Apple A9 SoC

There has been a few rumours milling around that we won't see a new iPad Air this year, which is surprising considering that you are beginning to get the iPad Air 2 on deals at various retailers and if they delay a new iPad Air until next year then surely those deals will get bigger and better. Considering the iPad Mini 3 was effectively just an iPad Mini 2 with a fingerprint home button, they certainly are in need of a refresh, although I'm not sure whether they will put an A8, A8x, A9 or A9x in it? I would be surprised if they just put an A8 in it, as I would expect at least the A8x or A9 to be housed in it. There have been benchmark results (whether fake or not) posted that show A9 and A9x results and if they are in fact genuine, then I wouldn't expect an iPad Pro/iPad Plus to have those results from the supposed A9x. Considering it's to run a higher resolution screen and run "PRO" applications.

Maybe we are thinking this all wrong? Maybe the supposed iPad Pro/iPad Plus will contain and Intel chip similar to the new 12" Macbook and will actually be a Macbook/iPad-cross? This is where I see things going.....eventually, and maybe it will happen this year, or at least be announced at the iPad event.

Very possibily, Apple is making the iPad move towards a true desktop/laptop environment with the ability to utilise iOS and OSX?

 

Maybe I'm getting old, but I find the A8/A8x's CPU & GPU performance to be more than adequate for 99.9% of mobile workloads, so even if the move to 14nm just brings about a minor bump in clockspeed, I'd be a happy punter.

What I really hope Apple delivers is a UFS 2.0 NAND interface, which will remove one of the biggest bottlenecks of mobile devices. A transition to LPDDR4 with its reported lower power consumption will offset the hopefully larger RAM allocation (2 GB), and boost memory bandwidth to 25.6 GB/s in the A9, and a chunky 51 GB/s in the A9X if they keep the 128-bit interface, which is close to low-end discreet GPUs

 

The iPad "Pro" can always be announced a few months earlier than its official release. That being said, I can't think of any particularly good reason for a September announcement over an October one, and I think there are a few good reasons for an October announcement. The October event is likely to have fewer announcements than the September event (ignoring the iPad "Pro"), or at least fewer "high-profile" ones. It may also be more "pro"-oriented with potential Retina iMac updates, which could better suit an iPad "Pro."

 

I'm willing to bet that I'm completely wrong again, but I don't expect more than an 4 cluster 7XT for the A9 and an 8 cluster 7XT GPU for the A9X. Assuming they manage to add another ~50% in something like Gfxbench3.0/Metal that'd be ~30+fps for the first and ~60+fps for the latter (offscreen 1080p always..).

 

Looks like I spoke too soon.

One advantage of an earlier announcement that I didn't think of earlier is that developers will have more time to create and modify apps for the iPad Pro. I think this factor will be particularly important if the iPad Pro is targeted at different markets than the other iPads.

9to5Mac also claims that there won't be an iPad Air update at all until next year. If that's indeed the case, then my guess is that the A9X is designed with essentially just the iPad Pro in mind.

 

Interesting. Could be related to volumes and yields in FF. Ship an expensive, premium tier device now that will probably have lower volumes and refresh the volume line next year when component supply is better.

 

Recently I have seen the Air 2 being sold for as cheaply as £265 here in the UK compared to the RRP of £399, if there isn't a refresh this year, then I fully expect prices to remain low or sales to slow even further. By all intents, the iPad Mini 3 didn't sell well because of the lack of additional features and I eventually picked one up as it was only £10 more expensive than the Mini 2 (which I thought was worth it for the touchID). Also if the new larger iPad is gonna be called iPad Pro, then it will be for the more professional and business customer and will more than likely cost considerably more money (probably starting at £599+), then the mainstream iPad's (Mini and Air) will be relatively low powered compared to the competition surely? As expected the Mini 4 will get an A8 upgrade from the A7. Unless, Apple is looking at reducing the starting price of both the Mini and Air?

