Just one question that I have no clear:
If I have on my PC 2 GPU cards with diferent VRAM each one (lets say 4gb and 11 gb) will Blender use the total amount of VRAM for rendering?
I have read that it will use only the VRAM of the smaller card (4gb in this case) … but I have read too that this is incorrect: https://www.daz3d.com/forums/discussion/267751/dual-gpu-vram-limits
Where is the truth?
Nope, in the example config your limit would be 4gb unless something has changed in recent versions. That doesn’t apply though, afaik if you are using cpu+gpu rendering.
So the URL in my example is completly wrong?
Honestly I’m not even sure what program is discussed in there. It’s hard to judge without knowing that.
Rendering with cpu only uses your system ram. Slow but large scenes might fit if you have a lot of ram.
Normally each gpu gets a complete copy of the scene’s resources in its memory. So the scene must fit in a card’s memory for that card to work on it.
I have read that some Nvidia cards can be connected with NVlink and that they can combine their memory into one big memory. I don’t know if that is right. It’s something you could research.
I also think, but not sure, that cycles has a limited ability to use system ram for something when gpu rendering. Maybe big textures?
When VRAM is filled data can be off loaded to system RAM. Depending on the application it may crash when out of VRAM or happily use system RAM with a speed penalty. I think blender will just crash if you run out of VRAM.
NVlink is only available on high end models and it does combine memory. 2x 3090 with an Nvlink bridge would have 48GB of VRAM to use in rendering rather than 24GB of the single card.
I have had for 3 years two GTX 970 in my system. Each one has 4 gb of vRam. Connected with NVLink. I never seemed that the amount of vRam available for Blender has been 8 Gb.
Now I have bought an second hand GTX 1080 Ti with 11 Gb of vRam.
Can not use the NVLink connector because diferent cards.
I just want to know if it is better mantain the 1080 ti with one GTX 970 or I should disconnect the GTX 970 and have only connected the 1080 ti.
And yes … re-reading the link I posted … its not possible to tell if they are talking about Blender.
The 970 and 1080ti uses an SLI bridge. NVlink is different from the SLI bridge. The SLI bridge does not share the VRAM of the cards, and its primary use was for gaming.
Nvlink was first supported on the 2000 series of cards, with the 2080 and 2080ti, from the information i found.
Keep both cards in your system. If you run out of VRAM on the 970, you can disable it in the preferences and just use the 1080ti.
If you know that you will regularly be exceeding the 4gb of VRAM, then you should remove it, since it will be hassle to constantly change preferences.
SLi Bridge =/= NVLink.
As for the 970. Assuming that you only use it for work not gaming and assuming you want to use the 1080 ti as the rendering GPU then you should switch the cards so that the 970 will be rendering your display output and the 1080 ti will be for compute only.
That is 970 in the top PCIe slot and the 1080 ti in a lower one. Display cable connected to the 970.
Well, so considering all your answers, I will sell both GTX 970 and will use the 1080 ti alone.
Thank you so much.