I was dumb enough not to cross reference iPad4-iPad3 and iPhone5. If you compare all three of them, the A6 does also somewhat better than A5X but obviously not to the same degree as A6X.
Next-Gen iPhone & iPhone Nano Speculation
- Thread starter Arun
- Start date
I was dumb enough not to cross reference iPad4-iPad3 and iPhone5. If you compare all three of them, the A6 does also somewhat better than A5X but obviously not to the same degree as A6X.
GLBenchmark 2.5 does appear to do a relatively decent job of roughly representing modern mobile game and some other graphics intensive workloads, and the A6X speed-ups reflected in the benchmark haven't overstated the frame rate improvement I've seen in the few apps I've tested on the fourth-gen iPad any more than synthetics almost always overstate the improvement (due to their unrealistic balance of a workload's limiting factors). So, Apple's decision to achieve the performance doubling primarily through more ALUs seems to show that they have their finger on the pulse of graphics requirements fairly well (although the relationship is two-way due to the influence on app design that Apple's SoC decisions have).
Using 554s instead of double-clocked 543s leaves headroom for the eventual G64xx powered SoC to make important improvements to pixel fill in addition to the other important metrics and functionalities. The die area analysis re-emphasizes the relative bargain ALUs can represent and the relative premium TMUs can cost, and one conclusion to be had in the bigger picture is that Apple isn't really close to bumping into limits that would prevent them from accelerating their focus on graphics in future SoCs. A6X is less of a beast than A5X, relatively speaking, even with the additional graphics functionality which will go unused.
Using 554s instead of double-clocked 543s leaves headroom for the eventual G64xx powered SoC to make important improvements to pixel fill in addition to the other important metrics and functionalities. The die area analysis re-emphasizes the relative bargain ALUs can represent and the relative premium TMUs can cost, and one conclusion to be had in the bigger picture is that Apple isn't really close to bumping into limits that would prevent them from accelerating their focus on graphics in future SoCs. A6X is less of a beast than A5X, relatively speaking, even with the additional graphics functionality which will go unused.
I've created a small comparison table of recent SoCs with Series5/5XT GPUs here: http://www.forum-3dcenter.org/vbulletin/showpost.php?p=9529636&postcount=806
It stands open to correction of course if anything should be wrong. Other than that I would post it here if me dimbass could figure out how to post a table in this forum properly.
And no I didn't use the 9th FLOP for the ALU throughput in those.
***edit: iPad mini results are up: http://www.glbenchmark.com/compare....ied_only=1&D1=Apple iPad mini&D2=Apple iPad 2
Break even with iPad2, which means most likely same 1000/250MHz frequencies. Nothing spectacular to see here, moving on.
It stands open to correction of course if anything should be wrong. Other than that I would post it here if me dimbass could figure out how to post a table in this forum properly.
And no I didn't use the 9th FLOP for the ALU throughput in those.
***edit: iPad mini results are up: http://www.glbenchmark.com/compare....ied_only=1&D1=Apple iPad mini&D2=Apple iPad 2
Break even with iPad2, which means most likely same 1000/250MHz frequencies. Nothing spectacular to see here, moving on.
I find it difficult to assess just how great the real-life improvement is. Everything retina-enabled I've got simply feels smooth on the iPad4.
You're completely right in pointing out that app design obviously is dependent on the SoCs that Apple uses. The SGX543MP will be used by an awful lot of devices, with the lower bound on graphics grunt/pixel defined pretty much by the iPad3 which was very low volume, with the more relevant lower bound being the iPhone4s/iPad2/iPadMini level. At present the iPhone5 is pretty much the only outlier, but it is bound to be a high volume one. A very console like environment in its consistency. I wonder if there is space for one more SoC generation using 5 series IP.
What I find difficult to assess in terms of graphics performance is real world fill rates - to what extent the TMUs/ROPs are limited by available bandwidth both in the theoretical case, but also when the bus is also used by the rest of the system in a "typical" way for game code. One reason Apple went for increasing ALU resources could for instance be that improved memory controller and faster LPDDR2 alleviated a bottleneck for TMU/ROP utilization and that this provided enough of a boost to these parts that the real life balance was pretty much maintained.
Low volume? I don't have exact sales figures and I'm pretty sure there are a lot less iPad sales than iPhone sales. That said, I think there have been around 30 million iPad 3s sold since its launch. That's just my rough guess, but I don't think I'm too far from the truth. It does mean that there are quite a lot of iPad 3s, not what I'd consider low volume. Could be that you meant something else, not sales.
It's probably less than a litre.
Apple sold 11.8 million iPads in the 2nd quarter, 17.0 million iPads during the 3rd quarter and 14.0 million iPads during the 4th quarter of 2012.
In the 1st quarter they sold 15.43 million iPads (these were mainly iPad 2 and perhaps still some refurbished iPads as the 3rd generation iPad wasn't released until march 16th).
Of course we don't know the ratio between the iPad 2 and the 3rd generation iPad.
In the 1st quarter they sold 15.43 million iPads (these were mainly iPad 2 and perhaps still some refurbished iPads as the 3rd generation iPad wasn't released until march 16th).
