Apple A8 and A8X

[Ailuros]

Why 100 MHz higher clocks ( + 22%) when the result is only 5% better?

I'd like to stand corrected but it seems that the high quality offscreen scores are even identical. No idea what to make of it, but it sounds increasingly like an as boring GPU as CPU upgrade in A8.

***edit:

so far we have the following theories for the A8 GPU:

* GX6450
* GX6650
* G6630 (doesn't make sense due to absence of ASTC)

Dumb question does the GR6500 support ASTC?

 

[wco81]

ARe the benchmarks corrected for the higher resolution?

 

[Ailuros]

ARe the benchmarks corrected for the higher resolution?

See post #78 above; even if not the 5S is at 1136*640 while the tested iPhone6 in question is at 1334*750; I wouldn't call that a significant difference http://results.rightware.com/devices/stock_model/apple/iphone_6

 

[Rys]

See post #78 above; even if not the 5S is at 1136*640 while the tested iPhone6 in question is at 1334*750; I wouldn't call that a significant difference http://results.rightware.com/devices/stock_model/apple/iphone_6

The devil's in the details there, since the 6 has 37% more pixels.

 

[anexanhume]

I had assumed it was an offscreen/ res indp bench when they didn't even bother listing the resolution of the phones. Perhaps not.

As for GR6500, I think Apple may have mentioned being the first ever consumer device to ship with accelerated ray tracing

 

[Ailuros]

The devil's in the details there, since the 6 has 37% more pixels.

True.

As for GR6500, I think Apple may have mentioned being the first ever consumer device to ship with accelerated ray tracing

Also true

 

[Lodix]

How does the Cyclone architecture compared to a normal Cortex A57 ? It is a variant for higher performance or for lower power consuption ?

 

[Entropy]

Well, I suppose we can't look forward to Anand's rundown of what the A8 SoC consists of.

Obviously Anand isn't going to report on it, but I wouldn't be so quick to dismiss Ryan Smith. The guy seems both knowledgeable and level headed, and I think he could do a very solid job of analysis if he puts his mind to it. And I think he will, partly because Anandtech has been the go-to place for technical reporting on Apple devices, and I don't think they want that to end, but also because I don't think he wants to give the Anandtech-after-Anand naysayers any fuel for their fire.
I have good hope.

 

[ltcommander.data]

Maybe Basemark X is memory bandwidth limited? If it's true that the A8 still uses 2x32-bit LPDDR3-1600 and relies on a larger L3 cache, games may need to optimize for cache residency to extract full performance from the A8 GPU. That's not going to be ideal for cross-platform games, although Apple probably feels that iOS games that need maximum performance will use Metal and if you're doing a custom Metal code path you can write it to take advantage of the L3 cache.

 

[Exophase]

How does the Cyclone architecture compared to a normal Cortex A57 ? It is a variant for higher performance or for lower power consuption
?

It's not a variant at all, it has nothing in common whatsoever other than supporting the same instruction sets. No more related than Haswell and Steamroller are.

From the look of things it's pretty significantly different, with a lot more perf/MHz and so far at least shipping with a lot lower MHz. It's hard to compare exactly with A57 since we don't have a full picture of how it achieves improved performance vs A15. But comparing Cyclone vs A15, the former is much wider decode/dispatch and has better memory reordering capabilities, among other things.

 

[Grall]

That's not going to be ideal for cross-platform games

What phone right now offer better bandwidth than this anyway? It's pretty much the state of the art right now, and mobile games frequently are not pushing the envelope anyway, so won't be a problem.

 

[ltcommander.data]

What phone right now offer better bandwidth than this anyway? It's pretty much the state of the art right now, and mobile games frequently are not pushing the envelope anyway, so won't be a problem.

LPDDR3-1866 is pretty widely available now although it's not a big upgrade from LPDDR3-1600. The Snapdragon 805 gets around the memory bandwidth issue by going to a 4x32-bit memory interface like Apple previously did with their X variants. But yes, easy memory bandwidth gains are stalled waiting on LPDDR4.

 

[RDGoodla]

The 6430@450MHz has 115 GFLOPs FP32 and/or 173 GFLOPs with a 13.0 fps score in Manhattan.

With a hypothetical 6650 clocked at 400MHz would have 154 GFLOPs FP32 and/or 307 GFLOPs. If you relate it to only FP32 GFLOPs you get 17 fps, but I figure that the additional FP16 ALUs would also contribute a bit more to the final score.

May I ask that is there any source which can confirm iPhone 5S has 115 Gflops? Because apple claim that GPU of iPhone 5S is twice as fast as GPU of iPhone 5. If iPhone 5s really has 115 Gflops than it may be 4x of iPhone 5. Thanks!

 

[Ailuros]

May I ask that is there any source which can confirm iPhone 5S has 115 Gflops? Because apple claim that GPU of iPhone 5S is twice as fast as GPU of iPhone 5. If iPhone 5s really has 115 Gflops than it may be 4x of iPhone 5. Thanks!

The Apple 7 GPU is a Rogue G6430. The only other unknown is frequency, but I doubt the estimated frequency of 450MHz is far from reality.

