Apple A12X: 4 Big, 4 Small cores

[PeterScott]

Starting when they released their first 64 bit part, and every iteration after that as their performance crept up to laptop/desktop performance you would have the usual cast of characters suggesting that they could never get high clocks or high core counts. The last couple of iterations those voices have been somewhat subdued.

So shifting goal post to include clock speed now. This I have seen discussed, as it is actually a design limitation. You can't simply take a low clock mobile chip and run it at 4-5 GHz. Architecture for high clock speed is different than for low clock speed.

Core counts on the other hand have no such limitation and I never saw anyone state this. If you did, what was the argument against doing it?

 

[scannall]

So shifting goal post to include clock speed now. This I have seen discussed, as it is actually a design limitation. You can't simply take a low clock mobile chip and run it at 4-5 GHz. Architecture for high clock speed is different than for low clock speed.

Core counts on the other hand have no such limitation and I never saw anyone state this. If you did, what was the argument against doing it?

I have no doubt that Apple can scale cores and clockspeeds however much they like, within the limitations of the process they are using. Currently they are targetting mobile devices, and are getting amazing results. If they wanted to target desktops, I'm certain they have the talent for that as well.

 

[Etain05]

So shifting goal post to include clock speed now. This I have seen discussed, as it is actually a design limitation. You can't simply take a low clock mobile chip and run it at 4-5 GHz. Architecture for high clock speed is different than for low clock speed.

Core counts on the other hand have no such limitation and I never saw anyone state this. If you did, what was the argument against doing it?

Have you really never seen someone state this? We had a discussion about this very thing on a different thread, and there were people who suggested exactly that: https://forums.anandtech.com/threads/apple-cpus-just-margins-off-desktop-cpus-anandtech.2554913/

 

[moinmoin]

But a True Console is needed. Nintendo would be outright killed by Apple if they do that.

Also.... Puting that on an Apple TV with some overclock would make it a brutal beast capable to play Console grade games.

Is it needed? Does Apple need it?

The dedicated gaming device market is a losing market. In Japan for example the mobile gaming market (where Apple devices dominate) is over five times the total revenue of the dedicated gaming device market, even though the latter includes hardware while the former is software only. Hence one could say Nintendo is already dead and Apple is already victorious.

 

[DrMrLordX]

Raven Ridge?

True, that's the oddball of the lot. Though it's lower volume than the non-iGPU products. AMD basically needed it to round out their APU lineup (to replace Bristol Ridge etc.). A12/A12x already has graphics IP so at least theoretically, Apple needs even fewer die designs than AMD. Not sure if they want to replicate the iGPU so many times over though. Maybe they do!

Fair enough. They could go full chiplet design. But is that even possible with big.little?

Not sure if it's possible. To the best of my knowledge, anyone dealing with A75 or later ARM designs now uses DynamIQ instead of "big.LITTLE", but I'm not sure what are the technological implications of that shift.

 

[Arachnotronic]

True, that's the oddball of the lot. Though it's lower volume than the non-iGPU products.

Most PCs sold have integrated GPUs.

 

[PeterScott]

Have you really never seen someone state this? We had a discussion about this very thing on a different thread, and there were people who suggested exactly that: https://forums.anandtech.com/threads/apple-cpus-just-margins-off-desktop-cpus-anandtech.2554913/

Yes, I really never saw that before. I don't read every post on every forum.

I have read these discussions on multiple forums and I have seen the clock speed argument come up a dozen times or more, and I have made that argument myself as well, but the core count argument I never saw before that link. To me it looks like a very small minority (one guy so far) arguing it would use too many transistors.

Best to take it up with him that indicate it is a widespread opinion.

 

[Thala]

Not sure if it's possible. To the best of my knowledge, anyone dealing with A75 or later ARM designs now uses DynamIQ instead of "big.LITTLE", but I'm not sure what are the technological implications of that shift.

DynamIQ is not a deviation from big.LITTLE but an optimzed variant of it. Where you had a big and a LITTLE CPU cluster with separate memory hierarchies before - with DynamIQ you have now a L1$/L2$ coherent cluster with shared L3$ where many combination of big an LITTLE cores are supported.

 

[Thala]

So shifting goal post to include clock speed now. This I have seen discussed, as it is actually a design limitation. You can't simply take a low clock mobile chip and run it at 4-5 GHz. Architecture for high clock speed is different than for low clock speed.

You cannot take the chip design itself, but the architecture should be very well scalable with respect to frequency - especially if you coming from mobile design.
Do not make the mistake assuming, that because the architecture targets mobile, it is not a high performance architecture allowing high clock frequency when paired with the proper high performance cell libraries and voltages.

