Apple A9X the new mobile SoC king

[Nothingness]

Aaron, you've been away from the ARM world for too long. You're basing your arguments on 15 years old experience, most of which is not applicable any more.

 

[stingerman]

The low hanging fruit for Apple in the PC market is the high performance Mac Pro which is sold to customers they mostly own, Graphics, Audio and Video. Not to mention the scientific market. It would make sense for Apple to release an A10x Mac Pro, reducing it's price considerably since a big part of the cost are the CPU and GPU from third parties. The Notebook and iMac markets would follow as important software is ported over. Though the transition from x86 to Arm using Xcode will be much easier than any other transition. The A9X was clearly marked as a desktop class processor by Apple with the iPad Pro marketed as faster than 80% of the notebook class CPU's actually sold. I don't know what clearer message Apple can give about the the direction they are headed.

The cost difference to Apple is significant. More than financial cost, the intellectual cost as well.

 

[simboss]

Neither will ARM help you get it right if you take out an architectural license.

Do you really think ARM only gives there architecture licensees the ARM architecture reference manual and nothing else?
Put it the other way, do you think ARM would let anyone put on the market an "ARM" cpu that has not even been through a validation that proves it can run ARM code properly?

The tools they sell won't get you that far. There are free tools that are basically the same thing for x86 as well, fyi.

What kind of tools are you talking about?
I don't think Intel provides to anyone anything comparable to what ARM provides to its architectural licensees.

And it is not impossible for anyone today to license the x86 ISA. Nor is it too complex, there are numerous examples of companies creating totally independent x86 designs.

Can you give examples of any company having taken an x86 license recently?

As said by someone else, you seem to have a very dated view of this, no one today can create a x86 CPU even remotely competitive.
You can of course read all the documentation about the x86 ISA and in theory come up with a x86 compliant implementation (forgetting about the patents), but the devil is in the details, a documentation will have flaws, gaps and ambiguities, and the only way to get it right would be to reverse engineer what Intel does in each of these cases.

I am pretty sure this is what AMD has been doing for a very long time, hence it has been at a disadvantage most of the time (apart from the x86-64 period where Intel got it so wrong that the tables turned).

If you have an ARM license, you just ask, and you get an answer.
The answer might be: this is not defined, do what you want, but at least it should be clearly stated in the documentation and the SW will have to deal with it.

Fine, I'll agree that neither is open.

Arguing about openness of ISA is indeed not very relevant, by definition an ISA is well defined (ie closed), but also always fully exposed (ie open) because it is the contract between the CPU and the SW.

At most you can add extensions to it, but unless the mechanism is well defined in the original ISA (which I believe is what RISC-V has done), it will always be "hacky" (eg: x86-64).

Also, its important not to confuse Open with you can do whatever you want. There are lots of examples of open standards with de-facto standards/restrictions. SPARC is technically Open but good luck adding instructions and getting adoption/support. Same for RISC-V.

The process to define this ISA can be more or less open, and in that regards, RISC-V is open, x86 and ARM are closed and happens behind closed doors.

You can define the openness by the possibility of someone else to implement the ISA in HW, which I believe was the starting point of the discussion. and in that regard, TODAY, ARM is open to anyone willing to pay for a license (and this seems to be an increasing number), whereas it is impossible to do it for a full x86 implementation.

AFAIK, there has only been 1 company that has received a full architectural license from ARM which was DEC. DEC's license allowed them to add to/modify the ISA, IIRC. After DEC's license it was over a decade before another architectural license was granted and AFAIK, it was not as extensive as the license that DEC originally got. I don't know if the license was modified when StrongARM was transferred to Intel. Nor do I know if Intel kept the license intact after selling StrongARM to Marvell.

You are mixing architectural license with the right to extend the ISA, these are 2 different things.
They are plenty of architectural licensees for ARM today, but afaik it is not possible to extend the ISA anymore (although it is possible to restrict it by not supporting optional features).

Again a very dated view, things have changed a lot since the DEC days.

 

[imported_ats]

The low hanging fruit for Apple in the PC market is the high performance Mac Pro which is sold to customers they mostly own, Graphics, Audio and Video. Not to mention the scientific market. It would make sense for Apple to release an A10x Mac Pro, reducing it's price considerably since a big part of the cost are the CPU and GPU from third parties.

