Apple A8 Benchmarked

[Eug]

I don't have an Android phone, but are the memory cards formatted as FAT32? If so, that's why. Not a hardware limitation.

Yes. Also, most flash cards come pre-formatted with FAT32, because it's so cross-compatible. Most people are putting pix and short videos on them for their phones, not 8 GB Blu-ray rips.

I really don't think this issue is as simple as it seems. It's really odd to me that, by pure coincidence, every 32 bit system I've used (very few) was limited to 4GB ram and had other weird 4GB limits (app usage or processor?). Nobody seems to agree as to why any of this happens.

Nobody seems to agree? Just about everyone has agreed in this thread. It's just that you didn't bother looking into it further.

Let's keep it simple:

FAT32 - Limited to 4 GB file sizes
Other common disk formats - Not limited to 4 GB file sizes

This has absolutely nothing to do with the OS or CPU being 32-bit or not.

However, FAT32 is ubiquitous and works (and is cheap). You're not going to find native NTFS write support on most phones, and you're not going to find exFAT support in most cameras. Nor are you going to find native ext4 support in OS X, or native HFS+ support on Android.

 

[ChronoReverse]

I really don't think this issue is as simple as it seems. It's really odd to me that, by pure coincidence, every 32 bit system I've used (very few) was limited to 4GB ram and had other weird 4GB limits (app usage or processor?). Nobody seems to agree as to why any of this happens.

That's because as a technical forum this stuff is already extremely trivial knowledge. As in "I learned this in grade-school" trivial.


I'm not trying to put you down though so I'll try to give a quick explanation:

You're right in that it's a 32bit counting limit but there are only a few things that are limited by the CPU's own bit'ness to 32bits. These are generally things that are tied very closely to the CPU's function.

The RAM limit on PC's actually isn't one them. As Rakehellion noted, PC's have had >4GB RAM capability long before 64bit CPUs. Even Windows had this capability although MS limited it to server versions (and yes I've worked with those type of systems but you could google it too I'm sure).

Simplified, the real limit is how much each process can see at once and how large a number the CPU can fundamentally (as in extremely low level) compute with. Basically (or rather, it's generally since extensions make it less hardcut), a 16bit CPU can only fit a number from 0-65,536 (unsigned) while a 32bit CPU can do 0-4,294,967,296 (unsigned) inside its registers. But you can clearly do math with numbers bigger than 4,294,967,296 on your 32bit Windows XP running on a Pentium 2 so generally these limits are only interesting to someone who is doing something extremely low-level.


In any case, a file system is so extremely high level that the CPU's bit'ness never comes into play at all. NTFS, introduced in 1993, has a theoretical file size limit of 2^64 or 16EB. So that predates x64 by over a decade!

 

[Rakehellion]

I really don't think this issue is as simple as it seems. It's really odd to me that, by pure coincidence, every 32 bit system I've used (very few) was limited to 4GB ram and had other weird 4GB limits (app usage or processor?). Nobody seems to agree as to why any of this happens.


Edit: apparently it's all a file system issue?


So no limit to ram? That's just propoganda.

Memory and storage space are two different things. You know that hard drives and RAM are different, right? One is limited by a 32-bit CPU and the other isn't.

 

[Rakehellion]

So have we come to any conclusions about the A8?

 

[Eug]

A7 was a HOLY CRAP THAT'S AMAZING jump for Apple. A8 is just a nice incremental upgrade, which remains quite competitive.

Or at least, that's my take on it.

 

[scannall]

BTW, when does Apple announce its new tablet? I thought they did that at the same event as the iPhone.

Probably early October. There was already too much jammed into this last event. Adding the iPad as well would have diluted the whole thing.

 

[Rakehellion]

A7 was a HOLY CRAP THAT'S AMAZING jump for Apple. A8 is just a nice incremental upgrade, which remains quite competitive.

Or at least, that's my take on it.

Going from the 4 to the 4S was a pretty huge leap too.

The A8 is underclocked to preserve battery life. Apple claims that it won't throttle after time like some phones (including the iPhone 5s) do. I guess we'll see once they ship.

 

[TreVader]

That's because as a technical forum this stuff is already extremely trivial knowledge. As in "I learned this in grade-school" trivial.


I'm not trying to put you down though so I'll try to give a quick explanation:

You're right in that it's a 32bit counting limit but there are only a few things that are limited by the CPU's own bit'ness to 32bits. These are generally things that are tied very closely to the CPU's function.

The RAM limit on PC's actually isn't one them. As Rakehellion noted, PC's have had >4GB RAM capability long before 64bit CPUs. Even Windows had this capability although MS limited it to server versions (and yes I've worked with those type of systems but you could google it too I'm sure).

