AVX-512 Performance With 256-bit vs. 512-bit Data Path For AMD EPYC 9005 CPUs
Running with the AVX-512 FP512 default mode didn't show any hit to the CPU peak frequency being achieved nor did it when operating in the 256-bit data path configuration.
There was also no significant difference in the CPU power consumption with either the AVX-512 FP256 or FP512 modes. The AMD EPYC 9755 across the assortment of benchmarks was consuming 292~305 Watts on average with a peak of 502~505 Watts, right in line with the default TDP rating.
There was also no real difference in the AMD EPYC 9755 core temperature across the different AVX-512 modes... Much better than the early days of AVX-512 on Intel Xeon where there was a significant impact to power/thermals from leveraging Advanced Vector Extensions 512. The AMD EPYC 9755 with the liquid cooling within the AMD Volcano reference server allowed the CPU to operate at 49~50 degrees (C) on average during these AVX-512 benchmarks and peaked at 63~67 degrees during the most harsh workloads.
When taking the geometric mean of all the raw AVX-512 performance benchmarks, AVX-512 in the default FP512 configuration yielded 1.45x the performance compared to disabling AVX-512 outright. Having the 512-bit data path allowed for 1.12x the performance compared to running the EPYC 9755 processor in the 256-bit data path mode, similar to how AVX-512 operates with Zen 4.
So all-in these latest benchmarks do show that the 512-bit data path mode for AVX-512 on AMD Zen 5 is valuable for servers/HPC and like with the FP256 mode there isn't any measurable impact to the CPU power consumption / clock frequencies / thermals. Similar to what I've seen with the Ryzen 9000 series desktop processors, AVX-512 with AMD Zen 5 appears to be very performant and robust. With AMD EPYC 9005 series for HPC servers this 512-bit data path is likely to be quite rewarding and a good design choice by AMD. The timing is important as well with increasing AVX-512 usage outside of HPC to speedy JSON parsing, AVX-512 optimizations for PostgreSQL, and other more conventional workloads.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.