So a small update: the machine is in our office already; we will be doing some tests/profiling next week.
Best regards,
Vlado
Hi Vlado
Do you have any updates?
Have you experienced the similar lower performance in vray with your EPYC system?
Even after a bios update, our machine is still performing far slower than expected.
Kind Regards
With all the hype going around Epyc, guess we should have an update soon 
Best regards,
Rapha
There is a serious problem with render time on AMD CPU. I have found one parameter which affects the render time significantly on AMD to compare with Intel CPU. Please check my post here.
The new Threadripper 3xxxx CPUs are much better for rendering than Epyc CPUs.
Best regards,
Vlado
Is this due to the lower average all-core frequency compared to the same core count Threadrippers or is it something else?
I’m afraid there is nothing we can do about the performance of Epyc CPUs through software updates.
Best regards,
Vlado
Some of the highest scores on the vray benchmark are Epyc setups…whats going on there?
Benchmarks don’t necessarily represents real world since difference between workloads may vary a lot.
Yeh, totally makes sense they wouldn’t translate completely. All scenes are different and a cpu could excel at single threaded tasks over multi, but to be in the top 5 of all tests, and then to hear there is a problem with Vray + Epyc, sounds strange to me.
This particular system has issues, and with the benchmark as well, so we can rule scene dependency out.
Clearly, those tested to be at the top of the benchmark list didn’t have issues.
I wouldn’t draw a general rule out of the one isolated case.
Still wondering if we should invest in Epyc vs Intel for our new render nodes. Last try was not convincing at all…
We have switched workstations to Threadripper (from dual Xeon) with huge benefits so I’m wondering if single Epyc CPU system perform well and multi CPU being the culprit… Anyone has a clue ?
It looks like memory access is often the bottleneck.
NUMA, f.e., will cripple performance.
I would assume multi-socket systems would not fare any better than single-socket of the same kind, and possibly worst.
I wonder if there’s any commercial solution to properly using these many-core beasts, via for example VMs with cores and Ram cherry-picked for optimal perfomance (f.e. 8cores x 8 VMs).
Sure, one’d have less overall resources (i.e. RAM and Cores) but then perhaps a DR would work better, and so would using the machine as a whole farm with more than one jobs active on it.
I am merely guessing though, based on the fact that my single-chiplet, non-NUMA threadripper has none of those issues.
Yes someone in another thread suggest a VM approach I might try to test that as I own some mutli Epyc config. Could be interesting..