Here's one of the first references from Anand for the G6430: http://www.anandtech.com/show/7335/the-iphone-5s-review/7

And here a detailed breakdown of capabilities and unit count on Rogue GPU IP: http://blog.imgtec.com/powervr/the-...ogue-gpus-specifications-features-api-support

Considering the 6430 has 128 FP32 SPs at 450MHz that's 115.2 GFLOPs FP32 and/or 192 FP16 SPs which equal to 172.8 GFLOPs FP16.

 

[RDGoodla]

The Apple 7 GPU is a Rogue G6430. The only other unknown is frequency, but I doubt the estimated frequency of 450MHz is far from reality.

Here's one of the first references from Anand for the G6430: http://www.anandtech.com/show/7335/the-iphone-5s-review/7

And here a detailed breakdown of capabilities and unit count on Rogue GPU IP: http://blog.imgtec.com/powervr/the-...ogue-gpus-specifications-features-api-support

Considering the 6430 has 128 FP32 SPs at 450MHz that's 115.2 GFLOPs FP32 and/or 192 FP16 SPs which equal to 172.8 GFLOPs FP16.

Thank you for these useful info.

However, according to wiki: http://en.wikipedia.org/wiki/Apple_system_on_a_chip

iPhone 5 only has 25.5 Gflops, then 2x performance should be 51 Gflops, which is very very close to the performance of G6430 @ 200 MHz (200 MHz 6430 can have 51.2 Gflops). Is it possible that apple A7 GPU operates at much lower frequency than 450 MHz?

 

[Ailuros]

Thank you for these useful info.

However, according to wiki: http://en.wikipedia.org/wiki/Apple_system_on_a_chip

iPhone 5 only has 25.5 Gflops, then 2x performance should be 51 Gflops, which is very very close to the performance of G6430 @ 200 MHz (200 MHz 6430 can have 51.2 Gflops). Is it possible that apple A7 GPU operates at much lower frequency than 450 MHz?

The A6 in the iPhone5 contains a SGX543MP3@325MHz. Without any SFU FLOPs the latter gets 31.2 GFLOPs arithmetic efficiency.

No it's not possible that the 6430 operates at a lower than 450MHz frequency only equal or higher than that: http://gfxbench.com/subtest_results_of_device.jsp?D=Apple+iPhone+5S&id=490&benchmark=gfx27

GLB2.7 fillrate at its highest from it is at 3464 MTexels/s divided by 8 TMUs gives = 433MHz and it's more than rare that GPUs have a 100% fillrate efficiency in such tests.

Other than that if you think that Apple while claiming 2x times the GPU performance between A7 and its predecessor meant arithmetic efficiency you're just plain and simple wrong. IMG GPU IP went from multi-core to multi-cluster scaling in Rogue and that means that you will get partially lower rates for geometry for instance:

http://gfxbench.com/subtest_results_of_device.jsp?D=Apple+iPhone+5S&id=492&benchmark=gfx27

http://gfxbench.com/subtest_results_of_device.jsp?D=Apple+iPhone+5&id=492&benchmark=gfx27

Apple gave probably for the first time some sort of average in performance increase.

 

[Lazy8s]

Fill rate benchmarking isn't exact; Kyro would, at times, test slightly higher than its theoretical max in some tests.

My theory of 433 MHz for the 5s's 6430 and 467 MHz for the Air's 6430 fits the data I was studying last year. Either way, ~450 MHz for the A7 GPU as you've stated is a good characterization.

I'm still waiting to see more indications for A8, but a GX6650 has only made sense to me from the perspective of what I'd expect Apple to pick. From the performance estimates, a GX6450 or some kind of custom G6630 inspired part makes more sense. Apple's marketing is nebulous enough, however, that it could easily be a GX6650 being promoted conservatively like with the A7 GPU marketing.

 

[Ailuros]

Fill rate benchmarking isn't exact; Kyro would, at times, test slightly higher than its theoretical max in some tests.

4 generations back (yes the count is accurate) and completely different fillrate tests.

My theory of 433 MHz for the 5s's 6430 and 467 MHz for the Air's 6430 fits the data I was studying last year. Either way, ~450 MHz for the A7 GPU as you've stated is a good characterization.

No chance; Manhattan is very ALU bound. If there would be a frequency difference you could detect it in those score differences. Performance even scores linearly with FLOPs; if you check Allwinner's A80 G6230 11.4 score it's most likely with a 780MHz frequency (albeit I've yet to see that kind of frequency in a shipping device).

I'm still waiting to see more indications for A8, but a GX6650 has only made sense to me from the perspective of what I'd expect Apple to pick. From the performance estimates, a GX6450 or some kind of custom G6630 inspired part makes more sense. Apple's marketing is nebulous enough, however, that it could easily be a GX6650 being promoted conservatively like with the A7 GPU marketing.

According to some happy go merry rumor mongerers it should be Apple's own GPU by now

 

[Lazy8s]

Won't be hard to distinguish an 8 from a 12 TMU part once some fill rate results are released.

I still can't agree with some of the sentiment here and from Ryan Smith that a GX6650 at a similar clock to A7's GPU would result in only a 50% performance increase nor that said configuration would score only around 20 fps in GfxBench 3.0 Manhattan offscreen. I am expecting a score in that range, though, from the A8, so I'm guessing Apple saved the die area and didn't pick a GX6650 for iPhone 6.

 

[mavere]

Geekbench details for iPhone 5s vs iPhone 6.