Apple A11 Bionic performance

[Verne Arase]

Unfortunately it is not the case anymore. iPhone 6s was smooth, iPhone 7 and now iPhone 8 are not smooth, they both have noticeable random frame drops on animations.

This is something phone specific though, may be 3D touch, because my iPad 10.5 Pro is much better in this regard: it can drop frames, but it only happens at the beginning (and, in addition, calling any 60Hz device "buttery smooth" looks like a joke after 120Hz of the iPad).

I think a lot of that buttery smoothness is the OS taking lists and imaging them into pixmaps and windowing them into the view regions of the scrollable panes on your display.

It's a tricky business, because for long lists it would be impossible to image all the list items into the pixmap (memory), and as you progress through the scroll you have to lop off portions which are out of view and add new rows realtime while the user is rapidly running his/her thumb over the scroll viewport.

A lot of that is tuning, especially with a newish OS and older devices. Expect better from 11.1.

 

[DrMrLordX]

Who'd have guessed back in the day that we'd have portable phones which had multiprocessing operating systems, and that these pocket devices would have the same problems with process flow as all that big iron ... and that some OS developers of those devices would fall into the same pitfalls as some of those early systems of yore?

I don't have the chops to discuss the pitfalls of early "big iron" systems and their problems, but I would think that most of those old systems - at least in MP configurations - had uniform processor distribution. That is, every CPU was basically the same as every other CPU. Or were there exceptions?

big.LITTLE chips seem to make things even more complicated, because you have your "fast" cores and your "slow" cores, and keeping everything running smoothly while dynamically switching execution between the core groups depending on TDP limits and battery level and other nonsense has got to be a headache.

 

[urvile]

Well. I didn't even know it was possible to benchmark mobile devices and to tell you the truth I don't really care.

Although it would be awesome to have a single mobile device that can be used like a pc for email, word processing etc.

 

[Verne Arase]

I don't have the chops to discuss the pitfalls of early "big iron" systems and their problems, but I would think that most of those old systems - at least in MP configurations - had uniform processor distribution. That is, every CPU was basically the same as every other CPU. Or were there exceptions?

Yeah, they were completely symmetrical.

big.LITTLE chips seem to make things even more complicated, because you have your "fast" cores and your "slow" cores, and keeping everything running smoothly while dynamically switching execution between the core groups depending on TDP limits and battery level and other nonsense has got to be a headache.

Yeah, unfortunately I don't know how the Apple dispatcher works.

My initial thought was that when a process came up in the ready queue that they'd give it a lightweight engine and step back and monitor what kind of utilization there was, and add additional resources should that prove insufficient.

But really, dispatching is probably done on an entire system basis, with the processor controller lighting up engines whenever something comes up in the ready queue, starting with the high efficiency engines, and proceeding upward through the big boys if needed to satisfy the processing load. Once an engine completes its task, if it remains idle long enough the dispatcher powers it down to save power (starting with the high power engines).

Really, the objective would be to minimize power use while satisfying the workload. This sounds crazy complicated but in a real time world, it may be self-regulating since large heavy threads may spend more time on their engines and ephemeral tasks would flit in and out as they were completed. This raises all kinds of questions like what are the internal channel paths like, and can I/O be initiated from any engine? Can I/O completion be seen by all engines?

Really, I'm not sure how critical it is to keep these processors "hot" every microsecond since we're no longer talking about a multi-million dollar investment; with faster MVNe storage, you might even want to loop on I/O completion to avoid adding complexity to the OS dispatcher and speed up the I/O without all the attendant problems of processing an asynchronous I/O completion.

Stuff like this is the "secret sauce" of the OS - in the low level guts of MVS's IOS, I can recall a 10 retry counter being used before reporting an I/O error .

 

[lobz]

Beginning of the end of x86. Well, maybe the beginning was with the original iPhone.

Bah, this nonsense again? ...

 

[Av9114]

I think its pretty certain a team as capable as they one they have realizes problems with CPU utilization. Regardless, they have a even better solution than SMT to improve multi-core performance.

The 4x Mistral low power cores work like extra logical threads in an SMT-enabled CPU. The fact that they made such a setup work is amazing. I assume because it uses entirely separate cores it would likely even avoid performance degradation due to resource contention as with SMT when the cores are fully utilized. Plus, they can even use those cores to lower power use during light-usage scenarios above and beyond(or is it below ) what Intel/AMD chips can do.

