I heard from a a hardware developer that the main wins for the M* architecture is really that they integrated the memory onto the cpu die and then the unified memory architecture between the GPU and CPU which was "holy grail" type stuff a decade ago.

I don't think we're going to see the performance increases past the ~10-20% each fab cycle now.

Performance per watt and having dedicated HW for video decoding / other tasks(AI?) I think might end up being the answer to the continuation of Moore's law. Especially considering that we are in the era of heterogenous computing.