Doug S (foo.delete@this.bar.bar) on September 15, 2022 8:46 pm wrote:
> --- (---.delete@this.redheron.com) on September 11, 2022 9:07 pm wrote:
> > Rayla (rayla.delete@this.example.com) on September 11, 2022 4:34 pm wrote:
> > > --- (---.delete@this.redheron.com) on September 11, 2022 12:51 pm wrote:
> > > > Eric Fink (eric.delete@this.anon.com) on September 9, 2022 5:26 pm wrote:
> > > > > me (me.delete@this.me.com) on September 9, 2022 4:41 pm wrote:
> > > > > > > I dint think there is much pointing to pessimistic conclusions. If Apples claims are correct, they have
> > > > > > > obtained a 7-10% improvement in single core at 20% lower
> > > > > > > power usage as well as 20% improvements in multi-core
> > > > > > > (presumably at the same power usage). These are rather significant a year-to-year improvements.
> > > > > >
> > > > > > Apple didn't claim anything other than it's faster.
> > > > >
> > > > > They explicitly stated “20% lower power”.
> > > > >
> > > >
> > > > It's never clear whether codenames have a deeper significance or not, but people
> > > > may find it interesting that the A16 core codenames are Everest and Sawtooth.
> > > > First obvious point is that we have a clear break. We had wind names till we ran out of winds,
> > > > then weather-related names, culminating in Avalanche+Blizzard for the A15 (which certainly
> > > > *sound* like they're broadcasting "these are chips designed with energy foremost").
> > > >
> > > > The new names could represent nothing, but this does not feel like when Apple pivoted
> > > > from cat names to California names; we weren't running out of weather-related words.
> > > >
> > > > So to my eyes this is another piece of evidence for the proposition that Apple is rethinking their
> > > > CPU lines (and differentiating them more – mountains for P-cores, tools for E-cores?)
> > > > Maybe someone will pick up another CPU codename in a macOS build, which might
> > > > clarify if there is a third separate line (stars? rivers? whatever)
> > >
> > > I'm skeptical that "Sawtooth" refers to a tool here; it seems more likely,
> > > at least to me, that it refers to the Sawtooth mountain range in Idaho.
> >
> > Nice catch! One learns something everyday!
> > I guess the whole point of the smaller core is that it gets smaller (less well-known) mountain ranges!
>
>
> Starting on Everest doesn't exactly leave a lot of room for future growth
> taking the biggest one first. Unless A17 uses the Olympus Mons core lol

There are multiple ways to sequence mountains, and I expect that Apple has (more or less deliberately, both for fun and to screw with the internet public) chosen a non-obvious way.
For example you could sequence by date of first summit. Or by climbing difficulty -- end of the mountain names comes with Dawn Wall!

> Especially considering that A16 big core appears to be for all practical purposes
> an A15 big core with little or no change, as all the speedup looks like it
> is coming from clock rate gain and the greater bandwidth of LPDDR5.

This is the amateur viewpoint that energy saving just, somehow, happens, along with everything else in a CPU apart from compute! Of course no-one knows for sure, but there are multiple tracks Apple could be on. For example
- huge core track [nothing shipped yet] goes for performance, large core goes for energy, and every year they cross-propagate ideas
- the phone cores have been used to test new scaling ideas [I've mentioned new coherence protocol and new interrupt protocol] in "safe" environments where a mistake is not a disaster

ALL benchmarks test (especially benchmarks by these non-AndreiF amateurs who know fsckall of what they are doing) is *what they look for*.
If a large feature of the A16 is a substantially faster hypervisor [yes, there is reason to believe this is so] none of the amateur benchmarks will show it.
If a large feature is faster core-core communication (especially across rather than within clusters) none of the amateur benchmarks will show it.
If a large feature is much better "hetero-core" efficiency (ie much faster to move compute between CPU, GPU, and NPU) none of the amateur benchmarks will show it.
If AMX has been substantially improved, none of the amateur benchmarks will show it. [We have already seen this with AMX on the A15/M2, not that anything in the mainstream tech press ever tests AMX or has a clue how it is evolving.]
etc etc etc