Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Confirmed: Apple's A7 is 1.3 GHz dual core CPU, quad cluster GPU. Samsung 28nm HKMG.
- Thread starter chrmjenkins
- Start date
- Sort by reaction score
These threads are simply amazing. Thank you for all the knowledge you share through these posts!
Think so.
Here's the silkscreen code on the A6: B8164B3PM
This corresponds to an internal DRAM die marking of B440ABA - you can read this using this PDF from Elpida
The new A7 has a silkscreen code of F8164A1PD
Assuming the silkscreen codes follow the same generic order (there's no PDF I can find), the "F" likely corresponds to RAM type. B was LPDDR2. F is probably LPDDR3. They both share "8164", which is likely the size of 1GB. B3PM vs A1PD and the parts after the dashes likely contain package type, speed and environment code. I can't cross reference them to anything easily in the PDF or existing products, so I'm stuck there. I think the speed is the first dash, "1D" on A6 versus "GD" on A7, but I'm grasping at straws there. The "F" at the end says they're lead free (RoHS compliant)
Last edited:
I know that I'm very excited about what the A7 will bring to the table. My next iPad should be insane. Made my day to be honest since we had no expectation of such a massive improvement. On the other hand, The 5c was a disappointment of monstrous proportions given it's price. Will do nothing for market share.
I do not think Apple's goal was market share with the 5C. I think it makes more sense that they are doing this to be able to keep their precious margins. They simply could not continue to offer the iPhone 5 at $100 less and keep their loved margins. It makes more sense to them to use a cheaper enclosure (steel and plastic) and refine it to use most of the existing iPhone 5 components and share as much as possible with the 5S. Economies of scale! Apple is the best company out there at this. Also, they are diversifying their lineup, much like they did with their iPod line back in the day.
I agree that market share was not Apple's goal as it turns out, but it appear now that they have no strategy to increase marketshare.
I just do not think Apple cares all that much about market share, they want the profit$. Also, they essentially have one model (now 2 with the 5c) and they still sell those phones by the millions. VERY impressive if you ask me. They will continue to make the best phone with the best software and the money will keep rolling in.
I've spent the last few days trying to poke holes in what I've written thus far.
First issue is transistor density. Disclosures on gate density between TSMC and Samsung are actually almost identical for 28nm. It's about 33% better compared to Samsung's 32nm. This fact surprised me quite a bit, as I originally claimed 20% at best. This stops being a point for TSMC being the manufacturer. That doesn't change the fact that the leaked A7 clearly had a marking that suggested it was no longer a Samsung part.
Second, there are a lot of potential new blocks that could increase the transistor count, taking away from either it being a quad core CPU or quad cluster GPU. The CPU cache likely doubled from 1MB to 2MB, which is a 50 million transistor increase in itself. The GPU caches also likely grew. There's also likely some sort of chip memory and circuitry to process the photos for all of the advanced camera features Apple announced. The circuit to process fingerprints also likely has significant memory and transistors for processing. Memory is inherently very transistor dense, which gets us closer to that 2x claim. For a counterpoint, it's possible that the M7 moved some parts off of the A6 that would have gone into the A7.
Finally, I've seen people suggest the 1 billion transistor claim is spread across the A7 and the M7. I don't think that's true, since they announced that number before they even talked about the M7 at all.
Interestingly, the early GPU benchmarks have the GPU reporting itself as an apple GPU instead of Imagination Technologies. We know for sure it's a rogue ImgTec part, but it sounds like it may have a lot of Apple customizations.
First issue is transistor density. Disclosures on gate density between TSMC and Samsung are actually almost identical for 28nm. It's about 33% better compared to Samsung's 32nm. This fact surprised me quite a bit, as I originally claimed 20% at best. This stops being a point for TSMC being the manufacturer. That doesn't change the fact that the leaked A7 clearly had a marking that suggested it was no longer a Samsung part.
Second, there are a lot of potential new blocks that could increase the transistor count, taking away from either it being a quad core CPU or quad cluster GPU. The CPU cache likely doubled from 1MB to 2MB, which is a 50 million transistor increase in itself. The GPU caches also likely grew. There's also likely some sort of chip memory and circuitry to process the photos for all of the advanced camera features Apple announced. The circuit to process fingerprints also likely has significant memory and transistors for processing. Memory is inherently very transistor dense, which gets us closer to that 2x claim. For a counterpoint, it's possible that the M7 moved some parts off of the A6 that would have gone into the A7.
