Implications of SGX543 in iPhone/Pod/Pad?

Entropy

Veteran
Link
Apart from the obvious that fill-rates increase to complement higher resolution screens, what else is there in this? It's no longer the A4 chip - could there be other changes? If so, which would be most likely/important? Is a multi-core implementation of the SGX543 likely? Et cetera.
 
Link
Apart from the obvious that fill-rates increase to complement higher resolution screens, what else is there in this? It's no longer the A4 chip - could there be other changes? If so, which would be most likely/important? Is a multi-core implementation of the SGX543 likely? Et cetera.

Multi-core extremely likely, I'll leave other more learned people to determine what that means in terms of performance improvements.
 
Given that they currently use the SGX535, I always thought Apple would choose the SGX545 for their next GPU. Both these models support full desktop APIs, DX9.0c and OpenGL 2.x in the SGX535 expanding to DX10.1, OpenGL 3.x, and full profile OpenCL in the SGX545, whereas all other mobile GPUs just seem to target OpenGL ES. This would give Apple greater flexibility with more functionality that they can expose over time as OpenGL ES extensions, which would be useful if Apple uses the SGX545 over 2 generations as was the case of the MBX Lite and SGX535. I haven't heard of Adreno 220 or Tegra 2 being DX10 class hardware and neither would a PSP2 with SGX543XT MP, so going with the SGX545 could have been a useful differentiator as the mobile gaming scene continues to heat up.

The SGX543XT is reported to support support OpenCL, but does it support the 1.1 embedded profile since it was presented as a new feature in the SGX544XT along with DX9.0 and OpenGL 2.x support.

Personally, I'm curious what process the Apple A5 will use. Is Samsung's 32nm process ready? Its certainly hard to see a pie in the sky dual core Cortex A9 with SGX543XT MP configuration on the same 45nm process as the Apple A4.
 
Personally, I'm curious what process the Apple A5 will use. Is Samsung's 32nm process ready? Its certainly hard to see a pie in the sky dual core Cortex A9 with SGX543XT MP configuration on the same 45nm process as the Apple A4.
Isn't Samsung offering 40nm, too? So I don't think a dual-core Cortex-A9 with a dual-core SGX543 would be unrealistic for the iPad with it's huge battery. But if Apple wants to reuse the same SoC (@40nm) in the iPhone and iPod touch like they did last year with the A4, that's where I see things getting complicated in terms of battery life, even if they clock down the dual-core Cortex-A9 by like 20% (as Apple did with the A4) and the dual-core SGX543 even more.
 
Given that they currently use the SGX535, I always thought Apple would choose the SGX545 for their next GPU. Both these models support full desktop APIs, DX9.0c and OpenGL 2.x in the SGX535 expanding to DX10.1, OpenGL 3.x, and full profile OpenCL in the SGX545, whereas all other mobile GPUs just seem to target OpenGL ES. This would give Apple greater flexibility with more functionality that they can expose over time as OpenGL ES extensions, which would be useful if Apple uses the SGX545 over 2 generations as was the case of the MBX Lite and SGX535. I haven't heard of Adreno 220 or Tegra 2 being DX10 class hardware and neither would a PSP2 with SGX543XT MP, so going with the SGX545 could have been a useful differentiator as the mobile gaming scene continues to heat up.

You might want to consider that there's no need yet for DX10.1 in mobile applications not for this or even next year. There comes up the dilemma whether a manufacturer wants to concentrate more on added functionality given a specific die area budget or additional performance.

According to IMG's whitepapers a SGX545 is at 12.5mm2 at 65nm/200MHz, whereby a single SGX543 is at 8mm2 at 65nm/200MHz.

As it looks like:

SGX545:

4 Vec2 ALUs
2 TMUs
16 z/stencil
40M Tris/s
12.5mm2@65nm/200MHz

SGX543 2MP:

8 Vec4 ALUs
4 TMUs
32 z/stencil
70M Tris/s
16mm2@65nm/200MHz

Outside of the 545 all other SGX variants are at least DX9/SM3.0. Compliance would be higher if they wouldn't had removed some functions to save die area like for instance max texture size which is at 2048*2048 for up to 540 and also 543 afaik. Any games in sight that are going for 4096*4096 textures and that's just one example. If you take the entire SGX package it's already ahead in terms of capabilities compared to the competition. USC ALUs are rare, any GPGPU capabilities are either non existant or limited, let alone some more advanced stuff like procedural geometry and what not.

