AMD 4th Gen EPYC 9654 "Genoa" AVX-512 Performance Analysis
Written by Michael Larabel in Processors on 19 December 2022 at 10:00 AM EST. Page 3 of 9. 65 Comments.
TensorFlow with ResNet-50 running on the AMD EPYC 9654 processors saw 1.73x the performance when AVX-512 was enabled compared to the forced-off baseline.
The CPU power consumption of the EPYC 9654 2P processors with AVX-512 enabled even was slightly more efficient than without AVX-512 and translated to having nearly twice the performance-per-Watt.
TensorFlow with AlexNet saw nearly a 3x advantage from AVX-512 on 4th Gen EPYC.
TensorFlow with GoogLeNet also saw compelling results with Zen 4's AVX-512 usage.
Leela Chess Zero as an AI-driven chess engine saw smaller but still practical improvements to both raw performance and performance-per-Watt with AVX-512 enabled for the flagship AMD Genoa processors.