Apple A9X SoC

[tangey]

Its expected that the iPad pro will become available to order end of his week. Although the volumes of A9X will be tiny compared to A9, I think the A9X will be every bit as interesting as the A9 from a technical viewpoint.

To start off, I wonder has IMG inadvertently revealed the GPU, in a slide pack from their recent USA summit.

https://imagination-technologies-cl...mmit-2015/Graphics-FromWearablesToServers.pdf

Slide 8 shows the "performance" variants used in different sectors of the market. Unless I'm mistaken,they have used the GPUs in each of the previous Apple socs for both the "high end smartphone" and "high end tablet" segments of that graphic. The most recent smartphone GPU is shown as the G7600 (in the A9). The most recent tablet GPU is listed as the G7800+, a variant that has not made it to market anywhere. It looks to me that this will be the iPad pro GPU, With the + implying it's a customised variant. The last time I remember the "+" designation being used in relation to an IMG GPU, was when Sony used a custom version of the 5 series in the PSP.

Less interesting, but also useful to note, is that as far as I am aware, that chart is the first instance of IMG formally accepting the existence of a GX6850. The Anantech article on the A8X suggested the GPU was two GX6450s which Apple had some major input in doing the customisation. Perhaps this chart might lean to more of a suggestion that it was a customisation done by IMG.

 

[Ailuros]

I'd be quite surprised if tha A9X GPU is anything else but a mirrored GT7600. With the 7800 representing 8 clusters and the 7900 16 clusters, there's not much else left to call it but 7800+ or 7850 etc. to imply that it has more than 8 and less than 16 clusters.

While I just read the pdf, Vulkan designed for TBDRs? Seriously? Can someone be so kind and clarify?

 

[iMacmatician]

So if the smartphone and high-end tablet GPUs on page 8 are in fact the iPhone and iPad GPUs from (late) 2012 through 2015, then we have the following GPUs:

iPhone 5   A6  544MP3     iPad 4      A6X  544MP4
iPhone 5s  A7  G6400      iPad Air    A7   G6430
iPhone 6   A8  GX6450     iPad Air 2  A8X  GX6850
iPhone 6s  A9  GT7600     iPad Pro    A9X  GT7800+

I see two differences between the above lineup and the guesses/claims from places like AnandTech. According to those sources, the A6 has the 543MP3 and the iPhone A7 has the G6430. Chipworks states that the iPad A7 die is the same to the iPhone A7 die, so I don't think that the two A7 variants use different GPUs. Perhaps the GPUs mentioned in the Imagination slide are representative of the perf/power/price in the various market segments and years, and are not necessarily identical to Apple's GPU choices.

 

[tangey]

Yes, it's generally felt that the A7 is the same across both iPad and iPhone. Might need to rethink the table and say that it was shown the highest implementation used in phones/tablets each year (but was there a 6430 in the field anywhere in 2013...Intel ?/ mediatek ?). Regardless of that, given the GT7800+ is listed as 2015, Apple is the only one that could have anything like that this year.

 

[Ailuros]

So if the smartphone and high-end tablet GPUs on page 8 are in fact the iPhone and iPad GPUs from (late) 2012 through 2015, then we have the following GPUs:

iPhone 5   A6  544MP3     iPad 4      A6X  544MP4
iPhone 5s  A7  G6400      iPad Air    A7   G6430
iPhone 6   A8  GX6450     iPad Air 2  A8X  GX6850
iPhone 6s  A9  GT7600     iPad Pro    A9X  GT7800+

I see two differences between the above lineup and the guesses/claims from places like AnandTech. According to those sources, the A6 has the 543MP3 and the iPhone A7 has the G6430. Chipworks states that the iPad A7 die is the same to the iPhone A7 die, so I don't think that the two A7 variants use different GPUs. Perhaps the GPUs mentioned in the Imagination slide are representative of the perf/power/price in the various market segments and years, and are not necessarily identical to Apple's GPU choices.

I wouldn't take it for granted that Kristof meant for all spots just Apple. The Samsung Exynos 5420 had a 544MP3 & INTEL Merrifield a G6400, which both appeared in high end smartphones.

 

[tangey]

For sale on the apple website from Wed. So not be too long before I have my hands on that GT7800 GT7600x2 or whatever it turns out to be.

Anyone seen any confirmed UK pricing ?

 

[Lazy8s]

Labeling the A8X's GPU as a "GXA6850" instead of the obvious GX6850 label, while only a subtle difference, implies a misunderstanding of Imagination's relationship with its licensees. We went through all of this last year when Anandtech's piece on the A8X was first discussed.

