CPU Performance

Like the A6, the 4th generation iPad's A6X integrates two custom designed Apple Swift CPU cores that implement the ARMv7 instruction set. The two cores share a 1MB L2 cache, just like they do in the iPhone 5's A6. I spent a lot of time characterizing the architecture and performance of Swift in our iPhone 5 review, so I won't go over a lot of that here. What I will say is that Swift manages to be the best balance of performance and power efficiency in an ARM based CPU core available today. Swift looks a lot like Apple's take on Qualcomm's Krait core. It's not a Cortex A15 competitor from a performance standpoint, but likely a more power efficient design.

Unlike what we saw in the previous iPads, the A6X only boasts a marginal increase in max clock speed (~7%) over the A6. The two Swift cores can run at up to 1.4GHz in the A6X compared to 1.3GHz in the A6.

Apple iPhone/iPad CPU Clock Comparison
  iPhone iPad iPad Frequency Advantage
Apple A4 Generation 800MHz 1.0GHz 25%
Apple A5 Generation 800MHz 1.0GHz 25%
Apple A6 Generation 1.3GHz 1.4GHz 7.7%

CPU performance improves substantially over the 3rd generation iPad thanks to the Swift cores. Unfortunately, big gains in CPU performance aren't always apparent in actual use. The iPhone 5 seemed to subjectively gain more from the move to A6, partially because of just how frequency constrained the A5 in the 4S was. The A5X CPUs in the iPad 3 were already running 25% faster than the A5's CPU cores.

I didn't notice appreciable differences in application launch times, but there are a few areas were the faster CPU cores definitely stand out. Auto enhance in iPhoto completes quicker on the iPad 4 vs. its predecessor. This is especially true for work on large (high MP) images. I don't think a pair of Swift cores is all it takes to make the iPad 4 a suitable photo retouching workstation but it's better than the previous model.

General Use CPU Performance Comparison
  iPad 2/mini iPad 3 iPad 4
iPhoto Auto Enhance 36MP JPG 11.7 seconds 12.0 seconds 9.1 seconds
Batman: The Dark Knight Rises Level Load Time 36.3 seconds 35.9 seconds 31.1 seconds

Scrolling around iOS Maps in 3D view is substantially smoother on the 4th gen iPad compared to its predecessor. We noticed something similar in our comparison of the iPhone 5 and 4S. Web pages do load quicker on the iPad 4, although the differences can be small depending on the complexity of the page in question. The real world responsiveness benefit from the faster CPU cores seems to be in the 10 - 30% range depending on what you're doing. The problem is at the lower end of the scale, it can be difficult to really feel if you're comparing the 3rd and 4th gen iPads.

We're still very limited in good, cross-platform CPU benchmarks. We, once again, turn to JavaScript tests run in the browser. For all of the tablets tested here we're using Chrome for Android and Mobile Safari for iOS.

SunSpider has been a part of all of our mobile performance testing for quite a while now. It's not the perfect benchmark, but the test ends up being a good measure of browser performance as well as cache latency and CPU performance.

SunSpider Javascript Benchmark 0.9.1 - Stock Browser

The Swift cores do an excellent job here thanks to their much improved memory performance. The Nexus 10 should be competitive however the current state of browser optimizations hold it back a bit here.

Google's Octane benchmark is a much larger test than SunSpider, but we're still looking at JavaScript performance. Octane includes all 8 of the tests from Google's older V8 benchmark but adds 5 new ones including a PDF reader, 3D bullet physics engine and portable 3D game console emulator all built in javascript.

Google Octane Benchmark v1

Octane tells a very different story. The iPad 4 does very well, but the Nexus 10's Cortex A15s are really allowed to shine here. The A15 is simply in a different league of performance.

Finally we have Kraken, a seriously heavy javascript benchmark built by Mozilla. Kraken focuses on forward looking applications that are potentially too slow to run in modern browsers today. The result is much longer run times than anything we've seen thus far, and a very CPU heavy benchmark:

Mozilla Kraken Benchmark

We see a similar story with Kraken. Apple improved CPU performance tremendously over the iPad 3, but the Nexus 10 is able to reach new heights.

Why don't we see a direct correlation between these CPU bound js tests and the real world performance numbers I mentioned above? The problem is that most of today's iOS apps are really built for the sort of performance that a pair of Cortex A9s can deliver. I see pockets of CPU bound activity but it's not consistent enough where you're going to necessarily see a dramatic improvement in responsiveness across the board. I do believe the gap is more apparent on the iPhone thanks to the huge difference in frequency between the 4S and 5, but on the iPad the range from subtle to noticeable differences is broader.

