Apple A9x chip discussion

[Nothingness]

Is that really true for the GPU? The Core M GPU scores lower in the GFX benches, but it scores much higher in 3dmark. I don't know if that is a driver issue or some other limitation of Intel's GPUs. And we don't really know if that is an issue that has been addressed with Gen 9.

Or an issue of 3dmark on iOS

And unfortunately we can't compare actual games, making it all the more difficult to compare the products.

Alas no.

 

[dahorns]

Or an issue of 3dmark on iOS

Oh sure, it could be that too. But it does seem that Intel GPUs have struggled with GFX in general comparative to their 3dmark scores (just do a cross compare of other products with similar 3dmark scores and see how they do on GFX benches). So, that leads me to believe that it is an Intel specific issue. What I don't know if it is a hardware limitation or a driver limitation. Given the substantial increase the HD 530 saw in the GFX benches, I'm leaning towards a hardware limitation of Intel's Gen 8 GPUs. I just wish I knew what exactly was being thrown at it in the GFX bench that it wasn't handling so well.

It looks like A9x vs Core M is going to be a close battle across the board. I imagine you'll end up seeing Core M ahead in single thread scores, behind in multi thread scores, and about even in most GPU benches (based on what we've been told from both sides).

It seems the real tests will be which processor does better sustained performance and which has better power efficiency.

 

[ShintaiDK]

They are not stuck with having that front end hack and associated hardware to translate 1980s vintage, patched on machine code into uOps back and forth.

Only 30 years legacy or so with ARM.

 

[Hugo Drax]

Only 30 years legacy or so with ARM.

So how many registers does ISA-x86 has vs AArch64?

 

[witeken]

So how many registers does ISA-x86 has vs AArch64?

The ISA is irrelevant for everyone but the CPU architect.

 

[Hans de Vries]

============================================
cpu according to the presentation: A9x vs a8x up to 1.8x
============================================

looks like a9x is quad core at ~2ghz versus a8x with triple
core at 1.5ghz plus ipc improvements. (in gb3 metric that
would be: Single threaded ~ 2400, multi threaded ~ 8000)

new core m3-6y30 geekbench score: St: 2485 mt: 4749
https://browser.primatelabs.com/geekbench3/3404121
this is a low-end core m, the fastest skylake-y version
operates at 40-45% higher turbo clocks. Can't wait to
see how the best core m7-6y75 retail implementations
perform.

this will be fun to watch.

In line with other results you would expect that the
Skylake core M M3-6Y30 score runs at the Turbo frequency
of 2.2 GHz. It has a sustainable base frequency of 900 MHz.

The top binning has a sustainable base frequency of 1200
MHz and a very high burst frequency of 3.1 GHz

Apple doesn't sell specially selected bins of the A9X with
higher frequency. The larger majority of all A9X produced
should fit in the frequency window of the Ipad Pro, so it is
probably fairer to compare with a mainstream Core-M3
as opposed to the fastest Core-M7 parts

 

[Hugo Drax]

The ISA is irrelevant for everyone but the CPU architect.

So how many transistors are tied to that whole conversion process between ISA-x86 and uOps And What the overhead is?

 

[R0H1T]

In line with other results you would expect that the
Skylake core M M3-6Y30 score runs at the Turbo frequency
of 2.2 GHz. It has a sustainable base frequency of 900 MHz.

The top binning has a sustainable base frequency of 1200
MHz and a very high burst frequency of 3.1 GHz

Apple doesn't sell specially selected bins of the A9X with
higher frequency. The larger majority of all A9X produced
should fit in the frequency window of the Ipad Pro, so it is
probably fairer to compare with a mainstream Core-M3
as opposed to the fastest Core-M7 parts

Which brings us back to the possibility of an A10(x?) replacing Intel at some point in time in the future, anything less than an MBP would be a likely candidate.

