Apple A8 and A8X

Now that also battery lifetime and long term performance tests have floated in for the A8X, there doesn't seem any noticable throttling on it and it seems to last a tad over 4 hours in T-Rex.

 

If I'm reading Kishonti's website correctly, they have, or are planning, support for comparing performance with Metal for both graphics and compute?

http://kishonti.net

 

No throttling in Gfx-Bench T-Rex. But T-Rex is also a pretty favorable Bench for Gpus with FP16 units as it's only using FP16 if i ain't wrong.

German Notebookcheck made it throttle, when first using "Relative Benchmark" and then Cpu/Gpu Benches.
http://www.notebookcheck.com/Test-Apple-iPad-Air-2-A1567-128-GB-LTE-Tablet.129027.0.html

CPU throttled 30% by using it afterwards in Geekbench and GPU 15% in 3dmark. But i don't really know whether Relative Benchmark is showing realistic Powernumbers or just numbers like Furmark, which won't happen normally. But at least with this Bench it's also getting pretty hot (43°C) and uses quite a bit of power.

 

Kudos to Kishonti then for creating a public benchmark that supposedly favours only PowerVR Rogue GPUs (where's the sarcasm smiley when you need it....) T-Rex is alpha test bound, that's something that someone would have to be blind not to notice.

If I take a G6430 Rogue and a G6400 Rogue (where the first has FP16 units, while the latter hasn't) and compare the TRex results, what would you suggest will I get?

Kishonti should have used both TRex and Manhattan for the long term performance tests, but that has been mentioned before.

All devices eventually throttle; the point is rather where and by how much for each case. Other than that I'd love to know where the hell they've substantiated that the A8X contains a G6650.

 

The reduction in Geekbench 3 performance due to throttling in the "stress test" is quite substantial in iPad Air 2. Do you have any idea how other tablets perform in this same "stress test"? My guess is that most thin fanless tablets will struggle with this type of test.

 

You're wrong.

 

Ah, tactful .

 

Someone should cap the iPadAir2 at 30 fps in TRex and nearly double the battery lifetime under 3D.....

 

Surprised that Anandtech hasn't reviewed the air2 yet, seems a bit slower than in anand's days.

 

No one's released a die shot yet either. They are probably working with Chipworks on that to have some physical evidence for their A8X analysis.

 

Anandtech review is out

 

The jump in performance on Google's Octane V2 benchmark for the Air 2 vs iP6 is surprising. As Octane is primarily a javascript benchmark, I assumed it was mostly single-core bound, does Octane benefit from the Air 2's 3rd core? The Air 2 was launched with iOS 8.1, whereas the iPhone was tested on an earlier version of iOS 8, which could be the difference. Anybody with a 6 or 6 plus running iOS 8.1 care to run Octane v2 and post their scores?, thanks.

http://octane-benchmark.googlecode.com/svn/latest/index.html

 

Want to be they'll have to revisit their review once chipworks is done with the A8X?

 

Perhaps on top of browser tuning, Octane is benefitting more than the other tests from the better memory performance?

 

Anandtech confirms from a die shot that the A8X contains an 8 cluster PowerVR 6XT GPU, not a 6 cluster as they previously surmised.
http://anandtech.com/show/8716/apple-a8xs-gpu-gxa6850-even-better-than-i-thought

 

A8X is a 6850. Helps account for a good part of those extra 1B transistors.

http://www.anandtech.com/show/8716/apple-a8xs-gpu-gxa6850-even-better-than-i-thought

 

It's a miracle!

For that die shot mock up: why would the block named "shared digital logic" in the middle of each quad core double with an 8 cluster config?

***edit: after actually reading through the text, it very well may be the case considering that Apple in a relative sense seems to have "doubled" somehow the GX6450.

If Anandtech estimates: total A8X die area at ~125mm2, ~3b transistors that means a rough transistor density of 24M/mm2@20SoC. At the estimated 38mm2 for the "GXA6850" that could mean over 900M transistors (holy shit!).....

 

There you go Since they just copied and duplicated the GPU part of the A8 die shot, the labeling is most likely not accurate.

 

Actually it is. If you were to look at the die shot, there's a strip of shared logic running right down the middle of the GPU, clearly different from the 8 USCs. I have to assume the various backends and frontends (ROPs, geometry, etc) were scaled up as well as the number of USCs.

 

Thanks for clearing that up