I maintain a 40+ node cluster of EPYCs with MI100s, MI210s, and MI250s and after much experimenting NPS4 is all we run.
We've tested all these apps:
https://www.amd.com/en/technologies/infinity-hub
Those are all GPU workloads, and we generally tie one rank to a GPU, a set of cores and one NUMA...
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.