Simplified, the real limit is how much each process can see at once and how large a number the CPU can fundamentally (as in extremely low level) compute with. Basically (or rather, it's generally since extensions make it less hardcut), a 16bit CPU can only fit a number from 0-65,536 (unsigned) while a 32bit CPU can do 0-4,294,967,296 (unsigned) inside its registers. But you can clearly do math with numbers bigger than 4,294,967,296 on your 32bit Windows XP running on a Pentium 2 so generally these limits are only interesting to someone who is doing something extremely low-level.


In any case, a file system is so extremely high level that the CPU's bit'ness never comes into play at all. NTFS, introduced in 1993, has a theoretical file size limit of 2^64 or 16EB. So that predates x64 by over a decade!

I think your grade schools must have been much more comprehensive than mine. My experience with computers up to 6th grade (outside of the Mac at home) was with some old gateway Celeron system playing sim city.


It sounds like a 64bit CPU might allow a processor to complete certain processes faster, because the larger registers would make it so you could compute with numbers greater than 4,000,000,000 in 64bit case. I assume you would have to find a way to work around the register problem for 32bit (say, if I had to multiply two very large numbers) which would take more clock cycles.

How is that a gimmick?

 

[TreVader]

Going from the 4 to the 4S was a pretty huge leap too.



The A8 is underclocked to preserve battery life. Apple claims that it won't throttle after time like some phones (including the iPhone 5s) do. I guess we'll see once they ship.

I'm betting the jump from 6 to 6S will be bigger than 5s to 6 CPU wise. They need to put more ram on it as much as I dislike samsung, I will give them credit for putting plenty of ram on the note 3.

And the jump from 4 to 4S will prob be the biggest jump.

 

[ChronoReverse]

How is that a gimmick?

For phones one reason it's a gimmick is because pretty much anything you would do doesn't need 64bit numbers. In fact, 64bit processors can be slightly slower if you're performing a task that falls within 32bits IF AND ONLY IF everything else is equal.

This last part is the key here. Usually with the jump to 64bits, the CPU designers take the opportunity to fix and improve other parts of the architecture and _that's_ where the large improvements usually come from. More registers is a big improvement for example.

This is the real reason the A7 is so much faster. Because it has improved architecture and not really because it's 64bits.

Let me repeat, marketing 64bits on mobile is a gimmick. The speed improvments isn't from that at all!

 

[Enigmoid]

Going from the 4 to the 4S was a pretty huge leap too.

The A8 is underclocked to preserve battery life. Apple claims that it won't throttle after time like some phones (including the iPhone 5s) do. I guess we'll see once they ship.

Its nice to see a phone not throttle.

But then there are (at least) two ways of looking at every problem and running at a flat constant clock can be seen as a failure to boost clocks when available.

 

[TreVader]

For phones one reason it's a gimmick is because pretty much anything you would do doesn't need 64bit numbers. In fact, 64bit processors can be slightly slower if you're performing a task that falls within 32bits IF AND ONLY IF everything else is equal.

This last part is the key here. Usually with the jump to 64bits, the CPU designers take the opportunity to fix and improve other parts of the architecture and _that's_ where the large improvements usually come from. More registers is a big improvement for example.

This is the real reason the A7 is so much faster. Because it has improved architecture and not really because it's 64bits.

Let me repeat, marketing 64bits on mobile is a gimmick. The speed improvments isn't from that at all!

Ok, it may be technically a gimmick.


BUT I think trying to explain why A7 is faster by saying it is 64bit is much simpler than the long paragraph you just wrote to explain it to me, right?

Edit: Hmmm controversy about this

 

[NTMBK]

For phones one reason it's a gimmick is because pretty much anything you would do doesn't need 64bit numbers. In fact, 64bit processors can be slightly slower if you're performing a task that falls within 32bits IF AND ONLY IF everything else is equal.

This last part is the key here. Usually with the jump to 64bits, the CPU designers take the opportunity to fix and improve other parts of the architecture and _that's_ where the large improvements usually come from. More registers is a big improvement for example.

This is the real reason the A7 is so much faster. Because it has improved architecture and not really because it's 64bits.

Let me repeat, marketing 64bits on mobile is a gimmick. The speed improvments isn't from that at all!

This was true for AMD64, as that is fundamentally the same old x86 ISA with 64-bit extensions and some new registers. ARMv8 (64-bit) is a whole new ISA, which is much more efficient than the old 32-bit ISA. It cleans up a lot of the cruft- it can genuinely allow for better performance.

 

[ChronoReverse]

This was true for AMD64, as that is fundamentally the same old x86 ISA with 64-bit extensions and some new registers. ARMv8 (64-bit) is a whole new ISA, which is much more efficient than the old 32-bit ISA. It cleans up a lot of the cruft- it can genuinely allow for better performance.

That's pretty much what I said. Architectural improvements is what makes the 64bit version so much faster, not because it's 64bit. I did say "IF AND ONLY IF everything else is equal" after all.