While Imagination does support scaling core counts and not just cluster counts in its recent GPUs and Apple's own GPU team are likely easily capable of piecing together a multi-core implementation of PowerVR GPU IP, Apple uses its extensive GPU engineering resources to produce industry leading implementations of Imagination-developed leading edge PowerVR GPU designs. They have no reason to try to "mirror" existing layouts and double the cluster count themselves for their own variants when Imagination would gladly produce a balanced part at the desired scale for them and any other licensee.

Considering Apple was the only practical customer for such a part last time and Apple demands their partners reveal as little info about their relationship as possible, Imagination has had little reason to publicly promote the GX6850. Still, it's an Imagination design and an Imagination part, not some hybrid Apple/Imagination IP design.

 

[Lazy8s]

As to the comment about Vulkan being designed for TBDRs, it has a few provisions with its render pass feature that allow the GPU/driver to understand how render targets will be used and to make scenarios where tile buffers might get flushed more predictable and avoidable.

Vulkan is not designed specifically for TBDRs but doesn't make them have to work around the confines of the API as much as past APIs have.

 

[Laurent06]

Don't know if this is genuine or not: http://browser.primatelabs.com/geekbench3/4149721

2 cores @2.16 GHz with 4 GB of RAM.

 

[tangey]

A basic preview of the A9X is up on anandtech

Sticks with the A9s dual CPU with around 22% clock increase. 128-bit bus width, giving double the bandwidth of the A9 @51GB/sec

GPU gives approx x2 perf of A9 and is suggested as either 10cluster with upclock, or 12cluster with similar clock to A9,

 

[Voxilla]

Dual core for the A9X at high clock of 2.26Ghz is unexpected and slightly disappointing.
The triple core A8X 1.5Ghz is reduced as being an 'anomality'
For single threaded applications it will be a nice improvement, but for multithreading not so much.

 

[Entropy]

2.25/1.85=1.216
22% over the A9 if no other architectural improvements are present. Which there may well be. The absence of other than the geekbench memory scores is a bit suggestive.
It sure seems to pack a punch.

 

[Entropy]

Having googled for the fanless Surface Pro 4 alternative, Intel Core m3-6Y30 and how it performs, it seems the A9x will be quite competitive with Skylake under similar conditions. Look forward to the upcoming reviews. I hope Apple did more than increase the clocks and memory bandwidth of the A9, although that would seem sufficient to validate their performance claims.

 

[Voxilla]

Look forward to the upcoming reviews. I hope Apple did more than increase the clocks and memory bandwidth of the A9, although that would seem sufficient to validate their performance claims.

The claim was 1.8x CPU performance of the A9X versus the A8X.
In some specific single threading tests that may be true, but for multi threading most definitely not.
Comparing 2 multi-core CPUs based on single core performance is very biased IMHO.

 

[asendra]

Or, looking at it this way, you can always add more core to improve multi core performance, single core performance, on the other hand..
It definitely makes the biggest impact in day to day usage. And I say this as an iPad Air 2, iPhone 6s owner.
I think is the right trade-off, specially considering how little usage that 3-4th core would have currently in iOs software.

 

[Ailuros]

GPU gives approx x2 perf of A9 and is suggested as either 10cluster with upclock, or 12cluster with similar clock to A9,

Anything but mirroring a GT7600 sounds way too complicated; how do you get to 10 clusters with a custom design anyway? Take a 7600 and glue a 7400 next to it? LOL

 

[Arun]

Anything but mirroring a GT7600 sounds way too complicated; how do you get to 10 clusters with a custom design anyway? Take a 7600 and glue a 7400 next to it? LOL

How do you get to 6 clusters? Take a 7400 and glue a 7200 next to it?

 

[tangey]

Anything but mirroring a GT7600 sounds way too complicated; how do you get to 10 clusters with a custom design anyway? Take a 7600 and glue a 7400 next to it? LOL

Don't ask me, ask the anandtech guy, they are his thoughts

Shipping status is indicating I might get my pro tomorrow. Is there any low level glbench test that would help in identifying how many cores are being used ?, fill rate ?

 

[Arun]

The low-level tests in newer GFXBench versions aren't very useful in an absolute sense (e.g. fillrate tests includes some trilinear so it's impossible to hit peak) but they are OKish to compare performance within the same architecture.

 

[zed]

http://arstechnica.com/apple/2015/1...-all-the-virtues-and-limitations-of-ios/4/#h2
happened sooner than I thought, but the fanless ARM chips have basically caught their fanned intel brethren. Like I said on these forums a couple of years ago, I wouldnt be surprised if we start seeing macbooks with ARM inside