I would like to use this thread also here in order to get an understanding about the shared memory behavior of the Apple Silicon design architecture if you agree @andrewc.
Did anybody was able to compare the memory behavior of having not sufficient VRAM avail on an nVidia GC and then tried this task to be done on a 128gb or 64gb shared memory Apple Silicon hardware instead? This question is not about speed, it is about using theoretically a higher VRAM limitation of an M2 Ultra for example and if this is managed sufficiently to work with it.