Finally, I've seen people suggest the 1 billion transistor claim is spread across the A7 and the M7. I don't think that's true, since they announced that number before they even talked about the M7 at all.
Interestingly, the early GPU benchmarks have the GPU reporting itself as an apple GPU instead of Imagination Technologies. We know for sure it's a rogue ImgTec part, but it sounds like it may have a lot of Apple customizations.
Last edited:
This is a massively huge and fallacious assumption.
The strongest argument that it is NOT quad core. Is Apple didn't advertise it as quad core. They have always been proud of their dual core A5. They would have said quad core if it was.
Fallacious assumption number 2. 3 core CPU's have been done, no technical barriers.
http://apcmag.com/amd_announces_threecore_desktop_cpu.htm
As I understand it those AMD parts weren't custom designed 3 core CPUs though, they were the quad-core parts with one of the cores disabled. I think there was even information circulating on the Internet about how to re-enable the disabled core which sometimes gave a perfectly workable quad core.
Just because a specific 3 core design hasn't been done before doesn't mean Apple couldn't be the first to do it of course. We're all guessing until we get the X-rays. There could even be some sort of asymmetric big.LITTLE type of architecture in there. I'm really looking forward to getting the results from detailed analysis of the A7 once the appropriately equipped experts get their hands on it.
The A6 was not advertised as dual core when it was. Apple doesn't always boast about GPU core count either (didn't for A6, A6X). It's not as simple as you think. Quad core designs are commonplace now. It's much easier to hang your hat on 64 bit as a marketing win.
Not for ARM cores - I should have clarified. IBM has done plenty of tri-core solutions too. But it's unheard of for ARM cores in mobile phones or tablets. Marvell has done it but they have no design wins.
Last edited:
did you see the load times of asphalt 7 in a hands on? oh....i think we WILL see a daylie basis difference, just like 4s to 5.
chrmjenkins,
Thank you so much for such an excellent thread, and for your continued responses and analysis. I for one very much appreciated it, and did NOT think it was TL. I DR. I also DR your previous roadmap thread. Please know that there are probably many more people who appreciated these threads but just didn't post and let you know.
Thank you again.
Curun,
As for any of the assumptions being "massively huge and fallacious", that's just not correct. They might be massively huge. And they might be wrong. But none of them are fallacious. None of them is a fallacy. An assumption like "a 64 bit CPU will be twice as fast at everything as a 32 bit CPU" is a fallacious assumption. But assumptions like "Apple won't do a 3 core CPU" or that they will keep their clock rates lower are not fallacious.
And, they have the great benefit of being explicitly stated assumptions in an excellent back-of-the-envelope-calculation done at a time when we have very little to go on. When you have very little to go on, you MUST make assumptions. As long as you state them clearly, and none of them is fallacious, that's not only fine, but essential.
Thank you so much for such an excellent thread, and for your continued responses and analysis. I for one very much appreciated it, and did NOT think it was TL. I DR. I also DR your previous roadmap thread. Please know that there are probably many more people who appreciated these threads but just didn't post and let you know.
Thank you again.
Curun,
As for any of the assumptions being "massively huge and fallacious", that's just not correct. They might be massively huge. And they might be wrong. But none of them are fallacious. None of them is a fallacy. An assumption like "a 64 bit CPU will be twice as fast at everything as a 32 bit CPU" is a fallacious assumption. But assumptions like "Apple won't do a 3 core CPU" or that they will keep their clock rates lower are not fallacious.
And, they have the great benefit of being explicitly stated assumptions in an excellent back-of-the-envelope-calculation done at a time when we have very little to go on. When you have very little to go on, you MUST make assumptions. As long as you state them clearly, and none of them is fallacious, that's not only fine, but essential.
