It would apply equally to GPU or RAM inference as those are also bandwidth constrained on decode, so people already try to optimize for it.

surely the supply of unified memory will rise to meet demand before this is needed