Apple A7 is now 64-bit

[Eug]

leaving that much clock speed on the table is a sign of a poorly balanced design.

Are you commenting in general, or specifically about the A7?

As a non-engineer and an outsider, I'm starting to wonder if Apple just feels it ships enough volume of their products that they can design to hit a certain feature set and performance target that it doesn't matter that much if every generation is quite a different design, as long as they achieve their targets.

Cuz in the last three iterations: A5/A5X, A6/A6X, and A7/A7X? we've had new designs. Personally I was predicting that the iPhone 5S would get an A6+, and then the iPad would get an A6X+ (meaning souped up versions of the A6 and A6X, on a 28 nm process). But instead, Apple gave us the A7 which is totally different.

 

[Khato]

Fair enough. You and Exophase have made it clear that it is not trivial. However, am I correct in assuming that it is still possible to design a core (a little different from the A7 obviously) which is nearly 2x the performance of the A7 cyclone. Maybe with 20nm (with performance gains of 20-30% over the current node), Apple (and/or the others) can theoretically reach core i7 levels of performance at a TDP of ~8W (or may be slightly higher power consumption).

Quite a bit is potentially possible. But far as I'm aware we don't even have hard confirmation of either A7 frequency or power consumption.

 

[TuxDave]

There's no guarantee you can double the peak clock speed of a uarch by throwing a lot of money at it. Normally you'd end up with a totally different uarch - leaving that much clock speed on the table is a sign of a poorly balanced design.

It's possible Apple can further optimize the timings by spending more time tuning things (including physical layout) without changing the behavioral specification of the uarch, but 2x seems like a lot given the experience of their design teams. Probably something is going to have to start giving in terms of cycle counts.

Agreed, a 2x frequency push with no change in voltage and process corner (aka 100% architectural) would be something that would even be a huge challenge for Intel. I would pretty much guarantee that pipelines will start growing left and right.

In my experience, if the project direction requires a 2x frequency push, you would not commit to doing it 100% architecturally. You would change your vmin and your process corner and get halfway there. Your power would scale super linearly. It won't be a simple 2x frequency = 2x power. That's why you see this huge gap between tablet TDP and server TDP.

 

[Ajay]

Yeah, it doesn't make sense to push that hard on the frequency as the cost of doing, especially in power consumption, don't make sense for the tablet or phone markets. If Apple wants to push ARM into it's Airbooks and Mac laptops, it'll be going quad core + allot more and will require a fairly large and separate design team to work on it for, say, the next four years (give or take). And I'm not sure if that would be enough, that may only get them to 14-16nm; Apple would need access to Intel's premium node to get enough performance (IMHO). ShintaiDK and other have made a good case for Apple probably not wanting to do that in the near future. ShintaiDK, thinks never, and may be correct - the design effort would be huge, just to come close to what Intel will be able to deliver in that time frame - then again, they do have the $$s, and are building the talent, to commit to such a project if the ROI was there). I'm beginning to wonder if the Airbook, over time, could be iOS powered, with some OSX features added as needed?? Then it could be crafted to work well on whatever Apple's top of the line ARM CPU is at the time. Maybe I'll do a little research and if there's anything there, start another thread.

From David Kanter over at RWT, it appears that Apple has at least two design teams and maybe a third (why would they need three - hmm....).

Also from RWT (the forum) it's rumored that Apple's design team may be around the size of AMD's CPU design team at present (apparently based off something Anand said).

Oh, FWIW, those who haven't taken a look at the Chip Works article, the A7 is made on Samsung's 28nm HKMG. Oh, and last little tidbit from MacRumors is that there is an A7X following hot on the heels of the A7 for a Q4 iPad/Mini update (there's more info if you look around the site).

 

[ancientarcher]

Yeah, it doesn't make sense to push that hard on the frequency as the cost of doing, especially in power consumption, don't make sense for the tablet or phone markets. If Apple wants to push ARM into it's Airbooks and Mac laptops, it'll be going quad core + allot more and will require a fairly large and separate design team to work on it for, say, the next four years (give or take).

