AMD Threadripper 3990X

Started by rolland1013, March 12, 2021, 12:32:23 PM

Previous topic - Next topic


Hey all,

For the past few months I've been lusting after AMD's Threadripper 3990X CPU, believing it would give my Terragen work a big performance boost over my current Threadripper 3960X.  But looking at the V.4 benchmark results I'm a little taken aback.  Reviewing AMD's specs on the Threadripper series procs, I would have thought it would have crushed everything other CPU in that series.  But the benchmark results tell a different story, with the 3970X holding the top spot.  Can someone explain this to me?

I have been digging through the forums looking for more info on this subject, and what I found I'm not sure I totally understand.  (Ok, full disclosure, I'm not a sys. admin or hardware guy.  I'm just a VFX artist working from home)  Back in April, Matt posted that beginning with build 4.4.62 he increased the number of detectable cores to account for cpu's like the 3990X.  He also mentioned build 4.4 wasn't NUMA aware.  I'm not sure what that means, but has that been fixed for the current 4.5.56 build?  Is it even a contributing issue?  I did notice in the benchmarks, that the results I'm speaking of were running version 4.44.44, and 4.44.46.  If those same cpu's were tested with the current version of TG, would there be a big difference in performance?

My current system

CPU: AMD Threadripper 3960X
Motherboard: ASUS ROG Zenith II Extreme
Memory 128GB DDR4
Video card: EVGA GeForce RTX 2080 TI
OS: Windows 10



The issue is NUMA Awareness. These are 64 core systems with 128 threads. To support more than 64 LPs TG needs to be aware of Windows processor groups and utilize them accordingly. As far as I know this functionality hasn't been implemented, so TG can only use 64cores/threads, which each render may have different performance due to what Windows allocates to TG. All LPs, or mix of LPs and PPs, etc. Depends on how the processor groups are configured and what's on group 0 which I believe is default for non-NUMA.


So this is a Terragen issue?  Matt, Oshyan, is there any plans or timeline to address this issue?


I know for a fact it's being worked on, it's been said a few times. I have no clue as for a timeframe. From what I've read online, and I think even Matt has hinted on it, the whole processor groups thing is a little janky right now.

In the meantime though, you could always load your project in multiple instances of TG and split the scene up into crop renders, then fire them all off at once. Pretty sure then windows will have to allocate from other processor groups so you'll end up utilizing all cores in theory. I think I've seen it done via screenshot evidence here (don't hold my word to that, so hard to find things via search)