SoC Analysis: Apple A9X

Diving into the heart of the iPad Pro, we have Apple’s latest generation tablet SoC, A9X. Like the other Apple X series SoCs before it, A9X is by and large an enhanced and physically larger version of Apple’s latest phone SoC, taking advantage of the greater space and heat dissipation afforded by a tablet to produce a more powerful SoC.


Apple's A9X (Image Courtesy iFixit)

That Apple has developed a new SoC to launch alongside the iPad Pro is in no way surprising, but just as how the iPad Pro has ramifications for the overall iPad lineup as Apple gets into the productivity tablet market, iPad Pro’s genesis is reflected in its component selection. Apple already needed a powerful SoC for the iPad Air 2 in order to keep performance up with the tablet’s high resolution of the screen, and iPad Pro in turn pushes Apple’s performance needs even harder. Not only is there an even higher resolution screen to drive – the 2732x2048 display has about 66% of the pixels of a 4K display – but now Apple needs to deliver suitable performance for content creation and meaningful multitasking. I don’t want to imply that the A9X was somehow specifically designed from scratch for the iPad Pro, as there are a number of more important engineering considerations, but I do want to highlight how the iPad Pro is not just another iPad, and that as Apple expands the capabilities of the iPad they need to expand the performance as well if they wish to extend their reputation for smooth UX performance.

Looking at the specifications of the A9X, it seems like Apple always throws us a curveball on the X series SoCs, and for their latest SoC this is no different. With A8X Apple delivered more RAM on a wider memory bus, a larger GPU, and surprisingly, three Typhoon CPU cores. To date it’s still not clear just why Apple went with three CPU cores on A8X – was it for multitasking, or as an alternative means to boost performance – and A9X’s configuration only serves to highlight this enigma.

Apple SoC Comparison
  A9X A9 A8X A6X
CPU 2x Twister 2x Twister 3x Typhoon 2x Swift
CPU Clockspeed 2.26GHz 1.85GHz 1.5GHz 1.3GHz
GPU PVR 12 Cluster Series7XT PVR 6 Cluster Series7
(PVR GT7600)
PVR 8 Cluster Series6XT
(APL GXA6850)
PVR SGX554 MP4
RAM 4GB LPDDR4 2GB LPDDR4 2GB LPDDR3 1GB LPDDR2
Memory Bus Width 128-bit 64-bit 128-bit 128-bit
Memory Bandwidth 51.2GB/sec 25.6GB/sec 25.6GB/sec 17.1GB/sec
L2 Cache 3MB 3MB 2MB 1MB
L3 Cache None 4MB 4MB N/A
Manufacturing Process TSMC 16nm FinFET TSMC 16nm &
Samsung 14nm
TSMC 20nm Samsung 32nm

Instead of continuing with a triple-core CPU design for A9X, for their latest X series SoC Apple has dropped back down to just a pair of Twister CPU cores. The catch here – and why two cores is in many ways better than three – is that relative to A8X and A9, Apple has cranked up their CPU clockspeeds. Way, way up. Whereas the iPad Air 2 (A8X) shipped at 1.5GHz and the iPhone 6s (A9) at 1.85GHz, the A9X sees Apple push their clockspeed to 2.26GHz. Not counting the architectural changes, this is 22% higher clocked than the A9 and 51% higher than the A8X.

The fact that Apple dropped back down to 2 CPU cores is unexpected given that we don’t expect Apple to ever go backwards in such a fashion, and while we’ll never know the official reason for everything Apple does, in retrospect I’m starting to think that A8X was an anomaly and Apple didn’t really want a tri-core CPU in the first place. A8X came at a time where Apple was bound by TSMC’s 20nm process and couldn’t drive up their clockspeeds without vastly increasing power consumption, so a third core was a far more power effective option.


A9X Die Shot w/AT Annotations (Die Shot Courtesy Chipworks)

Overall this means that iPad Pro and A9X will set a very high bar for tablet CPU performance. As we’ve already seen in the iPhone 6s review, the Twister CPU core is very potent and in most cases faster than any other ARM CPU core by leaps and bounds. Cranking up the clockspeed a further 22% only serves to open up that gap even further, as Twister is now reaching clockspeeds similar to the likes of Cortex-A57 and A72, but with its much wider execution pipeline and greater IPC. This is also the reason that an Intel Core CPU comparison is so interesting, as Intel’s tablet-class Core processors in many ways are the target to beat on overall CPU performance, and we’ll be touching upon this subject in greater detail a bit later.