I agree. The incentive for Apple to design a processor with a bigger power envelope will be for it to use it for macbooks/airs. The rationale will be money, IMHO, as long as they can the performance to nearly the same level as Intel. And that is a big incentive. The best intel low-power processors for the air and macbooks cost $550. Apple will be able to manufacture its PC processors at ~$30 (assuming a 50% price increase from the 28nm node). It already has design teams who have been working for sometime on the ARMv8 architecture. And a lot of those CPU design people have a long history in ARM design (see RWT). If it introduces its own processors, suddenly you will see Apple laptops in the mid-priced segment. Think what that will do to the Wintel ecosystem..

I don't think it will take 3-4 years for them to do CPU cores for laptops. If the 3-4 teams that David Kanter talks about started working at the same time, starting 2 years ago, its high power processor(s) might be coming to the market on the 20nm node mid next year.

Just because Intel charges $500 for something and says it takes a lot of time and expertise to do it doesn't mean it does. Maybe they are just hamstrung with carrying the x86 legacy/cruft/overhead. They are monopolist in that segment and charge what they can get away with.

Apple doesn't need its laptop processors to work with 32bit ARM. So, it can just shove that piece of overhead (backward compatibility) out of the way when designing processors for PCs, and thereby get some added performance.

For the ipad, yes, I agree with you. They will probably just double the GPU area and call it the A7X. And btw, on graphics performance, it will be nearly at par with ivy-bridge level processors. It will still be ~120mm^2 - around the area of snapdragon 800 (though that has the modem integrated in the same die).

I am also surprised that Apple put so much effort in designing the Swift cores on 32bit last year, given that they were moving to 64bit in 2013. It seems they have money/talent to burn!

 

[unon]

Wonder how it will do on fd-soi given a9 can clock to 3ghz...

 

[Exophase]

I am also surprised that Apple put so much effort in designing the Swift cores on 32bit last year, given that they were moving to 64bit in 2013. It seems they have money/talent to burn!

What of that effort towards Swift do you think didn't directly contribute towards Cyclone? Even if there were parts that were specific to ARM32 and not useful in ARM64 - parts which which would be very small - Cyclone is still very good at executing 32-bit code which is what it will spend most of its time doing for the foreseeable future.

 

[jdubs03]

I agree. The incentive for Apple to design a processor with a bigger power envelope will be for it to use it for macbooks/airs. The rationale will be money, IMHO, as long as they can the performance to nearly the same level as Intel. And that is a big incentive. The best intel low-power processors for the air and macbooks cost $550. Apple will be able to manufacture its PC processors at ~$30 (assuming a 50% price increase from the 28nm node). It already has design teams who have been working for sometime on the ARMv8 architecture. And a lot of those CPU design people have a long history in ARM design (see RWT). If it introduces its own processors, suddenly you will see Apple laptops in the mid-priced segment. Think what that will do to the Wintel ecosystem..

I don't think it will take 3-4 years for them to do CPU cores for laptops. If the 3-4 teams that David Kanter talks about started working at the same time, starting 2 years ago, its high power processor(s) might be coming to the market on the 20nm node mid next year.

Just because Intel charges $500 for something and says it takes a lot of time and expertise to do it doesn't mean it does. Maybe they are just hamstrung with carrying the x86 legacy/cruft/overhead. They are monopolist in that segment and charge what they can get away with.

Apple doesn't need its laptop processors to work with 32bit ARM. So, it can just shove that piece of overhead (backward compatibility) out of the way when designing processors for PCs, and thereby get some added performance.

For the ipad, yes, I agree with you. They will probably just double the GPU area and call it the A7X. And btw, on graphics performance, it will be nearly at par with ivy-bridge level processors. It will still be ~120mm^2 - around the area of snapdragon 800 (though that has the modem integrated in the same die).

