Apple iPhone 7 and 7 Plus are here

[astilo]

First of all, I would like to start with a nice comparison table borrowed from Anandtech:

View attachment 18132

the two phones incorporate a number of new features and feature upgrades compared to their immediate predecessors. This includes a new SoC – the A10 Fusion – new camera options, IP67 water resistance, and top-to-bottom support for both wide color gamut (DCI-P3) displays and photography

Let me say one thing. I personally find the small iPhone a joke for 650$. Yes I can put a Ferrari engine into a Fiat 500, but that doesn't mean I own now a Ferrari.
Seriously, that thing still has a 4.7-inch 1334 x 750 IPS LCD display. Thanks God, they increased at least the minimum storage up to 32GB. Most people would probably answer that it is enough, but if you only have to use it to txt and call, why do you need to buy a top range smartphone?
The Plus, makes already much more sense, I would say.

Anyway, what always drives my interest is of course the SoC. What a masterpiece. Chapeau Apple.

The A10 implements two types of cores; a pair of high performance cores, and a pair of low power cores. The slower cores are said to consume one-fifth the power of the high performance cores, though Apple has not indicated what performance is like. As is traditional for Apple, they haven’t said anything about the CPU cores themselves, but it’s a safe bet that the high performance cores are a direct descendant of the Twister cores used in the A9. More curious will be what the low-power cores are – given Apple’s fondness for developing their own ARM CPU cores and various technical considerations (such as the core interfaces), it may very well be that these are also Apple-designed cores, as opposed to an off-the-shelf solution like ARM’s Cortex-A53

Let's focus on the high performance cores, said to be 40% better than the A9 and its dual 1.85GHz Twister CPU cores.

View attachment 18135

I already spotted a couple of benchmarks on Geekbench, and based on the Apple's claim, they might definitely be genuine.

View attachment 18133

The Apple iPhone 6s plus single score was around 2400, so if we simply do the math, 2400*1.4=3360, and there we are.
Guys, this is a monster score, I'm truly impressed, moreover considering that the A10 is built on the same technology node as the A9, the 16nmFF+ (possible further optimized, but definitely with the same CDs).

To put this into a better contest:
View attachment 18134

Of course, there is nothing in the Android environment that can barely much such a terrific performance.
Truth to be told, Qualcomm and Samsung are more focused on multi core SoCs, so the core area is smaller than the Apple one, nevertheless, it would take them a couple of years at the very best to be in the same range of performance.

I would like to add a couple of considerations here.
In order to support such beasts, Apple had to add 2 small companion cores for the low computational tasks. I'm very curious to see if the high perf cores can sustain the loading for a long period without scaling down the frequency. They said that it was to improve the battery life, that of course it is also true, but I guess it was a kind of forced move, to avoid overheating issues. If you tweak the cores to achieve higher frequencies, probably they perform very bad as soon as you have to reduce the voltage, in terms of performance per power. This is probably also one of the reason why in the previous generation iPhones, the peak frequency was never too high.

The second point is indeed about the CPU clock. I truly believe that the new cores are not so different from the Apple A9 twister ones in terms of architecture. We also know that the tech node is the same, so how the hell they have been able to squeeze a 40% more performance from them?
Let's do the again the math. The twister core on the iPhone 6s was clocked at 1.85GHz, assuming the 40% increase in performance, we should end up having a 1.85*1.4=2.59 GHz clock for the A10.
This is indeed my guess, ~2.5-2.6 Ghz for the new high performance cores.

To conclude my quick SoC review, let's also briefly talk about the GPU.
Apple has confirmed that the A10 implements a “6 cluster” GPU design, most likely another PowerVR design, either a higher clocked version of the PowerVR GT7600 used in the A9, or perhaps a six cluster design based on the more recent PowerVR Series7XT Plus architecture. It is expected to provide a 50% higher performance than the A9's GPU. Again, wonderful.

Final comments
I was expecting even less from Apple about this new iPhone 7. I'm still quite disappointed overall (moreover because of the non plus version features), but when I look inside the iPhone, Apple was once more able to really impress me, and probably also the whole market, with a superb SoC design. The A10 is probably the only reason why I would ever buy the new iPhone.

 

[Speedfriend]

If the 40% performance improvement is a on geekbench single core, then it is truly incredible increase given no change in node. It is possible that they are talking about multicore, where 4022 *1.4 =5630 and where the companion core may boost the score?

Looking at the Kaby Lake Core M score, the i7-7Y75 4.5W gets SC 2801 MC 5701 and the i7-7Y75 7W gets SC 3375 MC 6799. So a smartphone chip has already surpassed Intel best tablet chip.

Having said that, the Kaby Lake benchmark look very good across the board, with the i7-7Y75 7W matching or outperforming the Surface Pro 4 i5.