GPU: Imagination PowerVR 12 Cluster Series 7XT

Meanwhile on the GPU side, as expected Apple has further increased the number of clusters on their SoC to drive the higher resolution display of a tablet. Whereas A9 used a 6 cluster design (PVR GT7600), A9X doubles this, giving us a relatively massive 12 cluster design.

In Imagination’s PowerVR Series7XT roadmap, the company doesn’t have an official name for a 12 cluster configuration, as this falls between the 8 cluster GT7800 and 16 cluster GT7900. So for the moment I’m simply calling it a “PowerVR 12 cluster Series7XT design,” and with any luck Imagination will use a more fine-grained naming scheme for future generations of PowerVR graphics.

In any case, the use of a 12 cluster design is a bit surprising from an engineering standpoint since it means that Apple was willing to take the die space hit to implement additional GPU clusters, despite the impact this would have on chip yields and costs. If anything, with the larger thermal capacity and battery of the iPad Pro, I had expected Apple to use higher GPU clockspeeds (and eat the power cost) in order to save on chip costs. Instead what we’re seeing is a GPU that essentially offers twice the GPU power of A9’s GPU.

However to put all of this in context, keep in mind that iPad Pro’s display is 5.95Mpixels, versus the 2.07Mpixel screen on the iPhone 6s Plus. So although Apple has doubled the number of GPU clusters for A9X – and I suspect clocked it fairly similarly – that increased performance will be very quickly consumed by the iPad Pro’s high resolution screen. Consequently even a 12 cluster GPU design is something of a compromise; if Apple wanted to maintain the same level of GPU performance per pixel as in the iPhone 6s family, they would have needed an even more powerful GPU. Which just goes to show how demanding tablets can be.

Memory Subsystem: 128-bit LPDDR4-3200, No L3 Cache

Responsibility for feeding the beast that is A9X’s GPU falls to A9X’s 128-bit LPDDR4 memory controller configuration. With twice as many GPU clusters, Apple needs twice as much memory bandwidth to maintain the same bandwidth-to-core ratio, so like the past X-series tablet SoCs, A9X implements a 128-bit bus. For Apple this means they now have a sizable 51.2GB/sec of memory bandwidth to play with. For an SoC this is a huge amount of bandwidth, but at the same time it’s quickly going to be consumed by those 12 GPU clusters.

Geekbench 3 Memory Bandwidth Comparison (1 thread)
  Stream Copy Stream Scale Stream Add Stream Triad
Apple A9X 2.26GHz 20.8 GB/s 15.0 GB/s 15.3 GB/s 15.1 GB/s
Apple A8X 1.5GHz 14.2 GB/s 7.44 GB/s 7.54 GB/s 7.49 GB/s
A9X Advantage 46.4% 101% 103% 102%

It’s also while looking at A9X’s memory subsystem however that we find our second and final curveball for A9X: the L3 cache. Or rather, the lack thereof. For multiple generations now Apple has used an L3 cache on both their phone and tablet SoCs to help feed both the CPU and GPU, as even a fast memory bus can’t keep up with a low latency local cache. Even as recent as A9, Apple included a 4MB victim cache. However for A9X there is no L3 cache; the only caches on the chip are the individual L1 and L2 caches for the CPU and GPU, along with some even smaller amounts for cache for various other functional blocks..

The big question right now is why Apple would do this. Our traditional wisdom here is that the L3 cache was put in place to service both the CPU and GPU, but especially the GPU. Graphics rendering is a memory bandwidth-intensive operation, and as Apple has consistently been well ahead of many of the other ARM SoC designers in GPU performance, they have been running headlong into the performance limitations imposed by narrow mobile memory interfaces. An L3 cache, in turn, would alleviate some of that memory pressure and keep both CPU and GPU performance up.