 

[PeterScott]

You cannot take the chip design itself, but the architecture should be very well scalable with respect to frequency - especially if you coming from mobile design.
Do not make the mistake assuming, that because the architecture targets mobile, it is not a high performance architecture allowing high clock frequency when paired with the proper high performance cell libraries and voltages.

Architectures do not automatically scale to higher clock speeds. It's not just process.

High IPC architectures designed for lower clock speed typically don't run at higher clock speed. Typically you need to re-architect completely for higher clock speeds, with shorter, more pipelined stages.

 

[name99]

Yes, I really never saw that before. I don't read every post on every forum.

I have read these discussions on multiple forums and I have seen the clock speed argument come up a dozen times or more, and I have made that argument myself as well, but the core count argument I never saw before that link. To me it looks like a very small minority (one guy so far) arguing it would use too many transistors.

Best to take it up with him that indicate it is a widespread opinion.

Oh I’ve seen plenty of twits engage in the core count argument. (Along with the equally ignorant, equally lame “but Apple doesn’t have HT” argument)
Apparently Apple is JUST smart enough to build a NoC and LLC that can support 4+4 cores, 7 GPU cores, the NPU, and various other (Secure Enclave, ISP, media etc) but they can’t quite ramp up the intelligence to expand that to, say 8+8 cores...

 

[DrMrLordX]

Most PCs sold have integrated GPUs.

If you remove Con cores from the mix, I'm sure you'll find that most of AMD's CPU sales have been CPUs, not APUs.

DynamIQ is not a deviation from big.LITTLE but an optimzed variant of it. Where you had a big and a LITTLE CPU cluster with separate memory hierarchies before - with DynamIQ you have now a L1$/L2$ coherent cluster with shared L3$ where many combination of big an LITTLE cores are supported.

Hmm. Theoretically those clusters could be connected with shared L4 cache . . .

 

[dark zero]

Is it needed? Does Apple need it?

The dedicated gaming device market is a losing market. In Japan for example the mobile gaming market (where Apple devices dominate) is over five times the total revenue of the dedicated gaming device market, even though the latter includes hardware while the former is software only. Hence one could say Nintendo is already dead and Apple is already victorious.

Indeed, they have their Apple TV but their gaming environment is still poor.

Just improving it to have the same friendly environment like the consoles would make a total difference. And most games are not that proffitable compared to the dedicated one.

They have already a selling HW (Apple TV), they just add some better friendly UI and well worked SW and they would do literally everything.

 

[Thala]

Architectures do not automatically scale to higher clock speeds. It's not just process.

High IPC architectures designed for lower clock speed typically don't run at higher clock speed. Typically you need to re-architect completely for higher clock speeds, with shorter, more pipelined stages.

Any architecture does scale in a wide range, because thats just a property of the CMOS process. The only question is, at which design point the current architecture is implemented. And knowing that it is a low power mobile design - it is very safe to assume that if Apple achieves 2.4 GHz with low leakage high-Vth cell mix at low voltage the upper headroom is large.

Your fallacy is assuming, that if you are doing an architecture for mobile, you have long critical paths in the architecture. However given with the slow performing cells you have to cope with at very low voltage coupled with the fact you often have to do timing closure for worst conditions you would end up well below 2.4GHz if you don`t have very deeply pipelined architecture - e.g. an architecture which allows high frequencies - if you relax your power target.

 

[Thala]

Hmm. Theoretically those clusters could be connected with shared L4 cache . . .

Of course you can add additional memory hierarchies yourself. In the DynamIQ implementation as licensed from ARM (aka DSU) each clusters contains shared L3$ and allowing up to 8 cores with private L1$+L2$ (any big and LITTLE combination) with system ports being either ACE or CHI. Using these ACE and CHI interfaces you can coherently connect several DynamIQ clusters together and also add shared L4$. This is how a many-core server system would look like.

 

[DrMrLordX]

Of course you can add additional memory hierarchies yourself. In the DynamIQ implementation as licensed from ARM (aka DSU) each clusters contains shared L3$ and allowing up to 8 cores with private L1$+L2$ (any big and LITTLE combination) with system ports being either ACE or CHI. Using these ACE and CHI interfaces you can coherently connect several DynamIQ clusters together and also add shared L4$. This is how a many-core server system would look like.

Which means a chiplet design is certainly possible from Apple. Not sure if they want to go that route, though.

 

[dark zero]

Which means a chiplet design is certainly possible from Apple. Not sure if they want to go that route, though.

They have money, but they need SW to use it and they are making it.

