Next-Gen iPhone & iPhone Nano Speculation

[Lazy8s]

I suspect the clock rate of the SGX535 was raised from 150 MHz to 200 between the A3 and A4, though my suspicion comes mainly from observations of comparative app performance.

My proposed solution to the display issue would be to sell iPhones with various screen sizes like they do with flash storage and case color yet not compromise their vision of the ideal 3.5" 960x640. As long as they continued to keep the focus on their ideal model, they'd be free to go crazy with a super-sized iPhone and have no real obligation to rebalance the platform's resolution/GUI to accommodate it.

Apple could maintain that the Classic model is perfect, yet throw out a 5 incher (with no difference in resolution) for those who have to have a big screen and would be willing to trade away the pixel density. 960x640 would still be respectable, and no extra work nor support would be needed by developers.

 

[silent_guy]

I am separating "apps" from "games" which rarely need to optimize anything but the worse performing lines of code since nothing runs in a time-sensitive 30/60fps loop but is event driven. I am also not arguing that apps are never multi-threaded, but most apps do not have a lot of incentive to write a ton of multi-threaded code whereas a game would like to squeeze out as much performance as possible.

Ok, fair enough.

That said, if I look at the most common apps I run on my phone, it's Safari, mail, messaging, an RSS reader and a forum reader. The latter two are very often used to open web pages.

'Executing' a web pages is probably one of the few tasks that can make efficient use of multiple cores (side loading assets, rendering, running JavaScript etc.) Since, for many people, it's such a big part of the daily usage of a phone, 4 cores may still be an ok trade-off.

 

[tangey]

I suspect the clock rate of the SGX535 was raised from 150 MHz to 200 between the A3 and A4, though my suspicion comes mainly from observations of comparative app performance.

glbenchmark indicates that the fill rate increased by 80% from 3GS to iphone4. Does this not suggest a similar difference between clock speeds ?

 

[Xmas]

Apple could maintain that the Classic model is perfect, yet throw out a 5 incher (with no difference in resolution) for those who have to have a big screen and would be willing to trade away the pixel density. 960x640 would still be respectable, and no extra work nor support would be needed by developers.

Except for those developers who actually care about the physical size of controls on the screen, not the pixel size. I would hope that most developers design the UI layout for physical size, though the question is obviously at which scale factor they would care enough to redesign. At least some games with on-screen controls would become quite hard to play at 1.43x.

 

[Ailuros]

I feel a bit iffy on the GPU scenario, but unless Apple changed the clock speed of the SGX 535 in the A4 compared to whatever is in the 3GS, it means they designed a chip to last two generations without anything more than CPU clock speed increases from die shrinks. Then again, none of the previous CPU/GPU designs could be used in a multi-core scenario, I could be totally wrong.

I don't really predict the iPhone screen to change radically unless they just decide "screw it, we're going with a 4.3" screen and 720p display". Highly unlikely, but you never know.

If Apple wouldn't serve also their tablets/iPads with the same SoC it wouldn't be much of a consideration. Instead since the advent of the iPad they seem to introduce a new SoC on a new generation tablet first and then later on re-use the same SoC with slightly lower frequencies for their smart-phones. Tablets come first, which also makes sense since the volumes aren't as big as for smart-phones obviously.

The iPhone screen estate/resolution might not change at all, but it's more than likely that the resolution will increase by 4x times on the iPad. Unless Apple starts developing two separate SoCs - one for tablets and one for smartphones - there's a real need for higher graphics performance as things look like at the moment for their next SoC.

For the time being since they're only serving smartphones and tablets with one SoC a 2nd lower end SoC doesn't make much sense as it'll just complicate things and raise cost unnecessarily.

 

[french toast]

The A6 the fastest? I can see that happening with regards to graphics performance, but not really in CPU performance. My guess would be something like a much higher clocked dual core Cortex-A9 and a higher clocked SGX543 MP2 or MP4 for the Apple A6. A well, we'll probably get to see what the A6 is like this quarter anyway.

I agree with this assumption, as some have previously mentioned, apples history dictates that they plan their die space for 2 generations, with a node shrink then just bump the clockspeeds up, faster ram(bandwith)/more ram.