One explanation may be that Apple deemed the L3 cache no longer necessary with the A9X’s 128-bit LPDDR4 memory bus; that 51.2GB/sec of bandwidth meant that they no longer needed the cache to avoid GPU stalls. However while the use of LPDDR4 may be a factor, Apple’s ratio of bandwidth-to-GPU cores of roughly 4.26GB/sec-to-1 core is identical to A9’s, which does have an L3 cache. With A9X being a larger A9 in so many ways, this alone isn’t the whole story.

What’s especially curious is that the L3 cache on the A9 wasn’t costing Apple much in the way of space. Chipworks puts the size of A9’s 4MB L3 cache block at a puny ~4.5 mm2, which is just 3% the size of A9X. So although there is a cost to adding L3 cache, unless there are issues we can’t see even with a die shot (e.g. routing), Apple didn’t save much by getting rid of the L3 cache.

Our own Andrei Frumusanu suspects that it may be a power matter, and that Apple was using the L3 cache to save on power-expensive memory operations on the A9. With A9X however, it’s a tablet SoC that doesn’t face the same power restrictions, and as a result doesn’t need a power-saving cache. This would be coupled with the fact that with double the GPU cores, there would be a lot more pressure on just a 4MB cache versus the pressure created by A9, which in turn may drive the need for a larger cache and ultimately an even larger die size.

As it stands there’s no one obvious reason, and it’s likely that all 3 factors – die size, LPDDR4, and power needs – all played a part here, with only those within the halls of One Infinite Loop knowing for sure. However I will add that since Apple has removed the L3 cache, the GPU L2 cache must be sizable. Imagination’s tile based deferred rendering technology needs an on-chip cache to hold tiles in to work on, and while they don’t need an entire frame’s worth of cache (which on iPad Pro would be over 21MB), they do need enough cache to hold a single tile. It’s much harder to estimate GPU L2 cache size from a die shot (especially with Apple’s asymmetrical design), but I wouldn’t be surprised of A9X’s GPU L2 cache is greater than A9’s or A8X’s.

Building A9X Big: 147mm2, Manufactured By TSMC

Finally, let’s talk about the construction and fabrication of the A9X SoC itself. Chipworks’ previous analysis shows that the A9X is roughly 147mm2 in die size, and that it’s manufactured by TSMC on their 16nm FinFET process.

At 147mm2 the A9X is the second-largest of Apple’s X-series tablet SoCs. Only the A5X, the first such SoC, was larger. Fittingly, it was also built relative to Apple’s equally large A5 phone SoC. With only 3 previous tablet SoCs to use as a point of comparison I’m not sure there’s really a sweet spot we can say that Apple likes to stick to, but after two generations of SoCs in the 120mm2 to 130mm2 range, A9X is noticeably larger.

Some of that comes from the fact that A9 itself is a bit larger than normal – the TSMC version is 104.5mm2 – but Apple has also clearly added a fair bit to the SoC. The wildcard here is what yields look like for Apple, as that would tell us a lot about whether a 147mm2 A9X is just a large part or if Apple has taken a greater amount of risk than usual here.

A9X continues to be the largest 16nm FinFET ASIC we know to be in mass production at TSMC (we’ll ignore FPGAs for now), and while this will undoubtedly change a bit later this year once the next-generation discrete GPUs come online, I don’t think you’ll find a better example of how the contract chip manufacturing market has changed in a single generation. 4 years ago it would be GPUs leading the charge, but now it’s phone SoCs and a rather sizable tablet SoC that are first out of the gate. After almost a decade of catching up, SoCs have now reached the bleeding edge for chip fabrication, enabling rapid performance growth, but also inheriting the risks of being the leader. I won’t dwell on this too much, but I’m immensely curious about both what A9X yields are like as the largest FinFET ASIC at TSMC, and just how much of TSMC’s FinFET capacity Apple has been consuming with the production of A9 and A9X.

