Apple A9 SoC

[Laurent06]

This document has some information about workload size. Mobile version has smaller workloads than desktop version.

http://www.primatelabs.com/geekbench3/workloads.pdf

Some subtests have very small workloads that fit even 1MB cache, such as BZip2 and Lua. Some are larger than 3MB, such as JPEG (1.6Mpix).
Lua test is particularly interesting because it's considered a "hard" test for CPU core (the data set is so small that cache generally does not matter much). A9 performs very well on this subtest and that suggests some improvement on CPU core other than cache (probably branch predictor).

I'm afraid you're confusing the size of inputs with the size of data that are being manipulated at runtime.

 

[pcchen]

I'm afraid you're confusing the size of inputs with the size of data that are being manipulated at runtime.

Unless they are using some particularly bad algorithms for Bzip2 and Lua, the working data size wouldn't be more than 10 times of the input data.

 

[SKYSONY]

I am far from being the tech experts you guys are. But I have to say, following as I do the benchmarks of the new phones, that the A9 seems to be more powerful than this leak suggested back in August:

http://appleinsider.com/articles/15...ies-chips-benchmarked-show-20-30-jump-over-a8

If the leaked Snapdragon 820 benchmarks are any close to the final ones, it seems the A9 will be way beyond the 820 in single-core performance, and very close in multi-core. That´s something amazing taking into account the A9 still uses two cores.

But I wonder, how do you think it will compete with the new Exynos? If those benchmarks are to be believed, it seems like a true beast. Not sure if they will really release it with 2,4 Ghz, but 8000 in multi-core geekbench is quite something.

 

[Laurent06]

Unless they are using some particularly bad algorithms for Bzip2 and Lua, the working data size wouldn't be more than 10 times of the input data.

For bzip2, I agree (there is even enough information in the pdf to reproduce something close to Geekbench), but for Lua IMHO it's much more difficult to guess.

 

[pcchen]

For bzip2, I agree (there is even enough information in the pdf to reproduce something close to Geekbench), but for Lua IMHO it's much more difficult to guess.

Well yes, but the JSON data used for the Lua test is just too small (just 169KB).
Another reference point is from the comparison between iPad Air 2 and iPhone 6 Plus (https://browser.primatelabs.com/geekbench3/compare/3550205?baseline=3550306). Despite with higher clock rate, a larger L2 cache and double memory bandwidth, iPad Air 2 does not perform much better than iPhone 6 Plus on these two tests.

 

[Grall]

L2 is bigger on A8X than on A8? I did not know that. What sizes are they, specifically? Thanks.

 

[iMacmatician]

2 MB and 1 MB respectively.

 

[Turbotab]

I wonder if Apple has come form of tiered NDA in place for early reviews? They have all been extremely light on objective data / benchmarking. Apart from GFXbench, I mostly want some form of decent javascript benches like Kraken or Octane 2.0, as they give an understanding of how well overall hardware / software integration is for a commonly used workload, i.e. web browsing. The benchmark examples I gave, are quick to setup and complete within minutes, so I'd be surprised if it was just journalistic ignorance or laziness that prevented their inclusion, but humans never fail to depress me.

 

[Turbotab]

I asked a NZer whom had just received their 6s, via reddit to run GFXbench, which they kindly did, but not all the way through.
Manhattan Offscreen - 37.6 fps vs iPad Air 2- 39.5 fps vs iPhone 6 - 19.2 fps

Texturing - 1080p offscreen fillrate - 6526 Mtexels/s vs iPad Air 2 - 6569 Mtexels/s vs iPhone 6 - 2320 Mtexels/s

http://imgur.com/a/x624w

 

[Turbotab]

I asked a NZer whom had just received their 6s, via reddit to run GFXbench, which they kindly did, but not all the way through.
Manhattan Offscreen - 37.6 fps vs iPad Air 2- 39.5 fps vs iPhone 6 - 19.2 fps

Texturing - 1080p offscreen fillrate - 6526 Mtexels/s vs iPad Air 2 - 6569 Mtexels/s vs iPhone 6 - 2320 Mtexels/s

http://imgur.com/a/x624w

NB - I believe that GFXBench has a glitch regarding iPhone 6 Fill Rate (Texturing) score, it's too low. 6 Plus is 3283 MTexels/s, even a 5s gets 1928 Mtexels, the real number is likely nearer the 6 plus's 3200.

