Using different machines for model and draft model across LAN? Chaining draft models?

[itch-]

These new announcements of APUs that can use much more vram got me wondering. I doubt I'd enjoy their tk/s speed, but what if I can still use my GPU to run a draft model while across LAN I have a (very) big model running on an APU? The overhead may not be pretty but there should surely still be a huge speed increase over using just the APU.

And if I do have all my vram to spend on draft models, what about chaining them? I would have enough to split between a big draft model plus a tiny draft model for the big draft model.

Maybe it's crazy but I gotta put it out there.