Does this project use CPU extensions such as AVX/AVX2/AVX-512 if the CPU has those features available?

---

Reno, NV
Team: SETI.USA

Does this project use CPU extensions such as AVX/AVX2/AVX-512 if the CPU has those features available?

Yes. The app is built on libTorch (the PyTorch C++ frontend, see https://www.pytorch.org), which detects and uses all CPU features.

I am not sure how where to look at that web site to understand your answer. But I have a machine with AVX-512, and the run times are not noticeably faster. So I don't think it's being used. How to tell?

---

Reno, NV
Team: SETI.USA

I don't record that information at the moment, and I'm not sure libtorch provides an API to do so. I'll have to look into it.

It seems like the code is using avx as the intels are generally much faster than the amds. Would be great to confirm this so we can plan accordingly on allocating our computing resources accordingly.

Understood. I'm nearly certain it is, since PyTorch uses the Intel MKLDNN lib, which is hyper-optimized for AVX2/FMA/SSE/AVX512. Even on AMD (the avx512, of course).

Oh. Exactly the information I was looking for.
I think the AMD 3000 series supports avx512, so should get pretty good runtimes with my 3600 and 3700x. Amazing what 2 years does - my 1800x using 1/3 more power if not more is trounced on by the 3600.
I thought AMD was beating Intel on the AVX front, or at least keeping right up with them. I don't have anything newer than a 4770 that I want to test that theory on apart from an i7 8750h. Not a fair comparison anyway.

AVX512 is only on (select) intel CPUs, mostly high-end Xeons, and also comes with a power and speed penalty.
AVX and AVX2 are supported on all modern CPUs
AMD CPUs before Ryzen 3000 (zen2) have a small penalty vs. intel CPUs when using AVX2.

I believe the current application uses all of these where available via the underlying libraries.

Very good to know.
Been looking at the Ryzen G series processors. Should be some interesting competition there especially if rumors are to be believed and we have an 8 c/16 t varient. I'd like to build a few intel nuc sized pcs - would be great for crunching in summer.