I fully agree that investing in performance is far more important at this stage. For one because Apple isn't obviously the only manufacturer opting for SGX54x MPs and on the other side SONY announcing sooner than expected something that will also contain a high amount MP config.

Finally Apple is most likely going to continue to scale in display resolutions. Here you don't get more pixel fillrate going from a 2TMU 535 to a 2TMU 545. On the other hand a 4 TMU 2MP does help and it's not like any other contenders in the embedded mobile space have GPUs or GPU IP that have more than 2 TMUs at a time unless I'm missing something.


The SGX543XT is reported to support support OpenCL, but does it support the 1.1 embedded profile since it was presented as a new feature in the SGX544XT along with DX9.0 and OpenGL 2.x support.
544 and 554 support DX9.0/L3 which means amongst others support for 4096*4096 textures. If you want to integrate such a core into a win7 environment under DX11 in order to get DX9.0/L3 in DX11 those are the added functionalities you'd need. Apple is working on it's own iOS; why should they even bother with windows stuff?

Personally, I'm curious what process the Apple A5 will use. Is Samsung's 32nm process ready? Its certainly hard to see a pie in the sky dual core Cortex A9 with SGX543XT MP configuration on the same 45nm process as the Apple A4.
The iPhone3GS (which had oversimplified a lower clocked "A4") had a quite large SoC too. The 2 A9 CPUs don't capture exactly a magnitude more in die area compared to a single A8 and as shown above a 2MP would on 65nm consume 3.5mm2 more die area than a 545 both at the same theoretical frequency, with the first delivering at least twice the performance. Shrink those values to 45nm and the 3.5mm2 die area difference doesn't even sound worth mentioning.
 
Isn't Samsung offering 40nm, too? So I don't think a dual-core Cortex-A9 with a dual-core SGX543 would be unrealistic for the iPad with it's huge battery. But if Apple wants to reuse the same SoC (@40nm) in the iPhone and iPod touch like they did last year with the A4, that's where I see things getting complicated in terms of battery life, even if they clock down the dual-core Cortex-A9 by like 20% (as Apple did with the A4) and the dual-core SGX543 even more.

I'd suggest (no I don't know) that the 535 in the iPad might be clocked at 250MHz; even if a 2MP is clocked at only 150MHz f.e. the difference between the two is still huge. Up to 540 cores are 8z/stencil as it seems. A 2MP would mean 4x times that, apart from all other values outside TMUs that have doubled.

I'd have a far more important question regarding the whole story: will Samsung really manufacture that SoC for Apple? On one hand we've seen Renesas presenting its own 2MP SoC and Samsung is either going for a 1.2GHz Hummingbird for its "tab2" or if I am to believe some funky recent rumors something based on Tegra2 with Orion (dual A9/Mali 400MP) appearing at the end of this year. Somewhere in all that mess I wonder if Samsung is really manufacturing for Apple in the end and if yes why they aren't using a modified version of it for itself.

After all apart from the SGX535/SGX540 difference the differences between A4 and the GalaxyS SoC shouldn't be that huge. Could be of course 100% wrong, but somehow that whole puzzle doesn't seem to add up for me.
 
32nm in volume production and end user products is still quite a ways off.

ARM targeted MPcore A9s primarily for the 45/40nm process generation in area, power, and heat, and dual A9s (or the equivalent) are standard for the new generation of mobile SoCs. Same for 543MP: two cores are around 10mm^2 or under.

The update cycle for SoC development takes about two years, so Apple will almost assuredly pack sufficient staying performance. That means dual A9s and 543MP2, though I've been very much hoping for an MP3.

The DirectX 10.1 focused SGX545 never made much sense for this generation's mobiles; it just trades off a lot of area that should've been used for performance in a primarily OpenGL ES 2.0 environment. The extra TMU of the 535 was at least a relevant asset to the iOS devices, and the push to get the core ready for Intel made it the first SGX available and the most developed back when Apple was choosing.

As for the first embedded profile of OpenCL, all SGXs support that.
 
