To much render elements

[Micha]

Hi,

I already knew it, but now I got it red on grey.

I hope it's a bug that so much unwanted render elements are shown by enabling denoising and it can be fixed. Is it possible and could it save memory?

-Micha

[vlado]

You can try using the bucket image sampler; it will use less memory. The denoiser needs all the render elements unfortunately.

Best regards,
Vlado

[Micha]

Hi @vlado,

at the moment I have a complex train model and tried to render it at 4000x2000, but I get the message:

D:\cgrepo\vraysdk\samples\vray_plugins\rt_private\ rt_opencl_pri(3390) : CUDA error 2 : unable to allocate enough memory to perform the requested operation (out-of-mem)

I disable the denoiser and the scene render at 5200MB VRAM usage. If I disable the progressive mode, than it needs 4700MB. This matchs my observations of the last weeks - disabling progressive rendering saves a little bit of memory, but the denoiser seems to be a memory killer. Or something is wrong with the memory estimation and the stopped render start isn't necessary.

For this test I used 2x2080ti for rendering and Rhino is fixed to a 1080ti, where the memory usage without rendering is approx. 4000MB. This setup allow to use the 2080ti for rendering and the 1080ti is used for the framebuffer. Is there a chance to get the memory usage of the frame buffer and the denoiser lower? It's a pity that rendering a complex scene isn't a problem, but the denoiser can't be used.

Best-
Micha

[vlado]

This is strange because the denoiser shouldn't take up extra GPU memory. Do you think you can get us a scene for this?

Best regards,
Vlado

[Micha]

I will try my best and look for a scene.

[Micha]

After a Rhino update and a restart of my machine anything works like expected - no additional memory by the denoiser, no error message anymore. If I see the problem again I will try to get a test file ready.

vlado Do you have an idea issue too? My HDRI environments are skipped for GPU on-demand mapping and I'm not sure about the recommended workflow.
https://forums.chaosgroup.com/forum/...onment-texture

[Micha]

No example scene to upload, only a new screenshot - this card is used for Rhino and the frame buffer only, for rendering the two 2080ti should be used. Denoiser is disabled. Output size is approx. 5000x3000. 5,4GB are not enough for the frame buffer?

OK, I closed an other Rhino task in the background with an other large model and tried it again. Now it works, also in progressive mode. The second screenshot shows the memory usage during the rendering now - 3,8GB only (system, Rhino, frame buffer). Why was the rendering no possible, if the memory was available? For me it looks like it's not a memory limit problem, it looks like the memory usage isn't right approximated.

Maybe this information helps to get an idea what could be the reason for the first stop.

[georgi.georgiev]

Micha

Could you please follow up on the private message I have sent you?

[Micha]

georgi.georgiev I got a test running. Last I send you a scene where you find six complex blocks for testing the NVlink. You open this scene and delete two of the moduls. Set the output size at 6000x... and noise threshold at 0.01. Now render this scene in GPU mode. Anything should be fine and quick done. Now open a second Rhino task and do the same there - after the LC pass the error message appears that there is not enough memory, but the GPU memories are quite empty. The memory usage during rendering is 7440MB and after the rendering is done 1100MB. So, there should be enough memory for the rendering at the second Rhino task. (2x2080ti are used for rendering only; system/Rhino/framebuffer are at the 1080ti)

[Micha]

Here a new example - the memory is free, but an other, at the moment not used task blocks the memory usage or cause a wrong approximation.

[Micha]

Maybe the same problem - I rendered a large, complex mode at Rhino task A. The rendering was finished and for a bug report I started a second Rhino task and rendered a very light test scene -> crash. The first task seems to block something so that the second task crashs.

[georgi.georgiev]

Hi Micha
Apologies for the late reply!

This issue is related to the previously discussed one (with the same scene), where 6 blocks consume ~7GB memory, but adding a 7th block the GPU engine fails even though there is memory still available. As reported, this has been resolved in V-Ray Core 4.1+ and it should not be reproducible in the next release (whether it is a stable nightly or an official build).