Freddie (freddie.delete@this.witherden.org) on February 2, 2023 3:39 pm wrote:
> —- (-.delete@this.redheron.com) on February 2, 2023 3:35 pm wrote:
> > So Apple’s answer to the OpenMP question would probably be to put the code
> > on P only and let the E cores handle whatever OS/IO work arises as they would
> > naturally, do t bother trying to squeeze out an extra few percent using them.
>
> Maybe. I think for the M1 Max series of chips there is an advantage to using the E cores
> if your code is memory bandwidth bound since the E core cluster can pull bandwidth above
> and beyond what the two P core clusters can. (Arguably, though, the fact the P cores in
> aggregate can only pull ~50% of peak bandwidth is a deficiency in the SOC design.)

The P-cores can pull in as much bandwidth as makes sense. Claiming that it's a deficiency is precisely the sort of obsessing over minor details rather than the big picture that has created the dichotomy between Apple and the other CPU vendors.
Yes, more CPU bandwidth is nice but it's not free. Would the additional bandwidth (which, let's remember is substantially beyond what any competitor can sport) beyond what's provided provide more value than using that energy, transistor, and interconnect budget elsewhere? Seems unlikely.

Like other unorthodox choices Apple has made, eg the SIMD latencies, or the single-ported D-TLB, it's easy to complain about these because the numbers are obvious. But mainly what that tells us is that the obvious numbers are not the full story; and obsessing over them rather than the full story is dumb.