Apple A9X the new mobile SoC king

[PricklyPete]

Those dumbed down slides are an insult to any smart brain interested in how stuff works.

This can be applied to nearly every marketing slide in the tech industry... We all realize that trying to distill technical intricacies in a marketing campaign is nigh impossible. The assumption is always that the company marketing their product is referencing a best case scenario

 

[Gikaseixas]

At this rate they'll catch x86 much sooner than we thought. A company as healthy as Apple can do this so it will be interesting to see Intel's reaction since they wont be happy if they loose Apple's business

 

[witeken]

But Intel isnt cheaper. There is no doubt that Apple is getting those huge SoCs for less than the cost of one tiny Intel skylake chip.

* Intel is an IDM. So wafers are dirt cheap compared to fabless companies.
* Intel has very good yields on a mature 14nm node, doubt anyone else comes close.
* As we know, 20nm + FF doesn't bring an increase in density (maybe move to SS will help, but there's still the increase in transistor count). So the SoC will be huge, which will add to cost (~100mm² for Skylake-Y vs 130mm²+ for A9X).
* Biggers dies have worse yields, increasing costs further.

 

[dahorns]

But Intel isnt cheaper. There is no doubt that Apple is getting those huge SoCs for less than the cost of one tiny Intel skylake chip.

What? Really? What do you base that on?

It is probably true that the direct cost to Apple for producing 1 A9X chip is less than it would cost to purchase a Skylake chip from Intel. Of course, that doesn't include the cost of hiring chip designers and other research involved in actually creating a chip.

 

[MikeA65]

I wouldn't be so sure its going to be the fastest tablet SOC, that's from Geekbench 3 a practically useless benchmark. Geekbench 3 shows the A8 with around Haswell level IPC while the far better SPEC2000 benchmarks shows roughly Core 2 level IPC.

Anandtech did a SPEC2000 run in its iPhone 6 review (http://www.anandtech.com/show/8554/the-iphone-6-review/3), the best SPEC2000 subtest is probably 176.gcc in which A8 got 1810 at 1.4 GHz, compared to 1954 for a 1.6 GHz Core 2 based Xeon 5110. That gives the A8 a bit more than Core 2 level IPC on this substest, certainly nowhere near the Haswell level shown by Geekbench 3. Until we get real benchmarks run (not Geekbench 3) we won't know where A9X really stands.

 

[IntelUser2000]

I wouldn't be so sure its going to be the fastest tablet SOC, that's from Geekbench 3 a practically useless benchmark. Geekbench 3 shows the A8 with around Haswell level IPC while the far better SPEC2000 benchmarks shows roughly Core 2 level IPC.

Spec2000 is an extremely outdated benchmark + latest Intel compilers have broken some subtests + more representative of Servers/HPC than mobile/consumers.

 

[Arachnotronic]

Spec2000 is an extremely outdated benchmark + latest Intel compilers have broken some subtests + more representative of Servers/HPC than mobile/consumers.

Is Geekbench 3 any better?

 

[MikeA65]

Spec2000 is an extremely outdated benchmark + latest Intel compilers have broken some subtests + more representative of Servers/HPC than mobile/consumers.

Yes it is, its still far far better than a toy benchmark like Geekbench 3. Spec2006 would be a lot better, but even that is outdated, a new Spec suite is long overdue. As for broken subtests, that is why I gave 176.gcc, that subtest has not been broken.

 

[dark zero]

At this rate they'll catch x86 much sooner than we thought. A company as healthy as Apple can do this so it will be interesting to see Intel's reaction since they wont be happy if they loose Apple's business

Actually I see that Apple managed to catch up the lowest Haswell BIG Core chip, which is a Celeron..., Apple ARM is no longer a joke and is becoming a real threat to Intel.

Seems that the next ARM uArch update would be a BIG one. And maybe AMD will be the first x86 company to suffer the power of ARM.

 

[Eug]

To put this in perspective, that means the performance of an SoC - A8 - in what is now a 2nd tier phone, the iPhone 6, beats my primary Windows laptop which is a 1.3 GHz Pentium SU4100 that's roughly a Core 2 Duo class CPU.

And their top iPad SoC - A9X - likely competes against my primary MacBook Pro CPU, which is a 2.26 GHz Core 2 Duo.

Yes these are old CPUs, but I still find this impressive actually.

 

[FwFred]

Before we get too excited, lets see how A9x does against the best Intel fanless SoC Core M7-6Y75 in Spec2k gcc or SpecInt 2006.

 

[Blitzvogel]

Sounds nice (though I'm one not to trust Geekbench), but can that performance be sustained for any kind of reasonable amount of time in a tablet form factor?

 

[jpiniero]

Well, Apple seems confident enough in the power consumption that they cut the battery on the 6S/6S+ by like 5% yet quote the same battery life. Granted that's the A9 and not the A9X.