The pricing for the Mac Pro is largely irrelevant of the CPU/GPUs included. The biggest factor in the price of the Mac Pro is apples 2-3x markups on the parts...

And if you think the A10x is going to compete with Xeon E5s...

 

[imported_ats]

Do you really think ARM only gives there architecture licensees the ARM architecture reference manual and nothing else?
Put it the other way, do you think ARM would let anyone put on the market an "ARM" cpu that has not even been through a validation that proves it can run ARM code properly?

They have a basic validation suite, but that pretty much the easy part of any validation. Get an ISS(instruction set simulator), run code. The majority of the validation for a new core design or heavy modification to an existing core design is at the unit level which ARM isn't going to be providing much if anything for.

What kind of tools are you talking about?
I don't think Intel provides to anyone anything comparable to what ARM provides to its architectural licensees.

ISS, frameworks, et al.

Can you give examples of any company having taken an x86 license recently?

no because pretty much no one wants to compete directly with Intel.

As said by someone else, you seem to have a very dated view of this, no one today can create a x86 CPU even remotely competitive.

That has little to do with the fact that there are numerous companies out there that have the full legal ability. AMD has and has had a full license to everything and they haven't been able to compete, but that is down to poor execution and strategy more than anything.

The same can be/will be said for many of the ARM licensees trying to compete with ARM...

You can of course read all the documentation about the x86 ISA and in theory come up with a x86 compliant implementation (forgetting about the patents), but the devil is in the details, a documentation will have flaws, gaps and ambiguities, and the only way to get it right would be to reverse engineer what Intel does in each of these cases.

Welcome to ANY ISA design from scratch. The same is true with ARM, there are going to be flaws, gaps, and ambiguities. HELL, that's still an actual issue with IEEE FP! So you do what most people do, you go with what every other device does.

I am pretty sure this is what AMD has been doing for a very long time, hence it has been at a disadvantage most of the time (apart from the x86-64 period where Intel got it so wrong that the tables turned).

If you have an ARM license, you just ask, and you get an answer.
The answer might be: this is not defined, do what you want, but at least it should be clearly stated in the documentation and the SW will have to deal with it.

The x86 ISA is at least as well defined in available public documentation as ARM is. In fact, x86 has overall a more thorough set of documentation at this point than pretty much every other ISA.

You can define the openness by the possibility of someone else to implement the ISA in HW, which I believe was the starting point of the discussion. and in that regard, TODAY, ARM is open to anyone willing to pay for a license (and this seems to be an increasing number), whereas it is impossible to do it for a full x86 implementation.

You keep using that word impossible, but it really doesn't apply here. I'm pretty sure that Intel would license x86 to anyone willing to pay the licensing fee as well...

You are mixing architectural license with the right to extend the ISA, these are 2 different things.
They are plenty of architectural licensees for ARM today, but afaik it is not possible to extend the ISA anymore (although it is possible to restrict it by not supporting optional features).

Again a very dated view, things have changed a lot since the DEC days.

Yes, they have gotten more restrictive.

 

[ShintaiDK]

x86=Make a new design, spend countless billions and years and hope it works out.
ARM=Get a license and start producing right away, the one who can make it the cheapest wins.

That´s essentially the difference between the 2 in terms of entry cost. Custom ARM designs is dying out fast for the same reason, just look at Qualcomm.

Nobody wants to compete in x86 because they can in reality only lose big money and time.

 

[NTMBK]

Custom ARM designs is dying out fast for the same reason, just look at Qualcomm.

Not sure how you came to that conclusion. Qualcomm did great with their custom core in the last generation. This generation they had to go generic with the A57, and lost lots of market share as a result. Next generation they are going back to custom. The moral I get from this is that custom CPU was a good differentiator for Qualcomm, and that they regret not having it this time around.

 

[Andrei.]

Qualcomm did great with their custom core in the last generation.

Much of that was due to excellent execution on Qualcomm's part (Good 28HPM designs) and failure of execution on the competition's part (Broken bL designs, mediocre physical implementation, disadvantageous process).

None of it having anything to do with the actual architecture of the CPUs.

