Blender new GPU(s) - does RAM size matter?

I want to purchase a 2nd or more GPUs for Blender and i want to know if the RAM size matters.
Do i get a CUDA error when i have a 2GB RAM GPU and a 4GB RAM GPU(e.g. when loaded too many textures). Does the smaller RAM size count or are they combined to 6GB RAM(which i think does not work) - or does only the 4GB RAM GPU matter?
Should i get GPU’s with the same RAM size as my main GPU?

The scene has to fit into every card’s VRAM, so the card with the least amount of VRAM becomes the limiting factor. If you have one card with 8 GB and one with 2 GB of VRAM, you can effectively only use 2 GB (unless of course you decide to render only on the larger card and use the other one for display only).

For Cycles what IkariShinji has said :slight_smile:

But as a side note thinking in single card and Eevee (OpenGL), in my test with Classroom scene for Eevee, it almost took 3.5GB of vRAM for 4k render. So thinking about Eevee and future, amount of vRAM also matters.
To be more precise, Blender with Eevee in 4k took almost 3GB. But you must add what the operating system takes of vRAM. I still do not know what happens in Eevee when vRAM does not fit (some error message? Does everything explode? :slight_smile: )

Edit 2:
I’ve opened two instances and rendered 4k at the same time with Eevee. Two instances render well without error messages. So now I’m not sure how vRAM usage is handled by Eevee/OpenGL. I’ll have to ask developers. For now you forget what I said about Eevee. Just pay attention to what IkariShinji has told you about Cycles.

You get the total vram of the card with the least vram when rendering on a set of differing cards.
Alternatively you can run the card with more vram as a render device and use the lesser card to drive your monitor(s). Doing this tends to lessen the effects of a laggy viewport in rendered view and locking up during rendering.

Thanks for the answers. Now i have a better idea of what i’m buying.