Thinking about buying the AMD Epyc CPU (single not dual) for rendernode.
Is the large one 7702 with 64 cores a good choice? Any problems with 64cores?
is the 7502 with 32 cores a better price/power ratio?
Any one using those? It would be my first amd so thats why i’m doublechecking 
Thx in advance!
check here to get an idea.
Clear the search term to verify where it is in the result list (edit: 14th.).
No issues with 64 threads, though, no.
thx @_Lele ! Seems this thing is a beast!
if you don’t want to use dual cpu there is probably a threadripper 3990 coming january with 64 cores - which will be cheaper than the epic I guess
I do not currently recommend using dual-CPU EPYC systems for rendering; due to their NUMA configuration, the memory becomes a major bottleneck when the many cores of one CPU attempt to access the memory attached to the other CPU which results in slow renders (sometimes slower than rendering on one CPU). Dual-CPU NUMA configurations are an area that AMD has traditionally struggled with.
On the other hand, single-CPU systems where the entire memory is attached directly to the one CPU seem to perform much better.
In any case, it is best if you get a chance to test the system before paying for it.
Best regards,
Vlado
While AMD has traditionally had issues with NUMA architecture, specifically using the old Opterons, of which we had a number of workstations, so observed this directly. The EPYC chips directly addressed this issue with the Infinity Fabric memory transport, and even more so with the latest generation of EPYC chips released this year, and is one of the reasons AMD is currently decimating Intel’s market share, so I’m not so sure the NUMA issue in dual CPU setups should have as much an effect any longer.
Besides if this was still an inherent issue with dual CPU configurations using EPYCs it doesn’t explain why it is only Vray that exhibits a speed issue relative to Intel CPU’s, compared to other render engines?
In house we have both dual EPYC 7451 workstations and older dual Intel Xeon E5-2680 v3 machines
Under Vray the EPYCs are, at best, on a par with the older XEON machines, often slightly slower.
Under Corona, or any other 3D benchmark for that matter, the newer EPYCs are 50% faster than the XEONs, as expected.
So there is definitely an issue with Vray.
We have an AMD representative coming to asses our workstations soon, so will report back on his findings, it would be great to get the performance out of these EPYC’s that they should be delivering under Vray.
We profiled both Corona and V-Ray on a variety of scenes and found out that both have issues, although it’s somewhat scene-dependent. From the profiles there didn’t seem to be something specific that can be done in the software. Therefore I don’t recommend dual EPYCs for rendering. A single Threadripper will perform better.
Best regards,
Vlado
Thanks for the reply.
It may well be that EPYC’s, or at least 1st Gen EPYC’s, may have issues. However it doesn’t explain why Corona and other 3D rendering packages indicate a 50% speed advantage over the older XEON’s, yet Vray doesn’t or worse can be slower?
I just received a duel EPYC 7702 system and can offer some insights.
Duel EPYC 7702 - 128 cores (256 threads)
128 Gigs RAM
M.2 OS drive
Supermicro H11DS1-NT-O motherboard
Over the last week we benchmarked over 5 different projects from an Audi commercial, to a Brother printer product video to a complex fluid sims and a high poly city made of wooden blocks.
Vlado is correct when he says result are very scene dependent.
Our existing rendering workstations at the office are 44 cores (88 threads) 68% XEON and 32% i7 that are around 4-8 years old. 6 workstations combined.
The EPYC system gets from 2.5X to 1.1X faster renders.
On average about 1.5X faster on most scenes.
1 scene of a high poly wooden city, from outside with brute force GI, rendered 2.5X faster
The same scene from inside a building looking outside through glass with baked Irradiance maps and Light cache took 2 times longer!!! That was a surprise. (256 threads twice as slow as old 88 threads)
V-Ray bench gives varying results, 40K - 60K with no Hyperthreading,
20K-40K with Hyperthreading, but the CPU’s don’t go to 100% with Hyperthreading on, they flicker from 50%-90% like a lightshow!
I’m not a power user that knows how to optimize the system to get 80K+ like the V-Ray benchmark leaderboards, They must have overclocked! No Idea. If I knew who they were I would call them up and ask how to optimize.
Cinebench r20 gives 27-28K consistently.
Interesting findings:
We tried to see how we could speed up the, twice as slow render, and what cut the time down from 30 minutes to 20 was increasing the bucket size from 24 pixels to the default 64 pixels.
Things to note:
Hyper threading gives about a 10% boost.
Windows 10 Enterprise won’t even boot with SMT (Hyperthreading) enabled in the bios, so Linux or Windows Server 2019 Data Center are the only OS’s that can utilize 256 Threads. We’re running Server 2019 in evaluation (180 days). With that may cores it’s $7000US to license.
According to our supplier who has contacts at AMD, Windows has some catching up to do with these new EPYC and 3990X CPU’s. Apparently they’re working on it for the next release. I’m hoping to see the our EPYC system get better results.
On a separate note:
Running Fluid sims or any high particle sims are insanely fast! We did a 300 frame, 3 million particle SPH sim that took less than an hour. We had to leave it overnight on a 6 core (12 Thread) system. And these sims can only use one CPU. We were using Maya with no slowdown while the sim was running. Nuke renders effects very quickly also, not much testing there yet.
It’s a hell of a workstation… for $24K US.
An update to my previous post…
I did some further benchmarks and scenes render faster when I only use one CPU for batch rendering, (64 cores out of 128 ) or (128 threads out of 256). I did this by setting the threads amount in render settings or with the command line -threads tag.
This makes me wonder if I can get better performance on both cores at the same time if I ran 2 Virtual Machines with a single 64 core CPU dedicated to each VM. I have never created VM’s so the plan will require a few tutorials on my end, but I would like to know if the idea has any merit or will the same bottleneck occur?
Any advice would be greatly appreciated.
Hi Bart !
Did you tried the virtual machine approach - sounds like a clever way to handle the problem.