AnandTech SPEC2006 scores for ARM SoC

By: Maynard Handley (name99.delete@this.name99.org), November 9, 2018 5:12 am
Room: Moderated Discussions
Beastian (no.email.delete@this.aol.com) on November 8, 2018 5:53 pm wrote:
> Doug S (foo.delete@this.bar.bar) on November 8, 2018 9:25 am wrote:
> > Beastian (no.email.delete@this.aol.com) on November 8, 2018 7:10 am wrote:
> > > Apple's big core design is likely to service future high end laptops of theirs (depending on
> > > their plans) where as the A76 seems like it's dedicated to traditionally smaller form factors
> > > mobile designs. ARM still has a bit of die space to play with for their future cores:
> >
> >
> > Apple has had bigger cores for years, they believe it is worth the investment for mobile devices.
> > Being able to use the same basic core (probably with some tweaks) in Macs would be appealing
> > to them, but the Mac is the low volume product by comparison so the needs of the iPhone would
> > likely be taken into account before the needs of the Mac when compromises must be made.
> >
> > ARM has such a wide range of customers it probably doesn't make sense for them to invest in a design
> > as complex as Vortex, because the number of high end Android devices that would be potential customers
> > is limited - basically Samsung is the only one with annual sales of high end products hitting 8 digits.
> > With Samsung still pursuing their own cores the potential market for a really high end ARM designed
> > core able to compete with Apple's is even more limited. Not only would ARM need higher licensing costs
> > for that core to amortize the design costs, the OEMs would be paying more for the SoCs (though less
> > each generation since the CPU/GPU is a smaller percentage of the overall die)
> >
> > Unless ARM decided on a "built it and they will come" philosophy hoping that a market for ARM
> > PCs not designed for the bottom end (i.e. Chromebook) will just appear out of thin air...
>
> Apple's large cores are quite remarkable. One of Apple's goals in recent years has been to try to control every
> last component that goes into their products, and I still think they planned their big core uArch from the beginning
> to give themselves some legroom in case they wanted a very high performance iPad (which they got) or to transition
> their Mac code base (as they've done before) to ARM so they wouldn't have to pay Intel their ridiculous ~60%
> x86 margins. Their audience is captive due to their closed platform; MacOS and iOS have never been about compatibility
> at all cost but rather an exceptionally good experience within the limited scope of their ecosystem, and their
> (very adult) management has always been about maximizing gross margins.
>
> It's notable that the current draw on their their big cores have forced them to throttle clock speeds in
> as little as a year to preserve battery longevity including 2017 models. I still think their extremely robust
> big cores weren't necessarily designed with exclusive mobile usage in mind first and foremost, but were
> meant to dazzle in the short term with excellent benchmark numbers for launches and also to provide some
> head room for future iPads and Macs. It is my impression the IPC gains were fairly modest year over year
> since the A7 when they first introduced the 6 wide design, and much of the gains were from substantially
> ramped clockspeeds 1.3 ghz in A7 to 2.5 ghz in a12 and the vastly improved memory subsystem.

This is incorrect. There's a plot (which I can't now find) showing that Apple's GB4 performance follows as close an exponential as you might expect for these things, with an annual improvement of about 30%. To make the point clear:
A12 is 4x the performance of A7 (GB 4800 vs GB 1200) at twice the frequency.
IPC and frequency have been more or less equal partners along the whole adventure.

In any particular year there are particular lump improvements (eg jump from 2 to three NEON units, or jump in uncore) but smoothed out over the benchmarks and the years, the pattern is as I describe. The biggest single jump (as measured by GB4 single thread) was from A8 to A9, the smallest from A9 to A10.