* Intel has very good yields on a mature 14nm node, doubt anyone else comes close.

Intel's 14 nm yields are still crap on anything that's not a tiny die, that's why they delayed 10 nm more than anything.

 

[dark zero]

To put this in perspective, that means the performance of an SoC - A8 - in what is now a 2nd tier phone, the iPhone 6, beats my primary Windows laptop which is a 1.3 GHz Pentium SU4100 that's roughly a Core 2 Duo class CPU.

And their top iPad SoC - A9X - likely competes against my primary MacBook Pro CPU, which is a 2.26 GHz Core 2 Duo.

Yes these are old CPUs, but I still find this impressive actually.

It could be... wait.. maybe an x86 S.O like Windows 7 can now compile on the A9X?

 

[freeskier93]

Last year many people scoffed at the idea of an ARM processor in a Mac, saying ARM could never match Intel in performance or it would be many many years before they even got close. I think it's clear that the only thing stopping Apple now is application support, I think the bigger purpose of the iPad Pro is to further drive development of ARM applications, specifically the pro apps. I still think an ARM Mac is a real possibility in the next couple years, I'm sure looking even further out Apple would love nothing more then to ditch Intel.

 

[freeskier93]

To put this in perspective, that means the performance of an SoC - A8 - in what is now a 2nd tier phone, the iPhone 6, beats my primary Windows laptop which is a 1.3 GHz Pentium SU4100 that's roughly a Core 2 Duo class CPU.

And their top iPad SoC - A9X - likely competes against my primary MacBook Pro CPU, which is a 2.26 GHz Core 2 Duo.

Yes these are old CPUs, but I still find this impressive actually.

If you put any credit towards Geekbench then Apple passed C2D performance last year with the A8. Aka my iPhone 6+ has more processing power then my 2010 MBP. The A9X is definitely way beyond C2D performance.

 

[IntelUser2000]

And their top iPad SoC - A9X - likely competes against my primary MacBook Pro CPU, which is a 2.26 GHz Core 2 Duo.

I'm absolutely baffled how anyone can deduce it like that when Core 2 Duo at the same clock as Silvermont Pentium is only about 40% faster in Cinebench(per Anand bench).

And mobile benchmarks show that 1.4GHz A8 beats 1.86GHz Silvermont by 20-60%: http://www.anandtech.com/bench/product/1352?vs=1212

 

[MikeA65]

I'm absolutely baffled how anyone can deduce it like that when Core 2 Duo at the same clock as Silvermont Pentium is only about 40% faster in Cinebench(per Anand bench).

And mobile benchmarks show that 1.4GHz A8 beats 1.86GHz Silvermont by 20-60%: http://www.anandtech.com/bench/product/1352?vs=1212

Where are you getting 40%? I see a C2750 does 1898 in single thread Cinebench R10 at 2.6 Ghz (its turbo) while a 2.6 Ghz E4700 does 2943, that's 55% more.

 

[StrangerGuy]

Well, Apple seems confident enough in the power consumption that they cut the battery on the 6S/6S+ by like 5% yet quote the same battery life. Granted that's the A9 and not the A9X.



Intel's 14 nm yields are still crap on anything that's not a tiny die, that's why they delayed 10 nm more than anything.

Apple sold 74.3M iPhone 6 in 2014 Q4, and there is no doubt they will sell at least as many iPhone 6S in the same period this year. My point is not only Apple has caught up with 14nm finfets, they are also going to easily outship Intel's 14nm chips in volume this year too.

 

[Idontcare]

Last year many people scoffed at the idea of an ARM processor in a Mac, saying ARM could never match Intel in performance or it would be many many years before they even got close. I think it's clear that the only thing stopping Apple now is application support, I think the bigger purpose of the iPad Pro is to further drive development of ARM applications, specifically the pro apps. I still think an ARM Mac is a real possibility in the next couple years, I'm sure looking even further out Apple would love nothing more then to ditch Intel.

I think you may have gathered the wrong impression from the ARM vs. x86 discussions.

That ARM could (and would) someday reach x86 class performance was never really in question. What was in question was whether they would do it without equally reaching x86 class prices, die sizes, transistor budgets, and power consumption.

This is not to be conflated with the entirely different discussion of whether or not an ARM based x86 replacement device would be economically viable - and that is where software ecosystem comes in.

Apple is being ridiculously intelligent here. They are going to let the market decide which they prefer on the basis of form factor and software.

Does the market want iPad pro's and the accompanying ARM itunes apps to replace their mac books and accompanying wintel x86 software?

Apple is making it such that they don't have to be the company that decides one for the other, people and their wallets will make that choice for Apple in the coming 3-4 years.