Yes, although I would think that big.little is not without it's own set of challenges/tradeoffs. But the A11 is the first design with a true big.little design (the A10 small cores weren't independent and seems like they may have shared resources with the big cores), so I still find it somewhat interesting that they choose not to implement SMT on earlier designs. I'm not making a judgement as to whether that is good or bad, just that it strikes me as interesting. It's actually probably simply a power issue, phones probably can't run two big cores at full bore for extended periods so there's no real point in adding more transistors for SMT.

If I remember correctly 2 big plus for 4 little is what Anand/Brian Klug were clamoring for ever since the introduction of big.little (I seem to recall a lot of rants about 4+4 in mobile).

Skill/Money/Tool are just a few things that can influence success. I think there's something just as important. You can have the smartest engineers, with best tools and funding, but they won't do much with it if they are in a depressed state, or no confidence in the future or for the place they are working. Apple is pretty much THE darling of the current tech industry consisting of mobile devices.

Agreed, if I we're a hardware guy the prospect of being able to design a chip that is basically a one off design for one product would be pretty appealing. Apple doesn't even share SoCs between iPad and iPhone anymore. Basically you can see precisely how far you can push things with a given power and maybe transistor budget.

 

[Eug]

Every time Apple releases their new AXX CPU it gets this attention and I always ask is there something else than geekbench?
Never got an answer.

Video editing bench? excel calculation? Large word or pdf work?

Cinebench?

Can it run crysis?

iMovie exports are reasonably fast on iOS, but it would be hard to compare since hardware encoding may be utilized.

As for Excel, I've long since said that for average business type spreadsheet usage, iPads are perfectly fine for performance. The main problem of using MS Excel on an iPad is NOT the performance. The main problem is the interface. It's a bit PITA to use Excel on an iPad. This is iOS's fault, not Microsoft's IMO. However, if they were to fix the interface to support this type of usage better, I betcha the performance of Excel on an iPad would be fine for 95% of the population with the current Axx chips.

BTW, slightly off topic but it's interesting you mention Geekbench vs Cinebench. I have run Cinebench sequentially on my 2017 MacBook m3-7Y32 vs other MacBooks of the same year, and it seems the performance of the m3 is very similar to the i5 and is pretty close to the i7 as well. The MacBooks are limited by heat since there is no fan, so running Cinebench repeatedly recapitulates what sustained load will do to them. After 10 runs (which is around 25 mins), my m3 actually beats the i5 and gets even closer to the i7, just a few % slower than the i7.

In my experience, Axx behaves the same way, not surprisingly. With software-based HEVC playback, the iPhone will start to overheat after say 10 minutes and the chip will throttle.

Why do I bring this up? Because the performance difference between the 2017 MacBook models is much bigger with Geekbench than it is with Cinebench. Geekbench's issue here is IMO the fact that it doesn't run for long enough and doesn't heat up the CPUs enough. So what you get is a measure of burst speed, not sustained performance. For a phone, that's actually often quite OK, but for other usage, it is more of a limitation for performance comparisons.

 

[cipurian]

This is why I wonder why we even use separate computers for basic work. Why is it that I can't just dock my iPhone into an iCase and use it like a computer. Now, I know this isn't new as ASUS did it years ago with the PadFone, and we have features that try to mimic this like Microsoft's Continuum. However, there's usually some sort of drawback. I'd be interested to see Apple's take on it as they usually focus heavily on polish.

Even though it failed I think ubuntu's unity 8 was best concept of phone and desktop integration yet.

However not that long ago there we reports of Apple filing for patent, that looks to be docking station/dummy laptop for iphone. It's supposed to have extra battery and gpu power for better performance, so idea is out there and good one I might add. However that would require iOS and MacOS convergence and I doubt that will happen any time soon.

Google just launched pixelbook with full android app support and if it proves to provide fully fledged desktop experience maybe there could potential with android-chrome hybrid as well.

 

[Zony Yu]

I am curious: why does ARM avoid developing a fast architecture like Apple's? Fewer-and-faster, instead of 8 'slow' cores, ex. 4 Cortex-A73 + 4 Cortex-A53. Or is Apple A6's introduction in 2012 still too soon?

A likely explanation is due to the fact that ARM, as well as Qualcomm are chip companies, therefore they make money selling chips. They cannot price the hardware too high otherwise the companies will look elsewhere. In comparison, Apple sells whole devices, so the profit margins are much higher, therefore they can invest more in advanced chip designs.