 

[LightningZ71]

I personally believe that it's just a matter of time before we start to see Dev systems released with A12X processors for prototyping and developing MacOS apps on their A series systems. I don't think we'll see actually shipping systems for a couple years though.

 

[IntelUser2000]

In the DynamIQ implementation as licensed from ARM (aka DSU) each clusters contains shared L3$ and allowing up to 8 cores with private L1$+L2$ (any big and LITTLE combination) with system ports being either ACE or CHI.

I wouldn't be surprised if the A11 Fusion chip has a much superior implementation of the big.little architecture than the one ARM has. Actually, its quite likely as the controller is sophisticated enough to use little cores to significantly boost performance in multi-threaded code.

 

[PeterScott]

I personally believe that it's just a matter of time before we start to see Dev systems released with A12X processors for prototyping and developing MacOS apps on their A series systems. I don't think we'll see actually shipping systems for a couple years though.

IMO if/when ARM Mac transition does, happen it will proceed slowly. They are now past the point where it seems (don't want to debate the merits of Geekbench as solid cross platform benchmark) that they have an iPad SoC that could replace pretty much all of the Laptop lineup, and much of the desktop lineup.

Unlike previous Mac transitions where they were under the gun, there is no hurry this time. In both previous transition 68K and PPC were both falling behind Intel x86 that the competition was using. Now they use the same thing as the competition, so there are no worries about falling behind, so they can take their time transitioning the Mac Ecosystem to a place where the transition will be simplified (Like all MacOS only supporting 64 Bit, and only using Metal as the GPU layer, etc...).

I expected for a long time that Apple would eventually have ARM chips in most if not all of its mobile devices, but I thought that on the lower end Apple would first add more features to iPad eventually giving it Mouse/trackpad support to make it more capable of being a fully laptop replacement, and admittedly, this is the direction I hoped it would go. I'd like a convertible iPad with touch/KB/Mouse support.

But Apple will do what Apple does, with little concern for the member of the peanut gallery.

 

[name99]

I personally believe that it's just a matter of time before we start to see Dev systems released with A12X processors for prototyping and developing MacOS apps on their A series systems. I don't think we'll see actually shipping systems for a couple years though.

Why delay? The moment systems are ready, what’s the win in not shipping them?
Intel was announced at WWDC, with hardware available for developers. Six months later the first consumer hardware shipped. Eight months after that, the entire line had swapped over.
I see no reason the ARM transition should be any slower.

 

[name99]

I wouldn't be surprised if the A11 Fusion chip has a much superior implementation of the big.little architecture than the one ARM has. Actually, its quite likely as the controller is sophisticated enough to use little cores to significantly boost performance in multi-threaded code.

Why are you hypothesizing about this? We KNOW the results. Look at GB4 numbers. An A12 small core is worth about a quarter of an A12 large core, and certainly can operate in parallel with it. The A12 small core is powerful enough that I think Apple should add (perhaps invisible to 3rd party software) a truly small core for use when the phone screen is off, to save more power. The existing small core is overkill for many low-power situations.

(Truth is, if we look at Andrei’s power-performance numbers for the small cores, I don’t know where Apple is going with them. They’re not a great low-energy solution. Part of me wonders if there IS still a hidden super low power core attached to each large core, like the A10? Or if the small cores are not really for low power, but to provide a testbed for both Apple and developers to start thinking about how to write code that scales to “many” (ie more than just 2 or 3) cores for the inevitable future?)

 

[USER8000]

Another Apple launch and another lot of overexcitement,now saying:
1.)Microsoft consoles are doomed
2.)Nintendo is doomed
3.)Intel is doomed
4.)Nvidia is doomed
5.)AMD is doomed
6.)HI silicon is doomed
7.)Media tek is doomed

That was on the first page of this thread. I didn't think this was Macrumours,maybe I was wrong.

So I expect all of you will be ditching your Intel and AMD CPUs now,plus your dGPUs,consoles,etc and Anandtech will soon become an Apple only phone/tablet only forum??

If you listen to the usual cast of characters,as the price goes up they shout louder and louder on forums,and then close ranks as much as possible. Its literally a handfull of the same posters who keep saying every tech company outside Apple is doomed,and then present the most positive technical reasons why Apple will destroy everyone else,whereas every other company is full of idiots and won't be able to fight them in anyway.This is a classic reality distortion sphere.

Yet despite all the doomage,Apple sales are slowing down and flatlining,only compensated by them increasing prices at each generation to compensate. So as the price increases so does the justification for why it is worth it.

