Radeon on Epyc - Windows 10 bare metal install

Hi everyone, I want to replace a Ryzen 3600 Windows 10 desktop with a high frequency Epyc Rome or Milan build based on 7F52 (16 cores) or 7F72 (24 cores) or Milan equivalent β€” also running Windows 10.

GPU is a Sapphire Radeon 5700 XT running four displays with occasional gaming.

Will be running WMware Workstation continuously in the background and connecting VMs to a vsphere domain running on another Epyc Rome server.

Would prefer to run Epyc over a Ryzen 5950X or 7950X or Threadripper 3000 series.

Ryzen has no ECC RDIMM and Threadripper seems like a poor value proposition. Epyc used ex-datacentre is great value and quality when sourced from a good supplier who tests well and has good QA. Curated used enterprise gear has typically been run with excellent power conditioning, good PSUs, good temp control, dust free, and typically fails in the first 3-4 years resulting in tried and tested refreshed components.

There are no more new X570 motherboards available for 5950X and X670 is a hot mess with the same disgraceful IO limitations as all previous generations.

Threadripper Pro 3960X is expensive for limited performance on desktop and has limited RDIMM support compared to Epyc.

Anyone have experience with Radeon GPUs on Epyc? Thanks.

I had no issues with a Radeon Pro Duo (Polaris) in my Epyc 7702 / 7713P system but that’s a bit older than your card.

Before you pull the trigger on those 16-24 core parts be aware that they are somewhat handicapped. Whether or not memory bandwidth matters is down to your individual use case.

Thanks for pointing that out about the 4-channel optimized parts. Explains their lower cache size and dramatically lower GB/sec memory bandwidth.

I’ll drop my 5700 XT into the Epyc server (H12SSL-i) and passthrough to a windows VM.

With DDR4 3200 a 4-channel part might specify approx 85GB/sec and an 8-channel part approx 205GB/sec. The 7F52 and 7F72 parts have 204.8GB/sec.

Interestingly, in a StorageReview comparison, the 7F52 exhibits only slightly better performance than the 7302p across five tests:

The 7302p is a cheaper SKU, is more common and has a much lower TDP and power draw. Not sure the 7F52 is going to bring any benefit to the desktop.

While neither seem to have a high enough frequency for the modern desktop they may not need to. My current desktop is a Ryzen 3600 which infrequently spikes above 50%.

Boost clock seems to be less important than cores and memory for multiple concurrent moderate workloads β€” persistent VMware Workstation VMs, a lot of browser windows, occasional game and a bunch of other applications.

1 Like