 

[Sweepr]

Well you can buy a full Windows convertible with Core M-5Y10c, 128GB and a nice IPS screen for as low $470 right now. iPad Pro 128GB WiFi + keyboard costs $1120. That's the pricing I expect from premium Core m7-6Y75 and Skylake-U convertibles with 8GB RAM and larger SSDs.

 

[dahorns]

In line with other results you would expect that the
Skylake core M M3-6Y30 score runs at the Turbo frequency
of 2.2 GHz. It has a sustainable base frequency of 900 MHz.

The top binning has a sustainable base frequency of 1200
MHz and a very high burst frequency of 3.1 GHz

Apple doesn't sell specially selected bins of the A9X with
higher frequency. The larger majority of all A9X produced
should fit in the frequency window of the Ipad Pro, so it is
probably fairer to compare with a mainstream Core-M3
as opposed to the fastest Core-M7 parts

I don't understand your conclusion. The Core-M3 parts will typically go in less premium devices (and less expensive) than the Ipad Pro. So, why should we ignore the top binning for Intel that would actually go into comparative products?

You could reword your statement to say "Apple doesn't sell specially selected bins of the A9X with [lower frequency]." The fact that Intel chooses to sell products that wouldn't meet the higher standard required for an M7 and the fact that Apple chooses not to sell products that wouldn't meet the standard required for an A9x doesn't seem to me to be critically important when comparing parts.

 

[Nothingness]

Only 30 years legacy or so with ARM.

AArch64 is only a few years old, and if Apple wants it they can get rid of the 32-bit support. A hint they might consider doing it is that all submissions to the app store must come with 64-bit support.

 

[Fjodor2001]

Quite impressive for Apple to be able to speed ahead like this, improving CPU performance as such a rapid rate. After all they are quite new to the CPU designing business, compared to e.g. Intel and AMD. And at least Intel cannot say it's due to lack of resources. So how come Apple can progress so fast, when Intel has ground to a halt?

 

[Hans de Vries]

This will be fun to watch.

I don't understand your conclusion. The Core-M3 parts will typically go in less premium devices (and less expensive) than the Ipad Pro. So, why should we ignore the top binning for Intel that would actually go into comparative products?

You could reword your statement to say "Apple doesn't sell specially selected bins of the A9X with [lower frequency]." The fact that Intel chooses to sell products that wouldn't meet the higher standard required for an M7 and the fact that Apple chooses not to sell products that wouldn't meet the standard required for an A9x doesn't seem to me to be critically important when comparing parts.

I think that to keep things simple we should compare SOC's and not
business models.

So, what we can can compare is:

How well do 80% or 90% of all working A9X chips perform compared
to 80% or 90% of all working Skylake-Y chips.

We will not get any information from Apple to compare the top 5%
of all A9X dies to the top 5% of all Skylake-y dies.


As a consumer you could certainly prefer a lower priced Core-M product
over an expensive Ipad pro of course, but here you're talking about
different business models, large margins versus huge margins....

 

[Khato]

So how come Apple can progress so fast, when Intel has ground to a halt?

Eh, let's wait to see how much performance gain is in the form of IPC versus frequency and core count? It's a lot easier to increase frequency on a sub 2GHz part than at 4GHz+, as evidenced by the performance gains on core m going from Broadwell to Skylake. And then adding more cores is by far the 'cheapest' way to increase 'performance'.

No question that Apple was progressing fast, but until we see get some actual information regarding A9 it's hard to say whether they're still on a roughly linear performance cagr or logarithmic (high gains to start which taper off to being similar to Intel.)

 

[dark zero]

I think that to keep things simple we should compare SOC's and not
business models.

So, what we can can compare is:

How well do 80% or 90% of all working A9X chips perform compared
to 80% or 90% of all working Skylake-Y chips.

We will not get any information from Apple to compare the top 5%
of all A9X dies to the top 5% of all Skylake-y dies.


As a consumer you could certainly prefer a lower priced Core-M product
over an expensive Ipad pro of course, but here you're talking about
different business models, large margins versus huge margins....

