determine mem usage of a scene

[s_gru]

hello. so, i use MSI Afterburner to keep track of my cards. I note that the one card used as primary monitor + rendering sometimes flat-lines at 16GB. I have the gpu mem graph set to 18GB max. my cards are 16GB. is there a log file entry or something that tells me how big the scene is. for example, i'm looking at changing to 24GB cards. but i'd kind of like ot know how much headroom i would have.

log would say "hey, this scene takes 21GB of video mem" regardless of which cards you have going right now.

thanks

[hermit.crab]

There's a Stats tab in the VFB's top right-corner.

[s_gru]

ugh. right there in front of me.

[pipe_mm]

I have one question regarting RAM stats.
while rendering it uses 2/3 of my GPU RAM.
Only at the end it goes up to almost 100%

Is this the denoiser or other Render Elements?
Which number should I use as a guide to avoid crashes during rendering?

[Muhammed_Hamed]

I have one question regarting RAM stats.
while rendering it uses 2/3 of my GPU RAM.
Only at the end it goes up to almost 100%

Is this the denoiser or other Render Elements?
Which number should I use as a guide to avoid crashes during rendering?

Hey!

This is not normal, is there a chance I can take a look at the scene under NDA?
On another note, using CUDA mode + Progressive image sampler will save you 2-3 GB of GPU memory

Best,

[pipe_mm]

Hello Muhammed

I already wanted to thank you for your very clarifying reply in this posting:
https://forums.chaos.com/forum/v-ray...90334-rtx-4090
really helps a lot to see your research on GPU performance.

I could hand over the scene, and I would love to see your possibilities on optimisation
Just have to consolidate it and upload it to my cloud.
I will PM you the link

[pipe_mm]

Hello,
I would like my question to be more precise:
How can I monitor the GPU memory usage?

The scene I built here is using a good amount of proxys and textures.
As rendering an animation some frames do render, but at a certain frame the memory is exceeded.

No problem, this is something I expected.
But how do I optimize my scene?

I already switched to different texture modes: "resize" and "compressed"
Now I am kicking out objects from the scene in the hope to get to a point the scene adheres to the limit.

How do you approach these kind of tasks?
Is there only try and error or can you proceed more scientifically here?

How do I read the stats in VFB to help me find the "victim"

Very much looking forward to your approaches to memory optimazation.

[pipe_mm]

here is the Log from the point where it exceeded.
The break seems to happen in the Light Cache process. (not in rendering)

[pipe_mm]

and here two screenshots from the stats.
Both are taken from rendering a reduced version of the scene (some geometry had to be deleted).
First screenshot is with texture mode "Resize all textures" (512mb).
Second is "Compressed"

as you can see there is only a minor difference.
Also we are not close to the limit. ~9,7Gig used from 16Gig

So I will slowly add more objects to the scene and see when it breaks.

[pipe_mm]

Bucket vs Progressive

the good old question
Muhammed you wrote I should use Cuda&Progressive.

I do so now, but when I switch to Cuda&Bucket the difference is not sigificant. (only talking but RAM usage)
the stats even show a bit less in Bucket mode.

[pipe_mm]

and last question, when bringing up the scene to its limit, which figure is relevant (GIB in Stats):
the number shown during rendering, or the one shown at the end of the process?

during rendering: 9
at the end: 11

[pipe_mm]

self answer:
- task manager and MSI hardware monitor are my best friends. straight forward view on the Ram usage.
- deadline render manager rocks - also on single machines. easy to install, easy to submit, render outside the GUI reduces the graphic load. if one frame crashed he just tries again (and doesnt stall the whole sequence)

[Muhammed_Hamed]

Hello Muhammed

I already wanted to thank you for your very clarifying reply in this posting:
https://forums.chaos.com/forum/v-ray...90334-rtx-4090
really helps a lot to see your research on GPU performance.

I could hand over the scene, and I would love to see your possibilities on optimisation
Just have to consolidate it and upload it to my cloud.
I will PM you the link

I'm glad you find this helpful, keep an eye on the announcement section. I will have a thread on GPU usage there soon

- task manager and MSI hardware monitor are my best friends. straight forward view on the Ram usage.

MSI After Burner or GPU-Z they are more reliable than task manager.
On another note we updated the stats tab to have reliable GPU memory reporting(in the nightly builds), right now it is not accurate at all

and last question, when bringing up the scene to its limit, which figure is relevant (GIB in Stats):
the number shown during rendering, or the one shown at the end of the process?

during rendering: 9
at the end: 11

Ignore both, look at the graphs in MSI After Burner. I will have a thread on GPU monitoring soon

I do so now, but when I switch to Cuda&Bucket the difference is not sigificant. (only talking but RAM usage)
the stats even show a bit less in Bucket mode.

In most cases Progressive should be taking less, I will take a look at your soon hopefully and tell you as well
Generally speaking, CUDA + Progressive is the most efficient when it comes to memory usage

How do you approach these kind of tasks?
Is there only try and error or can you proceed more scientifically here?

For me I would render some frames in random throughout the sequence to make sure nothing bad happens
If they render fine, the animation should probably render fine. What you described could be a memory leak, if individual frames render fine
I will be able to tell with a scene

Best,
Muhammed

[pipe_mm]

thank you so much Muhammed.
Vray is a great software, with so many wonderfull add on´s (library, scatter, ...) and on top we get this premium support by highly motivated professionals like you.
Keep the good work up
*M

[Muhammed_Hamed]

thank you so much Muhammed.
Vray is a great software, with so many wonderfull add on´s (library, scatter, ...) and on top we get this premium support by highly motivated professionals like you.
Keep the good work up
*M

Thank you for your kindness Martin!