About DS dropping GPU

Second Technician Rimmer

It has been mentionned before that if a scene exceeds what a video card can hold in its Vram, that card will be dropped by DS and render will go to CPU.

Now, here's a hypothetical situation. Let's say I have a gtx970 (4GB) and agtx 1070 (8GB) on my rig. Of course, Vram won't stack, that's a given. The cudas from both cards can still be used, from what I think I understood. However, let's say that my scene can fit in the 1070's Vram and not the 970's. Will the 970 be abandonned altogether, cudas and and ram, leaving the 1070 doing all the work, or will my render switch to 1070+cpu, or... or what else? O_O

Dim Reaper

Once your scene exceeds the vram of the gtx970, Iray will render on just the gtx1070 - it will not involve the cpu at this time as long as you don't have cpu ticked in the render settings.  If you are still working on your scene, and it later exceeds the vram of the 1070, then it will fall to cpu rendering.  This is actually better in most cases than going to 1070 GPU+CPU - on most people's machines, this is either slower than GPU alone, or only a few seconds faster.

I occasionally find both of my cards (6GB and 11 GB) getting dropped for cpu rendering when I've been working for a while on a scene.  I find that saving, shutting down DS, reopening DS, reloading the scene and rendering immediately usually allows me to render with both GPU.

Second Technician Rimmer

So in essence, rendering on both cards is useless unless the scene can fit on the vram of the most vram challenged card of both, right? Le's assume I have a scene that takes 6 or 7 GB and load it on the configuration I described. Based on what you said, is eems that the 970 will be dropped immediately anyway, leaving the 1070 alone to face the music. That's sad, I was hoping I could capitalize on the 970's cuda's even if the ram was too short.

Dim Reaper

Yeah, unfortunately, that's the situation.  One way to look at it though is that at least you have the use of that card for smaller scenes, and when you don't, it's your fastest card that is doing the work.

I don't have any direct experience of it, but apparently Octanerender allows the use of normal RAM for large scenes.  Not sure how it works though, but it might be a place to look if you do need a lot of memory for very big scenes.

The other alternative is to render a scene in parts using the spot renderer and put them together afterwards - but this is not as convenient as a single render.

There is a discussion / speculating thread in the Commons about the new 20xx series cards and whether the NV Link will allow memory pooling on the gaming cards.  It's looking like it will only be available on the Quadro cards unfortunately.