> It is ironic that Windows on Arm which would require such a beefy core is forced to use inferior ARM designs
> despite having the software stack to fully emulate x86, whereas Apple has yet to move MacOS to their own
> hardware. I would definitely buy an ARM Macbook provided they let me boot Linux or Windows on ARM there
> if I wanted, though ARM is finally realizing that the pizzazz of benchmarks do matter to their customers
> and that they could have clout in the data center if they tried. They will soon mainstream the A76 and
> its two successors that will have it scraping very close to Apple's big core performance.

Perhaps. But this assumes that 2018 is the year that Apple hit their performance limit. I don't think that's the case; rather ARM will continue to maintain THEIR 30% annual improvement, but always lagging two years behind Apple.

As for why Apple is doing this (or the more or less equivalent "is it worth it")
- I've mentioned before the difference between tech companies run by engineers [or if you prefer founders] and those run by finance. Apple is, bizarrely run by engineers, for the purposes of this classification. Maybe this reflects DNA Jobs injected into it, maybe it reflects Tim Cook, maybe it reflects Johny Srouji being young enough to have huge ambitions, and having delivered enough to be allowed to do whatever he likes.
But Apple are not, I think, carefully calibrating every year "what's the bare minimum we can get away with" ala Intel; rather it's "what's the maximum we can reliably deliver", and we'll figure out a way to turn it into customer value.

- you guys won't be surprised to know that I buy a new phone (and iPad and aWatch) basically every two years, and have friends who are on the opposite cadence from me. And I can tell you that the constant speed improvements really are noticeable and do matter. I have no idea how you would quantify them; but they are there. Yes, in some utilitarian sense, the iPhone 6 was "fast enough". But switching to the 7 and having more fluid gestures, fewer places where you had a noticeable (even if only 1/10th of a second pause) or occasional frame stutter, was very nice. And now, damn, the XS is a whole lot smoother than the 7.

It's perhaps hard to convince someone who doesn't use an iPhone of all this; these are milliseconds on every interaction, and if your starting point in an unfamiliarity with iOS (so that when you play with a friend's phone, what you're noticing is how everything you want to do is slightly different to how it's done on Android) you won't pick this up. But it is real, and it's as worth paying for as it's worth paying for slightly nicer coffee, or slightly nicer steaks, or a slightly nicer hotel room.

I'm sure that Apple had hopes from the very start that they could ultimately create their own PC (and then server) cores. BUT I don't think that it's reasonable to say that their current cores overshoot what's needed for phones, and are being improved to target the desktop.
Apple are very efficient about re-using IP (or, to describe it more realistically, IMHO, are very good at suppressing the sort of petty empire building that, in so many companies, leads to idiotic and destructive proliferation of very minor variations on an underlying theme). So as much as makes sense will be shared between the iPhone cores and upwards to the desktop, in the same way that we've seen what makes sense be shared from the iPhone cores downward to the watch.
But the desktop will not be the same product as the phone, just like the watch is not the same product as the phone. And the opportunities available on the desktop (people are willing to pay more, energy budgets are higher to a lot higher, device sizes are a lot larger) will be exploited aggressively; the goal will not be simply to match iPads, neither will it be to match Intel then relax...

< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
AnandTech SPEC2006 scores for ARM SoCGian-Carlo Pascutto2018/11/07 01:29 AM
  AnandTech SPEC2006 scores for ARM SoCFoo_2018/11/07 03:11 AM
    AnandTech SPEC2006 scores for ARM SoCAndrei2018/11/07 04:47 AM
      AnandTech SPEC2006 scores for ARM SoCGian-Carlo Pascutto2018/11/07 06:18 AM
        AnandTech SPEC2006 scores for ARM SoCAndrei2018/11/07 07:02 AM
      AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/07 08:16 AM
        AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/07 01:34 PM
          AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/07 02:05 PM
        AnandTech SPEC2006 scores for ARM SoCDavid Hess2018/11/09 07:43 AM
          AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/09 09:02 AM
            AnandTech SPEC2006 scores for ARM SoCDavid Hess2018/11/09 03:31 PM
      AnandTech SPEC2006 scores for ARM SoCdmcq2018/11/07 08:46 AM
        AnandTech SPEC2006 scores for ARM SoCnone2018/11/07 09:24 AM
          AnandTech SPEC2006 scores for ARM SoCWilco2018/11/07 02:47 PM
            AnandTech SPEC2006 scores for ARM SoCdmcq2018/11/07 03:02 PM
              AnandTech SPEC2006 scores for ARM SoCWilco2018/11/07 04:56 PM
                AnandTech SPEC2006 scores for ARM SoCAndrei2018/11/08 01:20 AM
              AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/07 05:20 PM
              AnandTech SPEC2006 scores for ARM SoCnone2018/11/07 10:12 PM
                AnandTech SPEC2006 scores for ARM SoCnone2018/11/07 10:15 PM
          AnandTech SPEC2006 scores for ARM SoCBeastian2018/11/08 08:10 AM
            AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/08 08:22 AM
              AnandTech SPEC2006 scores for ARM SoCBeastian2018/11/08 08:32 AM
                AnandTech SPEC2006 scores for ARM SoCBeastian2018/11/08 08:36 AM
                  AnandTech SPEC2006 scores for ARM SoCAndrei2018/11/09 02:06 AM
                    AnandTech SPEC2006 scores for ARM SoCpassing rando2018/11/09 10:10 AM
              AnandTech SPEC2006 scores for ARM SoCGian-Carlo Pascutto2018/11/08 08:58 AM
                AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/08 11:03 AM
                  AnandTech SPEC2006 scores for ARM SoCGian-Carlo Pascutto2018/11/08 12:29 PM
                    AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/08 03:39 PM
                      AnandTech SPEC2006 scores for ARM SoCWilco2018/11/08 04:12 PM
            AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/08 10:25 AM
              AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/08 11:18 AM
                AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/08 03:00 PM
                  AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/08 03:48 PM
                    AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/09 12:15 PM
                      AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/09 08:23 PM
                        AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/10 03:43 AM
                          AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/10 07:46 AM
                            AnandTech SPEC2006 scores for ARM SoCJukka Larja2018/11/10 08:41 AM
                              AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/10 09:25 AM
                                AnandTech SPEC2006 scores for ARM SoCJukka Larja2018/11/10 12:54 PM
                                  AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/10 02:17 PM
                                    AnandTech SPEC2006 scores for ARM SoCJukka Larja2018/11/11 12:24 AM
                          AnandTech SPEC2006 scores for ARM SoCBrett2018/11/10 01:24 PM
                        AnandTech SPEC2006 scores for ARM SoCJukka Larja2018/11/10 04:12 AM
                          AnandTech SPEC2006 scores for ARM SoCdmcq2018/11/10 06:21 AM
                          AnandTech SPEC2006 scores for ARM SoCMichael S2018/11/10 05:33 PM
                            AnandTech SPEC2006 scores for ARM SoCAnon2018/11/10 08:52 PM
                              AnandTech SPEC2006 scores for ARM SoCJukka Larja2018/11/10 11:56 PM
                  AnandTech SPEC2006 scores for ARM SoCDavid Hess2018/11/09 09:08 AM
              AnandTech SPEC2006 scores for ARM SoCBeastian2018/11/08 06:53 PM
                AnandTech SPEC2006 scores for ARM SoCMaynard Handley2018/11/09 05:12 AM
                  AnandTech SPEC2006 scores for ARM SoCdmcq2018/11/09 05:26 AM
                  AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/09 12:06 PM
                AnandTech SPEC2006 scores for ARM SoCdmcq2018/11/09 05:15 AM
              AnandTech SPEC2006 scores for ARM SoCFoo_2018/11/09 01:16 PM
                AnandTech SPEC2006 scores for ARM SoCDoug S2018/11/09 01:36 PM
      AnandTech SPEC2006 scores for ARM SoCDavid Hess2018/11/09 06:27 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?