Hi all, longtime lurker seeking some advice for a new(er) workstation within a reasonable budget. Below I’ve listed what I have now and the reasons why I’m looking at a (budget) EPYC build. I’m looking for any feedback on what a good CPU pick would be for my needs.
Current build:
I’ve outgrown my current X399 workstation and it’s also showing its age in various ways. I’ve run out of PCIe slots on X399, the board doesn’t play nice with more than 128gb memory, and the whole platform is a bit less stable than I’d like - couldn’t tell you how many times I’ve had to adjust timings and voltages or reseat the CPU over the past 6 years.
The current x399 system looks like this:
2950x, 128gb @ 2400
12TB spread across various SSDs
RTX A6000
2x RTX A4000
10gbe (X540 card)
1300w PSU
Dual boot with Win10 and whatever flavor of Linux I prefer on a given year
I run VMs and small test environments, do design work, video work, run local LLMs (and dabble in training), some rendering (albeit rarely), and as a daily driver.
Moving to EPYC?
Seeking any suggestions, but I’m mostly settled on a ROMED8-2T (assuming I am moving to EPYC). I’m eyeing it and a (which?) 3rd gen EPYC/Milan for the following reasons:
More PCIe (vs x399)
Stability
ReBar support (ROMED8-2T)
12x11" size fits spare cases I already have (ROMED8-2T)
I already have a pile of TR4/SP3 coolers
I already have a lot of 32gb RDIMMS (albiet 2666) that I can “borrow” from one of my servers.
Which EPYC makes sense (if any)?
Bottom line: I’ve never used EPYC in a workstation - love them at work (for VM hosts), but this is a very different use case. I’ve looked up so many obscure reviews and performance figures that I am firmly into analysis paralysis -mode. A big concern, one that I cannot answer, is whether a reasonably-priced 64 core (like a 7713 or 7C13) represents a big performance regression for any single/few threaded use cases? Milan is much newer than my trusty 2950X, but those clock speeds are awfully low under load (2.0ghz iirc). If I really want this as both a workstation and ‘daily driver’, should I avoid these and only look at the “F” models with higher clockspeed (or a lower core count / 7443)? Does the price premium for the Milan X models translate to a real difference? Is all of this a fool’s errand to be spending $2k on 3rd gen Epyc for a workstation (maybe don’t answer that one )?
Any sage wisdom, firsthand experience, and/or public ridicule is welcome
Following. Im learning to use the platform and i have interest in buildi g a WS soon too. If theres a better way to follow topics on mobile than commenitng im all ears. Good luck.
i would look for something in the 75x3 series. that would be pretty close single threaded performance with clock and IPC uplift, and a lot better multicore performance still.
EPYC makes a great workstation, but it is all budget vs performance calcs.
You’ll need to run Windows Server for driver support
single core performance between the 7713, 7c13, and 2950x is identical though the 7003 series will be substantially more power efficient given the performance and will run circles around your 2950x.
Milan X is less power efficient.
Make sure to populate all RAM channels with 1 DIMM and enjoy.
While I haven’t had an opportunity to use a higher clocked Epyc, from my experience with a 280W 64c I would avoid. With Milan, server 2022 is a must, but desktop experience for me was not very pleasant lol. Even with a faster chip, boot time is horrible. I’d recommend Threadripper honestly, I hear the Dell Precision is quite nice.
If you don’t turn it off you don’t have to worry about boot time. I’ve run two Epyc workstations for a couple of years now doing distributed computing on BOINC science projects. I’ve run both 240W and 280W processors, both Rome and Milan generations with no issues. 128GB of memory installed in all 8 slots always.
I never timed the boot from cold off before to Desktop, but I would bet its under 3 minutes max.
If you want high core clocks, either choose the high-clock parts or just run half or less of your cores and the clocks will boost up to the max boost under the available power budget with less than all cores occupied with work.
I appreciate the feedback. Was the main problem the long boot time, or was the workstation/desktop performance generally bad for you with Server 2022 and the 64 core? I ask since that is the planned setup I am targeting too.
This is excellent info, thank you. Efficiency / lack thereof is definitely a consideration and one of the reasons I was looking at the 7c13 in the first place.
If you want high core clocks, either choose the high-clock parts or just run half or less of your cores and the clocks will boost up to the max boost under the available power budget with less than all cores occupied with work.
I had been wondering about this. So if I were to turn off just a few cores I could, in theory, get on the higher end of the boost clocks with a lower TDP part like the 7c13 (225w) or 7713 (240w) - effectively turning them into a 7543 or similar?
Yes, depends on what kind of work you are running. If your applications allow you to limit the number of cores deployed in the app, that would be the ticket.
Or another way of reducing the number of cores deployed is to just turn them off in the /proc/cpuinfo interface.
You can run a script to turn off cores. For example on a 128 thread Epyc, you could do this to turn off half of them.
for x in /sys/devices/system/cpu/cpu[63-127]*/online; do
echo 0 >“$x”
and then check to see what cores are online afterwards with:
cat /sys/devices/system/cpu/online
Or you could just turn off SMT in the BIOS. Scripting 25% of the cores offline would allow the 75% online to boost higher because they get to divvy up the complete cTDP power budget instead of spreading that around to 100% of the cores.
Thank you for all the helpful info, the 64c definitely made the most sense in light of all this - I ordered a 7C13 and it should be arriving soon. I’ll have to post a build log once all the bits and pieces arrive.
)?