Epyc Build

Andrew_Carr

2[H]4U
Joined
Feb 26, 2005
Messages
2,777
So.... I saw someone selling a 7J13 AMD Epyc on ebay and couldn't resist. It seems to be basically a an Epyc 7763 chip built for OEMs. 64 cores, base clock of 2.45GHz, boosts to 3.52GHz, based on Milan (Zen 3). NKD hooked me up with a new GPU earlier so I decided to make a "workstation" for fun. Not sure if I'll keep it like this permanently or switch it out for something more sensible, but for now I'm playing around with benchmarks and such.


The build didn't start off great. The used 1600W T2 I bought was a little rougher than described. And then it wouldn't fit in my mid-tower Lian-Li case so I had to make a trip to Microcenter for a new case.
uhh.jpg

After that things proceeded pretty smoothly, if tediously. I picked up another Lian-li, the Mesh 2 this time, and while I like the design, it's pretty annoying to remove the drive bays and some of the other stuff. Here's how it looks with the PSU and motherboard installed. The cooler is massive, roughly the size of a NH-D15 but one giant chunk of metal. The only ECC memory I had lying around was a mixture of 2400MHz and 2666MHz (CL20 I think) so I'll have to put some 3200MHz in at some point. Right now that's a pretty big step down from the 3600MHz cl16 stuff I had in my desktop before.

1677723804878.png

After that I had to move the fans on the radiator and install the GPU. This leaves me with 1 case fan, 2 radiator fans, and the CPU fan. Not awesome, but there's not room for much more. I suppose I could possibly fit two on the top of the case, and maybe some extras on the back behind the CPU cooler and in front of the radiator, but that seems redundant. I also had to use extenders for the CPU power plugs (ugly yellow cables). A 12-pin GPU cable would've saved a lot of clutter here as well, but those new PSUs are expensive.

1677723910271.png

After everything was put back together it had to go under my desk for testing. I was worried about fan noise but so far it's basically silent, even when gaming or running cinebench for 10 minutes straight. Windows 10 pro had to reboot once to recognize it, and boot times are horrendous (almost like zen 4), but other than that not too many complications. Still playing around with BIOS settings but I think I'm just making performance worse. My 3DMark GPU score went from 35.8k to 15k, and the CPU score was basically unchanged from my previous 5800X3D. However, I'm not sure how well this translates to games since in-game performance doesn't seem to be nearly as bad as I expected.

side view.jpg



Some rough figures (1440P):
A Plague Tale: Innocence is running at between 120-230FPS with everything maxed except vsync. 2-3ms frametimes. Very smooth experience.
Witcher 3 Raytracing is running at 50-60 FPS with TAAU on (I turned off vsync, bloom, depth of field). Could be smoother. Feels similar to the 5800X3D performance.
Pathfinder: WOTR is running 50-60 FPS but sometimes I'm getting stutters.
Dota2 (the only one I logged before/after stats for using a standard benchmark): Pretty terrible, avg FPS went from 375 to 166. 1% lows went from 140FPS to 77FPS. About a ~50% reduction across the board.

Cinebench without any tweaks
1677742682111.png
 
Last edited:
To many threads and cores for the scheduler to handle.
There is some truth to this. Windows absolutely gave up when I ran it on my 72 core 288 thread phi chip.

Linux seamed to have a better handle on life
 
I would say, if you're going to move this build - at least put it up for some cpu (randomX) and gpu (dnx) mining. Very low wattage for either case and would buy back some form of value over time.

I feel cpu-wise would be quite profitable, randomX loves the cores and later(st) EPYC chips given the L3 cache increases.

Also, love the EVGA 1600w G/P/T units. Have them all of various efficiency flavors in my mining rig fleet - not a single one has failed in ~8 years now. Great choice.
 
To many threads and cores for the scheduler to handle.
Hmm.. Seems like windows 11 makes some improvements there, not sure if it's worth checking out. I've disabled SMT for testing so that should keep everything in 1 processor group it seems like.

I would say, if you're going to move this build - at least put it up for some cpu (randomX) and gpu (dnx) mining. Very low wattage for either case and would buy back some form of value over time.

I feel cpu-wise would be quite profitable, randomX loves the cores and later(st) EPYC chips given the L3 cache increases.

Also, love the EVGA 1600w G/P/T units. Have them all of various efficiency flavors in my mining rig fleet - not a single one has failed in ~8 years now. Great choice.
Yeah, going to put it to work at some point. I haven't messed around with dnx yet. Have everything on raptoreum for now. Should make about $1.50/day I guess.
 
I've disabled SMT for testing so that should keep everything in 1 processor group it seems like.

I don't know if it matters for Windows, but if your firmware has the setting, splitting the cores into NUMA groups helps out on NUMA-aware server use cases. May show up as 'L3 cache as NUMA Domain' or similar. While the L3 is shared, latency to the per ccx L3 is lower than cross ccx L3, so presenting that to the scheduler can help; but sometimes it hurts too. Worth a try.
 
Back
Top