Chaos Corona Forum
Chaos Corona for 3ds Max => [Max] Bug Reporting => Topic started by: webuilddreams on 2024-03-11, 13:23:52
-
Hi, i have a new pc and the geometry phase is very very slow, mostly 10x slower than my old pc & and the scene parsing is mostly twice slower???
The new pc is an AMD EPYC 9754 (9004 Genoa) (singel cpu) on a Supermicro board with 384 GB DDR5 and the old pc is an AMD Ryzen 5900X with 128 GB DDR4.
I testted it with 20+ scenes all the same result, and is see the following times:
new pc(max 2021 and corona 11):
scene parsing 52 sec, scene geometry 97 sec, render time 12 minutes @ 850,000 samples/s actual average
old pc (max 2021 and corona 11)
scene parsing 30 sec, scene geometry 13 sec, render time 95 minutes @ 115,000 samples/s actual average
I also tested with the corona benchmark at 31 832 636 (9th place in the ranking) and with Vray benchmark 132950 (5 th place) so both very very fast, I am lost what to do?
-
As far as I know the info displayed is not actually only scene/geometry parsing but also other things like like loading of maps, so it might be that your new PC is slower at loading maps from disk or through network. These are also single threaded so you may have a mix of different reasons - single thread speed of your new CPU might be slower than of the CPU of the old machine, disk/network access speed, and your new CPU might suffer from under-utilization of CPU threads (where only half of the threads are working on a task etc).
I'd look into other means of benchmarking (not 3dsmax) to measure different tasks (CPU single threaded vs multi-threaded, RAM speed, disk/network I/O) and compare how each of the PCs performs. This will probably give you a better picture of what exactly is different between the two PCs.
-
Thanks Pokoy, we load the 3dsmax scene on both pc's local from de D SSD M.2 drives, so that is not an issue, the new pc has a single core speed of 3.08 Ghz and the old pc 4.5 Ghz, that could explane the scene parsing perhaps (but still the old pc single core isnt twice faster only 50%).
I benchmarked all different computer parts (memory , cpu ssd, etc etc) and on all benchmarks the new pc is much faster than the 4 year old cheap pc,
But thanks for thinking with me!
-
The AMD EPYC™ 9754 goes to 3.1GHZ, and the 5900x turbo up to 4.8 GHZ, in my experience thats were that time is, the rendering is so much faster because its using all the cores, not all the parsing is multithreaded (and cant be )
-
thanks, i start to meassure this: in the scene parsing corona does in fact uses only a few cores, but in the geometry phase corona uses arround 70% of the cores. The screenshot was taken in the geometry phase.
-
Hi all, we are a step closer to identifing the problem, the AMD Epyc 9754 has a whopping 128 cores and so 256 threads, in the render phase it is fast but in the geometrie phase it is a lot slower compared to our oldest pc (all the scenes have the same issue so it's not scene releated bort pc corone 11.1 and max 2021.3 and max 2024.2):
old pc (ryzen 5900x) win 11 pro (128 gb), geometry phase takes 15 secs
amd epyc 9754 win 11 pro(384 gb), geomtery phase takes 60 sec!!
The difference is in the Embree calculation time (see corona log file in C:\Users\USERNAME\AppData\Local\Autodesk\3dsMax\2021 - 64bit\ENU\en-US\plugcfg\corona) and this does not make any sence : this cpu is one of the fastets where it comes to calculating Embree (https://openbenchmarking.org/test/pts/embree)
So we tested this (the Embree benchmark from github) and we got a nice fast score arround 130 frames per sec: conclusion our pc is fine
So is it possible that one of the core componens of corona render is a bit old?
And does any one have the same problem?
-
Really hoping one of the devs can check this out !
-
We are checking this out :) In fact we just took delivery of a new super high core/thread count dual EPYC machine this week so that we can better investigate any reported issues along these lines.
-
We are checking this out :) In fact we just took delivery of a new super high core/thread count dual EPYC machine this week so that we can better investigate any reported issues along these lines.
Great, that would be great, it is also very frustrating to work in interactive rendering(IR): before i can make any change (for example the color materials or move an object) i have to wait until IR finish rendering (max passes) so we put max passes at 5 (btw the thread override IR is set to -16)
We still can use Photoshop and other programs, just max hangs (test this with HF11.1 and 2021.3 and 2024.2)
-
I hope solution for your problem is found, but I will like to stress that if in some case you have non-retail stepping (ES/QS/etc..), the issue can be with on-chip memory controller which creates certain spiral of firmware/OS issues that can manifest under certain workloads (whether it's scheduler issues, CPU groups, NUMA nodes, etc..).
