AVX-512 Performance + Power Efficiency Shines With AMD Strix Halo
Kicking things off with some AVX-512-enabled CPU-based AI workloads, first up was the Intel oneDNN library that provides the building blocks for various neural network applications.
AVX-512 support really helps accelerate AI workloads on Strix Halo, in addition to the Ryzen AI NPU and 40 core Radeon 8060S graphics.
Having AVX-512 with Zen 5 really helps with AI workloads, compared to Intel Core processors lacking AVX-512. Unlike the early AVX-512 implementations on the Intel side, AVX-512 with Strix Halo was coming without any added power/thermal costs. Similar to the very energy efficient AVX-512 I've demonstrated with other Zen 4/5 desktop and server processors, with Strix Halo having AVX-512 makes for very capable performance without any increased power use.
AVX-512 paired with 16c/32t Zen 5 cores, optional NPU, and the 40 CU Radeon integrated graphics really makes for a slam dunk as a very well rounded platform for AI.
PyTorch also did a good job showing off the impact of AVX-512 on Strix Halo.
Simply toggling AVX-512 and repeating the tests from this HP ZBook Ultra G1a shows the enormous uplift provided by AVX-512 and without any additional power use.