[SOLVED] Since MKL is not optimized for AMD hardware isn't it better to have an Intel CPU for deep learning rig?

[s_quintanar]

In this article https://timdettmers.com/2023/01/16/which-gpu-for-deep-learning/ published in 2023-01-16 by Tim Dettmers, he states "AMD CPUs are cheaper and better than Intel CPUs in general for deep learning.".

I have chosen parts for a future 2023 deep learning rig build based on Intel only because I want to avoid the "MKL is not optimized for AMD hardware" issue. Now Mr. Dettmer claims AMD CPUs are better?!

Ok, I'm happy to pay less for better, but what about the MKL issue? Has this issue disappeared?

 

[Karadjgne]

The article is mostly about GPU deep learning and the advantage tensor cores have over matrix as far as latency goes. If he's saying that amd has a smoother range of end to end system latency where Intel has a more sporadic range, then amd will be better overall, especially if you can move up a class in cpu and still be cheaper than Intel. He also threw in the words 'in general' as applies overall to several classes of cpus, not just a specific cpu.

MKL isn't optimized for Amd, because it's basically Intels, not freeware. Intel isn't about to make it optimized for Amd too.

If you aren't deep learning for a year or more, you are wasting your time, you might get slightly better results with Intel flagships, but choosing that over the 65w Ryzen electric bill is somewhat foolish waste of money considering the length of time used. Context with statements apply.