prior AI/ML tuning guide benchmarks — I was using a Supermicro H13SSL-N retail board with the 64-core high frequency EPYC 9575F processor. The AMD EPYC 9575F has a default TDP of 400 Watts and a configurable TDP of 320 to 400 Watts. With the default TDP already being at the max cTDP, these results aren’t as impactful as some other AMD EPYC 9005 series parts when it comes to the tuning guide recommendations of resorting to “OPN Max” for the TDP/PPT power settings to maximize performance for processors where the default TDP is lower than the max cTDP. In any case this is just an initial look at BIOS tuning for Java workloads with the AMD EPYC Zen 5 platform.

Ubuntu 24.10 was running on this AMD EPYC 9575F 1P Supermicro server build. The CPU power consumption as well as the total server power consumption “wall power” were monitored during the testing to see the impact on power consumption and overall power efficiency for each of the Java benchmarks carried out. The stock OpenJDK Java 21 packaged on Ubuntu 24.10 was used for these Java performance tests.