As said above, the 7900XTX has both more memory, and also has more memory bandwidth, so it'll end up faster in most cases.
Software support is kinda meh and you won't be extracting 100% of the theoretical performance of that GPU, but you can get things done nonetheless, specially if you don't mind tinkering around. The 9070xt has no ROCm support still, but seems like some things can be done with the vulkan backend.
128GB of RAM would only be useful if you're offloading some layers from larger models into RAM, but then things would be really slow. If you already have 64GB, I'd say to continue with it until you feel the need to upgrade. Otherwise, you could try to start with 2x32/48/64GB and then add another pair of sticks down the road.
Eh, not really, each GPU has to keep the layers it's working on within its own memory. Having access to the memory of another GPU is not really useful in this case.
Those would be kinda slow, even slower than something like a RTX 3060.
Nonetheless, you can try to get those working, I've seen some reports of so in other forums, even if performance isn't all that great:
In this guide I’ll show you how to get stable diffusion up and running on your 100$ Mi25 on linux Cooling This thing does not come with a fan, you need to rig up your own cooling solution This thing is HOT and its heatsink is not that large, I had enough space to fit an entire blower...
forum.level1techs.com