 

[jhu]

Not sure this really matters to anyone not using Apple devices. Synthetic benchmarks can sort of gauge performance between the CPUs but since they're different platforms they'll never really be running the same software. Comparisons would be more useful if it was possible to run Android or Debian on a device with an A9X.

 

[Dresdenboy]

Where are you getting 40%? I see a C2750 does 1898 in single thread Cinebench R10 at 2.6 Ghz (its turbo) while a 2.6 Ghz E4700 does 2943, that's 55% more.

Why do you guys rely so much on a cache friendly fp heavy benchmark doing tasks which are relevant to <1% of the user base when discussing everyday use? The experience we often see is either GPU dependent or GUI related with lots of integer code, difficult to predict pointer handling and virtual function calls (all that OOP GUI app code).

 

[jhu]

I wouldn't be so sure its going to be the fastest tablet SOC, that's from Geekbench 3 a practically useless benchmark. Geekbench 3 shows the A8 with around Haswell level IPC while the far better SPEC2000 benchmarks shows roughly Core 2 level IPC.

Anandtech did a SPEC2000 run in its iPhone 6 review (http://www.anandtech.com/show/8554/the-iphone-6-review/3), the best SPEC2000 subtest is probably 176.gcc in which A8 got 1810 at 1.4 GHz, compared to 1954 for a 1.6 GHz Core 2 based Xeon 5110. That gives the A8 a bit more than Core 2 level IPC on this substest, certainly nowhere near the Haswell level shown by Geekbench 3. Until we get real benchmarks run (not Geekbench 3) we won't know where A9X really stands.

Wonder of it's because Geekbench's code/data is small enough to fit in cache?

 

[Stef123]

Yes it is, its still far far better than a toy benchmark like Geekbench 3. Spec2006 would be a lot better, but even that is outdated, a new Spec suite is long overdue. As for broken subtests, that is why I gave 176.gcc, that subtest has not been broken.

I find toy benchmarks much more useful to gauge every day performance than highly specific ones that really stress the CPU.

It's a bit of how SkyLake is often 100+% faster than SandyBridge in highly specific benchmarks yet the gains are much less impressive in every-day use or games.

So yeah in place of a better alternative I use Geekbench to gauge ARM speeds. I've been using ARM Linux for quite many years and the Geekbench scores correlate quite well to the capability of the CPU.

For example the new Exynos is definately i3-grade when installing Linux via chroot and certainly not C2D grade as those highly specific benchs would imply, simply because very few operations stress a CPU so far and if we talk about mobile OSes it's especially true as much of their software is relied on integer operations.

I trully and really expect the new A9X to be mobile-i5 Grade performance. I've almost no doubt about it even as we know little of the details of the SoC. It's not an accident that only now Apple is releasing their iOS laptop experiment, the performance has finally matured for that.

 

[raghu78]

I think you may have gathered the wrong impression from the ARM vs. x86 discussions.

That ARM could (and would) someday reach x86 class performance was never really in question. What was in question was whether they would do it without equally reaching x86 class prices, die sizes, transistor budgets, and power consumption.

This is not to be conflated with the entirely different discussion of whether or not an ARM based x86 replacement device would be economically viable - and that is where software ecosystem comes in.

Apple is being ridiculously intelligent here. They are going to let the market decide which they prefer on the basis of form factor and software.

Does the market want iPad pro's and the accompanying ARM itunes apps to replace their mac books and accompanying wintel x86 software?

Apple is making it such that they don't have to be the company that decides one for the other, people and their wallets will make that choice for Apple in the coming 3-4 years.

Apple is not doing that. They are clearly going to move to a single ISA in the not too distant future. That ISA is ARMv8. With A9X Apple has pretty much overtaken Core M and will continue to keep this relentless pace up. We will get confirmation form anandtech's testing. But the Apple custom ARMv8 core is already comparable to Intel's big cores like Haswell. There is a very good chance that we will see a high performance core from Apple in the not too distant future. A modified version of their custom ARMv8 core which can scale in frequency to 3.5+ Ghz and power their workstation, desktop and high end notebook requirements. Nothing stops Apple from doing that.

Intel's tick tock model is being bulldozed by Apple's model. Apple would love not to pay Intel those prices which they charge for Core M, Core i5 and Core i7. Every iPad Pro sold today is definitely a loss for Surface Pro and countless other Core M tablets as they are targetted at the same price segments, same purposes and same customers. Apple is going right after the x86 Wintel bastion and I think they will come out successful. Apple's margins will improve as they only have to pay a foundry like TSMC or Samsung for wafers rather than a chip manufacturer like Intel for finished CPUs. We are going to see Intel's x86 hegemony be destroyed over the next decade in all areas - from smartphone, tablet,notebooks,desktops and all the way to workstations and servers.