 

[Phynaz]

Put it the other way, do you think ARM would let anyone put on the market an "ARM" cpu that has not even been through a validation that proves it can run ARM code properly?

No, that's why the people with with architecture licenses don't call their CPUs "ARM" CPUs. Because they aren't ARM CPUs anymore.

 

[simboss]

No, that's why the people with with architecture licenses don't call their CPUs "ARM" CPUs. Because they aren't ARM CPUs anymore.

No, that's called marketing:
- QC choose not to say ARM because they are ARM customers as well as competitors
- Apple, not sure what their official line is, but basically it is irrelevant for their marketing target
- Applied Micro or Broadcom very happily say that their custom core is an ARM core

X-Gene World’s First ARMv8 64-bit Server on a Chip Solution

vulcan:

Quad-issue, quad-threaded 64-bit ARMv8-A core

 

[Dresdenboy]

No, that's why the people with with architecture licenses don't call their CPUs "ARM" CPUs. Because they aren't ARM CPUs anymore.

Define "ARM CPU".

And I don't agree with you about Kryo, Krait, A9 etc. not being able to run all designated ARM ISA variant's code. Because that's the actual meaning of your contra-validation answer.

 

[simboss]

They have a basic validation suite, but that pretty much the easy part of any validation. Get an ISS(instruction set simulator), run code. The majority of the validation for a new core design or heavy modification to an existing core design is at the unit level which ARM isn't going to be providing much if anything for.

Of course if you are literraly starting your design "from scratch", but noone does that anymore (because it would take years and mountains of cash for an very uncertain result), you take an existing design, tweak it and sometime switch it to a different ISA.
This is what Nvidia did with Denver, this is what Broadcom is doing with Vulcan, this is what Applied Micro did, this is what AMD is (supposedly) doing, ...

ISS, frameworks, et al.

Do you have any source for that?
This is the first time I see anyone saying this, although obviously this may not be public information, I don't see why Intel would make these tools available if they don't get anything out of it (ie a license fee).

no because pretty much no one wants to compete directly with Intel.

I am not arguing Intel does not do a very good job at doing x86 CPUs.
but Intel is clearly not encouraging anyone to compete (as opposed to ARM).
This may well be their demise, because if they miss, noone will keep x86 afloat.

That has little to do with the fact that there are numerous companies out there that have the full legal ability. AMD has and has had a full license to everything and they haven't been able to compete, but that is down to poor execution and strategy more than anything.

Funny that the only example you can give is AMD who has acquired the rights more than 30 years ago.

The same can be/will be said for many of the ARM licensees trying to compete with ARM...

Many, not all.

Welcome to ANY ISA design from scratch. The same is true with ARM, there are going to be flaws, gaps, and ambiguities. HELL, that's still an actual issue with IEEE FP! So you do what most people do, you go with what every other device does.

Again, the difference is that you have a contract with ARM that entitles you to do it and get support for it.

The x86 ISA is at least as well defined in available public documentation as ARM is. In fact, x86 has overall a more thorough set of documentation at this point than pretty much every other ISA.

Which is still one big text document very hard to translate into a CPU and validate. ARM won't help you with the design (but as said earlier, if you are starting from scratch today, you are clearly very optimistic), but will with the validation (in a non-trivial way).

You keep using that word impossible, but it really doesn't apply here. I'm pretty sure that Intel would license x86 to anyone willing to pay the licensing fee as well...

Chicken and egg problem.
It is even less likely to happen because ARM is the obvious solution if you need an ISA for your new CPU.
You can argue that "on principle" anyone could do an x86 cpu, in practice noone does.

Yes, they have gotten more restrictive.

And more attractive at the same time.

 

[stingerman]

There is a difference between architecture and micro architecture. Apple licenses the architecture, but designs their own micro architecture. Alternatively, they could have used ARM's reference designs which have already been validated. But, their acquisitions tell a different story. Their goal and they have achieved it was to build a desktop class processor within mobile phone constraints.

Can these processors be clocked to 3 to 4 Ghz if they enjoyed the cooling systems of an actual desktop? Why not? I'm sure they have these things water cooled just to see how far they can stress them back in the lab.

And, it would be interesting to learn what the new AppleTV processor is running at with it's larger heat sink.