I am also surprised that Apple put so much effort in designing the Swift cores on 32bit last year, given that they were moving to 64bit in 2013. It seems they have money/talent to burn!

Your assumption that Apple can design and have TSMC/Samsung manufacture a $30 chip that is competitive with core i3-i7 performance levels is off base. There is a reason why Intel charges the prices it does, yes some of that is to maintain its margins and doesn't have much competition in the x86 space (AMD has not innovated enough to drive Intel's chip prices lower. But most of that is to cover the vast expenses that are realized in the operation of their firm, that is R&D. It should be noted that Intel's ULV processors are in the $200 range.

I believe that Apple does have the potential to be performance/power competitive with Intel (core) based on the amount of resources (cash) that they have, but on the same node (which they won't have access to). However, potential doesn't equal results, and implementing that is much harder than you're assuming.

Also keep in mind that Intel is also innovating, in order for Apple to design a competitive solution when compared to Intel's core line they have to catch up, and they'd be catching up to a moving target. They will have ups and downs, but that is not to say they can't do it, its just unlikely that they would. Its all about cost/benefit, if they want to wean themselves off Intel for their Macs, they will and it will take time, but they'd be giving up top of the line performance, process node, and a valued supplier.

 

[ancientarcher]

Your assumption that Apple can design and have TSMC/Samsung manufacture a $30 chip that is competitive with core i3-i7 performance levels is off base. There is a reason why Intel charges the prices it does, yes some of that is to maintain its margins and doesn't have much competition in the x86 space (AMD has not innovated enough to drive Intel's chip prices lower. But most of that is to cover the vast expenses that are realized in the operation of their firm, that is R&D. It should be noted that Intel's ULV processors are in the $200 range.

I believe that Apple does have the potential to be performance/power competitive with Intel (core) based on the amount of resources (cash) that they have, but on the same node (which they won't have access to). However, potential doesn't equal results, and implementing that is much harder than you're assuming.

Also keep in mind that Intel is also innovating, in order for Apple to design a competitive solution when compared to Intel's core line they have to catch up, and they'd be catching up to a moving target. They will have ups and downs, but that is not to say they can't do it, its just unlikely that they would. Its all about cost/benefit, if they want to wean themselves off Intel for their Macs, they will and it will take time, but they'd be giving up top of the line performance, process node, and a valued supplier.

The quad-core Intel core i7 haswell processors are around 180mm^2 in area and have 1.5bn transistors. The A7 in iphone 5s is 100mm^2 and has 1bn transistors. The A7 already has higher transistor density but I guess that is because of the larger area given over to the GPU. The Intel core i7 transistor costs $450-500. The A7 costs $19 (as reported by IHS). These are on different process nodes - the Intel processor is on 22nm Finfet and the A7 is on 28nm planar.

So, on a 20nm scalar or 14/16nm finfet (which will have nearly the same transistor density) an Apple chip will have 2bn transistors in 140mm^2.

You think Samsung/TSMC will charge $200 for it instead of the $20 they were charging for 100 mm^2 in 28nm. I think Apple paying $30 for a chip of die size of 140-150 mm^2 is reasonable. The difference between $30 and $500 that Intel charges (or even $200 core i5 processors) are the monopoly profits accruing to Intel.

On the process node difference - both Intel and TSMC/Samsung will be on 14/16nm finfet in 2015. Yes, Intel will still be a year ahead in introducing 10nm (but then who knows, it might depend on whether/when EUV becomes available).

On moving targets - The ARM camp has been doubling its processing power every year in the last 5 years (that includes both architectural and process node improvements). In the last few years, Intel has done around 10-15% CAGR. So yea, both are moving targets but one is moving significantly faster than the other.

On R&D etc, the point is that it takes much less resources in terms of manpower and time to design on the new ARM architecture as AMD mentioned the other day. One of the reasons for that might be the burden of legacy that x86 code is carrying.

