POST A COMMENT

60 Comments

Back to Article

  • tipoo - Friday, September 21, 2012 - link

    Are these images enough for us to know if this is the first shipping A15 core, or if it's a custom core from Apple? I know a lot of the evidence points to custom core, but can we know with complete certainty now? Reply
  • Fleeb - Friday, September 21, 2012 - link

    Wasn't it already said in previous articles it is a custom A7? Reply
  • tipoo - Friday, September 21, 2012 - link

    As I said, but nothing was known with complete certainty afaik. Reply
  • EnzoFX - Friday, September 21, 2012 - link

    It was implied by their complete retraction of the A15 hype. Reply
  • bartoni - Friday, September 21, 2012 - link

    The large diffusion image clearly shows that the dual A6 ARM cores are unlike any other ARM processor core. The Cortex A15 may have been the starting point, but Apple's A6 cores have a full-custom stack-based design (like Intel or older AMD designs) as opposed to the more common standard-cell automated place-and-route sea-of-gates (like AMD Bobcat & Jaguar) approach.
    Roughly speaking a 1.2GHz Cortex A15 core with an improved bus and memory controller might reach about 2X the performance of the 800MHz ARM Cortex A9 cores in the A5.
    Reply
  • jameskatt - Saturday, September 22, 2012 - link

    So what you are saying is that Apple has achieved ARM A15 speeds using a custom dual A7 with vastly improved memory speed, which uses a lot less power and die space than the A15.

    That's an amazing feat from Apple!
    Reply
  • ImSpartacus - Saturday, September 22, 2012 - link

    Wait a minute, I don't understand where he implied that A7 was being used (in any capacity). Reply
  • Zink - Saturday, September 22, 2012 - link

    Probably just confused about Cortex-A7 vs the current ARM v7s instruction set. Reply
  • tipoo - Saturday, September 22, 2012 - link

    " Roughly speaking a 1.2GHz Cortex A15 core with an improved bus and memory controller might reach about 2X the performance of the 800MHz ARM Cortex A9 cores in the A5."

    I would think it's more than that, just going from 800 to 1200 is 1.5x the speed, add in better memory controllers(which is a big area A9 suffered) and you might just get to 2x without huge A15 cores.
    Reply
  • Zingam - Tuesday, September 25, 2012 - link

    Didn't they say it is a custom design according to requirements by Steve Jobs set 4 years ago? Reply
  • Jamezrp - Friday, September 21, 2012 - link

    3 GPU cores? That's been trouble in the past. Why the hell...I almost don't even want to know. The A5X has 4 GPU cores, basically the same as the PS Vita (different setup, same general build).

    There are so many problems with this whole iPhone 5 hardware thing that I can't even rap my brain around it.
    Reply
  • tipoo - Friday, September 21, 2012 - link

    Why would it be trouble? The SGX cores are built to scale to whatever number you want, this isn't like SLI or anything. What are these other problems you can't wrap (I assume you're not trying to write a song :P ) your head around? Reply
  • BrianTho2010 - Friday, September 21, 2012 - link

    What problems are those? Reply
  • DerPuppy - Friday, September 21, 2012 - link

    Indeed, what problems? Nvidia/AMD cut-down cores are regularly asymmetrical in some way without problems. I feel it would pose more of a problem if the memory channels were asymmetrical than if the actual GPU core count weren't some regular power of two/arbitrary happy number Reply
  • tipoo - Friday, September 21, 2012 - link

    I suspect he's thinking this is like a triple SLI or Crossfire implementation, but I'll not put words in his mouth. Reply
  • Blazorthon - Friday, September 21, 2012 - link

    A lot of Nvidia cards have an odd number of memory controllers. Anything with a 192 bit bus has three 64 bit memory controllers. Reply
  • Blazorthon - Friday, September 21, 2012 - link

    Also, the memory chips on Nvidia's cards with 1GB/2GB configurations on 192 bit buses have one controller that has 1GB of memory while the other two have 0.5GB of memory, so they don't even need to have the same memory capacity per controller, although it is more optimal. Reply
  • DerPuppy - Saturday, September 22, 2012 - link

    I believe it was mentioned that asymmetrical memory controllers for GPUs have deficits in certain areas depending on access patterns, whereas this is just the core (memory buses are still symmetrical) so what you mentioned isn't quite as relevant. Reply
  • KitsuneKnight - Friday, September 21, 2012 - link

    So many problems? Mind elaborating? Reply
  • Penti - Friday, September 21, 2012 - link

    It's three "compute units" nothing wrong with that and the SoC seems perfectly fine. Basically 3/4 of the power of the A5X just from the GPU-config, might have better clocks though. So roughly the same, but in phones. It's not three gpus or any SLI/Crossfire/Lucid configuration.

    They do stuff in parallel, they don't have three memory interfaces, the driver/software just schedules the work for the available units, but so does any other GPU. And it is tiled-based rendering regardless of how many cores you have here,1-16 is supported. Regardless of how many threads it's core can keep active. The tiles are handled in hardware. The scaling is pretty linear and don't have much overhead. It's not like having to copy frame buffers over the PCIe bus. Most mobile GPU's does do tiling as does the Xenos Xbox360 gpu. All have some type of multicore setup. Should make no difference that it is an odd number here. Rendering are already subdivided and multi-threaded even with one core here.
    Reply
  • ImSpartacus - Saturday, September 22, 2012 - link

    But remember that A6 uses a slightly newer GPU, so performance is more than 3/4 of the A5X's GPU. As mentioned in the performance article, the two are rather close together in GPU muscle. Reply
  • Death666Angel - Saturday, September 22, 2012 - link

    It doesn't use a newer GPU, it's the same SGX543 by all accounts, only with slightly higher clocks, which results in the 3 cores here being competitive with the 4 cores in the iPad or twice as fast as the 2 cores in the iPhone4s. Reply
  • ImSpartacus - Saturday, September 22, 2012 - link

    You're right. I don't know why I thought the A5 used SGX540MP2/4. Reply
  • Penti - Saturday, September 22, 2012 - link

    It's just clocked slightly higher, with a much faster cpu to accompany it. Making it roughly as fast or even a bit faster despite less bandwidth etc. Reply
  • Zink - Friday, September 21, 2012 - link

    The iPhone 5 is an amazing piece of hardware if nothing else (the OS is very basic). This is the first time in 6 years of iPhones that Apple has been able to double CPU performance two years in a row. This hardware is 4x faster than the iPhone 4 and Galaxy S etc. from 2 years ago with over 10x the 3D performance. The efficiency needed to get better battery life than the 4 and 4s is only possible currently with a custom optimized SOC and that is what this is.

    Battery life is the #1 feature in a phone. It wouldn't be hard to make the iPhone 5 3x as fast as the 4s but only getting 5h of battery life would make everyone hate it.
    Reply
  • jameskatt - Saturday, September 22, 2012 - link

    IOS is pretty much a UNIX OS with far more sophistication. It is essentially Mac OS X but customized for multitouch and tablet use. The power is hidden. But when jail broken and a terminal installed you will see that it is very sophisticated.

    Apple products are always the blend of custom software to custom hardware. They can't be teased apart from one another. And they are more than the sum of their parts.
    Reply
  • tipoo - Saturday, September 22, 2012 - link

    OP does not deliver Reply
  • Loki726 - Friday, September 21, 2012 - link

    The advances in automatic synthesis and place-and-route tools are really staggering.

    You can clearly see the distinction between the SRAM macros and the synthesized logic. Yet another incredibly hard problem falls to machine learning algorithms.

    I can only imagine how many engineers would have been up late into the night to do that layout by hand.
    Reply
  • Loki726 - Friday, September 21, 2012 - link

    Here is a paper that describes the underlying technology: www.princeton.edu/~chaff/publication/DAC2001v56.pdf Reply
  • Becherovka05 - Friday, September 21, 2012 - link

    I find it hard to see the difference between the GPU cores and what i am guessing is memory.
    Well its not my field so I leave it to the experts
    Reply
  • name99 - Friday, September 21, 2012 - link

    It's not memory, it's various other bits and pieces. There'll be hardware there to signal process images from the camera. There'll be an h.264 encode and decode engine. There'll be a block to play audio (so everything else can sleep while music plays). There'll be a memory controller. etc etc

    The actual 1GB memory will be on a separate chip mounted in close proximity to this chip, but fabricated separately, using different processes.

    What I am curious about is the green fungus on everything. Is that an imaging artifact, a consequence of how the cover was scraped off the chip, or what?
    Reply
  • Loki726 - Friday, September 21, 2012 - link

    The "green fungus" is automatically synthesized logic, perhaps from a high level model of video decoder engine. See my above post. Reply
  • Zink - Friday, September 21, 2012 - link

    Wow. That is truly amazing. So the computer optimized transistors are almost organic in organization? Reply
  • Loki726 - Friday, September 21, 2012 - link

    They use a combination of heuristics for constraint (SAT) solvers and machine learning algorithms (particle swarm, genetic evolution, or neural networks).

    They easily beat hand layout by people.
    Reply
  • Loki726 - Friday, September 21, 2012 - link

    The High-Level and Logic Synthesis track at The Design Automation Conference (DAC) http://www.dac.com/ gives a good overview of the current state of the art. Reply
  • KitsuneKnight - Friday, September 21, 2012 - link

    That seems unlikely. Look close at any of the GPUs and you can see the patterns of the 'green fungus' looking stuff isn't the same between the different GPUs/cores. It's most likely a product of how they removed the casing from the chip, doing it quickly to get a shot out ASAP instead of taking time to properly clear it all away. Reply
  • Loki726 - Friday, September 21, 2012 - link

    Maybe, but it does looks similar to what synthesis tools produce. See this figure of AMD steamroller for comparison: http://images.anandtech.com/doci/6201/Screen%20Sho... Reply
  • Loki726 - Friday, September 21, 2012 - link

    Also, if it is a product of the casing then why doesn't it cover the CPU cores? Reply
  • Zink - Friday, September 21, 2012 - link

    It is fairly similar between gpu cores. Image search "automatically synthesized logic silicon" and there are a few chips that have similarities. I think the CPU cores have an organic looking texture different from the fungus. Those for sure are optimized like this, maybe the fungus is just residue. Reply
  • Loki726 - Friday, September 21, 2012 - link

    You can explain the different pattern for different GPU cores by them just running the tools once for each core. The whole point of these tools is that you don't want to just design a component once and then replicate it a bunch of times: you want to do global optimization. Reply
  • stingerman - Friday, September 21, 2012 - link

    Considering maximum yields and power usage, I wouldn't be surprised if the iPhone 5 A6 only used two of the GPU cores. I can imagine that the next version of the iPad, will use a higher clocked A6 with all three cores turned on. A processor with a defective core would still work with the iPhone 5. I wonder if Apple will release a newer Apple TV with a single core A6 that would take up A5's with a defective CPU core... Reply
  • KitsuneKnight - Friday, September 21, 2012 - link

    That seems rather doubtful, simply because it would be very hard for them to achieve double the performance with just two cores (they would have to clock them very high). If the next generation of GPUs were ready, they'd have likely gone with one or two of those, but by all accounts, we should have seen an even larger performance jump if we went to Rogue. Reply
  • winterspan - Friday, September 21, 2012 - link

    very unlikely...

    1) Both iPhone 4s (MP2) and iPad 3 (MP4) used all the gpu cores.
    2) The performance being 2x the iPhone 4s is consistent with moving from 2 to 3 (higher-clocked) cores... To reach that with only two cores would require moving to the SGX 554 and/or a lot higher clocks
    Reply
  • Constructor - Friday, September 21, 2012 - link

    Nope. The iPhone 5 already needs all three at increased clock in order to boost the effective performance to the inficated level. All three would need to be live. They've also never used this approach in any of their flagship products (but the single-core A5 in the Apple TV may actually be production leftovers with one non-working core).

    For the iPad 3 it would still be a step down: It already has four of these same GPU cores (assuming the type is the same here) combined with a double-wide memory interface (4*32) to keep them busy and it definitely needs every bit of that infrastructure.
    Reply
  • tipoo - Friday, September 21, 2012 - link

    I have no doubt all three are being used in the iPhone, the iPhone and iPad have always gotten the fully functional chips, then marginal products like the Apple TV get the chips that didn't make that cut with some parts disabled. Like the new ATV is an A5 with only one core working. Reply
  • neo57 - Friday, September 21, 2012 - link

    Just got my new Iphone 5. First thing I did was download GLBench 2.5.

    Here are the results......

    C24Z16 4361 -39FPS
    C24Z16 offscreen 3009-27FPS
    C24Z24S4 4274 -38FPS
    C24Z24S4 offscreen FAILED
    C24Z16 fixtime 32857-34 FPS
    C24Z16 fixtime off screen 52352 22 FPS
    C16Z16 6739-60 FPS
    C16Z16 10330-91 FPS
    C24Z16 Fill test 1797670784 tex/sec
    C24Z16 Fill test offscreen 1791804544 tex/sec
    C24z16 tri TexTest 115217352 tri/sec
    C24z16 tri TexTest offscreen 96770984 tri/sec
    Reply
  • tipoo - Friday, September 21, 2012 - link

    For comparison

    http://www.glbenchmark.com/phonedetails.jsp?benchm...

    Not sure which tests match with which from your post but the ones I can tell do look significantly faster.
    Reply
  • 1008anan - Friday, September 21, 2012 - link

    Keep it coming Anand Shimpi. Looking forward to new specs soon.

    Some in the comment section said that Apple will use defective chips with one or more GPU shader cores or CPU cores disabled in other products. I don't agree because:
    --Apple hasn't traditionally done this
    --My understanding is that physical cost of manufacturing a 100 square millimeter die is very small. Most of the cost is in IP licensing. Apple doesn't have to pay licensing for defective die. Apple would rather save on the licensing fees and accept lower yields.

    What would the actual manufacturing cost of 100 square meters of die be? It probably costs Intel less than $1 in cost of goods sold. How much does Apple have to pay Samsung for the raw manufacturing cost per 100 square meter chip? Would anyone like to speculate?
    Reply
  • name99 - Friday, September 21, 2012 - link

    Apple is EXTREMELY disciplined about how they reuse their IP across products, which translates into being able to reuse dies across products. We have seen this in the way the same building blocks have been reused across iPhone, iPad, iPod Touch and Apple TV. Very little modification just to be different of the kind you see in other CE companies where even the different phones are in different groups and using different parts, let alone items as different as Apple TV and iPhone.

    We ALSO know that Apple has been happy to swap out some of the guts of a device if they can replace them with something equal or better (replacing the old A5 in the iPad 2 with the die-shrunk version).

    Since we don't know the yields, we have no idea whether the fraction of chips with only two working GPU units is high or low. But I expect Apple have plans either way; and if there is a large fraction of chips with broken GPUs, those chips are being stored for a future project for which they will make sense. Maybe the next iPod Touch? Maybe the next Apple TV? There's no real cost or difficulty to warehousing chips for a year or even more.
    Reply
  • tipoo - Friday, September 21, 2012 - link

    I have no doubt all three are being used in the iPhone, the iPhone and iPad have always gotten the fully functional chips, then marginal products like the Apple TV get the chips that didn't make that cut with some parts disabled. Like the new ATV is an A5 with only one core working. Reply
  • tipoo - Friday, September 21, 2012 - link

    iPhone 4S, 5, and GSIII in GLBenchmark 2.5

    Fill rate - Offscreen (1080p):
    773785 kTexels/s 1766182 kTexels/s 690334 kTexels/s
    Fill rate - Onscreen:
    777260 kTexels/s 1758664 kTexels/s 712236 kTexels/s
    GLBenchmark 2.1 Egypt Battery - 100% Brightness 30 FPS Max:
    5:03 [h] No result yet. 4:35 [h]
    GLBenchmark 2.1 Egypt Battery - 100% Brightness 30 FPS Max :
    508959 Frames No result yet. 495528 Frames
    GLBenchmark 2.1 Egypt Battery - 100% Brightness 60 FPS Max:
    3:19 [h] No result yet. 3:57 [h]
    GLBenchmark 2.1 Egypt Battery - 100% Brightness 60 FPS Max :
    703611 Frames No result yet. 652369 Frames
    GLBenchmark 2.1 Egypt Battery - 50% Brightness 60 FPS Max:
    4:05 [h] No result yet. 3:47 [h]
    GLBenchmark 2.1 Egypt Battery - 50% Brightness 60 FPS Max :
    862611 Frames No result yet. 708366 Frames
    GLBenchmark 2.1 Egypt Classic - Offscreen (1080p):
    4535 Frames 10253 Frames 6662 Frames
    GLBenchmark 2.1 Egypt Classic - Onscreen:
    6685 Frames 6776 Frames 6667 Frames
    GLBenchmark 2.5 Egypt HD Battery - 100% Brightness 30 FPS Max:
    2:49 [h] No result yet. 4:06 [h]
    GLBenchmark 2.5 Egypt HD Battery - 100% Brightness 30 FPS Max :
    196142 Frames No result yet. 216728 Frames
    GLBenchmark 2.5 Egypt HD Battery - 100% Brightness 60 FPS Max:
    2:47 [h] No result yet. 4:18 [h]
    GLBenchmark 2.5 Egypt HD Battery - 100% Brightness 60 FPS Max :
    190561 Frames No result yet. 194499 Frames
    GLBenchmark 2.5 Egypt HD Battery - 50% Brightness 60 FPS Max:
    3:21 [h] No result yet. 3:55 [h]
    GLBenchmark 2.5 Egypt HD Battery - 50% Brightness 60 FPS Max :
    253985 Frames No result yet. 215405 Frames
    GLBenchmark 2.5 Egypt HD C24Z16 - Offscreen (1080p):
    1252 Frames 3235 Frames 1708 Frames
    GLBenchmark 2.5 Egypt HD C24Z16 - Onscreen:
    2172 Frames 4499 Frames 1724 Frames
    GLBenchmark 2.5 Egypt HD C24Z16MS4 - Offscreen (1080p):
    Failed/Not supported Failed/Not supported Failed/Not supported
    GLBenchmark 2.5 Egypt HD C24Z16MS4 - Onscreen:
    2094 Frames 4245 Frames 1711 Frames
    GLBenchmark 2.5 Egypt HD Fixed Timestep - Offscreen (1080p):
    123980 ms 58097 ms 103347 ms
    GLBenchmark 2.5 Egypt HD Fixed Timestep - Onscreen:
    69569 ms 35291 ms 102272 ms
    Triangle throughput: Textured - Offscreen (1080p):
    34553 kTriangles/s 96764 kTriangles/s 8605 kTriangles/s
    Triangle throughput: Textured - Onscreen:
    44418 kTriangles/s 118977 kTriangles/s 16226 kTriangles/s
    Triangle throughput: Textured, fragment lit - Offscreen (1080p):
    29773 kTriangles/s 82395 kTriangles/s 7597 kTriangles/s
    Triangle throughput: Textured, fragment lit - Onscreen:
    37531 kTriangles/s 94779 kTriangles/s 8996 kTriangles/s
    Triangle throughput: Textured, vertex lit - Offscreen (1080p):
    32506 kTriangles/s 86461 kTriangles/s 7635 kTriangles/s
    Triangle throughput: Textured, vertex lit - Onscreen:
    39214 kTriangles/s 96884 kTriangles/s 9269 kTriangles/s
    Reply
  • Zink - Friday, September 21, 2012 - link

    Those numbers are hard to look at; I keep thinking I found a test where the GSIII is close but then I count digits and the iPhone 5 is an order of magnitude ahead. Reply
  • Aenean144 - Friday, September 21, 2012 - link

    Which GSIII do you have? Not that'll make too much of difference as the Mali in Exynos 4412 and Adreno 225 are somewhat competitive with the Mail being better right?

    Why so many "no results" for the iPhone 5?
    Reply
  • tipoo - Saturday, September 22, 2012 - link

    These aren't mine, they're pulled from the GLbenchmark results database, and I assume the rest will be filled in as they test them. Reply
  • Aenean144 - Friday, September 21, 2012 - link

    As I recall, the PowerVR SGX543MP2 in the Apple A5 SoC dwarfed the dual-core Cortex-A9 CPU. These A6 CPU cores look a quite a bit bigger. Reply
  • madmilk - Friday, September 21, 2012 - link

    As they should be, considering IPC has been nearly doubled. Reply
  • bartoni - Saturday, September 22, 2012 - link

    IPC hasn't doubled, it has increased by ~35% over the Cortex A9. The A6 cores run upto 1.2GHz while the A5 cores run up to 800MHz. Reply
  • tipoo - Saturday, September 22, 2012 - link

    The clock speed difference accounts for 1.5x change (1200 vs 800), the IPC isn't doubled. Reply
  • semiconDr - Friday, September 21, 2012 - link

    Let me see if I have this right. BUM TechInsights decapsulated the A6 chip and took a high resolution optical image of the die with a precision optical microscope, but they are estimating the die size to be 95.02 mm^2. How many significant figures to you get in the final measurement? Reply
  • mrtanner70 - Monday, September 24, 2012 - link

    http://semiaccurate.com/2012/09/24/an-update-on-ap...

    A8 or A9?
    Reply

Log in

Don't have an account? Sign up now