 

[witeken]

Their goal and they have achieved it was to build a desktop class processor within mobile phone constraints.

LOL!!

Do you know any ARM chip in the world that:

* Has similar IPC to Skylake
* Can be clocked at say 4.3GHz
* At 1.2V

Krait, for example, reaches something like 1.3V at half the clock speed :biggrin::biggrin:.

Can these processors be clocked to 3 to 4 Ghz if they enjoyed the cooling systems of an actual desktop? Why not? I'm sure they have these things water cooled just to see how far they can stress them back in the lab.

Why not? Because they haven't been designed for those frequencies. They have been designed for 4W or so, so why would they have bothered with higher frequencies than 2GHz something? It costs time and transistors and power and die area and money to optimize it for higher frequencies at a reasonable voltage. It just won't work. And never mind cooling the chip seeing how high Apple pushes the transistor density.

Without additional efforts, Haswell wouldn't go higher than 2GHz.

 

[Eug]

Who says "desktop class" must mean Skylake?

 

[Thala]

LOL!!

Do you know any ARM chip in the world that:

* Has similar IPC to Skylake
* Can be clocked at say 4.3GHz
* At 1.2V

Krait, for example, reaches something like 1.3V at half the clock speed

Clock frequency is much more a function of technology (e.g. low VT/high VT) and VDD than architecture. And then of course Skylake 4.6 GHz is binned from the FF corner.
There is no question at all that A9x can run at least double the stock frequency given the right parameters and then selecting the FF corner without adding a single flop with some active cooling.

Actually that is a big opportunity for Apple, using the chips from FF corner for desktop and from SS corner for iPhone/iPad. That is precisely what Intel is doing.

 

[stingerman]

LOL!!

Do you know any ARM chip in the world that:

* Has similar IPC to Skylake
* Can be clocked at say 4.3GHz
* At 1.2V

Krait, for example, reaches something like 1.3V at half the clock speed :biggrin::biggrin:.

Why not? Because they haven't been designed for those frequencies. They have been designed for 4W or so, so why would they have bothered with higher frequencies than 2GHz something? It costs time and transistors and power and die area and money to optimize it for higher frequencies at a reasonable voltage. It just won't work. And never mind cooling the chip seeing how high Apple pushes the transistor density.

Without additional efforts, Haswell wouldn't go higher than 2GHz.

Since you raised it, what is the average IPC for Skylake?

Overclockers have been breaking the rules for a couple of decades now. No reason not to think that Apple's clocks could not be overclocked to 3 to 4 GHz. After all the A9 is sustaining 1.8GHz at a low voltage, made possible by FinFet. A9X is probably at or higher than 2Ghz. Better yields come over time which means the A9 should eventually be cost effective at 2Ghz without increasing voltage.

I think the IPC for A8 was equal to Intel's current core at the time. It just might be the A9 surpasses Intel?

 

[imported_ats]

Of course if you are literraly starting your design "from scratch", but noone does that anymore (because it would take years and mountains of cash for an very uncertain result), you take an existing design, tweak it and sometime switch it to a different ISA.
This is what Nvidia did with Denver, this is what Broadcom is doing with Vulcan, this is what Applied Micro did, this is what AMD is (supposedly) doing, ...

Once you change the ISA of a design, you are pretty much validating it from scratch. So yes, Nvidia, Bcom, APM, and AMD, basically need to do a full validation path throughout the whole design. The front end has to change pretty radically when changing ISAs and that has repercussions throughout the design.

Do you have any source for that?
This is the first time I see anyone saying this, although obviously this may not be public information, I don't see why Intel would make these tools available if they don't get anything out of it (ie a license fee).

A simple example is QEMU and there are other. There are lots of reasons people would need/want an ISS outside of just microprocessor design.

Funny that the only example you can give is AMD who has acquired the rights more than 30 years ago.

IBM, Via, and there are others as well.

Again, the difference is that you have a contract with ARM that entitles you to do it and get support for it.

Which if you are designing your own core is basically useless. All they are going to do is say to do what the documentation states or what existing designs already do. The support is more useful if you are using one of ARMs designs, but they aren't and really can't handhold a vendor doing a unique CPU design.