So if Corona team received fully retail unit (which in 2024 means all-in-one system for most Epyc stuff), they might not be able to replicate your issue.
It's bit risky proposition to build custom multi-CPU servers today, too many variables with zero support. Super hard to investigate and find what can be culprit.
Not saying that's the case, just be prepared.
Nice to know that Corona devs got the machine though.
-
I hope solution for your problem is found, but I will like to stress that if in some case you have non-retail stepping (ES/QS/etc..), the issue can be with on-chip memory controller which creates certain spiral of firmware/OS issues that can manifest under certain workloads (whether it's scheduler issues, CPU groups, NUMA nodes, etc..).
So if Corona team received fully retail unit (which in 2024 means all-in-one system for most Epyc stuff), they might not be able to replicate your issue.
It's bit risky proposition to build custom multi-CPU servers today, too many variables with zero support. Super hard to investigate and find what can be culprit.
Not saying that's the case, just be prepared.
Nice to know that Corona devs got the machine though.
Hi we have a single retail Epyc 9754 (so no Numa nodes), we ruled out that the processor (and the whole system) is the culprit:
the slow geometry phase is caused by a particular step 'Embree scene calculation'.
The epyc is in the calculation of the 'Embree scene building' step very slow, 10x slower than for example a ryzen 5900x (12 cores)!
So to rule out that the Epyc cpu is slower when it comes to Embree calculation we want to test this slow Embree scene calculation and we tested it with Embree benchmarked software : phoronix-test-suite benchmark embree, the result was :
AMD Epyc 9754 : 127 frames/sec
Compared to
AMD Ryzen Threadripper Pro 7995WX : 116 frames/sec
Conclusion the Epyc is a bit faster in Embree calculations compared to a 7995WX (96 cores) so our system is not the issue, could it be the Embree version inside corona ?( Corona using old embree version?)Intel Embree is currently at version 4.3 do you know what version currently is used in Corona 11.1?
-
That is not what I meant, there is nothing faulty with CPU. But that doesn't mean all software can utilize it in same way. How Embree can utilize the hardware depends on many aspects of software stack, from motherboard firmware to OS scheduler. If those things are written for retail SKU, tiny discrepancy can cause misutilization.
Years ago, one of such solutions was upgrading Visual Studio redistributable used for compiling. Just saying it isn't as easy as updating libraries often, though I do hope it is for this particular issue.
-
That is not what I meant, there is nothing faulty with CPU. But that doesn't mean all software can utilize it in same way. How Embree can utilize the hardware depends on many aspects of software stack, from motherboard firmware to OS scheduler. If those things are written for retail SKU, tiny discrepancy can cause misutilization.
Years ago, one of such solutions was upgrading Visual Studio redistributable used for compiling. Just saying it isn't as easy as updating libraries often, though I do hope it is for this particular issue.
Hi Juraj, thanks for your input, we or I am a bit frustrated , spend a small fortune and now this shit, hope the devs solves this soon.
-
Yeah I can understand, but now that devs have latest high-end dual-Epyc, let's see what happens :- ). There are lot of after-market Epycs on market so that would be quite good for studios if their performance becomes bulletproof in Corona.
-
Any news or findings from the dev team?
-
News or findings will show up in a daily build as a "we hope this fixes...." ;) Or if something comes up where we'd appreciate a test "if you disable this in the BIOS..." or similar. So that is when it all becomes relevant for folks to test things - up until then while we are testing internally, there won't be any news.
-
There seems to be a pattern here with high core/thread machines -see my threadripper post. My cinnebench scores were great on a 64 core machine and disappointing real world performance with corona 11...
-
Yep, we are aware of the patterns - I say patterns as we do know there is more than one thing going on.
-
This is a frustrating but expected result, as @Pokoy stated, the more cores a system has, the slower single threaded operations will be.
I vaguely remember an old CPU single performance chart from Tom's Hardware website.
With a list of many high end products starting from 1 core 1 thread Pentium 4 3gHz, this CPU was the slowest of all processors in multicore benchmark and number fastest in single threaded.