Last generation A5 has plenty of performance left to use and is still the most powerfull graphics wise compared to tegra 3, 4470, and probably adreno 225(+drivers opt).
So all they would need to do is simply improve clocks, ram, bandwith to keep performance parity with A5 generation matched to the increased screen resolution.

 

[french toast]

As you guys are the pros, and im a bit of a noob! would you think that to increase bandwith easilly, they could adopt the LPDDR3 ram mated to their duel channel memory controller without any changes to the architecture/much difficulty?

 

[sebbbi]

if their next iPad truly goes up to 2048*1536, which is quite a brutal increase in resolution with the exact same GPU block it sounds more like they'll end up with less performance than on iPad2.

The 7x GPU performance improvement from iPad -> iPad2 was a huge jump (bigger than expected). There has been speculation if Apple really wanted to have 2048x1536 resolution already in iPad2, but that plan somehow failed late in the process (display price or availability wasn't where they wanted). 7x GPU improvement along 4x more pixels to render would have made more sense.

The current GPU should be enough for fluid web browsing and generic applications at 2048x1536. You would only want something beefier to run games... and I expect games to start rendering at non-native (lower) resolution if 2048x1536 and 2560x1600 (rumoured next gen Galaxy Tab resolution) become the norms in tablets. You don't need that much resolution in games (consoles are still rendering games at 720p).

 

[ltcommander.data]

The 7x GPU performance improvement from iPad -> iPad2 was a huge jump (bigger than expected). There has been speculation if Apple really wanted to have 2048x1536 resolution already in iPad2, but that plan somehow failed late in the process (display price or availability wasn't where they wanted). 7x GPU improvement along 4x more pixels to render would have made more sense.

Apple's performance claim was actually a 9x performance improvement for the GPU in the iPad 2 over the iPad. Anand's benchmarks show ~5x improvement is achievable in the real-world. 7x was iPhone 4S over iPhone 4.

Taking some slightly contrived numbers, a 2048x1536 iPad 2 actually wouldn't fare that well on a performance per pixel measure. Using the iPhone 3GS as a baseline with a 480x320 resolution and an assumed 150MHz SGX535, the iPad was 1024x768 and say a 250MHz SGX535 so 5x pixel count and 1.7x GPU power. Obviously performance per pixel fell behind with the iPad. Now a 2048x1536 iPad 2 with a theoretically 9x faster SGX543MP2 equivalent to a 2250MHz SGX535 (9x250MHz) would have 20.5x the pixels of the iPhone 3GS but only 15x the GPU power. So a Retina iPad 2 on the A5 would still have had less GPU power per pixel than a 2 year old device. For reference, for a Retina iPad to maintain the same GPU power per pixel as the iPhone 3GS would require a SGX535 clocked at 3075MHz which equates to a SGX543MP2 clocked 1.37x faster (3075MHz/2250MHz) than currently implemented in the Apple A5. A pure die shrink, clock bump Apple A6 with say 1.6GHz 2xCortex A9 and 400MHz SGX543MP2 (assuming the Apple A5 has a 250MHz SGX543MP2) would appear to do quite well. However, it would still mean that the iPad is falling increasingly behind the iPhone in GPU power per pixel at native resolution.

The current GPU should be enough for fluid web browsing and generic applications at 2048x1536. You would only want something beefier to run games... and I expect games to start rendering at non-native (lower) resolution if 2048x1536 and 2560x1600 (rumoured next gen Galaxy Tab resolution) become the norms in tablets. You don't need that much resolution in games (consoles are still rendering games at 720p).

When the Retina iPhone 4 was first released there was some speculation between the tradeoffs of going 480x320 + AA or full 960x640. Clearly developers seemed to have settled on going Retina. It'll be interesting to see if this is the case for a Retina iPad too. PowerVR mentioned in a gamesindustry.biz interview that Series5XT specifically optimized AA performance so the tradeoffs could be different on the SGX543MP than on the SGX535. Would 1024x768 + AA be cheaper from a performance perspective on a SGX543MP than 2048x1536? Would it look significantly worse?

 

[Ailuros]

The 7x GPU performance improvement from iPad -> iPad2 was a huge jump (bigger than expected). There has been speculation if Apple really wanted to have 2048x1536 resolution already in iPad2, but that plan somehow failed late in the process (display price or availability wasn't where they wanted). 7x GPU improvement along 4x more pixels to render would have made more sense.

