Apple A8 and A8X

You can keep the 8 cluster theory if you want, but at around 450 MHz then. My former layman's math shouldn't be completely worthless.

so with the fixit teardown [check .. X] .. where's the chipworks die photo of the A8X ?

if they put 3 cores into the A8X is it safe to assume with a 14nm/16nm FinFET process next year perhaps, that the A9 will get 3 CPU cores as well ?
 
so with the fixit teardown [check .. X] .. where's the chipworks die photo of the A8X ?

if they put 3 cores into the A8X is it safe to assume with a 14nm/16nm FinFET process next year perhaps, that the A9 will get 3 CPU cores as well ?

I have this feeling that Apple is targeting a particular usage model on the iPad to warrant that 3rd core, perhaps the rumoured split-screen apps or an updated multi-tasking model. Will most iOS apps /iPhone benefit much from a 3rd core, compared to bumping the MHz that will benefit every app. If Finfets really improve leakage, perhaps Apple will be more tempted to chase clockspeed next round?
 
The PC gaming industry moved away from mixed precision rendering many many years ago. Mixed precision created a mess for game developers who had to create custom mixed precision code paths for different architectures.

Power efficiency, power consumption, heat, noise, etc. matter in all areas now, not just mobile. The GPU performance and GPU power efficiency is good enough nowadays where there should not be an overwhelming need to render pixels with reduced precision IMHO.

Is there any specific instance where full FP32 would provide a discernible image quality increase? I have own iPhones for years, all with powerVR chips and all running FP16, and I've never seen any problems...
 
so with the fixit teardown [check .. X] .. where's the chipworks die photo of the A8X ?

if they put 3 cores into the A8X is it safe to assume with a 14nm/16nm FinFET process next year perhaps, that the A9 will get 3 CPU cores as well ?
Errrr one step at a time. Even if we get a die shot I'm awful with them. Someone had estimated from an early pic ~117mm2, but my gut feeling says over 130mm2 for sure.
 
I have this feeling that Apple is targeting a particular usage model on the iPad to warrant that 3rd core, perhaps the rumoured split-screen apps or an updated multi-tasking model. Will most iOS apps /iPhone benefit much from a 3rd core, compared to bumping the MHz that will benefit every app. If Finfets really improve leakage, perhaps Apple will be more tempted to chase clockspeed next round?

I hear a lot of people say this but I doubt it. Every implementation of split screen multitasking on a tablet (that I've seen) has been clunky and personally I've never had an instance where I thought "darn, if only I had split screen". It sounds great as an idea, but I think it would be too hard to implement successfully. The extra core is likely just there to increase CPU performance across the board. With all the new photo editing apps, they can really benefit from a 3 core CPU.
 
The increase in fill rate suggests a cluster count and/or clock speed that should have pushed the assumedly ALU bound Manhattan test to higher frame rates, even accounting for a much higher degree of theoretical max fill rate being achieved. This is in the context of current assumptions about the configuration and performance characteristics of the TMUs in announced PowerVR cores, of course.

Maybe the GX6650 or some custom PowerVR six cluster variant they're using picked up an extra four TMUs over the expected twelve (for sixteen in total) and is being clocked somewhere around 500 - 550 MHz (I realize that's a big range)? Or maybe the performance characteristics of each TMU significantly improved from Series 6 to 6XT?
 
This is indeed shown by Anandtech SPECint2000 run: 181.mcf is putting a lot of pressure on the memory subsystem (the article wrongly hypothesizes integer multiplication is the source of the speedup).


Having studied some non benchmark code (e.g., browsers displaying web pages as opposed as running some JS benchmark), the data cache hit ratio is quite high, much higher than what 181.mcf shows (about 10 misses/kinst vs 100 misses/kinst). So even SPEC shows some distortion (SPEC CPU 2006 mcf version has slightly lower D$ misses, but TLB misses are higher).

Yes. Of course, while I think the current almost exclusive focus on mostly L1 cache resident benchmarks is an abomination, the truth is of course that benchmarking that is good enough to have useful predictive value is really difficult. I actually think that outside of SPEC, the Geekbench approach of having a multitude of small benchmarks complemented by a memory subsystem test is a reasonable approach, the problem being that the tests are too few and simplistic - I'd like memory footprint vs bandwidth/memory footprint vs latency graphs to get a handle on the memory subsystem. It is far from impossible, it just isn't done much. (I run LinPack with increasing matrix sizes on all new systems I come across for the first.) It could still be boiled down to a useless Single Figure of Merit for the impatient, anything can. :)

My opinion is probably strongly affected by the fact that pretty much any code I run across in my field is ultimately limited by the memory subsystem rather than lack of ALU capabilities, and has been for very long time.
 
The increase in fill rate suggests a cluster count and/or clock speed that should have pushed the assumedly ALU bound Manhattan test to higher frame rates, even accounting for a much higher degree of theoretical max fill rate being achieved. This is in the context of current assumptions about the configuration and performance characteristics of the TMUs in announced PowerVR cores, of course.

Maybe the GX6650 or some custom PowerVR six cluster variant they're using picked up an extra four TMUs over the expected twelve (for sixteen in total) and is being clocked somewhere around 500 - 550 MHz (I realize that's a big range)? Or maybe the performance characteristics of each TMU significantly improved from Series 6 to 6XT?

It's the fillrate that made me re-think the possibility of more clusters. Onscreen it gets 7626 MTexels/s.

7626 / 12 = 635.5
7626 / 16 = 476.6

Scenario B sounds more like "Apple". On A8's GX6450 each dual cluster (TMUs included) is roughly under 6.5mm2, we just don't know how everything else in the middle would scale.
 