 

[astilo]

Both the singlecore and multicore scores, have the same scaling factor, between 1.3 and 1.4, depending on what reference benchmark you consider (for the 6s we can go from 2400 to 2500 in the single core one). That means that the companion cores, either have really poor performance or they cannot simply be used at the same time with the high performance cores. My guess is that they are no real companion cores, and the SoC switches between one couple or the other depending on the computational loading requested.
Yes, I have seen those Intel numbers and that is the reason why I was so impressed by this A10 SoC.
Can you imagine what is going to be the A10X in the next iPad?

The 2.5GHz frequency estimation I did about the A10, it is of course the theoretical maximum we can expect from this chip, otherwise we would have seen a greater than 40% boost over the previous generation.
Most likely, if I do a scaling factor vs the top iPhone 6s scores, assuming no other improvements, the high perf cores should be clocked around 2.35GHz. Let´s wait and see.

 

[lilo777]

If the 40% performance improvement is a on geekbench single core, then it is truly incredible increase given no change in node. It is possible that they are talking about multicore, where 4022 *1.4 =5630 and where the companion core may boost the score?

Looking at the Kaby Lake Core M score, the i7-7Y75 4.5W gets SC 2801 MC 5701 and the i7-7Y75 7W gets SC 3375 MC 6799. So a smartphone chip has already surpassed Intel best tablet chip.

Having said that, the Kaby Lake benchmark look very good across the board, with the i7-7Y75 7W matching or outperforming the Surface Pro 4 i5.

I believe I read somewhere that Geekbench benchmarks for desktops and smartphones are different and comparing the scores from the two does not make sense.

 

[astilo]

I believe I read somewhere that Geekbench benchmarks for desktops and smartphones are different and comparing the scores from the two does not make sense.

Maybe and maybe not. It is supposed to provide comparable results across platforms, but of course the code must also run on different OS, and some minor differences might be expected because of that.

Geekbench 4[FONT=&quot] scores are calibrated against a baseline score of 4000 (which is the score of an Intel Core i7-6600U). Higher scores are better, with double the score indicating double the performance.[/FONT]

 

[name99]

Both the singlecore and multicore scores, have the same scaling factor, between 1.3 and 1.4, depending on what reference benchmark you consider (for the 6s we can go from 2400 to 2500 in the single core one). That means that the companion cores, either have really poor performance or they cannot simply be used at the same time with the high performance cores. My guess is that they are no real companion cores, and the SoC switches between one couple or the other depending on the computational loading requested.

It's also possible that
- the companion cores CAN run at the same time as the primary core, but Apple couldn't get the scheduler for that optimized in time for iOS 10.0, but it will come with iOS 10.1 OR
- they were forced to make some compromises to work around some bug in the fast.SLOW core switching for the A9, but those bugs will be squashed by the time the A10X ships.

Ming-Chi Kuo, who did pretty damn well in his A10/iPhone7 predictions, reckons A10X (and so new iPads) will ship only in 2017, and on TSMC 10nm. That provides time for the (possible) issues I described to be fixed, and allows for A10X to run (assuming reasonable scaling) at something like 3.4GHz.

I also suspect that, at some point, all those cores will run simultaneously simply because of the way Apple announced this as a quad-core CPU. These Apple events are very public, and speak to a totally ignorant population that includes a litigious fraction eager to sue for any stupid reason. If they don't ship a version of the system that can, one way or another, plausibly claim to be running 4 threads simultaneously, there will be some form of stupid quad-core-gate, with the attendant bad publicity, and some class action lawsuit of people claiming to be tremendously hurt by the fact that they really really thought they were buying the ability to run four threads, and Apple lied to them about this. (Look at the AMD lawsuit a year ago...)

Apple did not have to use the language they did, of quad-core. They could have used more technical languages (sounds more impressive, and doesn't lead to lawsuits) of something like "We've introduced a heterogeneous core that's split into two parts, one that can go very fast, one that uses very little power". To say "we've introduced a four-core CPU" is to make a statement that I think they have to back up.

My personal hypothesis is that writing an OS scheduler that handles all the possible cases well is not easy, and that Apple's first attempts led to occasional UI glitching. It is much easier to write a scheduler that just migrates from fast to slow as necessary, and that we'll see the better scheduler in a few months. ARM went down the same path, with the first schedulers ONLY running a big or a LITTLE core and toggling between the two; but after some experimentation and thought, that restriction was relaxed.

 

[name99]

I believe I read somewhere that Geekbench benchmarks for desktops and smartphones are different and comparing the scores from the two does not make sense.

That's no longer true for GeekBench 4.
For Geekbench 3 it was true -- if you want to deny reality. The workloads were different (to match the actual RAM available on early phones, and to not run the benchmark for too long) but the scores were supposed to be scaled so as to be comparable. You can imagine how different tribes interpreted this fact; but the GeekBench 4 results make it clear that the GeekBench 3 scaling was pretty damn correct.