It was the same with the core count arguments - people were talking about Apple totally replacing Intel and AMD in all markets,ie,the high core count Xeons,etc and about the massive discrepancy in process nodes,ie,7NM against 12NM/14NM.

So apparently AMD and Intel should not bother,as the "experts" here have said Apple making a 4+4 core SOC(which other companies have done on a lower level),means Epyc,etc is instantly pointless and the Perlmutter supercomputer and the other projects,should switch to Apple.

The tech press is hyping up Apple products,basically as a form of advertising and you are just all reading into it,since Apple launches make more ad revenue.

 

[tempestglen]

What's the price of A12X chipset? 40 USD?

 

[DrMrLordX]

Another Apple launch and another lot of overexcitement,now saying:
1.)Microsoft consoles are doomed
2.)Nintendo is doomed
3.)Intel is doomed
4.)Nvidia is doomed
5.)AMD is doomed
6.)HI silicon is doomed
7.)Media tek is doomed

That was on the first page of this thread. I didn't think this was Macrumours,maybe I was wrong.

So I expect all of you will be ditching your Intel and AMD CPUs now,plus your dGPUs,consoles,etc and Anandtech will soon become an Apple only phone/tablet only forum??

If you listen to the usual cast of characters,as the price goes up they shout louder and louder on forums,and then close ranks as much as possible. Its literally a handfull of the same posters who keep saying every tech company outside Apple is doomed,and then present the most positive technical reasons why Apple will destroy everyone else,whereas every other company is full of idiots and won't be able to fight them in anyway.This is a classic reality distortion sphere.

Yet despite all the doomage,Apple sales are slowing down and flatlining,only compensated by them increasing prices at each generation to compensate. So as the price increases so does the justification for why it is worth it.

It was the same with the core count arguments - people were talking about Apple totally replacing Intel and AMD in all markets,ie,the high core count Xeons,etc and about the massive discrepancy in process nodes,ie,7NM against 12NM/14NM.

So apparently AMD and Intel should not bother,as the "experts" here have said Apple making a 4+4 core SOC(which other companies have done on a lower level),means Epyc,etc is instantly pointless and the Perlmutter supercomputer and the other projects,should switch to Apple.

The tech press is hyping up Apple products,basically as a form of advertising and you are just all reading into it,since Apple launches make more ad revenue.

Damn, someone's got some sour grapes.

Those GB4 numbers are turning heads. Personally I don't like GB4, especially since it supposedly allows for the CPU to "cool off" between bench elements to avoid throttling. Which is why I would love to see something like PovRay or Blender compiled for iOS and run on the chip instead, just to get an idea of what would happen when the A12x is forced to run a sustained load for more than a few seconds.

Taking a step back, though, we still have to realize that:

Apple won't (currently) sell this chip to anyone outside of tablet buyers
Apple won't ship this chip with anything more than 4 GB of RAM (bleh)
Apple apparently has no interest in the server market at all, and very little interest in the workstation market. Incidentally, Intel makes most of their money in the server market, and AMD is looking to chew up a lot of that market.

Apple is a $1 trillion+ company that could potentially stomp flat any CPU market they wanted with 2-3 years of dev effort built on expanding the scope of the A12x (or a successor). That is a ton of performance/watt - probably. The question is a matter of whether or not they will bother. Apple is having a lot of fun making a ton of money selling platforms and services, not moving in on other markets. If they wanted to, they could leverage their ability to produce entire platforms to roll out laptops; desktops; and even servers based on some modified A-series CPUs with expanded RAM and I/O. They have the developer community, they have the internal resources, they have the power. It's just a matter of having the will. If Apple keeps pushing forward with technological process, the software will follow. You can't deny it, and neither can I.

For the time being, I can't/won't toss my 1800x for an iPad Pro because it doesn't run the software that I want it to run. I also hate the closed nature of iOS. It can't work as my daily driver. It won't. And for that reason I'm going to continue to buy AMD products for the time being, and continue running Win10 so long as MS survives. It will take Apple killing the x86 development community for me to switch to anything else. Hopefully, by that point, there will be ways to circumvent the Apple bootloader locks and put something like Linux on whatever A-series chip finally puts the nail in the x86 coffin - or I'll have to look at some competing design from Huwaei or (preferably) Qualcomm.

As for your list of who is or isn't doomed, I have no idea how you came to those conclusions. Only Intel and AMD are really threatened by Apple - the rest, not so much. Apple may actually blaze a trail for Qualcomm, Mediatek, Huawei, and anyone else following in their wake. They're all a part of the ARMy. Nintendo is already technically using an ARM design (er, sort of), and MS can follow suit if they so choose. MS makes sense since they're now in bed with Qualcomm.