Guessing an eight cluster and normal sixteen TMU configuration has the advantage of not having to speculate about custom cores (just unannounced configs), helps to better explain some of those unaccounted transistors, does definitely sound a lot more like Apple with their preference for wider designs and lower voltages, and finally fulfills Imagination's suggestion of cluster scaling to eight (though, as mentioned, would not be a max limit... probably more like a customer interest limit at the time.)

One implication of that configuration is that there's probably several percent of untapped performance potential that could be unlocked with a possible future driver update.
 
I'm really curious about the dysmal physics scores in 3Dmark. A8X kills every other test, but on this specific test they perform at snapdragon 800 levels.


Maybe this test isn't as relevant to real world conditions. Or maybe 3Dmark is biased against apple
 
Their benchmark is allowed to run code that the A8/X doesn't like... as long as they provide a reasonable connection to wide-spread, "real-world" coding practice.

Also interesting is the Air2's relatively poor 3DMark Graphics score scaling (~1.3x) over the iPhone 6Plus.
 
3DMark tests, even the graphics specific tests, have always exhibited more CPU dependency than GFXBench's.

Kishonti's GFXBench is still one of the best to use if you're trying to isolate GPU performance only.
 
Their benchmark is allowed to run code that the A8/X doesn't like... as long as they provide a reasonable connection to wide-spread, "real-world" coding practice.

Also interesting is the Air2's relatively poor 3DMark Graphics score scaling (~1.3x) over the iPhone 6Plus.


I thought it scored 29k?
 
Please don't mention that I caught Ailuros in the exact same trap a couple of weeks back.
You did and it happens to me all the time. It doesn't hurt my feline pride *meowww* :LOL:

Lol :LOL:
so with the fixit teardown [check .. X] .. where's the chipworks die photo of the A8X ?

if they put 3 cores into the A8X is it safe to assume with a 14nm/16nm FinFET process next year perhaps, that the A9 will get 3 CPU cores as well ?

As I mentioned in my last post..FF does not really increase density so that will only happen if they are willing to increase die size (FF is going to be more expensive as it is). Though I have to admit that in a phone, Apple's approach of having an extremely high performance dual core seems to be more than competent enough. So I dont even think it will be necessary, especially since there should be a decent clock increase and possibly further architectural changes.
I have this feeling that Apple is targeting a particular usage model on the iPad to warrant that 3rd core, perhaps the rumoured split-screen apps or an updated multi-tasking model. Will most iOS apps /iPhone benefit much from a 3rd core, compared to bumping the MHz that will benefit every app. If Finfets really improve leakage, perhaps Apple will be more tempted to chase clockspeed next round?

Apps will not benefit unless they are specifically coded to take advantage of the extra core. Will definitely help for future apps and once devs make the changes for existing apps. The OS could also make use of the cores for scheduling tasks more efficiently and for the rumoured split-screen feature as you mention.
 
Last edited by a moderator:
The apps Apple highlighted to demonstrate the performance of A8X really were well chosen. Pixelmator and Replay were claimed to show gains of 2x to 4x in certain basic functions over the A7 powered iPad Air.

Accessing what may be a 500 MHz GX6850 + 1.5 GHz Cyclone v2 tri-core through Metal opens up compelling performance for computational photograph/video editing, computer vision, augmented reality, and other applications that require a good measure of high performance, sustained, heterogeneous coordinated processing.
 
Regarding 3DMark Ice Storm, you really want to look at the individual scores, including the GT1 & GT2 scores for graphics. The GT1 test is geometry-heavy while the GT2 test is fillrate/post-processing heavy.

GT1 / GT2
iPad Air 1: 103 / 71
iPad Air 2: 147 / 128 (+43% / +80%)
Shield Tablet: 209 / 126

i.e. K1 is ~40% faster for the geometry heavy GT1 test but the iPad Air 2 is ~2% faster for the pixel-heavy GT2 test. Note that this is for Ice Storm Unlimited *WHICH RUNS AT 720p offscreen*. If you ran GT1 at 1080p as in the original Ice Storm Extreme (now VSync limited) it would be less geometry limited.
 
Bought my first piepad just the other day, and it's been a pretty nice experience so far. However, Apple still needs to work on their integration I'd think, setting it up by restoring an iphone icloud backup installs iphone apps on the pad when there is an ipad version of the apps available in the app store for example. Also, it says settings do not copy over, even though the vast majority of settings are identical between ipad and iphone. This seems an unnecessary and arbitrary limitation, especially when a person like me moves from having had an iphone only and thus don't have any ipad backups yet to restore from.

Also, damned if I know how to make my photos share over to the ipad. Photo stream is activated in icloud, but nothing actually copies over.

The air2 pad is still somewhat heavy in a way. Probably not compared to the pre-air pads (the original retina ipad was almost obnoxiously heavy after extended hand-held use I recall), but it feels a little hefty still all the same. The weight is probably unavoidable though, if one wants sufficient rigidity in the product... You wouldn't want the screen cracking because you hold the pad solely by one corner for example. :)

The built-in speakers feel very weak. My iP5S makes LOTS more noise and its single speaker is smaller than the dual drivers in the pad. Weird. Maybe is because of THX lawsuit against Apple using resonance chambers, I dunno.

I haven't noticed any undue heating yet by the way. No app I've used so far has made the pad get more than even slightly warm. Even the initial setup with massive simultaneous app downloads, installs and battery charging going on at once didn't heat the device much. In contrast, my iP5S got really warm during that stage.
 
B

Also, damned if I know how to make my photos share over to the ipad. Photo stream is activated in icloud, but nothing actually copies over.
Once you enable photo stream, going forward it makes photos that you take, available to other devices (is it for 30 days ?). It won't make photos available that were taken prior to photostream being activated. I think apple have a new cloud storage for that in IOS8, but I haven't played with it.
 
Back
Top