I'd suggest (no I don't know) that the 535 in the iPad might be clocked at 250MHz; even if a 2MP is clocked at only 150MHz f.e. the difference between the two is still huge. Up to 540 cores are 8z/stencil as it seems. A 2MP would mean 4x times that, apart from all other values outside TMUs that have doubled.
Personally, I would expect something like this for the iPad 2 and I think it's not totally unrealistic:

- 2048x1536 display (maybe 2x is wishful thinking but the latest rumors point to a higher resolution display in the next gen but with a lower dpi (in this case 264) than the iPhone 4 (326) and I don't see Apple going to something in between like 1400x1050 or 1600x1200 etc.)
- Apple A5 @40nm Samsung
- 1GB LPDDR2 RAM (512MB if Apple continues to be very (too) conservative with RAM?)
- 1GHz dual-core Cortex-A9 with a 200MHz SGX543MP2

and for the iPhone 5 with a 960x640 display they clock the CPU down to 800MHz and the GPU down more than 20% (to like 100-150MHz). (off topic: IMHO at some point Apple needs to change the size and aspect ratio of the iPhone display, but I don't think that's gonna happen in 2011, more likely in 2012 in combination with a major GUI revision.)
 
So Samsung is going to produce both Orion and A5, one with Mali 400 and the other with SGX543?
 
Samsung's development cycle for their own SoCs can't change just because they're a fab for hire.

Apple and Samsung develop their processor updates in alternating years.
 
It wouldn't matter even if they developed at the same time, taped out on the same day and even produced effectively identical chips. Samsung fabbing for Apple has nothing to do with anything else at Samsung in the Semiconductor or Mobile groups.
 
According to IMG's whitepapers a SGX545 is at 12.5mm2 at 65nm/200MHz, whereby a single SGX543 is at 8mm2 at 65nm/200MHz.
I didn't think there was such a large die area difference between the SGX543 and SGX545, in which case I can understand why that makes the SGX545 unattractive.

Finally Apple is most likely going to continue to scale in display resolutions. Here you don't get more pixel fillrate going from a 2TMU 535 to a 2TMU 545. On the other hand a 4 TMU 2MP does help and it's not like any other contenders in the embedded mobile space have GPUs or GPU IP that have more than 2 TMUs at a time unless I'm missing something.
I thought the speculation was that the SGX554 would be a 8 ALU, 4 TMU chip, but I guess it's a bit new for to already be implemented.

The DirectX 10.1 focused SGX545 never made much sense for this generation's mobiles; it just trades off a lot of area that should've been used for performance in a primarily OpenGL ES 2.0 environment. The extra TMU of the 535 was at least a relevant asset to the iOS devices, and the push to get the core ready for Intel made it the first SGX available and the most developed back when Apple was choosing.
I wasn't aware of the time scale, but it does makes sense that time to market was the reason Apple chose the SGX535 rather than additional functionality.

Personally, I would expect something like this for the iPad 2 and I think it's not totally unrealistic:

- 2048x1536 display (maybe 2x is wishful thinking but the latest rumors point to a higher resolution display in the next gen but with a lower dpi (in this case 264) than the iPhone 4 (326) and I don't see Apple going to something in between like 1400x1050 or 1600x1200 etc.)
- Apple A5 @40nm Samsung
- 1GB LPDDR2 RAM (512MB if Apple continues to be very (too) conservative with RAM?)
- 1GHz dual-core Cortex-A9 with a 200MHz SGX543MP2

and for the iPhone 5 with a 960x640 display they clock the CPU down to 800MHz and the GPU down more than 20% (to like 100-150MHz). (off topic: IMHO at some point Apple needs to change the size and aspect ratio of the iPhone display, but I don't think that's gonna happen in 2011, more likely in 2012 in combination with a major GUI revision.)
Apple did warn developers that future resolution changes may not be an easy 2x increase so in between resolutions are possible. 1600x1200 seems like a reasonable choice.

Supply constraints were supposed to be the reason why Apple went increased the memory bus to 64-bit in the A4 and stuck with LPDDR rather than LPDDR2. Has that cleared up significantly? Otherwise, Apple will probably stay conservative and go 512MB of LPDDR2 on a 64-bit bus across all devices this generation. Going directly from 256MB to 1GB would seem like a huge jump for one generation.

The thing about clock speed is that since Apple controls everything from transistor layout to SoC integration and up the software stack, it'd be great to see them implement dynamic clocking between the CPU and GPU like Turbo Boost in Sandy Bridge. A 1GHz dual core Cortex A9 and 200MHz SGX543MP2 seems reasonable, but something that can shut down the second core and Turbo the remaining one to 1.4GHz for existing single threaded apps or downclock both cores to 800MHz and overclock the GPU to 250MHz for GPU limited situations would be great for performance and the flexibility would allow Apple to work out their usual battery life wonders.

