AMD 3990X massive slowdowns

[FS-Marc]

Hi there,

We configured a Workstation with a AMD Threadripper 3990X hoping it would be a blast to work with in C4d + Corona which it is if you do rendering only. But as soon as you use the Interactive Renderer do this the whole System becomes unusable - you cant even move explorer windows without huge lags. So every speed advantage is completely lost since you cant work in your scene while IR is running.
Also Material editing seems to be extremly slow due to (I guess) rendering the preview for the material manager before updating the Viewport.

# threads override seems to work if you use a very small count like 12 - which then is too slow in rendering. If you use anything high the system breaks down again.

Has anyone any ideas or had a similar experience and knows a fix? Please let me know

Faithfully,
                 Marc


Workstation-Specs:
OS:Windows 11
Processor: AMD Threadripper 3990X
Motherboard: ASUS ROG Zenith II Extreme Alpha
Ram: 32GB ECC UDIMM DDR4 2666 x8
GPU: Geforce RTX 390
Storage: Samsung 980 Pro NVME 1 TB x2
Cooling: Noctua NH-U14S TR4-SP3
PSU: Be Quiet! Straight Power 11, 1000W

Everything runs in Stock-Settings. No OC.

[Juraj]

No advice. I've built countless of workstations over the years, I have both 3990X and some of the highest Intel Dual-Xeons (provided to me by Intel itself) like 8280/8380, etc..

This is just how 3dsMax always behaved to me, but you use C4D which is interesting. Window 10 Pro/Workstation/Enterprise, Windows 11, etc.. doesn't matter.
Project Lasso, every possible setting, doesn't matter.
IR + Material Editor = Fucking hell.

Maybe it's just expectations but I've never seen other people complain about it or maybe other people just don't use heavy scenes I don't know.

It's little bit better for less-threaded machines because they keep better clocks during multi-threaded use. IR is very demanding because you're using your CPU fully multi-threaded, but still need high speed and availability of best single-core.
Windows is not exactly fantastic with process scheduling, which is finding the best available single-thread and funneling resources to it (so for example, memory access isn't hogged).
Windows 11 is little bit better in this because the scheduler has been optimized for complex architecture like LittleBIG for Intel Alder-Lake. That ultimately benefits AMD as well as it will stop Windows giving 3dsMax (or C4D or other app) to slowest thread during heavy multi-threaded use.

But I have go as far as only giving Corona 50perc. of CPU during IR. Still the same, heavy lag. Multi-seconds to edit materials in editor.

[FS-Marc]

Maybe it's just expectations but I've never seen other people complain about it or maybe other people just don't use heavy scenes I don't know.

Do you experience it only in "heavy scenes"? I made the experience that it does not matter how heavy or light the scene is.

It's little bit better for less-threaded machines because they keep better clocks during multi-threaded use. IR is very demanding because you're using your CPU fully multi-threaded, but still need high speed and availability of best single-core.
Windows is not exactly fantastic with process scheduling, which is finding the best available single-thread and funneling resources to it (so for example, memory access isn't hogged).

Youre definetly right with that! A colleague has the 3970X and and it works like a charm. Just disappointing that its "bigger brother" is handicaped in such way.

But I have go as far as only giving Corona 50perc. of CPU during IR. Still the same, heavy lag. Multi-seconds to edit materials in editor.

Well I guess if there is no way to resolve this we will use it as another renderslave and have to invest a bit more to get a more flexible system...


I just thought that it might be some kind of issue with the way Corona Renderer is handling things since I dont have any problems using the built in renderer in interactive mode which should also be multithreaded AFAIK.

[Nejc Kilar]

So we have identified a similar issue not so long ago and are currently looking into improving things. We definitely appreciate your post here as the more configurations we know about the better.

For the time being what you can try doing is disabling the SMT in your BIOS. That will effectively be disabling multi-threading on your CPU which means you will be running with 64 cores and 64 threads instead of 64 cores and 128 threads. Without a question that is not an ideal solution because you'll lose a good chunk of performance (roughly ~20-30% or there abouts) but it beats limiting the threads inside Cinema 4D to 12 as you mentioned. The UI and everything should work quite butter smooth once you do that.

Do note that messing with the BIOS is something you're doing at your own peril. That said, it shouldn't be too difficult to find the SMT toggle :)

Just to make this super clear though, we have identified the issue and are working to fix it so the above should be a temporary solution if you're running into the issue you're mentioning.

[Juraj]

Interesting.. I am curious what that could be :- ).

I think 3990X suffers from this somewhat more since the memory Controller is probably overwhelmed during Multi-threaded use, enough to create "traffic jam" for any further single-core use.
2990WX suffered from this because only half of dies had direct memory controller access, but this is no longer the case with 3990X.