A6 had load time improvements too. I think they're substantially improving their memory subsystems again, amongst other things.
edit: I should say at this point that I'm having serious doubts about it being TSMC after reading that Samsung and TSMC transistor densities are theoretically identical at 28nm, but the A7 leak with chip identifier that didn't match Samsung prevents me from discarding that all together. Quad core CPU seems less likely as we come up with explanations for increased transistor budget, but I think it has a fair chance. Quad cluster vs dual cluster GPU is simply a matter of how Apple wants to clock it.
edit 2: heh, iFixIt fell for the photoshopped package pictures of the A7 in the keynote. All the date codes are the same from their A6 picture. More here.
Last edited:
so refreshing to see a well reasoned intelligent post on MR, a rare occurrence these days..
64 bit surprised me as well, but all things considered it turns out to be quite a logical step for Apple to take. They have obviously spent considerable design effort on ARMv7 Swift and no doubt they will capitalize on it across their lineup into the future. So for the next ground up design, it only makes sense to start with ARMv8 which happens to be 64 bit, but that's mostly relevant to the marketing department at this point. ARMv8 is a better, cleaner architecture regardless of 32/64 bit. Still, after Swift's IPC improvements, increased clock speed and 32nm process, it was difficult to imagine what can be done to significantly outperform it outside of going quad core. 28nm vs. 32nm is not major enough of a step and 20nm is still out. Clock speed may go up slightly, but Apple will always undershot to keep reigns on the consumption. I thought perhaps stretching everything a little bit, Apple will squeeze at most 50% gain over A6 while keeping it dual core. Then came the keynote claim of doubling CPU performance. I can't imagine we're still talking about dual core chip, but if it turns out I'm wrong - hats off to Apple. Quad core at similar if not slightly lower frequency with slightly improved IPC seems more plausible from where I stand.
The writing is on the wall, one day we will see Apple designed ARM processor in a laptop form factor. Performance won't be an issue, power will be saved, battery size, dimensions and weight will shrink as well as the pricetag. MBP will keep using x86, but MBA could easily switch to ARM before we know it.
64 bit surprised me as well, but all things considered it turns out to be quite a logical step for Apple to take. They have obviously spent considerable design effort on ARMv7 Swift and no doubt they will capitalize on it across their lineup into the future. So for the next ground up design, it only makes sense to start with ARMv8 which happens to be 64 bit, but that's mostly relevant to the marketing department at this point. ARMv8 is a better, cleaner architecture regardless of 32/64 bit. Still, after Swift's IPC improvements, increased clock speed and 32nm process, it was difficult to imagine what can be done to significantly outperform it outside of going quad core. 28nm vs. 32nm is not major enough of a step and 20nm is still out. Clock speed may go up slightly, but Apple will always undershot to keep reigns on the consumption. I thought perhaps stretching everything a little bit, Apple will squeeze at most 50% gain over A6 while keeping it dual core. Then came the keynote claim of doubling CPU performance. I can't imagine we're still talking about dual core chip, but if it turns out I'm wrong - hats off to Apple. Quad core at similar if not slightly lower frequency with slightly improved IPC seems more plausible from where I stand.
The writing is on the wall, one day we will see Apple designed ARM processor in a laptop form factor. Performance won't be an issue, power will be saved, battery size, dimensions and weight will shrink as well as the pricetag. MBP will keep using x86, but MBA could easily switch to ARM before we know it.
You might be surprised about 28nm
Samsung claims a 33% density increase, which is pretty impressive for a half-node transition. As for the 2X claim, it's really hard to get there with IPC improvements alone unless they have another significant clock bump like A5 to A6 did (800 MHz to 1.3 GHz). A similar percentage bump would put you in the 1.9 GHz range. Apple has tended to be on the low side of what other vendors are averaging, so I have a little difficulty believing that number.
However, ARM claims A57 implementations can vary from 4.1 to 4.8 DMIPs/MHz. If Apple went for a really beefy custom implementation, they may be able to get away with dual core still and a modest clock bump. However, I'm slightly averse to that scenario given how similar Swift was to Krait. I'm guessing it's a situation where similar design goals have created similar solutions to the problem. Given Krait cores have lagged behind vanilla A15 IPC claims in even their most aggressive implementations, I'm guessing that's a continued philosophy on the part of Apple too, rather than a fluke.