I also experience this a lot and makes me feel pessimistic about our profession. That's why I work with my 4 Cores laptop until it is rendering time. It is definitely more responsive during viewport operations than the 24 Core Xeon one.
-
I am also checking on this. It is crucial to get info on whether dual CPU systems will be better optimized or not.
-
News or findings will show up in a daily build as a "we hope this fixes...." ;) Or if something comes up where we'd appreciate a test "if you disable this in the BIOS..." or similar. So that is when it all becomes relevant for folks to test things - up until then while we are testing internally, there won't be any news.
Hi TomG, is it possible that the whole series 9004 is affected (the Zen 4) by the problem 'ít takes 10 x longes to make the embree calculation' ? And is there any news from the dev team?
-
It is possible the entire series is affected (when run in dual CPU set ups), as they all fit into the "dual CPU, high thread/core count". It's still under investigation, tracking down the actual cause is tricky here as the "obvious" candidates were ruled out some time ago and that leaves it as something much more unusual.
-
It is possible the entire series is affected (when run in dual CPU set ups), as they all fit into the "dual CPU, high thread/core count". It's still under investigation, tracking down the actual cause is tricky here as the "obvious" candidates were ruled out some time ago and that leaves it as something much more unusual.
Thanks, but i run a single AMD Epyc 9754 and am also affected.... any update on when the dev team expect this to solve and e.g. is the threadripper pro 7995wx also affected since this is an overclocked 9654?
If this is not solved soon i sell my rig with a write off of atleast 4K!
I consider to sell my current setup since i dont get any responds from the helpdeks on when this is solved , very very unhappy with the lack of responds from the helpdek!!
And the question remains what cpu's are functioning properly for corona, is there any list?
Sorry for this bit hard words , am a bit frustrated, i sepnd 20K on a new pc and cant really use it, it is so slow especially in IR.
-
For single CPU high thread core counts, the problem was isolated some time back and the fix moved into the dailies https://forum.corona-renderer.com/index.php?topic=41598.msg221510#msg221510 Those machines are now scaling in performance as we would expect. Only dual CPU systems remain with some extra issue.
The list is, with Corona 12 dailies, every CPU is working unless you are running dual CPUs with more than 64/128 cores/threads.
PS, I have also confirmed with support that they are regularly responding to you, so not sure why you believe there is a lack of response?
-
For single CPU high thread core counts, the problem was isolated some time back and the fix moved into the dailies https://forum.corona-renderer.com/index.php?topic=41598.msg221510#msg221510 Those machines are now scaling in performance as we would expect. Only dual CPU systems remain with some extra issue.
The list is, with Corona 12 dailies, every CPU is working unless you are running dual CPUs with more than 64/128 cores/threads.
PS, I have also confirmed with support that they are regularly responding to you, so not sure why you believe there is a lack of response?
sorry tom, but the issue is NOT fixed, they geometrie phase (and in particularly the Embree calculation) takes way to long (up to 5 minutes in simple scenes), the rendering phase is as fast as i expect
-
I have left a note in the support ticket. I suggest we continue over there. Of course, when there will be some important information to share (e.g. the solution), sharing it on the forum is a good idea.
-
I am watching this with interest and I have some partial off-topic note. The up-coming Windows build, that will be public in September I think, but is already in fast-track beta public test (not sure which channel that is but probably ISO is available to everyone), 24H2 comes with completely new (process/core) scheduler. I do wonder if this might help our situation as well? At least a bit?
I am debating whether to test this myself, but I don't have much time or patience currently with very busy life.
The worst performance with high-core machines in general work is 3dsMax/Corona while doing IR. The rendering performance is as expected (the Ray/s metric), but every other process (Material editor refresh, Precomp phasing, etc..) is perceivably slower at same time.
Once again, not to de-track from above issues, just wondering if it could be additional piece of puzzle. Windows Scheduler was always big weak point for high-core machines.
-
"The worst performance with high-core machines in general work is 3dsMax/Corona while doing IR. The rendering performance is as expected (the Ray/s metric), but every other process (Material editor refresh, Precomp phasing, etc..) is perceivably slower at same time."
Yes that is also an additional problem, but my problem is that the calculation of the Embree takes extemly long(up to 5 mins in older machines only seconds),the rendering itself is very fast