I think the main benefit of joining two dies together via some advanced interconnect is the latency. Memory lookups are much quicker on Apple Silicon which, in turn, avoids many wasted clock ticks waiting for data.
TBH I prefer AMD's 7900 series approach of pushing the memory controller + cache out onto little memory controller dies. This allows for a monolithic GPU with a really wide bus, without all that wasted space on the controller/cache/pins. And those dies don't care about being split up since the memory access is interleaved anyway.
There is the little, medium, and big die, and the 2x big die.
Dual die chip is the ultra.
That said I think it's pro and ultra, and I can't recall which (if not both) use the magic interconnect