So, Apple hires a ton of talented chip designers (which it started doing a few years ago, and also bought PA Semi and Intrinsity) and gets them to work on the new ARMv8 64bit. Given the same number of transistors, they will take some time to get to where Intel is, but they are also starting from a cleaner slate and have the learnings of x86 to fall back on.

So, if you are standing in mid-2015, with both Intel and TSMC/Samsung/GloFo on 14/16nm Finfet, it should be theoritically, and I am betting practically, possible for chip designers from Apple to design a chip equal to or better than the best Intel processors with a lower power consumption and 1/5 to 1/10th the cost of the core i7s. But we will see what happens...

 

[jdubs03]

So, if you are standing in mid-2015, with both Intel and TSMC/Samsung/GloFo on 14/16nm Finfet, it should be theoritically, and I am betting practically, possible for chip designers from Apple to design a chip equal to or better than the best Intel processors with a lower power consumption and 1/5 to 1/10th the cost of the core i7s. But we will see what happens...

This part just doesn't make sense economically. Intel has a low-power and high-power architecture for a reason. And keep in mind Bay Trail and A7 are on par in terms of CPU performance. Also, TSMC as a BEST case scenario will be on 16nm FINFET in 2015. Most likely they'll be there in 2016, they will only be ramping on 20nm planar in the 2nd half of next year.

If Apple were to introduce a higher performing chip at a higher power envelope it would probably be cheaper than a comparable Intel offer (they're producing it for themselves), but not by 1/5. There is a cost to produce a higher-performing chip. Just like any product, compare two different Ferrari's (maybe not the best analogy but sufficient for my point). A F458 is around $270,000-$300,000. The new LaFerrari is likely to be over $1 million.

Each is a different beast targeted toward a certain clientele, the same is true of Bay Trail and Core, and would be of A8/9/10, and a high-power version. Like some others said x86 reached a point where scaling and other features no longer allowed the architecture to boost performance by 100%, ARM will follow the same path, its physics.

Intel won't stay still in either the low or high-performance space, and I expect Airmont to be a significant overhaul, new architecture (or at least a big refresh) and new process node will likely entail at least a 60% increase in performance at the same power envelope. Eventually I would think that Intel will merge both Core and Atom, but that is at 7nm or beyond.

But you're right we will see.

 

[superrockie]

How do you think the apple a7 compares to x86 chips. Is it comparable to a pentium 4? Or does it already perform better then that.

 

[Arachnotronic]

How do you think the apple a7 compares to x86 chips. Is it comparable to a pentium 4? Or does it already perform better then that.

Try Core 2 Duo.

 

[jhu]

How do you think the apple a7 compares to x86 chips. Is it comparable to a pentium 4? Or does it already perform better then that.

Cortex A9 would be Pentium 4 level performance.
Cortex A15 is about 2/3 performance of Core 2/Phenom II

 

[RadiclDreamer]

You talk as if no one uses RAM quantities that aren't powers of 2. 6GB laptops are common. I'm sure 3GB modules cost less to make than 4GB ones, and could still be really pushing the limits for what they can stack in a PoP.

Those 6GB laptops are using one 4gb stick and one 2gb. I've still yet to see a 3GB stick of ram anywhere

 

[mrmt]

The quad-core Intel core i7 haswell processors are around 180mm^2 in area and have 1.5bn transistors. The A7 in iphone 5s is 100mm^2 and has 1bn transistors. The A7 already has higher transistor density but I guess that is because of the larger area given over to the GPU. The Intel core i7 transistor costs $450-500. The A7 costs $19 (as reported by IHS). These are on different process nodes - the Intel processor is on 22nm Finfet and the A7 is on 28nm planar.

So, on a 20nm scalar or 14/16nm finfet (which will have nearly the same transistor density) an Apple chip will have 2bn transistors in 140mm^2.

