Switching over from CPU to GPU, any technical way to calculate memory usage on heavy scenes?

[Psithur]

Hey

I'm wondering if there is a way to calculate if a large scene will fit into the GPU memory other than random testing.
Is there a rule of thumb to go by when calculating if a scene will fit in the GPU memory? We use a lot of Forest Pack / Rail Clone. There's a lot of technical information on this subject all over the place but it's hard to understand everything in life :P!

All I know is that the way everything is packed in GPU VRAM differs from normal RAM so you can't just do a straight comparison.
I also understand that NVLink can be a potential win but it's new tech an there's a not a lot of info on it.

The situation is.... the place I work for are potentially interested in the power of GPU rendering. The scenes we render are often in 20-30GB RAM usage area.... whilst not the heaviest of scenes they are still pretty big.
We are wondering how to calculate if our scenes will work sufficiently if we switch to GPU & if it's worth switching over the farm to GPU rendering i.e investing in lots of more hardware. Theoretically, if it worked it could be a big boost for a small cost.

Is there a good resource out there that explains this? Or are we still in the Wild West where there are too many unknowns and it's just trial and error and the users end?

Cheers
J

[glorybound]

The VRay log will give you an error saying out of memory.

[idnok]

Last time i was concerned about scene fitting into gpu memory i read that 1m polys = 200mb (although that's seems pretty exaggerated), forest instances takes 16kb eatch, and normal instances takes almost nothing.
Other than that, gather all your maps together and you have an idea of what you're looking at.
Now, that being said.. V-Ray uses mipmap, adaptive lighting and a WIP adaptive dome that all together do an EXCELLENT job into reducing your scenes memory occupation without any visible impact!
Other than that, how do you check if your scene fits in your system ram (specially with vray that, will render whatever you throw in it)?

[glorybound]

My 20+GB interior rendered just fine in hybrid mode with one 1080ti. Sorry, I’m not much help

[squintnic]

complex 80gb+ scenes fail miserably lol on GPU as expected...il be sticking with CPU for these until the technology matures or there is some redshift style out of core rendering.
but it does shines on smaller product type stuff or abstract 'instagram' scenes that kids seem to make these days.

[idnok]

I would love to change your mind so.. can you show me a scene you aren't able to fit into 11gb GPU?

[matanov]

You can try your scene on CUDA C++ and check the memory consumption in the V-Ray log.

[squintnic]

I would love to change your mind so.. can you show me a scene you aren't able to fit into 11gb GPU?

allright let me see about getting a scene together! i currently have 2 x 2080 with NVlink so 16gb of VRAM (2080ti founders editions were gonna take 8 weeks to get in store....)