I might be wrong but I wouldn't think that Apple is he type of company that would take a few very important factors like display availability as lightly. In fact it seems that they've made quite a sizeable investment to secure high resolution display availability. It would be awkward at least if they'd plan a high resolution device and someone "forgot" to do some proper research considering availability for such an important device part. Even more so considering that such displays are nowhere near mainstream nor is Apple dealing with any pee wee volumes.

As ltcommander pointed out above Apple claimed an up to 9x times graphics performance increase for the iPad2, while in reality the texel fillrate increase between iPad2 and iPad is just a tad above a factor of 2x. SGX535 has 2 TMUs which should be clocked at 200MHz and SGX543MP2 has 4 TMUs probably clocked at 250MHz.

In a pure floating point case the latter of course can process up to 9x times the FLOPs than the first.

Take another case example: the ULP GeForce in T20 is clocked at 333MHz and in T30 at 520MHz. Tegra3 gets in 1080p comparable performance as Tegra2 in 720p in Egypt standard/GL Benchmark2.1 or else a tad over 20 fps.

The current GPU should be enough for fluid web browsing and generic applications at 2048x1536. You would only want something beefier to run games... and I expect games to start rendering at non-native (lower) resolution if 2048x1536 and 2560x1600 (rumoured next gen Galaxy Tab resolution) become the norms in tablets. You don't need that much resolution in games (consoles are still rendering games at 720p).

If small form factor devices would get a significant amount of console ports, the latter would make sense. I fully agree that even 1080p is a too high resolution for mobile games, but if Apple intends to have future mobile games to run in super high resolutions, the need for additional hw resources is definitely there.

A5 is quite large already, somewhere over 120mm2 at Samsung 45nm, which doesn't sound like Apple cared about die area up to a specific point and in extension related costs. They'll most likely go with their next SoC to a smaller manufacturing process, likeliest scenario being 32nm. Assuming they still care as much about die area as with A5 and want to have one or more advantages against their future tablet competition, there's a quite a bit of headroom to increase performance in several blocks CPU and GPU blocks included. Just frequency is one case scenario, just more cores another one and a more o a middle of the road scenario both core and frequency increases or else whatever makes more sense to reach N% more performance without necessarily increasing overall average power consumption of the system.

My speculation doesn't render in absolutes here. You'll get the same performance out of a SGX543MP2@500MHz vs. a SGX543MP4@250MHz for example. If the target is higher there are quite a few other combinations (even possible odd amount of cores) with different frequencies that would make sense.

 

[french toast]

Quote:As ltcommander pointed out above Apple claimed an up to 9x times graphics performance increase for the iPad2, while in reality the texel fillrate increase between iPad2 and iPad is just a tad above a factor of 2x. SGX535 has 2 TMUs which should be clocked at 200MHz and SGX543MP2 has 4 TMUs probably clocked at 250MHz. Quote:

Wasnt SGX535 in ipad & iphone 4 clocked higher than that? i thought performance increased from 3gs-4...

 

[Rys]

3GS -> 4 performance increase wasn't down to SGX in particular, and clocks are v.similar. 9x in iPad 2 is peak shader performance.

 

[ltcommander.data]

3GS -> 4 performance increase wasn't down to SGX in particular, and clocks are v.similar. 9x in iPad 2 is peak shader performance.

I guess the move from a 32-bit to 64-bit memory controller was very helpful in this regard, beyond just clock speed increases.


http://9to5mac.com/2012/01/06/ios-5...soon-ship-ipads-iphones-with-quad-core-chips/

iOS 5.1 beta analysis yields references to quad core support so perhaps the A6 will be quad core.

 

[Helmore]

To be honest, I'd be really surprised if Apple went with a quad core design for their A6 chip. There's no real reason why they Apple won't got with a quad core chip though, that's just my gut feeling and that's not worth much when it comes to hardware speculation.

 

[Pressure]

Well, the rumor mill is going and it looks quite possible that the A6 would contain a quad-core design.

Guess we will know in due time.

