someone (someone.delete@this.somewhere.com) on September 23, 2013 12:57 pm wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on September 23, 2013 11:02 am wrote:
> > someone (someone.delete@this.somewhere.com) on September 23, 2013 10:25 am wrote:
> > > Doug S (foo.delete@this.bar.bar) on September 23, 2013 10:03 am wrote:
> > > > someone (someone.delete@this.somewhere.com) on September 23, 2013 9:43 am wrote:
> > > > > jose (1.delete@this.2.com) on September 23, 2013 4:43 am wrote:
> > > > > > A while ago Charlie Demerjian was almost eaten alive when he stated that
> > > > > > Apple wanted/was planning to migrate to ARM for their laptop/desktop products.
> > > > > > Is it a coincidence that Apple is the first one with an ARM64 chip?
> > > > > >
> > > > > > Regards,
> > > > > >
> > > > > > Jose
> > > > > >
> > > > >
> > > > > The latest Macbook air with Haswell's single thread performance and multithread
> > > > > throughput measured with Geekbench 3 (sigh) are both more than double that of
> > > > > an A7 running 64b and triple A7 running 32b.
> > > > >
> > > > > Apple has a *long* way to go before it can replace Intel in its Macs without a huge
> > > > > drop in performance.
> > > >
> > > >
> > > > That doesn't suggest Apple has all that far to go. If they can get nearly 50% of Haswell's performance
> > > > on a 2 watt power budget, surely given Haswell's power budget they could design something to if not match
> > > > it, come close enough that it doesn't matter. I wonder how it compares on the graphics side?
> > >
> > > Let's say A7 was designed with an Fmax of 1.5 GHz while the Haswell core was designed
> > > with an Fmax of 4.0 GHz. This difference means that A7 designers can get away with
> > > shorter pipelines, higher skew/lower power clock system, bigger L1s with lower latency,
> > > far fewer global signal repeaters etc. compared to what Haswell designers had to deal
> > > with. This means compared to a Haswell voltage/frequency scaled down to Macbbook air
> > > levels A7 enjoys both a non-trivial uarch IPC bonus and a significant Watt/GHz bonus for
> > > operating near its design point. Despite this Haswell is still 2-3 times faster.
> > >
> > > OTOH this means you can't scale an A7 up closer to scaled down Haswell performance
> > > because your frequency gains with higher voltage rapidly hits a brick wall imposed by
> > > all those really nice design shortcuts permitted by a 1.5 GHz original Fmax target that
> > > were taken advantage of.
> >
> > You analysis has huge flow. It's called turboboost. Intel i5-4250U in MacBook Air has
> > so much higher single-core scores than Apple A7 almost exclusively due to turboboost.
> > So, it's ability to run at higher frequency (not 4 GHz, but 2.6 GHz is also pretty high relatively
> > to 1.3 GHz) is not a dis-advantage as you seem to claim, but very essential part of its advantage.
> > Here how it looks without turbo:
> > http://browser.primatelabs.com/geekbench3/compare/78462?baseline=56197
> >
>
> I am not sure of your point. You think it is hard to go from 1.5 to 2.6 GHz but
> the path from 2.6 to 4.0 is less painful?

No, I think, it is less painful than the path from 1.3 to 4.0

> This is pushing a boulder up a hill
> that gets sharply steeper with frequency not shallower.

It was true 15 years ago when transistor switching was still a dominating limiting factor. Since than we got them accompanied by wire delays, thermal dissipation and current supplies. And the last two got leakage as significant component.
So, may be it's still true or may be not. Only those designing modern CPUs know for sure.

>
> If Intel was designing for Fmax of say 2.8 GHz it'd have likely made many design
> choices differently. It could still turbo up to 2.6 and its perf/GHz and Watt/GHz
> would be somewhat better than existing Haswell at every point up to 2.6 GHz.

I am not sure that Haswell was not designed with 2.5-3.0 GHz as a primary target frequency and, as far as its designers are concerned, the fact that with a little higher supply voltage and a lot higher power budget it works well at 4.0 GHz is not just a bonus.