Which is still one big text document very hard to translate into a CPU and validate. ARM won't help you with the design (but as said earlier, if you are starting from scratch today, you are clearly very optimistic), but will with the validation (in a non-trivial way).

Besides a pretty basic set of tests and an ISS, they aren't going to be able to help your validation in any meaningful way. They aren't even going to understand what your various sub-units are or what tests are appropriate for those unit tests. If I'm designing anything complex anyways, the ISS and associated test suite may in fact be completely useless.

 

[imported_ats]

Actually that is a big opportunity for Apple, using the chips from FF corner for desktop and from SS corner for iPhone/iPad. That is precisely what Intel is doing.

Um, no, devices from the FF corner tend to be used in the lower power spaces, not the higher performance spaces, fyi.

 

[erunion]

Are imported_stingerman and imported_ats the same person?

 

[DrMrLordX]

Has Internet Strongman Juanrga destroyed you so completely that you have totally lost hope with AMD and out of desperation have now jumped onto the ARM bandwagon?

ahahahah don't say that name! He might come here!

But maybe AMD should hire an Internet Strongman . . .

Who says "desktop class" must mean Skylake?

Clearly it means Bulldozer.

. . .

(okay, maybe not)

 

[antihelten]

Not sure how you came to that conclusion. Qualcomm did great with their custom core in the last generation. This generation they had to go generic with the A57, and lost lots of market share as a result. Next generation they are going back to custom. The moral I get from this is that custom CPU was a good differentiator for Qualcomm, and that they regret not having it this time around.

Not only that, but other major players are also getting into the custom core game. Namely Samsung with their Mongoose core, which is allegedly set to make an appearence in the GS7 in the form of Exynos 8890 (should be arriving around the same time as Snapdragon 820 equipped phones). Leaked benchmarks show 8890 getting fairly close to A9 as well (slightly lower single threaded performance and much higher multi threaded performance, although the latter would largely be down to having 4 cores vs. 2 cores)

LG is also getting into the SoC design business with their NUCLUN SoC, although they appear to be sticking to off the shelf designs for now.

Clearly it means Bulldozer.

. . .

(okay, maybe not)

I would actually be kinda curious to see how an octacore ARM chip (with one of these new fancy custom cores) did against Bulldozer.

 

[thunng8]

For A8X, I get just under 2340 ms in Kraken 1.1 on iOS 9.1.

For A7 in the iPhone 5s I get 3542 in iOS 9.1, which is a huge improvement from before. It was 5905 in the AnandTech review.

Meanwhile, for A5 in my iPad 2 with iOS 9.1, I get just over 25000, at 25024.

Just ran kraken 1.1 on the 6s and got 1739ms.

So just over double the speed of the 5s which scores 3540ms using same ios9

Impressive speed increase in just 2 years

 

[IntelUser2000]

Just ran kraken 1.1 on the 6s and got 1739ms.

So just over double the speed of the 5s which scores 3540ms using same ios9

Impressive speed increase in just 2 years

First it was Geekbench, now its Kraken. 1700ms in Kraken 1.1 rivals the Core M in Yoga 3 Pro. Now, this is for the A9. Core i7 5600U gets 1173ms. Core i5 5200U gets 1660ms. That means A9X is probably going to beat the Core i5 5200U.

The scores suggest in Kraken, the Core M performs quite close to U chips since its a light and short benchmark. It seems that Intel designed the Core M to be "ARM competitor Core chip", except they aren't doing such a good job at it and ends up being a bad PC chip.

This is quite amazing. A 15W Core i7 is only 25-30% faster than A9X. Skylake is going to add 10% to that.

If this turns out to be true in more tests, I see a harsh wind blow against Intel for which will start their decline, and they won't have a Pentium M style fallback to save them this time.

 

[Eug]

Just remember though iDevices do exceptionally well in JavaScript tests like Kraken because Apple optimizes the hell out of Safari. Great for the iOS browsing experience, but not great for cross platform comparisons... except for JavaScript.

As evidence of that, as I've illustrated with the test result I posted with A8X, the results improve dramatically with each iOS version release using the exact same SoC.

OTOH it's good for intra-platform comparisons. A9 does exceptionally well when compared to A8(X). Impressive generational improvement.

Oh and congrats on getting your 6s already.