Apple A14 and A14X SoCs

[wco81]

Someone posted benchmarks showing that the Snapdragon 865 beat the A14 in some areas and the A14 overall wasn't much faster than the A13. Then the claim that Snapdragon 875 should leave A14 well behind.

https://www.notebookcheck.net/Leake...ass-the-Qualcomm-Snapdragon-865.494548.0.html

 

[Ailuros]

I can't be bothered to count how often the author warns in the above writeup that results should be taken with a grain of salt.

 

[snc]

https://twitter.com/x/status/1312254432571080705

 

[Entropy]

Which is in line with previous leaks/rumors, and for that matter straight extrapolation from A12 scores given Apples presentation. What is still up in the air is memory technology. A shift to LPDDR5 should help graphics performance in particular.

The major silicon investment Apple has done with the A14 lie in beefing up their machine learning blocks, and that isn’t reflected in Geekbench at all. (Nor is memory bandwidth, which for SoCs is particularly unfortunate.)

 

[Entropy]

Compute score is higher than expected (12571). A12z is 11628. Hmm. We'll see what's behind that, if accurate.

 

[Pressure]

It matches the 8-core GPU in the A12Z with only a 4-core GPU.

 

[Entropy]

Yes, and I immediately got suspicious. Probably says more about my personality than anything else though, because I doubt a doctored benchmark would show oddities there, when simply scaling the A13 + add noise would be so much more believable.

I made a link to the subtest scores compared to the A13 here.
I can’t say what this reveals regarding underlying changes.

 

[snc]

I expected bigger memory points difference

 

[snc]

 

[wco81]

Are there any benchmarks which track the performance of the "neural engine" from generation to generation?

Seems like Apple is touting it more and more every year. Does that translate to them devoting more silicon budget to it every year like they do the other cores?

Since they're also hyping the computational photography modes too, it would be interesting if there's a way to quantify those changes.

 

[PSman1700]

So, how good is this compared to say the latest SD865+ found in the S20 ultra etc? Observe im kinda new to all this mobile SoC stuff. Many say Iphone 11 is faster then anything else, but many benchmarks point to Snapdragon being faster, while some show ip11 being much and much faster.

 

[pcchen]

So, how good is this compared to say the latest SD865+ found in the S20 ultra etc? Observe im kinda new to all this mobile SoC stuff. Many say Iphone 11 is faster then anything else, but many benchmarks point to Snapdragon being faster, while some show ip11 being much and much faster.

It can be quite difficult comparing benchmarks between different OS. Geekbench is probably the more famous cross-platform benchmark, with reasonable transparency on what it's testing. Its test sets are also from widely used algorithms.
It's much harder to find serious CPU benchmarks unfortunately. The best we have is probably the SPEC CPU tests by Andrei from AnandTech.
There are also various Javascript benchmarks, which is fairly depending on which browser you use, and unfortunately there's no common Javascript engine available on both platform for better comparison.

 

[Entropy]

Are there any benchmarks which track the performance of the "neural engine" from generation to generation?

Seems like Apple is touting it more and more every year. Does that translate to them devoting more silicon budget to it every year like they do the other cores?

Since they're also hyping the computational photography modes too, it would be interesting if there's a way to quantify those changes.

I don’t know of any such benchmarks on iOS. They’d have to use CoreML, obviously, so some tinkering would be necessary if you wanted to do cross platform compares.

If I were to create an application benchmark, it would have to use an app that had at least a free demo, and then make available image files or video files that required at least a minute execution time on the faster hardware for reasonably accurate timings. That’s not super easy to achieve since most processing is quite fast, and capabilities of the Neural Engines already span a factor twenty (!!).

A better alternative may be to make a really simplistic synthetic benchmark, with the usual caveats regarding relevance. There are a number of industry standard ML benchmarks, but again, they’d need to call CoreML.

Benchmarking is difficult. The Neural Engine and matrix block of the A14 are as large as CPU+GPU, but don’t show up in either Geekbench or SPEC.

 

[Laurent06]

So, how good is this compared to say the latest SD865+ found in the S20 ultra etc? Observe im kinda new to all this mobile SoC stuff

https://browser.geekbench.com/v5/cpu/compare/4208577?baseline=4183824

Note SD865+ will likely be replaced by something faster early next year so the gap will reduce.

Many say Iphone 11 is faster then anything else, but many benchmarks point to Snapdragon being faster, while some show ip11 being much and much faster.

From a CPU point of view I don't think there's much to discuss: Apple currently has the lead. For the rest of the SoC and the phone, including software, that's another discussion.

 

[PSman1700]

Apple is much and much faster yes, SD might reduce it abit but i dont think they can even come close. Apple also uses NVME whereas android usually UFS.

 

[Rootax]

Very impressive. But kind of useless too for most users ? I mean in what app will you feel the difference in power ?

 

[pcchen]

Very impressive. But kind of useless too for most users ? I mean in what app will you feel the difference in power ?

A lot, actually. For example, the computations done by a modern photo app is quite a lot. With more computation power, you'll be able to take more photos, or even doing image processing when taking video. These weren't possible with slower CPU.
AR is another example. Those effects in Snapchat do require a lot of processing power. With faster CPU it's possible to make more accurate AR effects. It's also now possible to do real time language translation from an photo directly on the device, which also needs a fast CPU.

 

[Rootax]

Processing videos, yes.


But photos, Google step back with the Pixel 5 because a 765G was good enough, and I agree. 1 sec instead of 0.x is not a big deal for most people. And serious photographers have other devices anyway ?

About translation/image/text processing, Google Lens is using the cloud or local cpu ?

But with that said, I kind of get it now, and anyway more processing power will create new ways to do existing things, and create new ones.

 

[PSman1700]

Very impressive. But kind of useless too for most users ? I mean in what app will you feel the difference in power ?

Not now maybe for most users, but it sure does after many years, the iphones wont slow down as much in games etc.
For the A7/iphone 5s this isnt really true though, did a test vs a SD800 equiped phone (xperia z1 compact), and they still trade blows. Apples chips must have gotten much better since then as compared to the competition?

Otherwise, the S20 ultra and 11 pro max basically are about as fast in most applications. Probably because both are much faster then most apps need right now.
Ive been using android for a very long time, but im more and more changing direction to apple. Nothing beats 5+ years of full system updates and 2 to 3 years of security updates after that. Also the perfect integration of software, never any problems with iphones.

 

[mavere]

Early results for A14 and 3DMark Wild Life have been around the net. Scores are slightly lower than A13, but the actual framerate timeline is weird, no? It's unreasonably even, almost like there's a bottleneck somewhere. Driver bug?