Hacker News new | ask | show | jobs
by klelatti 2029 days ago
This is a really interesting point. I wonder if there is another route though. If we assume that Intel (for process reasons) are out of it, then could AMD produce an 'x86 version' of the M1 on TSMC 5nm using Zen3 and successors which would help other firms compete with Apple.

I guess the extent to which M1 is great because its ARM isn't clear - there are some advantages in instruction decoding, having bigLITTLE etc - but the impact of all these together hasn't been quantified. Plus there is the deeper integration with Apple's software.
1 comments
rsynnott 2029 days ago
> then could AMD produce an 'x86 version' of the M1 on TSMC 5nm using Zen3 and successors

I mean, they could, but the A12Z in the devkits, produced on TSMC's 7nm process, also showed very impressive results. It's not clear how much gain the M1 is getting from 5nm, but I'd be _extremely_ skeptical that it's the whole story, especially given limited improvements between the directly comparable A13 (7nm) and A14 (5nm, some microarch changes).

> Plus there is the deeper integration with Apple's software.

This is relevant to some OS niceties (eg the disconcertingly instant wakeup and resolution shifts) but should have no bearing on, say, SPECInt2017.

link
klelatti 2029 days ago
Agree that they couldn't match - but if they got within say 15% on performance and battery life would that be good enough?

I think the software integration goes a bit further than wake-up etc e.g. they seem to have speeded up handling of reference counting through architectural changes (not sure I could explain how though and even Apple's software engineers seem a bit in the dark!)

link
rsynnott 2029 days ago
> but if they got within say 15% on performance and battery life would that be good enough?

I mean, depends what you mean by 'good enough'. It would, obviously, be good for AMD; I don't see that it would make it competitive in perf per watt terms, though.

> e.g. they seem to have speeded up handling of reference counting through architectural changes

They have, but that wouldn't be relevant to (most) synthetic benchmarks, I wouldn't have thought. SpecInt2017 won't be doing much if any reference counting, for instance :) The M1's advantage in ObjC and Swift should be expected to be even greater than its advantage in 'normal' (ie C) code, but beyond refcounting microbenchmarks I don't think there are many tools to demonstrate that (and the vast bulk of performance-sensitive code running on MacOS is C/C++ anyway).

link