CUDA has had managed memory that pages between VRAM and system RAM for a decade. Problem is doing so is unusably slow for AI purposes. Seems like an unnecessary layer here.

[dead]