Been planning a dev/ML workload server for a few weeks, with some help of the folks here (thank you!) and almost ordered an WRX80 + 5995wx setup, when a buddy of mine brought up : “wait, but if you are using it as a server, headless, with a hypervisor layer, why not go the EPYC route ?”
My response was " 1. I don’t know, 2. arent these things terrible to cool in a home office tower setting?" but I left it there, and decide to come here and get some further advice.
So for a devops/ml system that doesn’t have any toys in it other than nvmes and GPUs, and is meant for remote development and ML workloads, and some DB stuff… would you and why go any of these routes ?
EPYC Milan-X 7773x + whatever is the right board (help me here please)
ThreatRipper Pro 5995wx + WRX80 creator/sage
for the possible “why 64 cores and will you saturate all ?” question, I would say … I usually build systems with the intention to keep using them with minimal upgrades for ~5+ years… also, i don’t overclock.
One reason I was thinking of getting the 5995wx is that i can always decide at a later stage to make it into a workstation where a server grade board/cpu would not be great… there is also a slight price benefit right now to get epyc over top, but this computer helps me make money… so its not about savings … 512gb ecc btw. if that matters.
I don’t think I’ve missed something here. I mean, I guess servers don’t typically give respectable fan control… but other than that, what’s it matter?
My SuperMicro server board seems like it has it’s own buggy bullshit going on, but that’s just SuperMicro’s incompetence that era or with this specific model/revision (or in general for all I know). My intel server board… features not bugs. Does the things the decent documentation promises, except IPMI as far as I can tell.
What is your power budget and do you plan on doing any underclocking should be your question …
Epyc is not meant to slow down, so it doesn’t do any Sx suspend/sleep and the ‘all cores slow’ setting on my 7402P is 1500Mhz/130W at idle, the 7773 with 64 cores will idle at or around 200W+, methinks
As for cooling, the GPUs will generate far more heat than the processor … depending on the MB you choose you may be ending up with a weird orientation of the SP4 cooler, but that has never caused any cooling issues to me so far, using an NH14S-TR4, temps never go over 60C.
Fans on my Asrock Rack can be controlled from the IPMI interface, and honor PWM
It’s the ages old question, which you can only answer for yourself: price optimized or upgrade ability. You’re maxing out a dead end platform. That isn’t necessarily something bad by itself, but something to consider.
With Genoa you probably could probably upgrade to whatever the next generation after it if you needed to. Or more cores. Those CPUs go up to 96 cores, with 12 memory channels. And with the differences between Zen3 and Zen4, 48c Genoa seems like a sane comparison to 64c Milan.
Summing up, whether to get Genoa or Milan is up to you, I just wanted to point out that Genoa is releasing tomorrow.
I would also say that if Zen4 is launching tomorrow, non OEM motherboards and sku availability will happen in a year or so … at least that’s what happened with Milan …
@jaskij EPYC Zen 4 won’t be available on Thursday, November 10. AMD is only announcing what they will have in the Zen 4 EPYC line on Thursday, November 10. Also, EPYC Zen 4 will only be available as a complete system meaning the EPYC Zen 4 won’t be available as a separate part. EPYC Zen 2 will reach EOL (end of life) on December 31. I believe EPYC Zen 4 complete systems will become available in the first or second quarter of 2023. Maybe @sectorix can purchase an EPYC Zen 3 for a DIY build sometime in 2023.