You think Samsung/TSMC will charge $200 for it instead of the $20 they were charging for 100 mm^2 in 28nm. I think Apple paying $30 for a chip of die size of 140-150 mm^2 is reasonable. The difference between $30 and $500 that Intel charges (or even $200 core i5 processors) are the monopoly profits accruing to Intel.

You are comparing two different things. $19 is just manufacturing costs for the MPU, which includes basically TSMC costs and margins. All the design costs will be included in the final ipad/iphone price.

Intel chips OTOH have included manufacturing costs plus manufacturing margins on top of the costs, so all the money they spend must be recouped on the processor price.

In the end Apple may still get a price advantage, but they still get an inferior processor. And the difference is far smaller than the order of magnitude you are trying to imply here.

 

[ancientarcher]

You are comparing two different things. $19 is just manufacturing costs for the MPU, which includes basically TSMC costs and margins. All the design costs will be included in the final ipad/iphone price.

Intel chips OTOH have included manufacturing costs plus manufacturing margins on top of the costs, so all the money they spend must be recouped on the processor price.

In the end Apple may still get a price advantage, but they still get an inferior processor. And the difference is far smaller than the order of magnitude you are trying to imply here.

The fixed costs overheads are already there. What do you think, Apple doesn't spend anything on designing the A7 chip? But Apple won't spend something like Intel ($10bn) on R&D, though you can make the point that the R&D spend of Intel includes a lot of other stuff...

As a comparison, the Qualcomm snapdragon s600/800 processors are $20-25.

So, the only way Intel can charge something like $500 or higher on ultra lower power x86 chips is because it is a monopoly. There cannot be any other reason.

On Apple getting an inferior processor, we will find out soon enough. Apple will have more transistors to play with than Intel, a cleaner architecture and no legacy/overhead/burden of carrying stuff from 20 years ago. They also have arguably the best ARM design team and get learnings from the ARM designed cores as well. On the GPU front, they dont have to reinvent the wheel (remember the very successful Intel project Larrabee) and get cores from IMG. GPGPU will help improve the CPU power. Given all this, I am betting they will be able to design a medium- and high-powered chip (10W - 30W TDP) by mid-2014 for use in their notebooks and possibly desktops.

And yes, it will cost apple 1/5th to 1/10th that of the best of breed low power Intel chips.

 

[mrmt]

The fixed costs overheads are already there. What do you think, Apple doesn't spend anything on designing the A7 chip? But Apple won't spend something like Intel ($10bn) on R&D, though you can make the point that the R&D spend of Intel includes a lot of other stuff....

Fixed costs are *not* inside the $19 number. Manufacturing costs do not include IC design costs.

 

[Arachnotronic]

The fixed costs overheads are already there. What do you think, Apple doesn't spend anything on designing the A7 chip? But Apple won't spend something like Intel ($10bn) on R&D, though you can make the point that the R&D spend of Intel includes a lot of other stuff...

As a comparison, the Qualcomm snapdragon s600/800 processors are $20-25.

So, the only way Intel can charge something like $500 or higher on ultra lower power x86 chips is because it is a monopoly. There cannot be any other reason.

On Apple getting an inferior processor, we will find out soon enough. Apple will have more transistors to play with than Intel, a cleaner architecture and no legacy/overhead/burden of carrying stuff from 20 years ago. They also have arguably the best ARM design team and get learnings from the ARM designed cores as well. On the GPU front, they dont have to reinvent the wheel (remember the very successful Intel project Larrabee) and get cores from IMG. GPGPU will help improve the CPU power. Given all this, I am betting they will be able to design a medium- and high-powered chip (10W - 30W TDP) by mid-2014 for use in their notebooks and possibly desktops.

And yes, it will cost apple 1/5th to 1/10th that of the best of breed low power Intel chips.

ASP of PC client chips (that's CPU + PCH) is ~$120.

While manufacturing costs for Apple may be cheap, designing such a sophisticated IC is not so.