Finally, it's also interesting to note just how large A9X is compared to other high performance processors. Intel's latest-generation Skylake processors measure in at ~99mm2 for the 2 core GT2 configuration (Skylake-Y 2+2), and even the 4 core desktop GT2 configuration (Intel Skylake-K 4+2) is only 122mm2. So A9X is larger than either of these CPU cores, though admittedly as a whole SoC A9X contains a number of functional units either not present on Skylake or on Skylake's Platform Controller Hub (PCH). Still, this is the first time that we've seen an Apple launch a tablet SoC larger than an Intel 4 core desktop CPU.

Introduction and Design SoC Analysis: On x86 vs ARMv8
POST A COMMENT

408 Comments

View All Comments

  • MathieuLF - Monday, January 25, 2016 - link

    Obviously you don't actually work in a real office where they require lots of specialized software. What's the point of having one device to complement another? That's a waste of resources. Reply
  • LostAlone - Tuesday, January 26, 2016 - link

    Totally agree. For a device to really be useful on a professional level then it needs to be something that is useful all the time, not just when it suits it. If you already wanted a tablet for professional stuff to begin with (a pretty shaky assumption since pro users tend to be working in one place where proper keyboard and mouse are usable) then you need a tablet that can be the only device you need to work on. You need something that you can put in your bag and know that whenever you arrive somewhere you are going to have every single tool you need. And that is not the iPad Pro. It's not even close to replacing an existing laptop. It's certainly very sexy and shiny and the big screen makes it great for reading comics and watching videos on the train but it's not a work device. It's simply not. Even in the only field where it might have claim to being 'pro' (drawing) it's not. It's FAR worse than a proper Wacom tablet because the software is so hugely lacking.

    It's an iPad dressed up like a grown up device laying in the shadow of actual professional grade devices like the Surface Pro. You get a Surface then you can use it 24/7 for work. You can buy a dock for it and use it as your primary computer. You can get every single piece of software on your desktop plus anything your employer needs and it's easy for your work sysadmins to include it in their network because it's just another Windows PC. Just dumb stuff like iPads refusing to print on networked printers (which happens ALL the time by the way) exclude it from consideration in a professional space. It's a great device for traditional tablet fare but it's not a pro device.
    Reply
  • Constructor - Tuesday, January 26, 2016 - link

    Your straw man scenarios are just that. In reality most users don't need every last exotic niche feature of any given dinosaur desktop software as a praeconditio sine qua non.

    Which you also might be able to deduce from the fact that most of those features had not been present on these desktop applications either when "everybody" nevertheless used them professionally even so.

    In real life the physical flexibility and mobility of an iPad will often trump exotic software features (most bread-and-butter stuff is very much supported on iOS anyway) where the circumstances simply call for it.

    Actual professionals have always made the difference notably by finding pragmatic ways to make the best use of the actual tools at hand instead of just whining about theoretical scenarios from their parents' basements for sheer lack of competence and imagination.
    Reply
  • kunalnanda - Wednesday, January 27, 2016 - link

    Just saying, but most office workers use a LOT more than simply notetaking, email, wordprocessor and calendar. Reply
  • Demigod79 - Friday, January 22, 2016 - link

    Software companies only create crippled, lesser versions of their software because of the mobile interface. Products like the iPad are primarily touch-based devices so apps must be simplified for touch. Although the iPad supports keyboard input (and have for years) you cannot navigate around the OS using the keys, keyboard shortcuts are few and far between and and it still has touch features like autocorrect (and of course the iPad does not support a trackpad or mouse so you must necessarily touch the screen). The iPad Pro does not change this at all so there's no reason why software companies should bring their full productivity suite to this device. By comparison, PC software developers can rely on users having a keyboard and mouse (and now touchscreens as well for laptops and hybrids) so they can create complex, full-featured software. This is the primary difference between mobile and desktop apps. Just like FPS games must be watered down and simplified to make them playable with a touchscreen, productivity apps must also be watered down to be usable. No amount of processing power will change this, and unless the iPad Pro supports additional input devices (at the very least a mouse or trackpad) it will remain largely a consumption device. Reply
  • gistya - Sunday, January 24, 2016 - link

    This is just wrong. There are many fully-fledged applications available on iPad. It comes with a decent office suite, plus Google's is free for it as well. I haven't touched MS Word in a couple of years, and when I do it seems like a step backwards in time to a former, crappier era of bloatware.

