GPU memory usage shown at the frame buffer?

[Micha]

How can the display of the true GPU memory usage be enabled?

-Micha

"Knowing this, we have modified the memory statistics shown in the V-Ray frame buffer so you can track actual GPU memory usage there. "
https://www.chaosgroup.com/blog/prof...idia-rtx-cards

[Zahari.Ivanov]

Hello Micha,

This is a feature not currently available in V-Ray for Rhino. One of our developers will be tasked with assessing a possible implementation.

[Micha]

Hi Zahari,

the feature will be very welcome. For now do you know an other tool that allow to see the memory usage during NVlink usage?

[a7az0th]

If you are rendering on Windows or Linux you can use the NVSMI tool that comes bundled with the NVidia GPU Driver.
On windows it is located in "C:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe"
Starting this exe with the -l (this is the lowercare letter L) command line switch will tell it to loop and reprint the data every second.
It looks like this:
https://drive.google.com/a/chaosgrou...CMtm0t4ubQPEot

[Micha]

a7az0th Looks like something is wrong here - NVlink is enabled, but both GPU are filled with the same amount of data only. Do you have an idea?

[a7az0th]

Do they use the same amount of memory if you disable NVLink? We try to balance memory allocations on both devices when NVLink is on, so having close (or in some cases identical numbers) is not an indication of a problem.
However if you disable NVLink and the numbers are the same, it would be best to send a scene to support@chaosgroup.com for investigation.

[Micha]

a7az0th I did some further tests.

(A) I disabled NVlink and started a medium complex scene to render. You can see three cards and a different VRAM usage at each. The scene needs approx. 8GB to render.
(B) I enabled NVlink, restarted Rhino and started the scene again (only the two 2080ti enabled for rendering) - second screenshot. The VRAM usage at both cards is the same and the amount is in the range like before, approx. 8GB for the scene and 1GB+ for system use. I don't see an advantage of the NVlink, it looks there is not more VRAM available.

I hope there is a bug for VfR and this kind of scene isn't the limit for two linked 2080ti.

[a7az0th]

Hmm.. this is really odd. I will take a look at it. I believe I have seen this problem before and it is related to the driver memory reporting when there is NVLink. Maybe that bug has resurfaced. If it has, then the data outputted here can not be trusted, But the NVLink still works, just the statistics looks messed up.

[Micha]

I can render the test scene without NVlink at approx. 10GB. If NVlink is enabled and I open a few large images at Photoshop than this shouldn't cause a crash, but it cause a crash. The problem isn't so much that a wrong VRAM usage is displayed, the problem is that it looks like I get crashes by not enough VRAM. There is not big difference between NLink enabled or not enabled. If two RTX2080ti per NVlink are not enough to render my old interiors scenes, than my investment in GPU power was wrong.

Could you test a complex scene at VfR per NVlink too please?

[georgi.georgiev]

Hi,

Could you please send the project to support@chaosgroup.com so we can test in our environment and advise accordingly?