 

[krumme]

We will see. I think Intel dreams of mobile just went dead with apple a7 using armv8. And with a bang. Bt Dead. Period.

You can have other interpretation - fine. But pretending this kind of performance was anything like general anticipated looks like denial to me. For me it was a huge surprice.

 

[ancientarcher]

Fixed costs are *not* inside the $19 number. Manufacturing costs do not include IC design costs.

What about the Qualcomm snapdragon pricing then? that is still ~$20
My point is that, no matter how you parse the costs, the best of breed ARM chip costs $20-25. Mediatek, Rockchip sell much cheaper chips for single digit $s.

And yes, Intel PC client group ASP includes desktop chips which are priced lower, because AMD is still relevant there, at least in the mid-tier segment.

The majority of core series sales is still i5. the share of i3 is larger than i7. If you take these 2 factors into account, that explains the ASP of only $120.

And yes, Apple doesn't do core i3. They get the best only (as with everything else) from Intel. If Apple moves its desktops/notebooks away to its own chips, the decline of ASP's for Intel PC client group will be a little more!

 

[Exophase]

Those 6GB laptops are using one 4gb stick and one 2gb. I've still yet to see a 3GB stick of ram anywhere

A PoP module isn't analagous to a stick of RAM. You don't see individual DIMMs push unusual capacities on RAM because there's little consumer demand for maxing out RAM capacity by loading every DIMM slot. But it's different on phones where you have to shove all the RAM in a little chip that goes on top of the SoC.

We will see. I think Intel dreams of mobile just went dead with apple a7 using armv8. And with a bang. Bt Dead. Period.

You can have other interpretation - fine. But pretending this kind of performance was anything like general anticipated looks like denial to me. For me it was a huge surprice.

For this to even get off the ground as a theory A7 would have to be available for Android or somehow directly reflect the quality of other SoCs deployed on Android (both using ARM doesn't matter). That, or you'd have to expect that Apple would take back most of the market share Android won.

 

[Arachnotronic]

We will see. I think Intel dreams of mobile just went dead with apple a7 using armv8. And with a bang. Bt Dead. Period.

You can have other interpretation - fine. But pretending this kind of performance was anything like general anticipated looks like denial to me. For me it was a huge surprice.

On a per-core basis, excluding the sketchy crypto benchmarks, Silvermont holds its own against Cyclone in integer performance (perf/watt, not IPC) is all that matters). Also, compile GB3 with a proper compiler instead of GCC for the Intel chip (if Intel spends all this $$ developing a compiler to give its chips an edge, it should be used) and you'll see FPU and INT scores skyrocket.

But let's just say that A7 is better than Bay Trail and that Geekbench is the be-all, end-all of chip benchmarking tools. Who gives a hoot? Is Apple selling chips to Intel's customers now? When I buy a Windows 8.1 tablet or Android tablet, will I have an Apple A7 option? No.

 

[Nothingness]

Also, compile GB3 with a proper compiler instead of GCC for the Intel chip (if Intel spends all this $$ developing a compiler to give its chips an edge, it should be used) and you'll see FPU and INT scores skyrocket.

BT scores we have were obtained with Visual Studio compiler since all we have are Windows scores. This is the compiler that most of the software industry uses, because it works (I don't want to remember the number of times icc bailed out with some internal error message) and because it targets equally well (or bad) AMD and Intel.

 

[Arachnotronic]

BT scores we have were obtained with Visual Studio compiler since all we have are Windows scores. This is the compiler that most of the software industry uses, because it works (I don't want to remember the number of times icc bailed out with some internal error message) and because it targets equally well (or bad) AMD and Intel.

We also have Android scores. Or, at least, I do

 

[Nothingness]

We also have Android scores. Or, at least, I do

Rooooh want to look at them :biggrin:

BTW did you notice this: http://browser.primatelabs.com/geekbench3/compare/93819?baseline=93813

(we're getting off topic, probably should move...)