    Sure, I still use Photoshop and Maya and Pro Tools on my Mac, but guess what? iPad has been part of my professional workflow for four years now, and I would not go back.

    Ask yourself: is a secondary monitor a "professional tool"? Heck yes. What about a third or fourth? What if it fits in your hand, runs on batteries, and has its own OS? Now you cannot find a professional use?

    Give me a break. People who are not closed-minded, negative dolts already bought millions of iPads and will keep buying them because of how freaking useful they are, professionally and non-professionally. They will only keep getting more useful as time goes on.

    The main legit critique I've heard is a lack of availability for accessories but that's a production issue, not a product issue. iOS 9 has been out for only a few months and the iPad Pro much less than that... companies have to actually have the device in-hand before they can test and develop on it. Check back in another year or two if it's not up to speed for you yet though.

    That's what I did, waited for iPad 3, iPhone 3gs before I felt they were ready enough. 3gs for its day was by far the best thing out, so was the 4s. I don't think the iPad Pro is nearly as far back as the iPad 1 or iPhone 1 were when they launched but give it time...
    Reply
  • xthetenth - Monday, January 25, 2016 - link

    If you consider features to be bloat, I guess you could say that the iPad has full-fledged applications. It's not true, but you could say it. Actually trying to do even reasonably basic tasks in google sheets is horrifying next to desktop Excel. Things like straightforward conditional formatting, pivot tables, the ability to dynamically order the contents of a table and so on might strike you as bloat but they're the foundation of the workflows of people who make the program the cornerstone of their job. An extra monitor is a much better professional tool if it doesn't have its own special snowflake OS with different limitations and way of moving data. Reply
  • Relic74 - Saturday, February 27, 2016 - link

    The fffice suits available aren't fully fledged applications, they are still just mobile apps. I can't open 80% of my Excel sheets on the iPad Pro simply because it doesn't support Macros. Visual Basic and Databases. Stop trying to convince everyone that the iPad Pro is a laptop replacement, it's not. It's just a bigger iPad, that's all. Which is fine and has it's uses, but the only people who would use the iPad Pro as an actual computer are the same ones who could get by using a ChromeBook. A professional person could use the iPad Pro, yes but they would have a focused purpose, which means only a few specific apps, the rest of time would be spent on a desktop or laptop computer. It's a secondary device. Reply
  • Relic74 - Saturday, February 27, 2016 - link

    Their not crippled, just mobile versions and they have their uses. The problem I have with these comments is when someone says that the iPad Pro is immensely better than the Surface Pro 4. These are completely different devices intended for completely different tasks. These comparisons honestly need to stop, one doesn't buy a Surface Pro 4 to use as a tablet and vice versa. The iPad Pro is a content consumption device first and foremost. Yes there are some productivity apps and certain professions like a musician or an artist could take advantage of the iPad Pro's capabilities however it is not and I cannot stress this enough, is not a laptop replacement . Those that can use the iPad Pro as a laptop are the same types of people who can just as easily get by using a ChromeBook. The iPad Pro is a secondary device where as the Surface Pro is a primary device.

    IOS is just to limited in it's capabilities to be even considered as a standalone professional device. No, an Architect would not use the iPad Pro to design an house with, the Architect might use one to show off the plans to a client, mark down corrections with the Pencil. No, a programmer would not use the iPad Pro to develope on, he however might use one to create an outline of what needs to be done or even as a second monitor for his laptop. No, a musician would not use one to create his album with, he might use one as the brain for one of his synthesizer, a recording device, instrument, etc. it's a companion device, it's an iPad, nothing wrong with that but stop trying to convince everyone that it's some super computer with fantasy powers. Just the file system issue alone should be enough to tell you that the iPad Pro isn't really meant for connect creation.
    Reply
  • FunBunny2 - Friday, January 22, 2016 - link

    -- Modern software is very bloated memory consumption wise, especially software relying on managed languages, the latter are also significantly slower in terms of performance than languages like C or C++.

    well, that was true back in the days of DOS. since windoze, 80% (or thereabouts) of programs just call windoze syscalls, which is largely C++.
    Reply

Log in

Don't have an account? Sign up now