And taking a step back from talk of the iPad 2, I wonder if Apple will keep a version of the original iPad and sell it at a lower price point like they do for the iPhone and once did for the iPod Touch? A lot of Android iPad competitors seem to be landing around the iPad's existing price points, but the RIM Playbook for one seems to be heavily marketed on being priced under the iPad. If Apple retains the 16GB WiFi iPad at $399 or even $349, they can probably cut out a lot of the lower price competitors before they even get established. Certainly the iPad 2 will probably be announced by the time the Playbook is actually available to steal some of the thunder, but a $349 iPad 1 will make things even more difficult. The Playbook may have better performance, but it doesn't have the maturity or the iTunes/App Store level of integration. It's just a matter of whether Apple thinks a cheap iPad 1 will cannibalize sales of the iPad 2, which it may not for consumers but may for line of business applications which probably won't need a faster CPU or GPU although the cameras would be useful. If the goal is to grab as large a market share and user base as possible early on as the tablet market is still getting established, a cheaper iPad 1 will certainly allow that to happen.
 
Right, that's what I was saying with my first comment about their fab for hire being completely separate from their own development.

My second comment was addressing the surprise people had over the possibility Samsung could fab a chip for a "competitor" that was superior to their own SoC design, and I was adding the perspective that, while that's totally possible since the businesses are separate, it's more the effect of the update cycle happening in alternating years (iPhone 3GS using the SGX535 in 2009, Hummingbird Galaxy S using the SGX540 in 2010, iPhone 5 using the 543MP2 in 2011, and then Orion using Mali400mp4 in 2012 -- though that last one will make for an interesting comparison.)
 
My second comment was addressing the surprise people had over the possibility Samsung could fab a chip for a "competitor" that was superior to their own SoC design, and I was adding the perspective that, while that's totally possible since the businesses are separate, it's more the effect of the update cycle happening in alternating years (iPhone 3GS using the SGX535 in 2009, Hummingbird Galaxy S using the SGX540 in 2010, iPhone 5 using the 543MP2 in 2011, and then Orion using Mali400mp4 in 2012 -- though that last one will make for an interesting comparison.)

And where do you fit A4 in there?
The Hummingbird and A4 were made at the same time, both by Samsung.

It doesn't sound like Samsung is limited to making one SoC at a time.


Plus, it seems obvious to me that there will be Orion-equipped products in 2011:
Samsung said:
Samsung's new dual-core application processor, Orion, will be available to select customers in the fourth quarter of 2010 and is scheduled for mass production in the first half of 2011.



So it's possible that Samsung might end producing a SoC for Apple (A5) that's slower than its own (Orion), not the other way around.
 
Last edited by a moderator:
They're not offset by exactly one year nor are updates exactly every two years, but Orion devices will release almost a year after the A5 iOS devices.

A4 was just the refresh from the perspective of the processors: still using an A8 and a 535 yet with enhancements like improved clock speed. The second generation iPod touch was the same thing, refreshed the same processors with an increased clock rate and such.
 
Last edited by a moderator:
I don't know why everyone keeps thinking about it this way. Samsung is working as a FOUNDRY for Apple. Steve Jobs could decide to switch to TSMC tomorrow if he wanted. There is no reason whatsoever to compare the A4/A5 with Samsung's in-house SoCs except that they use the same process.

The original iPhone 3G SoC was a bit different (mostly designed by Samsung with specifications from Apple iirc) but those times are long gone - this is a foundry relationship and nothing more.
 
Even IMG pointed that out in the press release for the IP licensing to the mystery "international electronics systems manufacturer".

"The SoCs to be developed under this license agreement will be produced for this new partner by Imagination's existing semiconductor partners and/or new chip manufacturing partners."

... and then Samsung had to sign their own fab licensing agreement.
 
Personally, I would expect something like this for the iPad 2 and I think it's not totally unrealistic:

- 2048x1536 display (maybe 2x is wishful thinking but the latest rumors point to a higher resolution display in the next gen but with a lower dpi (in this case 264) than the iPhone 4 (326) and I don't see Apple going to something in between like 1400x1050 or 1600x1200 etc.)
- Apple A5 @40nm Samsung
- 1GB LPDDR2 RAM (512MB if Apple continues to be very (too) conservative with RAM?)
- 1GHz dual-core Cortex-A9 with a 200MHz SGX543MP2

and for the iPhone 5 with a 960x640 display they clock the CPU down to 800MHz and the GPU down more than 20% (to like 100-150MHz). (off topic: IMHO at some point Apple needs to change the size and aspect ratio of the iPhone display, but I don't think that's gonna happen in 2011, more likely in 2012 in combination with a major GUI revision.)

