I have code that utilizes numba prange extensively. I’m able to achieve near-complete compute saturation on all my CPU threads when leveraging Intel MKL. I have tested my code with OpenBlas and am not able to achieve the same performance unfortunately.
My question to the community:
I’m thinking about investing in a 32-or 64-core AMD Threadripper system to significantly speed up my computation time. I’ve also read that at one time there was a crippling of MKL on AMD CPU’s, a provided workaround flag, and subsequent removal of said flag by Intel. All of this documentation on the web is nearly 3 to 4 years old, and I don’t see much that has been written since then. I’m wondering if anyone in the community has recent experience running numba on AMD CPUs (Zen 3 architecture) with Intel MKL (specifically leveraging prange). Any noted experience/suggestions would be greatly appreciated.
Thanks in advance for any help/guidance.