Apple A12, A12X, and A12Z SoCs

Apple has announced the A12 Bionic in the new iPhone Xs, iPhone Xs Max, and iPhone Xr during today’s September event.





The A12 Neural Engine can perform 5 trillion operations per second, up from the A11 Neural Engine’s 600 billion. (I’m watching the livestream and there was no mention of what precision was used for those numbers.)

(From the AnandTech live blog.)

 

Plus didn't they claim they use less energy on those increased operations?

 

Isn't the name XS Max a little on the nose? Maximum excess?

Are they taking the piss and saying that this version is more than necessary?

 

Yeah, Roman numerals can be hard.

 

Nobody will call it 10-ess.

 

So they're citing only 15% increase in CPU performance, 8x performance in their neural engine and 50% increase in the GPU over the A11?

 

Yes "only"

 

Well compared to the other metrics and compared to the deltas in previous years ...

 

Using Geekbench official iOS values for iPhone the delta between A7 and A8 was only +12%. So the claimed +15% would be indeed less than other increases (A11 over A10 was +23%) but not the smallest one.

A6 -> A7 : 1.68
A7 -> A8 : 1.12
A8 -> A9 : 1.70
A9 -> A10: 1.43
A10 -> A11: 1.23

It's like for everything else in this universe: nothing can indefinitely increase

Joke apart I think they are getting close to what makes sense without getting too bad efficiency.

 

I think there was some thought that whatever gains they got from 7 nm, they were using a lot of it for more power efficiency at the cost of greater performance gains.

But yes, diminishing returns each year.

The fact that they touted some of these metrics and the fact that it's the first 7 nm chip for mobile devices means that they know they have an advantage, at least for now.

They haven't always emphasized their silicon but really, they haven't come up with tentpole features to exploit this leading performance either. They showed a lot of games, which aren't going to be huge sellers because they appear to be games which will carry higher price tags.

 

I look forward to seeing some estimate of how this translates into performance consistency. There is something of a balance to be struck between race-to-sleep and being able to provide consistent performance for those cases that need it.
Be that as it may, the CPU cores doesn't seem to be the big story here, even though they are quite performant. The GPU (and to some extent the "neural engine") is where the juicy tidbits are hiding.

 

Geekbench 4 results for the A12 are starting to appear.

According to Bloomberg,

• The iPhone Xs Max is codenamed D33,
• The iPhone Xs is codenamed D32,
• The iPhone Xr is codenamed N84.

Then from the Geekbench results,

• The iPhone11,2 has motherboard D321AP and is likely to be the iPhone Xs,
• The iPhone11,6 has motherboard D331pAP and is likely to be the iPhone Xs Max,
• The iPhone11,8 has motherboard N841AP and is likely to be the iPhone Xr.

I have not seen any other iPhone11,x results in Geekbench that were uploaded after the Apple event. Taking the geometric means (I don't know what averaging method is used in the Geekbench benchmark charts) of the scores on or after September 12 gives the following numbers:

EDIT: New Geekbench scores have appeared since my post that have changed my conclusions drastically.

• A12, iPhone Xs Max: 4786 single-core, 10934 multi-core (20 results as of this post),
• A12, iPhone Xs: 4783 single-core, 10996 multi-core (9 results as of this post),
• A12, iPhone Xr: 4783 single-core, 11010 multi-core (5 results as of this post).

As with the three A11-based iPhones, the three A12-based iPhones have virtually identical Geekbench scores on average and I think it's safe to assume that they have the same clock speeds.

Code:

                                2017 iPhone GB4   2018 iPhone GB4 and change from 2017
2017 iPhone     2018 iPhone        ST      MT        ST             MT

iPhone X        iPhone Xs Max    4209   10116      4786 (+14%)   10936 (+ 8%)
iPhone X        iPhone Xs        4209   10116      4783 (+14%)   10996 (+ 9%)
iPhone 8 Plus   iPhone Xr        4218   10164      4783 (+13%)    11010 (+ 8%)
iPhone 8        iPhone Xr        4220   10107      4783 (+13%)    11010 (+ 9%)

(The A11 numbers are from the benchmark charts.)

Old numbers here:

Spoiler

• A12, iPhone Xs Max: 4773 single-core, 10745 multi-core (12 results as of this post),
• A12, iPhone Xs: 4749 single-core, 11290 multi-core (5 results as of this post),
• A12, iPhone Xr: 4754 single-core, 9367 multi-core (1 result as of this post).

The single-core scores are basically the same, but the multi-core scores for the Xs Max and Xr are 5% and 17% lower than the multi-core score of the Xs. Assuming that the lone result for the iPhone Xr is representative of Xr performance, the A12 in the Xr actually has lower multi-core performance in Geekbench compared to the three A11-based iPhones, which have nearly identical Geekbench scores.

Code:

                                2017 iPhone GB4   2018 iPhone GB4 and change from 2017
2017 iPhone     2018 iPhone        ST      MT        ST             MT

iPhone X        iPhone Xs Max    4209   10116      4773 (+13%)   10745 (+ 6%)
iPhone X        iPhone Xs        4209   10116      4749 (+13%)   11290 (+12%)
iPhone 8 Plus   iPhone Xr        4218   10164      4754 (+13%)    9367 (– 8%)
iPhone 8        iPhone Xr        4220   10107      4754 (+13%)    9367 (– 7%)

(The A11 numbers are from the benchmark charts.)

 

I wonder if Apple are doing more (too much?) aggressive clock reduction when multithreading.

 

Just my luck, four more Geekbench results for the Xr were uploaded after my post and they have multi-core scores in the ~11500 range.

With this new information I now think the A12's in all three iPhones are identical.

 

GFXbench comparison between the iPhone XS and the iPhone X. Improvements are remarkable.
Edit: And the Antutu gfx score is almost a factor two higher for the iPhone xs than for the x, reinforcing the impression that something rather major has happened in terms of capabilities. I really look forward to better data and analysis.

 

I swear I had seen more results for the A12 the other day when I first saw them. Anyway data and results are way too sparse, but I haven't the slightest clue why you are comparing A10 to A12 and not A11 to A12, where of course so far data don't show anything unexpected: https://gfxbench.com/compare.jsp?be...&hwtype2=GPU&hwname2=Apple+Inc.+Apple+A11+GPU

Looking at the other page here: https://gfxbench.com/device.jsp?ben...pi=metal&D=Apple+iPhone+XS+Max&testgroup=info and seeing leftover extensions such as those:
GL_IMG_read_format
GL_IMG_texture_compression_pvrtc
Leave a rather weird question mark as to what we are really looking at.

 

What do you mean leftover extensions? Apple couldn't possibly abandon pvr (the texture format) anytime soon. There's too much software in the wild that depends on it. In fact it's still the only compressed texture format Apple exposes with webgl.

 

The link I gave was to a comparison between the iPhone xs and the iPhone x. The iPhone x uses the A11 chip just as your A8+, not the A10 that is indicated by the Kishonti database. (Which is horrible.)

Yes, we need better data. Anyone up to taking one for the team?

 

I'm hoping Anandtech comes with an iphone review this year after skipping the it and the ipad pro last year. None of the reviews I've seen does much performance testing on the A12.