Could we be seeing a significant drop in RRP for both....that would make sense, then if they let both models run on a year old chip and be, effectively less powerful than the current iPhone range?

Mini 4 starting at £249 and Air 2 starting at £299 with the Mini 3 starting at £199 and Air @ £249 and Mini 2 dropping out.

Could this happen?

 

From WCCFTech: "iPhone 6s Benchmark And Resolution Details Leaked; Future Hardware Does Not Stand A Chance."

I wouldn't have linked it if it weren't for my observation that the single-core and multi-core Geekbench 3 scores actually seem reasonable for once given the claimed clock speed.

Code:

                  Clock    Single-core  Multi-core

A8 (iPhone 6)     1.4 GHz  1610         2883
A9                1.8 GHz  2248         4036

Ratio             1.29x    1.40x        1.40x
Per-clock ratio            1.09x        1.09x

Also, the display resolutions are supposedly increasing.

 

That'd be a ~10% IPC increase, solid but not what I call impressive Also note there might a large distortion in the memory score of Geekbench if memory went from LPDDR3 to LPDDR4.

That's exactly what I thought too... OTOH it looks like the ratios are almost identical for A7 to A8.

 

Or they boosted crypto hash instructions, or they added a better image compression/decompression block. Geekbench has a fairly low correlation coefficient with the real world.

Cheers

 

At least its correlation is better than SPEC 2006 quantum when compiled by Intel SPEC compiler

Joke aside I agree there's too much of crypto thing, but OTOH compression/decompression isn't accelerated.

 

SPEC 2006 has alas a drawback: since it's there to measure both CPU and compiler performance, compiler writers have targeted it too much. Just look at libquantum results for any icc (Intel compiler) result. This reminds me of AnTuTu being broken due to the use of icc (again...) which has detection for one of the nbench tests (which AnTuTu uses).

IMHO apart from 403.gcc, SPEC has become useless to get a hint for end-user type of application.

I agree, let's wait and see where the 9%, if real, come from

 

That's also the case for the A6 to A7. I'm actually a bit more suspicious of the A9 MC score being 1.400x the A8 MC score, literally to the nearest integer. But maybe that was just by luck (real scores or not).

Code:

                 Clock    Single-core  Multi-core

A5 (iPhone 4S)   800 MHz          215         405
A6 (iPhone 5)    1.3 GHz          708        1271
A7 (iPhone 5S)   1.3 GHz         1396        2514
A8 (iPhone 6)    1.4 GHz         1610        2883
A9               1.8 GHz         2248        4036

A5-to-A6 ratio     1.63         3.293       3.138   0.953
   (per clock)                  2.03        1.93

A6-to-A7 ratio     1.00         1.972       1.978   1.003
   (per clock)                  1.97        1.98

A7-to-A8 ratio     1.08         1.153       1.147   0.994
   (per clock)                  1.07        1.06

A8-to-A9 ratio     1.29         1.396       1.400   1.003
   (per clock)                  1.09        1.09

For each n, 5 ≤ n ≤ 8, the last column takes the multi-core An to A(n+1) ratio and divides it by the single-core An to A(n+1) ratio.

 

For the sake of getting my hopes/expectations written out, I'm expecting a bigger performance jump for the A9 vs the A8 compared to the A8 vs A7. Part of it is assuming that "S" iPhone refreshes are still focused on speed and part of it based on evidence in the developer previews that the next iPad Mini will include full multitasking support including Split View which I presume will mean the use of an A9 with performance similar to the A8X. Apple promoted the A8 vs A7 as having a up to 25% CPU and up to 50% faster GPU and I'm expecting a 50% faster CPU and 2x faster GPU in the A9 vs the A8.