Remember the customer experience...
And the legión of fans behind the company.

Also... Despite Windows 10 improved the touchscreen experience... Unless you go full metro

You are Depending on keyboard and mouse.

 

[Fjodor2001]

It's a lot easier to increase frequency on a sub 2GHz part than at 4GHz+, as evidenced by the performance gains on core m going from Broadwell to Skylake.

But did we really see that much performance gains for Core M going from Broadwell to Skylake? 5Y71 (1.2/2.9) vs 6Y75 (1.2/3.1). So the only difference is that the turbo frequency is 200 MHz higher (and Skylake can also turbo a bit higher in dual core mode). Apart from that's it's only the usual ~5-7% general IPC improvement we saw going from Broadwell to Skylake.

So I'd say that is a lot less than how much Apple appears to have improved performance on A9X.

 

[Headfoot]

I highly, highly, highloy doubt the 80% figure is average performance increase. 99.9% chance that is the maximum performance advantage in a perfect scenario. I dont know why anyone is buying marketing material from one of the best marketing companies in the world. Wait for real measurements by independent reviewers before we start going on about how skylake core M is dead.....

 

[dahorns]

I think that to keep things simple we should compare SOC's and not
business models.

So, what we can can compare is:

How well do 80% or 90% of all working A9X chips perform compared
to 80% or 90% of all working Skylake-Y chips.

We will not get any information from Apple to compare the top 5%
of all A9X dies to the top 5% of all Skylake-y dies.


As a consumer you could certainly prefer a lower priced Core-M product
over an expensive Ipad pro of course, but here you're talking about
different business models, large margins versus huge margins....

Yes, let's compare chips, and let's compare like chips. Apple, by only producing the A9X chips is intentionally releasing only the top x% of chips. We have no idea what x is because we have no idea what functional yield Apple gets on its chips.

Intel could choose to only produce M7 chips just like Apple chooses to produce only A9X chips. Instead, Intel produces M3, M5, and M7 chips in order to use a larger percentage of the available dies. Again, we have no idea what the comparative numbers of A9X vs M7 chips is for a single wafer. It could be that only 1 in 4 potential A9X chips is actually usable as an A9X. And the numbers could be similar for the M7. The only difference here is that Intel chooses to use those non-qualifying chips for something else.

What we do know is that the M7 and the A9X will compete in similar products. That is the most rational and reasonable way to compare the chips. Your choice to artificially limit Intel to only the M3 is simply a handicap in Apple's favor. And it is a completely arbitrary handicap and quite frankly one that Apple doesn't need.

 

[Eug]

Speaking of the A8X. Here's a reminder from last year.
http://www.anandtech.com/show/8716/apple-a8xs-gpu-gxa6850-even-better-than-i-thought

So Apple claims a 2X performance increase from that. I didn't get last years iPad Air 2(or the year before it), but there are several things going through my mind:

1. Does Apple need this? The only logical conclusion I have is if they are going with a 4K display, but Apple has often been good at ignoring fads. Look at their core count. Apple completely sidestepped the MOAR CORES hypetrain and with good reason. Do people really need 4K at 10-13 inches? So if not 4K, why would they need this massive GPU increase? They talked vaguely about how this will "unlock new applications" but from what I saw being demoed, I doubt that you couldn't do that with an A8X quite well.

It's 5.6 Megapixels for the MaxiPad and I'd say that resolution is most definitely needed. The pixel density happens to be exactly same as the iPad Air 2, at 264 ppi, which is Retina. If it went much lower, it would lose its "Retina-ness" IMO. The Retina MacBooks are lower ppi, but in general laptops are used with a further working distance.

Now given that the MaxiPad can be used like a laptop, some could argue that the working distance might be extended for a lot of people so a lower resolution would be OK, and while that's true, there's bound to be a lot of people who would still use it like they would an iPad Air 2. So, Apple needed to keep the PPI up on this.

So while not 4K it's still pretty high, and add in gaming and you get the need for such a fast GPU.

