It is not that as pointed out in the end of the document, but whether this can help large memory model as accessing >16gb memory is expensive. Whilst not as far it is the memory that one is thinking about.
It is the alternate path some may go if this is true.
Toyed with similar. It works, but “toy” is the operative word here. It’s tangibly slower than a proper gpu.
Think the only one that has pulled this off in workable fashion are the M1/M2