 

[Lazy8s]

The original iPhone included a VGP Lite when some people figured such a prospect unlikely since the co-processor would benefit primarily just games. The 3GS went straight for the SGX535, and a nice clock rate at that, when the 530 was the obvious fit for the market, showing a real focus by Apple on extra performance. Including a 543MP2 in the next core update then came as no surprise to those paying attention.

Seeing Apple's push for high performance graphics as signifying anything more than their obvious acknowledgement that "the more GPU the better" on a visual platform like iOS is just ignoring a focus/agenda that Apple makes no attempt to hide.

 

[Ailuros]

Wasnt SGX535 in ipad & iphone 4 clocked higher than that? i thought performance increased from 3gs-4...

See answers above. In any case a legitimate point I forgot to add for the SGX535 vs. SGX543MP2 TMU comparison, is that the latter SoC has higher bandwidth and I think the system level cache for each 543 in A5 should be as high as it can get. Else the MP2 obviously has most likely more headroom to stretch its legs in terms of fill-rate.

However in strict hw terms all other conditions identical the 535 vs. MP2 TMU comparison is still valid.

Well, the rumor mill is going and it looks quite possible that the A6 would contain a quad-core design.

Guess we will know in due time.

There's more that speaks for it then against it. IMHO:

1. Too early for A15 and Apple's roadmap.
2. If the performance increase target is high enough it's again a matter of frequency vs. power consumption scheme than anything else. A9 cores are quite cheap in hw anyway.

Maybe Apple should be considering custom CPU designs like Qualcomm in the future.

 

[french toast]

Have any manufacturers ever included some dedicated graphics memory outside of handheld consoles?..im assuming were going to get to that point within the next 18 months or so.
How about either adding some exotic memory or slapping in some high clocked lpddr3 linked up to that duel core memory controller.. that alongside significant higher clocks would surely provide a cost effective way to match performance to that new screen.

 

[french toast]

*duel channel memory

 

[Exophase]

Have any manufacturers ever included some dedicated graphics memory outside of handheld consoles?..im assuming were going to get to that point within the next 18 months or so.
How about either adding some exotic memory or slapping in some high clocked lpddr3 linked up to that duel core memory controller.. that alongside significant higher clocks would surely provide a cost effective way to match performance to that new screen.

My old Dell Axim x50v had dedicated VRAM. That's a PocketPC, not a phone but also not a gaming handheld. The GPU was a rebranded MBX, like what was on pre-3GS iPhones. Another example is the OpenMoko Freerunner phone, which has an external GPU ("Glamo") with embedded RAM. The usage of this external GPU is pretty universally considered a design mistake in this phone.

Dedicated VRAM on mobile devices can be a double edged sword. I remember the Gizmondo handheld, which used an old GoForce with dedicated VRAM. It was a commercial flop and had hardly any real games released from it, instead more popularly bought to run emulators. These emulators relied on pure software rendering and were hindered by a very slow interface to the external VRAM. Freerunner was even worse. There were no 3D drivers at all and only minimal X11 2D acceleration and here the connection to VRAM was REALLY slow. It was impossible to update the entire 480x640 framebuffer at 16bpp at more than a few FPS, all while eating a bunch of CPU time.

This use case may be going down in need but you don't really want to end up in a situation where you can't update the whole native LCD resolution at 30+FPS in software if you have to.

PSP has separate VRAM where the CPU can access it as quickly and directly as its main memory, maybe even more quickly. But I suspect this is because they used eDRAM. If you're going to memory that the CPU and GPU can both access quickly that's also external to the SoC then you'd may as well just make it another DDR channel.

I do think bandwidth requirements are not a huge concern, and I find Vita's decision for external VRAM puzzling. Even huge resolutions like 2560x1536x32bpp only need 900MB/s to update at 60Hz. Most mobile GPUs are tilers (save Tegra, which has a color cache that you could probably coerce into software tiling with) that can usually get away with writing the whole framebuffer once per frame, then have it read once per frame to be sent to the screen. Of course they then need bandwidth for textures, but this can be a lot less thanks to compression. Shared memory bandwidth should be able to scale well enough to handle this.

Of course, in the real world games will make due with lower resolutions and even look nearly as good with 16bpp dithered output (with tilers anyway). 4:2:2 YUV output may also be an option, which is probably really annoying to implement with direct framebuffer rendering but straightforward with a tiler.