I tend to think its putting the memory on the package. Putting the memory on the package has given the M1 over 400GB/s which is a good 4x that on a usual dual channel x64 CPU and the latency is half that of going out to a DRAM slot. That is drastic and I remember when the northbrige was first folded into the CPU by AMD with the Athlon and it had a similarly big improvements in performance. It also reduces power consu…
On package memory increases efficiency, not speed.
However, most of the speed and efficiency advantages are in the design.