Ok, it may be technically a gimmick.

BUT I think trying to explain why A7 is faster by saying it is 64bit is much simpler than the long paragraph you just wrote to explain it to me, right?

That is precisely it. That's why I'm not upset that the marketers just say "64bit so it's faster". It's only on a more technical forum like this one that it's worth making the point clear.

 

[cytg111]

... Let me repeat, marketing 64bits on mobile is a gimmick. The speed improvments isn't from that at all!

- And a move to mature the architecture(and software eco system) for desktop-class usage?

 

[TreVader]

Different topic, but related to A8. May want to make a new thread?
OK so this is about the S1 processor in the apple watch


http://arstechnica.com/gadgets/2014/09/moto-360-review-beautiful-outside-ugly-inside/2/


^ This Ars article talks about how the processor in the moto 360 is an old (2010 era) Texas Instruments ARMv7 processor, comparable to the iPhone 4. Will apple pawn off the S1 to Ti? And what if the real constraints (shock resistance) mean Apple will be forced to use an SoC that's been tested to handle all that movement, which I believe would mean something similar?


Really hope apple didn't stick an ARM A5 800mhz in it and call it a day. They've never done that before. My guess is they will be tight lipped about specs, even moreso than usual.

 

[ams23]

Apple claims that it won't throttle after time like some phones (including the iPhone 5s) do. I guess we'll see once they ship.

The iPhone 5S GPU performance does not throttle in the GFXBench T-Rex HD Onscreen test (at least based on the "long term" performance test where the benchmark is looped several dozen times). That said, in that particular benchmark, do note that most ultra mobile GPU's (such as PowerVR GPU's and Adreno GPU's) are only rendering at partial [FP16] precision and not full precision.

The iPhone 5S CPU performance does throttle with Anand's custom CPU benchmark, but the throttling is only significant after > 100 seconds of benchmark run time.

 

[Essence_of_War]

BUT I think trying to explain why A7 is faster by saying it is 64bit is much simpler than the long paragraph you just wrote to explain it to me, right?

Yeah but the nice thing about written paragraphs is that someone only has to write them once if others can link to them. So you don't have to give the wrong answer, you can show someone a well-written correct answer.

 

[mavere]

most ultra mobile GPU's (such as PowerVR GPU's and Adreno GPU's) are only rendering at partial [FP16] precision and not full precision.

Is that a bad thing? By definition, realtime graphics engines use tricks and make sacrifices to trade fidelity for speed. There's no such thing as "full precision" here.

Edit:

From one of Unity's developers:

• basically, always use FP16 when possible. Can be faster; but even when isn't it's less power
• there are certain things where FP16 is clearly not enough ofc. World space positions; UVs etc.
• (...) all I'm saying is that using FP16 is preferred on *all* mobile

 

[Rakehellion]

Its nice to see a phone not throttle.

But then there are (at least) two ways of looking at every problem and running at a flat constant clock can be seen as a failure to boost clocks when available.

Boosting clocks when available is the same thing as throttling clocks when not available. It results in inconsistent performance.

 

[MarkizSchnitzel]

Boosting clocks when available is the same thing as throttling clocks when not available. It results in inconsistent performance.

It's not just "when available", I suspect more often it's gonna be "when needed". Surely lower power usage when doing light tasks is preferable?

 

[ams23]

Is that a bad thing?

Well not necessarily (although do note that virtually all modern day console, PC, and notebook graphics architectures and games are designed with full FP32 precision in mind). But the point I was trying to make is that on the T-Rex HD test, due to significant differences in rendering precision, the workload and rendering quality is not always the same when comparing one GPU to another. For instance, the render quality reported by the T-Rex HD benchmark is much higher for an SoC such as Tegra K1 vs. an SoC such as A7 or S800.

 

[Enigmoid]

This isn't confirmed and is GPU related but it appears that preliminary GPU benchmarks of the iphone 6 are out.

http://www.tomshardware.com/news/apple-iphone-6-gpu-benchmarks,27668.html

Merely speculation but its possible that the 50% figure may have been under a best case senario.

 

[liahos1]

soft, does not bode well for ipad

 

[Khato]

Merely speculation but its possible that the 50% figure may have been under a best case senario.

Given that it's almost certainly a transition from PowerVR Series 6 to Series 6XT that's quite possible. From http://www.anandtech.com/show/7629/...rchitecture-available-for-immediate-licensing

Meanwhile when it comes to performance Imagination has made several improvements and optimizations for Series6XT, with Imagination claiming performance gains of up to 50% versus Series6. Among other things, Imagination has specifically targeted front end and back end performance on Series6XT, nothing that they’ve made changes to improve sustained polygon and pixel fillrate performance.

That would be the case if, for some reason, Apple didn't make use of the die shrink to go from the 4 cluster design to 6... which is what those benchmarks imply is the case.