 

[anexanhume]

2GB LPDDR4 is confirmed https://www.ifixit.com/Teardown/iPh...social&utm_source=app.net&utm_campaign=buffer

 

[Ailuros]

I asked a NZer whom had just received their 6s, via reddit to run GFXbench, which they kindly did, but not all the way through.
Manhattan Offscreen - 37.6 fps vs iPad Air 2- 39.5 fps vs iPhone 6 - 19.2 fps

Texturing - 1080p offscreen fillrate - 6526 Mtexels/s vs iPad Air 2 - 6569 Mtexels/s vs iPhone 6 - 2320 Mtexels/s

http://imgur.com/a/x624w

The Driver Overhead 2 results are interesting in that one compared to the A8; an increase of almost 80%. For those not aware what the low level test does:

This is an enhanced version of the original Driver Overhead test found in GFXbench3.0 and approximates the graphics driver's CPU load when running Manhattan high level test. It renders the same amount of geometry twice to add a glow effect, using lots of draw calls, changing rendertargets, vertex formats and shaders as well. The test also alters graphics states like depth test and blending between blocks of draw calls. Required minimum API: OpenGL_ES3.0 / OpenGL4.1.

 

[tangey]

I asked a NZer whom had just received their 6s, via reddit to run GFXbench, which they kindly did, but not all the way through.
Manhattan Offscreen - 37.6 fps vs iPad Air 2- 39.5 fps vs iPhone 6 - 19.2 fps

Sept 9, Apple said A9 was X1.9 A8 on the gpu.
19.2x1.9=36.48.

If that test above is accurate, then looks like it's very consistent with Apples claim.

Apple also said A9X was x2 A8X. 80 fps in Manhattan looks likely

 

[Voxilla]

Looks like the A9X is everything the A9 is but x2.
So 4 CPU cores, 6 GPU blocks.
Probably higher clock too, like 2 Ghz.

 

[Ailuros]

In order to avoid mistunderstandings for each GPU block there's 2 GPU clusters

 

[Ailuros]

Sept 9, Apple said A9 was X1.9 A8 on the gpu.
19.2x1.9=36.48.

If that test above is accurate, then looks like it's very consistent with Apples claim.

Apple also said A9X was x2 A8X. 80 fps in Manhattan looks likely

It's 43.5 fps under Metal for the Air2. I don't expect though the 1st ULP SoC spot to last for long, since I'd be very surprised if NV won't pre-announce Parker.

 

[Entropy]

Looks like the A9X is everything the A9 is but x2.
So 4 CPU cores, 6 GPU blocks.
Probably higher clock too, like 2 Ghz.

I don't think we will see four cores. Apple indicated a 70% improvement in CPU performance for the A9 vs. the A8, effectively inside the same thermal/power envelope. The A9x is supposed to yield an 80% improvement, presumably using the same metric as the A9, over the A8x. Given that the thermal/power envelope probably has increased a bit over the environment the A8x is in, due to both a larger body to act as heatsink and a larger battery to supply power yielding the same overall battery life, it is not unreasonable to assume that we are simply looking at a slightly higher increase in clock for the A9x over the A8x. Something like 2.0GHz would be my guess as well. (1.8/1.4=1.286 2.0/1.5=1.33)

Judging from the A9, this might allow the score to increase roughly 1.6 times in Geekbench terms, which would yield 2900(single)-7400(multi). We'll see, once the iPad Pro starts trickling out. This would also allow some rather straightforward enhancements for future generation products.

 

[Lazy8s]

With that texel fill, the GPU target might be 567 MHz to 600 MHz.

PowerVR's effectiveness in GFXBench's fill tests have gone from extremely high to so-so as the benchmark has evolved. Given these results, I'm guessing the A9 SoC, in at least the Plus version of the 6S, is GT7600@600MHz + 2xTwister@1.8 GHz.

Perhaps the GPU in the regular 6S could be clocked a bit lower; maybe not.

 

[Ailuros]

With that texel fill, the GPU target might be 567 MHz to 600 MHz.

PowerVR's effectiveness in GFXBench's fill tests have gone from extremely high to so-so as the benchmark has evolved. Given these results, I'm guessing the A9 SoC, in at least the Plus version of the 6S, is GT7600@600MHz + 2xTwister@1.8 GHz.

Perhaps the GPU in the regular 6S could be clocked a bit lower; maybe not.

I'm not so sure the frequency is higher; it's IMHO either on the same level (533MHz) or slightly lower. Have a look here:

https://gfxbench.com/compare.jsp?be...U&hwname1=Apple+A9+GPU&D2=Apple+iPhone+6+Plus

That's not a pure texel fillrate per se with the healthy portion of alpha blending it contains. Look at the alpha blending results vs. the fillrate results. Also if you take the A9 GPU fillrate results and parallelize them to the A8X GPU fillrate results you'd be scratching your head even more considering that you have 16 TMUs in the latter and 12 in the GT7600. Maybe you should think pixels from the back end * frequency instead of TMUs * frequency for the 3.0 and 3.1 fillrate tests.

 

[iMacmatician]

From Chipworks: "Inside the iPhone 6s."

The A9, named APL0898, is about 8.7 mm x 10.7 mm = 94 mm^2.

The APL0898 size fits with an 80% shrink on the A8, and seems to have 8 MB cache - and our first look leads us to believe that our sample is from Samsung.