Memory Bandwidth

A look at Chipworks' die shot reveals the same memory interface arrangement as we saw in the A5X:


Apple A6X die via Chipworks

The A5X/A6X place the memory interface blocks (and perhaps the controllers themselves?) adjacent to the GPU, while the A5/A6 more tightly integrate the CPU and memory controller. This highlights a pretty substantial difference in priority between the A5/A6 and A5X/A6X SoCs. The latter really do prioritize memory bandwidth delivery to the GPU, and for good reason. The Retina Display equipped iPads have over 4x the number of pixels as the iPhone 5.


iPad 4 DRAM via iFixit

The 4th generation iPad that iFixit took apart had 1GB of Hynix LPDDR2-1066 on-board, which would indicate a small increase in memory bandwidth. Just as we saw with the A5/A5X comparison, the additional memory bandwidth doesn't appear to be available to the CPU cores - only the GPU cores.

Display Analysis GPU Performance
POST A COMMENT

113 Comments

View All Comments

  • darkcrayon - Friday, December 07, 2012 - link

    Ok that was pretty funny. "When you need paper that gets stuff done".

    Most documents in existence electronically are closer to a 3:4 aspect ratio. I'm sorry, but it's true. And they will fit an iPad sized screen more conveniently. But please, feel free to show me the treasure trove of ready made documents that are 9:16 or even 10:16 long.

    Also, show me the benchmarks where the Nexus 10 GPU is better than the iPad 3 - not that that disproved my point. The one in the iPad 4 is quite a bit more powerful, so it's a legitimate area where the Nexus 10 is *not* a "much better tablet". And of course don't even get us started on 10" Android tablet apps.
    Reply
  • ltcommanderdata - Thursday, December 06, 2012 - link

    Any estimate on the SGX554MP4 clock speed? Since the 16% pixel fill performance difference is supposed to be largely clock speed related and the iPad 3's SGX543MP4 was clocked at 250MHz, that works out to 290MHz. Previous Apple SoC used a 4x multiplier between CPU and GPU. If the A6X changed to a 5x multiplier, then a 1.4GHz CPU works out to a 280MHz GPU, which seems to match what is being seen here.

    And process improvements for SoC power and thermals don't have to wait for 20 nm. Samsung has a 28 nm process, which they claim uses the same design rules as their 32nm process allowing a quick and easy transition, while offering power or performance benefits over their 32 nm process. I'm guessing 2013 iDevices will use a combination of shrinks and new SoC on the 28 nm Samsung process, which with all the TSMC rumours, may be the last SoC Apple makes with Samsung. 2014 iDevices would then use TSMC 20 nm.
    Reply
  • Krysto - Thursday, December 06, 2012 - link

    So it seems that in terms of gaming performance (which is really what matters when you evaluate a GPU), the devices compare as it follows:

    Ipad 4, Nexus 4 and iPhone 5 have the same performance in games - compared as devices, running at their own native resolution.

    http://images.anandtech.com/graphs/graph6472/51760...

    As for the GPU's themselves, it seems the A6X GPU is 42% faster than Mali T604, while Mali T604 and Adreno 320 are about 10-20% faster than the A6 GPU (iPhone 5), and than the A5X GPU (iPad 3)

    http://images.anandtech.com/graphs/graph6472/51761...
    Reply
  • coder543 - Thursday, December 06, 2012 - link

    from one perspective. Honestly, benchmarks are just benchmarks. I'd agree with that assessment as far as these benchmarks go though. I'm not sure how much it matters -- suffice it to say they're all very fast. Reply
  • Alucard291 - Sunday, December 09, 2012 - link

    Very slow surely? As this is STILL about 10 years behind desktop gpu's Reply
  • ltcommanderdata - Thursday, December 06, 2012 - link

    Well the Nexus 4 having the same gaming performance as the iPad 4 and iPhone 5 depends on how often people take their phone into a freezer to play games. Reply
  • michal1980 - Thursday, December 06, 2012 - link

    amazing. Apple must be paying for this coverage, and it was a week or two ago I was told that something was coming soon.

    should have figured, another sligthly updated apple product
    Reply
  • darkcrayon - Thursday, December 06, 2012 - link

    Just as we should've figured you'd be in the comment section whining about it ;) Reply
  • vision33r - Thursday, December 06, 2012 - link

    Why would Apple be interested in marketing to PC enthusiasts on a site like this? It's a waste of money if they did pay Anand.

    In case you ever opened a male or female fashion mag or financial magazine, Apple advertisements are everywhere. Clearly they are more interested in marketing towards rich casual computer users.

    So your theory of Apple paying for this coverage as usual trolling and Apple hating. That's ok, they don't mind you hating them on a website they have little interested marketing on.
    Reply
  • pliablemoosethebanned - Thursday, December 06, 2012 - link

    Yep, Anand needs to focus more on crap products like the Surface tablet. Reply

Log in

Don't have an account? Sign up now