This approach is slicing a finite resource. It can only extend so far, and it sounds like they extended it about that far in one step. The amount of information the kernel keeps about each core has to be drastically reduced, for the next order of magnitude, or else cache hardware and behavior will need to change in comically-parallel chips.