BTW, the PPI on an Apple iPhone is actually much higher, and that's also needed, because people typically hold their phones closer than they do their tablets.

By the way, It's almost confirmed it has 4GB ram.
https://blogs.adobe.com/conversatio...ivity-apps-during-apple-ipad-pro-keynote.html

I actually care more about this (to a certain extent) than the SoC specs. Yeah we knew the MaxiPad's CPU/GPU was going to be fast, but since I'm not an avid gamer and won't be doing content creation on my tablet either, I was more interested in other things, such as RAM. RAM is of ultimate importance for longevity of an iDevice, and it's nice to see they didn't do what I was predicting, which was limit it to 2 GB and then upgrade it next year or so to 4 GB.

Apple has a long history of being very stingy with RAM, and the iDevices suffered this in spades. Apple compensated by being extremely aggressive with memory management in iOS, but it was still an issue. However, it seems that 2015 is the year when targeted a decent baseline. All the main iDevices are at least 2 GB now, or at least I'm predicting they are:

iPad Pro: Known to be 4 GB RAM and A9X
iPad Air 2: Known to be 2 GB RAM and A8X
iPad mini 4: Likely 2 GB RAM and known to be A8
iPhone 6s: Likely 2 GB RAM and A9
iPhone 6s+: Likely 2 GB RAM and A9
Apple TV: Known to be 2 GB RAM and A8

The only outlier here is the new iPod touch with 1 GB RAM, but that came out months ago, and has always been considered a 2nd tier device in Apple's product line in terms of specs, which makes sense given its lower cost.

 

[Space69]

Is that really true for the GPU? The Core M GPU scores lower in the GFX benches, but it scores much higher in 3dmark. I don't know if that is a driver issue or some other limitation of Intel's GPUs. And we don't really know if that is an issue that has been addressed with Gen 9.

And unfortunately we can't compare actual games, making it all the more difficult to compare the products.

No, you can't really compare. Most (all?) benchmark/games etc. use FP16 (shader) and very little FP32 (Vertex) on IOS/Android. Most benchmark/games on PC use FP32.

If we assume Apple use Series7XT with 16 clusters at 400MHz (which is roughly 88% faster than A8X) you'll have

FP16: 128Flops * 16 * 400MHz = 819.2GFlops
FP32: 64Flops * 16 * 400MHz = 409.6GFlops
FP64: None

It doesn't support simultaneous FP16 and FP32 operation

Single Intel EU FP32 16Flops

FP32: 24 EU at 1100MHz: 16 Flops * 24 * 1100MHz = 422.4GFlops
FP64: 24 EU at 1100MHz: 4 * 24 * 1100MHz = 105.6GFlops

 

[Edrick]

Quite impressive for Apple to be able to speed ahead like this, improving CPU performance as such a rapid rate. After all they are quite new to the CPU designing business, compared to e.g. Intel and AMD. And at least Intel cannot say it's due to lack of resources. So how come Apple can progress so fast, when Intel has ground to a halt?

Apple controls the CPU engineering, the OS design (and drivers), and on a large part, the app development. All for a specific small set of devices.

Fairly easy to advance design when you control all aspects of the process.

Intel, on the other hand, does not have that luxury and has to maintain legacy compatibility for multiple OSes, drivers, platforms, busses, applications, etc. which are mostly out of their control. I'm fairly confident that if Intel only had to design a CPU that would run Windows 10 with a small set of in house designed drivers, with only Intel designed hardware, that it would easily beat the yearly performance jumps we see today.

 

[kimmel]

Apple controls the CPU engineering, the OS design (and drivers), and on a large part, the app development. All for a specific small set of devices.

Fairly easy to advance design when you control all aspects of the process.

Intel, on the other hand, does not have that luxury and has to maintain legacy compatibility for multiple OSes, drivers, platforms, busses, applications, etc. which are mostly out of their control. I'm fairly confident that if Intel only had to design a CPU that would run Windows 10 with a small set of in house designed drivers, with only Intel designed hardware, that it would easily beat the yearly performance jumps we see today.