But since 3970X fares better with identical IO and 280W (this 280W is shared not only by 3/4/8 respective die, but also by IO, so the IO can get starved by juice as well).

I never actually disabled SMT myself, now I will experiment as well. Def not want to dedicate my 3990X into node yet :- ).

[Juraj]

Maybe it's just expectations but I've never seen other people complain about it or maybe other people just don't use heavy scenes I don't know.

Do you experience it only in "heavy scenes"? I made the experience that it does not matter how heavy or light the scene is.

No, I probably should have worded that differently, I speak too much stuff at once. The lag is there even in empty scene IR.

[Nejc Kilar]

Interesting.. I am curious what that could be :- ).

I think 3990X suffers from this somewhat more since the memory Controller is probably overwhelmed during Multi-threaded use, enough to create "traffic jam" for any further single-core use.
2990WX suffered from this because only half of dies had direct memory controller access, but this is no longer the case with 3990X.

But since 3970X fares better with identical IO and 280W (this 280W is shared not only by 3/4/8 respective die, but also by IO, so the IO can get starved by juice as well).

I never actually disabled SMT myself, now I will experiment as well. Def not want to dedicate my 3990X into node yet :- ).

This is just my personal observation, the devs will probably know more because they are the real brainiacs but my impression is that whatever is happening in Cinema 4D is a bit more complex and could be related to the way the threads are being scheduled or that kind of complex stuff I honestly don't really understand perfectly myself.

On the responsive or not front, I personally haven't noticed my 3970x (~3.6 all core) being any faster UI wise than for example the 5995wx (~3.2 all core) with the exception of whatever happens in Cinema 4D. That's subjective though, I haven't measured anything, just my impression. It's also true that it is comparing Zen 2 vs Zen 3 so there might be differences there at play such as the IPC prowess.

[FS-Marc]

For the time being what you can try doing is disabling the SMT in your BIOS. That will effectively be disabling multi-threading on your CPU which means you will be running with 64 cores and 64 threads instead of 64 cores and 128 threads. Without a question that is not an ideal solution because you'll lose a good chunk of performance (roughly ~20-30% or there abouts) but it beats limiting the threads inside Cinema 4D to 12 as you mentioned. The UI and everything should work quite butter smooth once you do that.

Do note that messing with the BIOS is something you're doing at your own peril. That said, it shouldn't be too difficult to find the SMT toggle :)

Thanks for the tip will try that as soon as I am back in our office and give some feedback.

Just to make this super clear though, we have identified the issue and are working to fix it so the above should be a temporary solution if you're running into the issue you're mentioning.

Awesome to hear!

[FS-Marc]

Do note that messing with the BIOS is something you're doing at your own peril. That said, it shouldn't be too difficult to find the SMT toggle :)

Just to make this super clear though, we have identified the issue and are working to fix it so the above should be a temporary solution if you're running into the issue you're mentioning.

Thanks you very much for the advice. I can confirm this method resolves the problem (with a hit to performance . This makes my work process much faster so the tradeoff is worth it for now. Hopefully one day we can make use of the cpu as supposed to.

[4b4]

Hi, chiming in over here as a 5995WX user with exactly the same experience in C4D. I've disabled SMT so I've got a usable machine at least but it's a painful move to make having just lost a kidney to pay for the machine 😀

[ONeill]

I'm running a 3995wx Threadripper Pro. I've been experimenting with various scene files, and I have not experienced the issue you are describing.

I was able to open up a scene file with 40 million polygons. The viewport was extremely slow, but the Interactive Render worked without any problems. I did get the low memory warning, but the system was not lagging.

I opened up another file with scattered trees (20 million polys) and did not have any problems with rendering.... although I did have a crash (corona dll).

My system specs are:
Windows 10
AMD Threadripper Pro 3995wx
Asus WRX80 Sage mobo
64GB RAM
RTX A5000

On a side note... a few years ago I was using Vray (VrayforC4D) on a dual Xeon machine. I was experiencing the same identical problem you are describing. The system would freeze for 10-15 seconds every time I tried to make a change to the materials, or try to render a preview. It was unusable, and the only solution was to disable hyperthreading in the BIOS. I eventually gave up on it and moved over to Arnold. Now, I'm thinking of switching to Corona.

[rubensabou]

I've noticed the same issue on both the 3990X and the 5995WX; they behave similarly, though it seems the 5995WX is even worse. I’m working with an almost empty scene, just a single box, and when I start interactive rendering, the PC becomes nearly unusable. Interestingly, it actually performs better with more objects in the scene rather than just a single box or an isolated object.

In contrast, my 3960X workstation doesn’t have this problem at all. It seems this issue is specific to these high-end workstations with 64 CPUs and 128GB of RAM.