blue (blue.delete@this.blue.com) on August 12, 2020 11:25 am wrote:
> Maynard Handley (name99.delete@this.name99.org) on August 11, 2020 2:59 pm wrote:
> >
> > I expect them to ship around end of 2021. Whats the competition?
> > - AMD next? That's a legitimate unknown. But seems unlikely to be more than ~20% faster generically
> > (probably quite a bit faster for particular FP workloads), at the same power levels.
>
> Presumably if it is around the end of 2021, that would AMD next+1.
>
> So call it ~20% from whatever their Zen3 server CPU is, and another compounding 20%?
>
> iirc, servethehome found 7462 to use only a bit more power than the previous flagship 7601.
>
> AMD's competition for such a late 2021 part could be ~45-50%
> higher performance with about half the power per core.
>
> This would still put them below Nuvia's Pheonix cores perf/watt for this workload (assuming the different
> core implementations have the same pref/watt for Picasso versus chiplets, and various other things).
>
> So Nuvia's claims are not actually that insane (based on my understanding that
> they had people who have worked on very good architectures, like Apple's A series)
> and they're not insanely ahead of a decently optimistic gain for AMD.
>
> I think I had a conclusion in mind at some point, but lost it.

Do you believe that AMD can ship a substantial boost over Zen 3 by end of 2021?
Not to knock AMD, but that seems, uh, ambitious. Even when Intel was firing on all cylinders, the best they could do was Tick Tock.
I'd expect Zen 3 (7nm, IPC boost) to have minor tweaks for 5nm, the easy stuff (eg larger LLC) but no serious IPC boost. MAYBE they can pick up a few percent in frequency, but more likely, I'd expect, is they take the process win as lower power/able to sustain today's maximum frequency longer/over more cores simultaneously.

Sure, I see Apple as being able to do this and not AMD. Am I being hypocritical? No, I'm looking at the track record in both spaces. Like I said, even Intel at its best couldn't do what you're suggesting -- something about the combination of x86 and/or reaching for frequency (lotsa custom circuits, trying to tweak *any* feature impacts clock) and/or trying to target a single design from server to mobile seems to make that sort of pace impossible.