You forget the most important part, control what benchmark you are using to claim the gen on gen performance gain and not publicly disclose what benchmark you are claiming that gain on. I'm fairly certain I could find a benchmark to tell me whatever I wanted on whatever architecture I wanted.

 

[IntelUser2000]

It doesn't support simultaneous FP16 and FP32 operation

Single Intel EU FP32 16Flops

FP32: 24 EU at 1100MHz: 16 Flops * 24 * 1100MHz = 422.4GFlops
FP64: 24 EU at 1100MHz: 4 * 24 * 1100MHz = 105.6GFlops

Starting with Skylake's Gen 9, FP16 is supported and its double the Flops of FP32.

Looks like A9X is quad core at ~2GHz versus A8X with triple
core at 1.5GHz plus IPC improvements.

(In GB3 metric that would be: Single threaded ~ 2400, Multi
threaded ~ 8000)

Now I think about it, I don't agree multi-thread scores being 8000. They just quoted "80% faster". With A8X, they claimed 40%, but that turned out to be 20% ST and 60% MT. It could be that 80% turns out to be 60% in ST and 100% in MT. Anyway the graph where they claim that iPad Pro is 22x faster than the 1st gen iPad correlates to their claim of 40% if you extrapolate it.

Any chance of future Intel chips landing on Apple is completely busted. If anything, soon they'll be able to emulate x86 and run at comparable performance.

 

[Edrick]

You forget the most important part, control what benchmark you are using to claim the gen on gen performance gain and not publicly disclose what benchmark you are claiming that gain on. I'm fairly certain I could find a benchmark to tell me whatever I wanted on whatever architecture I wanted.

Exactly

- Apple releases new CPU with more cores and new instructions that only show up in certain benchmarks. Apple fanboys response: "Apple is great. We get 80% performance increase almost every year. Who cares if it is only on specific benchmarks or marketing slides. Who cares if developers have to update their code to utilize it. Intel sux!"

Meanwhile

- Intel releases new CPU that doubles the FP throughput (AVX2) that shows dramatic increases in all software which utilizes the new instructions. Apple/AMD fanboys response: "Intel sux! Developers have to re-write code to take advantage of these new instructions. We want 50% IPC increase every year! ...and without any code re-compiles."

 

[itsmydamnation]

The ISA is irrelevant for everyone but the CPU architect.

i love how experts come along and tell everyone how it is.....

Now do you remember the days when x64 first came out, i do.
Do you remember changing nothing in code and recompiling to x64 and getting a perf boost, i do.
you want to guess where that perf boost came from? :awe:

So if registers only matter to CPU architectes what do compile devs do, people who write in assembly or CPU optimisations guides matter for then.

Oh thats right your statement is so fundamentally wrong its not funny.

here is another important "ISA is irrelevant for everyone but the CPU architect" FACT:
taken from realworldtech
For 176.gcc on integrate.i compiled with gcc 4.8.3 for x86-64 and gcc linaro 4.9 2014.09:
x86-64. 7250M instructions
AArch64 6690M instructions

AArch64 ISA has 3 operand for many more instruction then x86, lots of moves and copies saved. But i forgot 3 operand instruction are only the second coming of christ when intel implements them.

Exactly

- Apple releases new CPU with more cores and new instructions that only show up in certain benchmarks. Apple fanboys response: "Apple is great. We get 80% performance increase almost every year. Who cares if it is only on specific benchmarks or marketing slides. Who cares if developers have to update their code to utilize it. Intel sux!"

thats your confirmation bias right there because you don't know that. Second AVX2 did not double fp throughput , AVX2 doesn't even have FP instructions but int SIMD instructions, so dont let facts get in the way.........

What increased AVX throughtput from IB forward has been FMA and actually being able to load store 512/256bit a cycle. 512bit AVX isn't coming to consumer so is irrelevant.