Its also worth noting that a majority of android phones are mid-rangers/budget phones, so that is why the chip makers have a bigger focus on the mid-range chipsets and budget chipsets.

 

[Phynaz]

A likely explanation is due to the fact that ARM, as well as Qualcomm are chip companies, therefore they make money selling chips.

ARM doesn't sell chips.

 

[Zony Yu]

Well I know that the a10x in my iPad Pro is much faster in Adobe Lightroom compared to the i5 broadwell in the 2015 MacBook Pro (about 1.3-1.5x in raw export and more than that in applying/previewing filters)

It is also much faster in exporting edited 4K video. About 3x faster. But I believe the video export is bound by the video encode dsp inside the a10x not the cpu.

the hardware enabled HEVC decoder does play quite a big role in aiding video export. im just not sure if it takes up its own chunk of silicon, or its integrated within the architecture of the GPU

 

[Zony Yu]

They make very impressive CPU's, no one can deny that. I am not a fan of Apple, I don't really like a lot of their business practices (dongles for dongles) but no denying they know how to make a good chip. Those new Ipad Pros have amazing performance.


I have no idea how well the A11 or A10X scales, but imagine a 95W desktop class CPU of that.

there definately wont be as big of a performance jump as you might think. as the performance increases linearly, the power draw increases exponentially. its a good example of the laws of diminishing returns

 

[Zony Yu]

Apple A11 Performance Review with the iPhone 8 Plus: Taking on Desktop?

https://www.pcper.com/reviews/Processors/Apple-A11-Performance-Review-iPhone-8-Plus-Taking-Desktop

"The fastest smartphone CPU even beats a current-gen Intel mobile part (in Geekbench, anyhow). The A11 Bionic is slightly higher in single-threaded performance, with the large multi-threaded advantage clearly demonstrating all six cores at work (compared to the four threads of the Core i5)."

"The A11 Bionic provides significantly higher graphics performance to edge out the i5-7300U (Intel HD Graphics 620) in the overall score, though the physics test shows a clear advantage from Intel. The Snapdragon 835 lags far behind in all categories here. Clearly a new standard for mobile platforms has been set by Apple, and it will be interesting to see how the industry responds"

Apple's chip design prowess is outstanding. Its no surprise that Apple is looking to move its Macbook lineup to their custom A series. Intel simply cannot keep with this juggernaut.

https://asia.nikkei.com/Business/Companies/Apple-A-semiconductor-superpower-in-the-making?page=2

The big winner here is TSMC when Apple moves Macbooks to the custom A series chips sometime in the next 2 years.

The Android competition like Snapdragon 835 is left eating the dust in CPU performance while it holds up much better in GPU performance against the A11.

Its also worth mentioning that qualcomm is taking a graphics-centric approach for its Adreno GPU architecture, while Apple is pursuing a General Purpose GPU (GPGPU) /ML approach. What that means is that Qualcomm has put in more texture units and shader units into the GPU, allowing for better 3D graphics rendering, while Apple prioritized compute units for rapid floating point calculations.

Looking at benchmarks, the A11 still outperforms the S835 in terms of graphics processing despite the S835's graphic-centric design, as running GFXBench manhattan 3.0 1080p offscreen yielded 60fps on the S835 and 83fps on the A11 running the same load.
When it comes to GPGPU tasks such as video rendering, photo editing, 3D modelling and such, the A11 smokes the S835, delivering 15,300 points of compute power in GB4 compute compared to 7,800 achieved by Qualcomm. Do keep in mind, however, that for GB4 compute, the A11 is running on Metal, where as the S835 uses Open GL ES 3.x. Im not sure if that affects the results in any certain way.

 

[Zony Yu]

ARM doesn't sell chips.

I meant to say Architectures.

 

[Zony Yu]

Why it is faster is interesting but in the end result is the most important metric.

Lightroom is faster on an iPad Pro than the MacBook Pro

Having hardware dedicated for a certain task is much faster than running the same task on software. Thats the difference between software enabled image decoding and Hardware Enabled image decoding. Its almost like running a graphically intensive game on a CPU with software that allows the CPU to calculate floating point ops vs running it on a GPU. The MacBook Pro 2015 simply doesn't have hardware enabled image decoder, while newer models (2016 and up) supports it.