I'm still expecting a dual core and the CPU performance gains would come from both clock speeds, perhaps 1.6-1.8 GHz, and architectural improvements. At WWDC this year Apple introduced Bitcode allowing apps to be stored as LLVM intermediates on Apple's servers so that they can always be compiled with the latest compiler on distribution to users. They said one of the motivations for this is to allow apps to automatically take advantage of new processor capabilities they might be adding. I suppose that want to avoid the flip-flop situation where the armv7s target for Swift (A6/A6X) was added as and then later removed as a default target in Xcode. Seeing Apple is making all iOS 9 apps use Bitcode by default the A9 may well come with an evolved ISA and bigger microarchitecture changes (relative to Typhoon vs Cyclone although I'm not expecting a complete redesign), perhaps armv8.1 which ARM said would see silicon by early adopters later in 2015 or maybe Apple's own set of customizations.

For the GPU, I'm expecting a PowerVR GT7600. A more aggressive 6 cluster configuration vs a 4 cluster configuration better suits the iPad Mini 4 and would also give a useful performance boost on a per pixel basis for the iPhone which didn't see much improvement in the iPhone 6/6 Plus generation due to the increased resolution.

Memory will hopefully be 2 GB LPDDR4, which is probably the most important requirement for Split View in the iPad Mini 4. It would be interesting to see an increase in the L3 cache if it'll allow them to more aggressively power down the memory controller.

For the A9X, I'm expecting a 1.8-2.0 GHz tri-core CPU with a 2 x GT7600 implementation. Higher clocked in the iPad Pro and lower clocked in the iPad Air 3. Perhaps 4 GB of memory in the iPad Pro and 2 GB of memory in the iPad Air 3.

https://www.semiwiki.com/forum/content/4824-tsmc-apple-update-q2-2015-a.html

If Semi-Wiki is to be believed, that the A9 will be Samsung 14 nm and TSMC will get the A9X on 16 nm FF+, then I'd think there must be an iPad Air 3 in order to justify the volumes to not only make a A9X, but put it on a different process, since an iPad Pro isn't likely to do it alone and there's no indication that the new Apple TV will use the A9X even though it's supposed to be gaming focused. Given the lack of iPad Air 3 rumours, perhaps it'll just be a low-key refresh like the iPad Mini 3, in this case with just an SoC update, or perhaps it'll launch in spring 2016 as an addendum to the inevitable Apple Watch update.

As to the rumours above of higher resolution screens for both iPhones, I hope Apple isn't going to play the dpi game and waste GPU performance driving higher resolution screens that don't provide much user benefit. I don't believe there have been many complaints about the 401 dpi iPhone 6 Plus screen and working from 3x graphics before scaling down seems to have alleviated issues from non-native scaling. I could see them switching to 3x graphics scaling and 401 dpi displays in the iPhone 7 if only to simplify their supply chain, but native 3x displays in the iPhone 7 and 7 Plus seems like a waste.

 

While I respect the above speculation, I don't necessarily agree with the latter. In T-Rex offscreen the iPhone 6 Plus ends up at 48,30 fps, while the iPhone5S is at 27,60 fps. Given how much impact z/stencil has in that test (amongst others) a 75% increase in performance obviously cannot come without any serious back-end related improvements. I can sense that you might be referring to texel fillrates, but if you think that 8 TMUs at 533MHz aren't enough to drive a modern UI in a relatively modest resolution like 1080p you're more than wrong IMHO. In a contrary case the iPad Air (1) GPU at 2048*1536 would be in "serious trouble" even for the UI to start with. Even more so an even more humble G6230 at the same frequency I'm having on the tablet at 1536p.

Any quad cluster Rogue could easily handle way higher resolutions than the above; whether Apple, their marketing or whatever else considers it a necessity or not is besides the point. On a pure technical level at least to think of a supposed necessity for a crapload of more TMUs is absurd IMHO.

As I said above I'm most likely on the wrong foot again with my own speculation, however if Apple would want to stay borderline competitive in terms of GPU performance 4 S7XT clusters are enough. Anything above that (as long as consumption stays reasonable) is always welcome.