But this is why this is all exciting, because none of that is for sure and we get to be surprised by how Apple decided to tackle the problem this time
Swift being hand-placed and all I did not expect them to come up with a whole new core in just one year's time. Would be easier to double up cores, perhaps die shrink them, but no, the guys are out with a whole new design!
Its hard enough for us armchair chip designers to wrap our heads up around what they did to double the performance, even after the keynote. The sick thing is those guys had this chip on a drawing board for the past 3 years or so! Talk about foresight and planning abilities!
I also don't like the quad core scenario, although perhaps the easiest to explain away. Single threaded performance still beats multithreading for responsiveness IMO and Apple generally shows sound judgement. The Swift cores did not look that big on the die shot, with 33% density increase at 28nm, I can see them fitting double serving on the same plate. Not sure about Rogue performance gains vs. die area, but perhaps it's feasible to achieve doubling of performance thanks to a better architecture alone?
A beefier-than-A57 core does not sound likely. It's a matter of trade-offs and ARM themselves gunning for server sockets with A57 no doubt did not hold themselves back with chip complexity. I don't think Apple can outperform them with a mobile targeted architecture.
But perhaps we are really looking at triple-core? Easier to fit, wouldn't require dramatic clock increases and could easily hit 2x overall performance? There's nothing fundamentally odd about odd-core configurations, just some lack of symmetry, but that's the least of anyone's worries on SoC where the CPU is just a small part of the die.
Can't wait for the Chipworks analysis...
Its hard enough for us armchair chip designers to wrap our heads up around what they did to double the performance, even after the keynote. The sick thing is those guys had this chip on a drawing board for the past 3 years or so! Talk about foresight and planning abilities!
I also don't like the quad core scenario, although perhaps the easiest to explain away. Single threaded performance still beats multithreading for responsiveness IMO and Apple generally shows sound judgement. The Swift cores did not look that big on the die shot, with 33% density increase at 28nm, I can see them fitting double serving on the same plate. Not sure about Rogue performance gains vs. die area, but perhaps it's feasible to achieve doubling of performance thanks to a better architecture alone?
A beefier-than-A57 core does not sound likely. It's a matter of trade-offs and ARM themselves gunning for server sockets with A57 no doubt did not hold themselves back with chip complexity. I don't think Apple can outperform them with a mobile targeted architecture.
But perhaps we are really looking at triple-core? Easier to fit, wouldn't require dramatic clock increases and could easily hit 2x overall performance? There's nothing fundamentally odd about odd-core configurations, just some lack of symmetry, but that's the least of anyone's worries on SoC where the CPU is just a small part of the die.
Can't wait for the Chipworks analysis...
Even if they get that full 33% density improvement, they still have to get 50% more density to hit twice the transistors in the same area. That's why we're probably talking about significant amounts of SRAM.
To be fair, if you're doing full custom layout, going triple core doesn't seem that much a challenge anymore to me
edit: Pretty significant update. Ben Bajarin (analyst) claims to have a source that backs up A7 being dual-core, not quad core. https://twitter.com/BenBajarin/status/378567815969525761
If his source is accurate, that lends further support to a lower clocked quad cluster GPU as opposed to a high clocked dual cluster GPU to spend some of those 500M extra transistors from A6 to A7.
edit 2: I'm hearing 1.7 GHz clock from multiple sources. This meshes with the 31% faster rumor that came along with the 64 bit (which turned out to be right). What this means is that to get to the 2X claim, the processor has to have 50% better IPC (unlikely) or it's merely better optimized for some scenarios because of new instructions/registers for specific tasks.
edit 3: Ben Bajarin is claiming Apple may be doing custom additions to the ARMv8 (ISA) to get a performance boost. That would be a very big deal because it would make them almost equivalent to AMD in that AMD and Intel both contribute to the x86 ISA. Apple wouldn't have to disclose these to ARM though because they don't sell chips themselves.
https://twitter.com/BenBajarin/status/378585813887356928
edit 4: I've updated the thread title to reflect the most recent news. CPU is likely dual core with custom ISA extensions to hit their 2x performance claim. GPU prediction unchanged. TSMC claim slightly less likely given my misinformation about transistor density, but the silkscreen code on the leaked A7 still suggests it may not be Samsung.
