Multiple GPUs... do people have a workflows?
retained text
Ok so my humble rig has 2 cards. One 16 GB of VRAM and another with 12 gb . I know you can't combine them for a pool of 28 GB of unified VRAM but shouldn't their be a way to offload one of the major items like the CLIP or the VAE to my other card? I looked into it and doesn't seem to be supported natively in Comfy UI. There are some custom nodes I found but they don't seem too popular. Is this just a silly idea? Normally my second card just sits around at 20% memory usage running my GUI and a few bits and bobs. I could certainly afford to use 8 or 9 GB of my second card towards something instead of it offloading that 8 or 9 GB during a generation to my system ram (32 gb ) which is what comfy does now when using H3 or MOSS. I have no idea if this would be faster than dumping everything to the system ram but it would definitely give me more breathing room and less OOM errors on those borderline gens and upscales. **Is this idea loco or am I on to something?** EDIT: I appreciate all the people who are saying just use the second card for an llm or to run other types of generations on my smaller card, but that's not what I am proposing. I know you can do that. In fact I do it quite often. Let me explain this a different way that might make more sense. What I really am asking is there a way, when doing a single generation, let's say a 15 second h3 generation at over 1 MP, for comfy to prioritize memory allocation in order of 1) the 5080, 2) the 3060 3) the system ram 4) SSD swap file . Like a ladder of priority. For example the minimax H3 audio and video VAE are over 6 gigs at fp16. That would certainly fit on the 3060s available space. I know it might be slower but that extra 6 gigs of free space would be quite helpful for higher resolution gens.
document fetched 2026-11-15 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1xvp4qu/multiple_gpus_do_people_have_a_workflows