As we all know, Intels MKL is still playing this funny game and falls back to using the SSE Codepath instead of AVX2 if the vendorstring of the CPU is AMD. This is of particular horror, if you are using Matlab. So now I came across this in the www: Note that by default, PyTorch uses the Intel MKL, that gimps AMD processors. In order to prevent that, execute those lines before starting the benchmark: "export MKL_DEBUG_CPU_TYPE=5" You can find many of these if you google for it, not only for PyTorch. Apparently, this is an undocumented Debug Mode that forces the MKL to use AVX2 and overrides the vendor string result. Any of you cracks got a https://9apps.ooo n idea how to test this in Matlab? It would surely help many users out there.