Can they really fit that into a $500 product? The resolution jump seems a big step. Apple has invested a lot of money, maybe in the billion dollar range, to help a couple of display manufacturers build new factories. Maybe a high resolution display is more likely when those factories come on line.

They just made developers support new resolutions this year. Can they make them support new resolutions every year?

OTOH, if they were to continue to make iPad1s as lower-cost SKU, they would have to differentiate the new models in a very visible way (the faster SOCs may not necessarily wow a lot of buyers). A lot of people may be more wowed by high-res display than a SKU said to offer faster CPU and GPU.


As far as high-performance graphics games, Apple doesn't tip its hand so it would be hard to develop for newer hardware, especially if they're putting more capable HW every year?

PSP2 comes out and takes the performance crown but every year, new mobile devices are coming out. The situation could be analogous to consoles and PCs, where the latter offers increasing performance every year. Now Sony could lock up developers for PSP2 games, as console games outsell PC games. But unless Sony pays for exclusivity, why would these developers not make games for smart phones and tablets, which will likely sell in much greater numbers? iOS alone may provide a larger gaming market than the PSP2, despite Apple's indifference to "core" games.
 
OTOH, if they were to continue to make iPad1s as lower-cost SKU, they would have to differentiate the new models in a very visible way (the faster SOCs may not necessarily wow a lot of buyers). A lot of people may be more wowed by high-res display than a SKU said to offer faster CPU and GPU.
Front and back facing cameras plus a SD card slot as rumored might do it.

As far as high-performance graphics games, Apple doesn't tip its hand so it would be hard to develop for newer hardware, especially if they're putting more capable HW every year?
Games are finally starting to drop support for older devices and focus on OpenGL ES 2.0. Infinity Blade gets all the visual praise, but it's really just the first generation of OpenGL ES 2.0 only games. The graphics push in iOS might well come by bringing the bottom up rather than pushing the top end. With iOS 5, if not iOS 4.3, looking to require 3rd gen devices and up, games that push the Cortex A8, NEON, SGX535, and 256MB of RAM to their limits should still make for an impressive show on the Apple A5. Hopefully with 30fps on 3rd gen devices translating to 60fps on 5th gen devices.

PSP2 comes out and takes the performance crown but every year, new mobile devices are coming out. The situation could be analogous to consoles and PCs, where the latter offers increasing performance every year. Now Sony could lock up developers for PSP2 games, as console games outsell PC games. But unless Sony pays for exclusivity, why would these developers not make games for smart phones and tablets, which will likely sell in much greater numbers? iOS alone may provide a larger gaming market than the PSP2, despite Apple's indifference to "core" games.
I think the latest speculation was that the PSP2 could be available in time for the holidays at which point the complete Apple A5 device lineup will already be available. A SGX543MP2 vs. SGX543MP8 comparison heavily favors the PSP2, but by that time the Apple A6 will be right around the corner. Smartphones could probably offer comparable performance to the PSP2 within 2 years of it's release. It was just reported that Crytek was looking for iOS developers. The applicability of the phrase "But can it run Crysis?" might be expanding.

It'll be interesting if both Apple and Sony are in fact using SGX543MP based products since it would presumably be helpful for cross-platform development. iOS will probably benefit from the ease of porting for AAA titles from the PSP2 to make the platform more appealing to hardcore gamers, while conversely the PSP2 should benefit from ease of porting of iOS games to Playstation minis to make the platform more appealing to casual gamers. Perhaps the relationship will be partially symbiotic in addition to competitive?
 
Is a 1920x1280 resolution at all likely? Eschewing 2x2 iPad in favor of 2x2 iPhone 4 (or 4x4 original iPhone)

I don't know why everyone keeps thinking about it this way. Samsung is working as a FOUNDRY for Apple. Steve Jobs could decide to switch to TSMC tomorrow if he wanted. There is no reason whatsoever to compare the A4/A5 with Samsung's in-house SoCs except that they use the same process.

Probably because A4 uses Hummingbird and an SGX535, making it look like an S5PC110 clone. That is of course ignoring all the other parts of the SoCs, but those aren't very interesting to people (and who knows how similar the rest of it actually is). On paper the two SoCs seem closer than 3GS's SoC was to S5PC100, despite lots of people still claiming 3GS used it.

Of course, with Apple owning Intrinsity now that decouples that relationship.
 
Last edited by a moderator:
Back
Top