Of course we don't know the ratio between the iPad 2 and the 3rd generation iPad.
ltcommander.data
Regular
With 4x the GPU performance over the iPad 2 and a 4x pixel count increase, even though the iPad 4 and SGX554MP4 are new, the GPU performance (at native resolution) of the iPad 4 would be part of the iPhone 4S/iPad 2/iPad Mini bracket which really does make it the high volume performance sweet-spot to target development for. With the iPad 3 having half the GPU performance at native resolution of the iPhone 4S/iPad 2/iPad Mini/iPad 4 and the iPhone 5 having twice the GPU performance at native resolution, the entire Series 5XT iDevice family only has a 4x performance spread from top to bottom, which is pretty good considering the different years and form factors those constitute.
If the TMU/ROPs weren't a bottleneck for the iPad 4, it begs the question why Apple didn't go with a SGX554MP2 instead of the SGX543MP3 in the A6 for the iPhone 5 where resolution is even less of an issue? The iPhone 5 would now appear to have excessive TMU/ROP resources. Die area may have been a slight concern since at 8.7mm2 per core a SGX554MP2 would be larger than the 5.4mm2 per core SGX543MP3 at 17.4mm2 vs 16.2mm2.
I'm guessing the bigger issue may have been risk. The A6 came with a new CPU architecture and the 32nm process would have been new for much of it's design phase, plus with a yearly update cycle, the A6/iPhone 5 had to ship on time so sticking with the known SGX543MP made sense. Given there's only a month between release, the A6X would have been developed alongside the A6 rather than developed from it and would have been a high risk chip given the simultaneous shift in CPU and GPU and the new-ness of the 32nm process especially for a larger chip. However, that risk was likely acceptable since the iPad 4/A6X probably didn't have to ship this year if things didn't work out.
Relating back to the possibility of one more Series 5XT chip, I think the 2013 flagship iPad and iPhone will most certainly be an A7 with a Series 6 GPU. Especially with the usual March launch window for the iPad being unlikely due to it being less than 6 months since the last refresh. However, it's possible that rather than using the A6 or a straight 28nm shrunk A6 with SGX543MP3 for the 2013 iPad Mini and 2013 iPod Touch they go with a 28nm A6 with SGX554MP2 (an A6S?). Although if Apple really does introduce a retina iPad Mini next year rather than wait for the 3rd generation, they'll probably have to go with a 28nm shrunk A6X with low volumes forcing the iPod Touch 6 to stick with a regular SGX543MP3 A6 shrunk to 28nm. I'm guessing that the 28nm process will be a prerequisite to getting the A6/A6X into the smaller form factor/battery of the iPad Mini and iPod Touch while maintaining the same clock speeds as the iPad 4 and iPhone 5 in order to reduce fragmentation.
Why would they?
No idea; otherwise I wouldn't ask.
ltcommander.data
Regular
Along the vein of wco81, the differences between OES2.0 and DX9L3 may be easier to identify, but is it clear what the difference in functionality between a base Series 5XT and a DX9L3 capable Series 5XT is? Even the base Series 5XT incorporates additional functionality above OES2.0 which may or may not yet be exposed in OES extensions. There seem to be extensions for the base Series 5XT for expanded texture support (floating point, RG, depth, more NPOT texture filtering and mipmapping, etc), vertex array objects, sync objects, multi-sample renderbuffers, buffer subrange mapping, boolean occlusion queries, etc. (I'm just listing functionality already enabled by extensions in the Series 5XT compared to the new features list for OES3.0.) Multiple render targets is supported in Series5/5XT, but isn't enabled in drivers yet. Presumably there is still other untapped functionality there so I wonder how big the feature gap between the SGX543MP and SGX544MP/SGX554MP really is?
It appears that Apple felt a strong need to get the CPU performance of the ipad in line with the new iphone (which is understandable), and while they were at it, they improved the GPU performance too but certainly not to the extent of a true doubling of SGX 543MP4 performance that many were expecting. Based on Anandtech's measurements, the SGX 554MP4 (used in the A6X SoC) has the following onscreen performance advantages over the SGX 543MP4 (used in the A5X SoC):
Fill Test: 15.6%
Triangle Texture Test: 65.4%
Triangle Texture Test - Fragment Lit: 83.5%
Triangle Texture Test - Vertex Lit: 44.0%
GLBenchmark 2.5 Egypt HD: 99.5%
GLBenchmark 2.5 Egypt Classic (offscreen): 52.4%
So while this is a very significant increase in GPU performance, a true 2x improvement would only be in the most optimistic of cases. And as usual, Apple is using an SoC die size that is substantially larger than A5X when normalized for the 45nm fabrication process. I suppose that Apple could have simply die shrunk the A5X SoC and given consumers a thinner/lighter/cooler tablet with even better battery life, but that was not a priority for them due to the competitive threat of the Exynos 5 Dual SoC, due to the release of the A6 SoC, and due to the need to drive a relatively high resolution display.