To be fair, if you're doing full custom layout, going triple core doesn't seem that much a challenge anymore to me
edit: Pretty significant update. Ben Bajarin (analyst) claims to have a source that backs up A7 being dual-core, not quad core. https://twitter.com/BenBajarin/status/378567815969525761
If his source is accurate, that lends further support to a lower clocked quad cluster GPU as opposed to a high clocked dual cluster GPU to spend some of those 500M extra transistors from A6 to A7.
edit 2: I'm hearing 1.7 GHz clock from multiple sources. This meshes with the 31% faster rumor that came along with the 64 bit (which turned out to be right). What this means is that to get to the 2X claim, the processor has to have 50% better IPC (unlikely) or it's merely better optimized for some scenarios because of new instructions/registers for specific tasks.
edit 3: Ben Bajarin is claiming Apple may be doing custom additions to the ARMv8 (ISA) to get a performance boost. That would be a very big deal because it would make them almost equivalent to AMD in that AMD and Intel both contribute to the x86 ISA. Apple wouldn't have to disclose these to ARM though because they don't sell chips themselves.
https://twitter.com/BenBajarin/status/378585813887356928
edit 4: I've updated the thread title to reflect the most recent news. CPU is likely dual core with custom ISA extensions to hit their 2x performance claim. GPU prediction unchanged. TSMC claim slightly less likely given my misinformation about transistor density, but the silkscreen code on the leaked A7 still suggests it may not be Samsung.
Last edited:
Apple's in charge of their compiler (clang/LLVM), sure they can make extensions they deem useful. Swift appeared as a different ARMv7s architecture target than earlier ARMv7 based Cortex-A9, they probably did it already. Apple's engineer posted on LLVM mailing list that Apple works with an internal fork of LLVM's AArch64 front-end and that they will be commiting it later as it's going to take plenty of work to re-integrate to the open source project. Clearly they have been holding something close to their chest - something like support for custom architecture extensions..
Perhaps desktop class performance means desktop class complexity of the core getting desktop class IPC? When you take cache away there's surprisingly little die area left of each core. Is it possible Apple went all out with die area to make a monster of a number cruncher? It's not difficult to achieve, it's been done on desktop before, it's a matter of choosing the right compromises. Aiming for best performance per core is the right thing to do, not just for mobile where it's way more useful than multiple slower cores that can only be reasonably utilized with careful multithreaded coding, but it also paves the way for Apple to then take their high performance mobile dual core and turn it into say laptop quad core with a clock speed bump!
So 1.3 to 1.7 is roughly 31% which is entirely believable target moving from 32nm to 28nm. What it means is that we're looking at about 5 DMIPs/MHz to get 53% ahead from 3.3 DMIPs/MHz of Swift - which is actually slightly more than Intel's Core 2! Desktop class alright.
P.A.Semi were hitting 4.4 DMIPs/MHz with their PowerPC design at the time Apple bought them - however relevant that may be.
http://en.wikipedia.org/wiki/Instructions_per_second
Really exciting - for A7 as well as for the future of other product lines..
Perhaps desktop class performance means desktop class complexity of the core getting desktop class IPC? When you take cache away there's surprisingly little die area left of each core. Is it possible Apple went all out with die area to make a monster of a number cruncher? It's not difficult to achieve, it's been done on desktop before, it's a matter of choosing the right compromises. Aiming for best performance per core is the right thing to do, not just for mobile where it's way more useful than multiple slower cores that can only be reasonably utilized with careful multithreaded coding, but it also paves the way for Apple to then take their high performance mobile dual core and turn it into say laptop quad core with a clock speed bump!
So 1.3 to 1.7 is roughly 31% which is entirely believable target moving from 32nm to 28nm. What it means is that we're looking at about 5 DMIPs/MHz to get 53% ahead from 3.3 DMIPs/MHz of Swift - which is actually slightly more than Intel's Core 2! Desktop class alright.
P.A.Semi were hitting 4.4 DMIPs/MHz with their PowerPC design at the time Apple bought them - however relevant that may be.
http://en.wikipedia.org/wiki/Instructions_per_second
Really exciting - for A7 as well as for the future of other product lines..
Last edited: