With their iPhone keynote behind them, Apple has begun updating some of their developer documentation for iOS to account for the new phone. This of course is always a fun time for tech punditry, as those updates will often include information on the hardware differences in the platform, and explain to developers the various features that different generations of hardware can offer developers.

To that end we have compiled a short analysis of the A8 SoC based on these documents and other sources. And we believe that at this point we have a solid idea of the configuration of Apple's latest SoC.

Apple SoC Specifications
  Apple A6 Apple A7 Apple A8
CPU Swift @ 1.3GHz(x2) Cyclone @ 1.3GHz (x2) Enhanced Cyclone @ 1.4GHz (x2)?

A8’s GPU: Imagination PowerVR Series6XT GX6650 GX6450

On the GPU front this year appears to be especially bountiful. After being tipped to an update for Apple’s Metal Programming Guide, we can now infer with near certainty that we know what the A8 GPU is.

New to this edition of the Metal Programming Guide is a so-called iOS_GPUFamily2, which joins the existing iOS_GPUFamily1. We already know that the iOS_GPUFamily1 is based around Imagination’s PowerVR Series 6 G6430 GPU, so the real question is what does iOS_GPUFamily2 do that requires a separate family? The answer as it turns out is ASTC, the next generation texture compression format is being adopted by GPU vendors over the next year or so.

Imagination’s PowerVR Series6 family of GPUs predates ASTC and as a result iOS_GPUFamily1 does not support it. However we know that Imagination added support for it in their Series6XT designs, which were announced at CES 2014. Coupled with the fact that Apple’s documentation supports the idea that all of their GPUs are still TDBR (and thus PowerVR), this means that the GPU in the A8 must be a Series6XT GPU in order for ASTC support to be present.

This leaves the question of which of Imagination’s 4 Series6XT Apple is using. Imagination offers a pair of 2 core designs, a 4 core design (GX6450), and a 6 core design (GX6650). Considering that Apple was already using a 4 core design in A7, we can safely rule out the 2 core designs. That leaves us with GX6450 and GX6650, and to further select between those options we turn to Apple’s A8 performance estimates.

Apple SoC Evolution
  CPU Perf GPU Perf Die Size Transistors Process
A5 ~13x ~20x 122m2 <1B 45nm
A6 ~26x ~34x 97mm2 <1B 32nm
A7 40x 56x 102mm2 >1B 28nm
A8 50x 84x 89mm2 ~2B 20nm

A8 is said to offer 84x the GPU performance of the iPhone 1, while last year Apple stated that the A7 offered 56x the iPhone 1’s performance. As a result we can accurately infer that the A8 must be 1.5x faster than the A7, a nice round number that makes it easier to determine with GPU Apple is using. Given Apple’s conservative stance on clockspeeds for power purposes and the die space gains from the 20nm process, accounting for a 50% performance upgrade is easily done by replacing a 4 core G6430 with the 6 core GX6650. At equal clockspeeds the GX6650 should be 50% faster on paper (matching Apple’s paper numbers), leading us to strongly believe that the A8 is utilizing a PowerVR Series6XT GX6650 GPU.

Once the iPhone 6 is out and Chipworks can photograph the SoC, this should be easy to confirm. If Apple is using a GX6650 then the die size of the GPU portion of the A8 should be very similar to the die size of the GPU portion of the A7. Otherwise if it is the 4 core GX6450, then Apple should see significant die size savings from using a 20nm fabrication process.

Update: The Chipworks die shots have confirmed that there are only 4 GPU cores present, not 6. So our earlier speculation was wrong; A8 is powered by GX6450, not GX6650

A8’s CPU: A Tweaked Cyclone?

Though we typically avoid rumors and leaks due to their high unreliability, after today’s presentation by Apple we have just enough information on A8’s CPU performance to go through the leak pile and start picking at leak. From that pile there is one leak in particular that catches our eye due to the fact that it matches Apple’s own statements.

On Monday night a supposed Geekbench 3 score of the iPhone 6 was posted. In this leak the iPhone 6 was listed as having a single-core score of 1633 points and a multi-core score of 2920 points. Curiously, these values are almost exactly 25% greater than the Geekbench 3 scores for the iPhone 5S (A7), which are 1305 points and 2347 points respectively.

The fact that ties all of this data together is that in their iPhone 6 presentation, Apple informed viewers that the iPhone 6 is 25% faster than the iPhone 5S. This data was later backed up with their latest CPU performance graph, which put the iPhone 6 at a score of 50x versus a score of 40x for the iPhone 5S.

Given Apple’s data, it looks increasingly likely that the leaked Geekbench 3 results for the iPhone 6 are in fact legitimate. The data leaked matches Apple’s own performance estimates, and in fact does so very well.

In which case we can infer a couple of points about the A8’s CPU, starting with the clockspeed. Given no other reason to doubt this data at the moment and given Apple’s preference for low clocked SoCs, the 1.4GHz reading appears legitimate. In which case this would be a 100MHz increase over the 1.3GHz A7 found in the iPhone 5S.

However the fact that it’s a 100MHz increase also means that clockspeeds alone cannot account for the full 25% performance gain that Apple is promoting and that these Geekbench results are reporting, as 1.4GHz is only a roughly 8% clockspeed increase over 1.3GHz. This in turn means that there must be more going on under the hood to improve the A8’s CPU performance other than clockspeed alone, which rules out a straight-up reuse of Apple’s Cyclone CPU.

Since Apple already had a solid ARMv8 architecture with Cyclone, there’s no reason to believe that they have thrown out Cyclone so soon. However this does strongly suggest that Apple has made some unknown revisions to Cyclone to further boost its single-threaded (Instruction Level Parallelism) performance. What those tweaks are remain to be seen as we would need to be able to benchmark the A8 in depth to even try to determine what Apple has changed, but for the moment it looks like we’re looking at an enhanced or otherwise significantly optimized version of Cyclone. And given Apple’s already high ILP, squeezing out another 16% or so would be a significant accomplishment at this time, especially for only a year’s turnaround.

1GB of RAM

Last but not least, the apparent validity of the Geekbench 3 leak means that one last piece of information on the A8 can apparently be confirmed: the earlier rumors about it being paired with 1GB of RAM are true. Unfortunately Apple’s official product image of the A8 is of no help here – it’s clearly a doctored version of the A7 image based on the product numbers attached – but this information is consistentwith earlier rumors based on leaked images of the real A8, which had also suggested the SoC contained 1GB of RAM. Again this is based on what we believe is a sound assumption that the Geekbench 3 leak is accurate since it so closely matches Apple’s own CPU performance estimates, but at this point we don’t have any substantial reason to doubt the data.

Image Courtesy Macrumors

The good news is that this is going to be the easiest aspect of the iPhone 6 to confirm, since diagnostic apps will be able to query the phone for the RAM amount. So one way or another we should know for sure come September 19th.



View All Comments

  • D V - Monday, September 15, 2014 - link

    I wouldn't be surprised that any future product for which they want to turn on multitasking in two windows concurrently will have 2GB RAM. Best candidate: iPad Air and future large iPad... and just maybe the iPhone 6 Plus. Reply
  • Tangey - Wednesday, September 10, 2014 - link

    According to IMG, series 6 -> 6XT gains "up to 50%" improvement on a clock to clock basis, from architectural improvements. Apple also referred to "up to 50% in the presentation.

    So its possible that it is GX6450 with a clock bump. Alternatively it might indicated a G6650 clocking slower than the GPU in the A7.
  • darkich - Wednesday, September 10, 2014 - link

    It is the GX6650.
    The sustained performance comes from the GX6650, the Imagination themselves wrote about that GPU keeping its cool even after 50 consecutive benchmark runs.
  • anexanhume - Wednesday, September 10, 2014 - link

    The 50% boost is a result of the changes to the ALUs to have 50% more FP16 performance, so if Apple is using theoretical FP16 numbers, then yes, it could be GX6450 at the same clocks. Not sure why this is being ignored. I don't think we have a precedent in their claims that show they always use FP32. Reply
  • lucam - Thursday, September 11, 2014 - link

    I do agree, still don't get why it can't be the GX6450. Reply
  • dj_aris - Wednesday, September 10, 2014 - link

    What is the cost of adding another gig of RAM I wonder? After 4 or 5 tabs in Safari, each visit in every pages reloads it from scratch. That's unacceptable. So much for the "user experience". Reply
  • MonkeyPaw - Wednesday, September 10, 2014 - link

    Well, sometimes there's more too that than just available RAM. The OS can also aggressively free RAM to conserve power. I had a Galaxy Note 10.1 with 3GB of RAM, and Android Chrome reloaded tabs all the time, even after I disabled as many unused Samsung services as possible. I imagine iOS is pretty aggressive at this too, since even the iPads are limited to 1GB. Reply
  • lilo777 - Thursday, September 11, 2014 - link

    I have Galaxy Note 10.1 2014 and I have never seen tabs reloading (some pages do reload but that's because the web site forces an update). And reloading the page probably requires more power than storing it. I believe that looking at across the specs where we see iDevices having less RAM, less storage (at the same price), less antennas, less sensors, smaller batteries than high end smartphones it is obvious that it's all about money. Reply
  • ninjaquick - Wednesday, September 10, 2014 - link

    http://www.dramexchange.com/WeeklyResearch/Post/2/... 8 to 9 dollars per unit. Reply
  • JDG1980 - Wednesday, September 10, 2014 - link

    1GB of RAM on a flagship device in 2014 is just not OK. I don't want the flash memory (or worse, data plan) to get thrashed every time I switch tabs. They should have moved to 2GB at least a year ago - failing to do so now is inexcusable. I like the iOS ecosystem and the clean layout and lack of crapware compared to Android, but such a small RAM capacity completely rules it out for me. Reply

Log in

Don't have an account? Sign up now