The Mali T604 was not able to keep pace with the SGX 554MP4. It will be really interesting to see what NVIDIA, Qualcomm, and ARM bring to the table over the next few months. Anand suspects that NVIDIA is really the only one who can challenge Apple in the very near future. On a side note, Anand posted a video a few weeks back (?) where AMD's Hondo was benchmarked using GLBenchmark 2.5 HD. The onscreen result was 36 fps, but no information was given about screen resolution.
Last edited by a moderator:
A table from Microsoft showing the differences between DX9 and DX9L3 should show the majority if not all differences. The majority should be texturing related from the top of my head.
http://www.glbenchmark.com/compare....tified_only=1&D1=Apple iPad 4&D2=Apple iPad 3
The difference in GLBenchmark2.1 between iPad4 and iPad3 is at 1.5x and in GLBenchmark2.5 at 1.94x. Whether upcoming games will concentrate more on texel fillrates or on floating point performance for example remains to be seen, but I have severe doubts that Kishonti is standing on the wrong foot with its estimates with 2.5. But even if things would be the other way around in that example there aren't any SFF GPUs at the moment available that yield over 2.0 GTexels/s fillrate in real time. The SGX554MP4 in the A6X is IMHO clocked at 280MHz which gives it a theoretical peak fillrate of 2.24 GTexels. The Mali T604 is clocked at 533MHz which means 2.13 GTexels/s peak with an onscreen (2560*1600) fillrate measured in 2.5 of 1.35 GTexels/s.
Whether T604 remains at the point it is right now or might increase performance via future driver updates is something we'll see in due time. However that should not mean that there might not appear performance improvements for other solutions via drivers in the meantime. In other words T604 is not a closed case yet in my book.
As for ALU:TMU ratios they'll continue to grow in upcoming SFF GPUs and the 554 is just a first sign of where things will be leading. IF Apple should go with a Rogue GPU in its next tablet iteration that's the solution Wayne and others will have to compete with an not the current solution.
On an unrelated note for AMD's Hondo I honestly hope that there's much more than just 36fps under it's hood, otherwise it will be more than disappointing even if it's at 2560*1600.
Here is what arstechnica had to say about the ipad 4 when comparing it to the ipad 3 (with both using the newest iOS software):
So at least at this point in time, it is possible that the improvement in CPU performance is a bit more noticeable than the improvement in GPU performance when moving from ipad 3 to ipad 4.
Regarding Hondo, I read somewhere that AMD is focusing on 1080p tablet resolutions, so that would imply that the screen resolution is 1920x1080 in the GLBenchmark 2.5 onscreen test video posted by Anand. 36fps is a bit less than A6X, but still not too shabby. Moving to 28nm fabrication process and a newer GPU architecture should make for a significant improvement, whenever that happens...
arstechnica said:
So at least at this point in time, it is possible that the improvement in CPU performance is a bit more noticeable than the improvement in GPU performance when moving from ipad 3 to ipad 4.
Regarding Hondo, I read somewhere that AMD is focusing on 1080p tablet resolutions, so that would imply that the screen resolution is 1920x1080 in the GLBenchmark 2.5 onscreen test video posted by Anand. 36fps is a bit less than A6X, but still not too shabby. Moving to 28nm fabrication process and a newer GPU architecture should make for a significant improvement, whenever that happens...
No surprise there; if you consider that the 4 gets already around 133fps offscreen in 2.1 the dilemma would be if you'd want to invest hw resources in that direction or into the direction future games will go. Albeit not directly comparable what was one of the keys for the success of G80 about 6 years ago? It was the fastest GPU for anything DX9 by far at the time while at the same time was on the right track for anything DX10.
In the given case once you're at N performance point for current applications it would be nonsense to invest resources for getting =/>200 fps in 2.1, when you can achieve twice the performance in 2.5 with the lowest possible hw resources.
I am not even bothering to compare it to A5X or A6X in the back of my mind. It just seems to be an indication for something I said in the past: for DX11 class SFF GPUs anything above 28nm sounds like nonsense to me. Despite some AMD fans always liked to rant how AMD would kill anything in sight if AMD would had entered the tablet market earlier.
DX11 costs a crapload of additional transistors and there's still a vast difference in that regard even to the DX9L3 SGX554MP4. And since I personally prefer apples to apples comparisons (no pun intended LOL) I'm afraid that if the Hondo scores reflect reality and AMD hasn't anything vastly faster in store for its successor, it's going to be a very tough battle against Rogue.
Hondo uses rather simple x86 cores at 1GHz, a 64-bit memory system, and run the 80 "radeon cores" at undisclosed (presumably quite low) clock speeds in order to achieve its 4,5W TDP. Essentially, it is a very cut down PC part.
There is every reason to believe that a SoC like the A6x will walk all over it in terms of performance.
Then again, Hondo doesn't compete directly against the ARM SoCs, it exists as a low cost alternative to Intels offering in the x86/Windows space. In that respect it's a godsend for the hardware manufacturers in that segment, since otherwise they would be completely at the mercy of Intel. Remains to be seen how many are interested in running x86 legacy code on tablets - to me it seems like the netbook crippleware experience all over again - but Hondo is the lowest cost ticket into that market you can find.