Implications of SGX543 in iPhone/Pod/Pad?
- Thread starter Entropy
- Start date
You could change that allocation (up to a point) on a frame-by-frame basis though.
Yeah, efficiency never seemed to be an issue for Naomi2. It performed a lot more than 2x Dreamcast.
While the extra RAM helped alleviate the storage limitation for the display lists and the dedicated T&L granted more-than-proportional gains to sustained performance in the context of the extra silicon ELAN added, the two CLX2s also scaled the rendering together a lot better than desktop multi-processor implementations.
While the extra RAM helped alleviate the storage limitation for the display lists and the dedicated T&L granted more-than-proportional gains to sustained performance in the context of the extra silicon ELAN added, the two CLX2s also scaled the rendering together a lot better than desktop multi-processor implementations.
IMO the point of OpenCL is to write parallel code without caring about the exact number of devices/cores/whatever. The API and hardware handles workgroup distribution.
If no one means me then you're correct. Otherwise I disagree.
PC GPUs work the same way. It's only when you have multiple GPUs and no shared memory that things differ.
Same thing here. If the double density screen turns out to be true, my iPad will hit eBay the next minute. 1280x800 would be only a very incremental increase and you'd still see the pixels (in additional to be impractical wrt existing apps). Given the slightly larger distance when using an iPad compared to an iPhone, 260ppi would be perfect. Hoping, but not counting on it...
http://en.wikipedia.org/wiki/Sega_Naomi#Sega_NAOMI
IMG formerly known as Videologic sold IP back then too. http://en.wikipedia.org/wiki/PowerVR
To be accurate desktop GPUs typically scale from highest to lowest end, so it isn't like they're scaling up but rather down. It might sound like a stupid remark but without modifications they won't scale as easily with those all the way down to the embedded space as much as IMG or any other IHV with embedded GPUs can scale that easily from embedded to high end desktop.
If hypothetically desktop GPUs would scale from say 1 cluster to say 10 there's no chance that with each added cluster the result will have everything duplicated. It would generate quite some redundancy and exactly such a redundancy is there on embedded GPU MPs. It's IMHO difference enough to call the latter multi-core, - processing or whatever.
If in theory IMG would decide to design tomorrow a high end desktop GPU it would consist of what? 32 RGX cores or IP for one single GPU chip? Of course could they if needed also have mGPU configs with 2 of the latter cores on a single PCB, yet not a wild collection of a gazillion of mini-core IP on one single chip.
ltcommander.data
Regular
http://www.koreatimes.co.kr/www/news/tech/2011/01/129_80374.html
It appears Samsung has reached an agreement with Apple to quadruple their allocation of "sheets", which I'm taking to mean wafers, from 5000 sheets to 20,000 sheets. So it appears Apple will be sticking with Samsung as their fab partner for the A5 after-all. I guess some of that increase is the assumption of increased demand and some is presumably because the A5 will be a bigger chip. I wonder how many iOS devices that translates to that Apple thinks it'll be able to sell in 2011? Samsung's total capacity at 45nm is supposedly 40,000 sheets according to the article, so Apple is claiming half of Samsung's capacity at that node if that is what Apple will actually be using for the A5. I don't know if Samsung's 32nm process is ready for volume production yet. Either way Apple will still need 45nm capacity to keep the A4 in production for another year for low-end models. Perhaps capacity constraints are also partially responsible for Samsung looking to use Tegra in future Galaxy Tabs. Certainly one way to constrain the competition is to lock up the capacity for flash memory, SoC, and displays.
The article also speculates that Samsung's Fab 2 Phase 2 expansion in Austin, Texas may be geared toward supplying Apple. Perhaps, this fab is one of the investments that Tim Cook was talking about? Intrinsity is in Austin Texas as well, so it would be sensible for Apple to pay to lock in guaranteed access to a local fab for prototyping and future production. Cook said deals with three suppliers which could then be Sharp and Toshiba for LCD production and Samsung for SoC production.
It appears Samsung has reached an agreement with Apple to quadruple their allocation of "sheets", which I'm taking to mean wafers, from 5000 sheets to 20,000 sheets. So it appears Apple will be sticking with Samsung as their fab partner for the A5 after-all. I guess some of that increase is the assumption of increased demand and some is presumably because the A5 will be a bigger chip. I wonder how many iOS devices that translates to that Apple thinks it'll be able to sell in 2011? Samsung's total capacity at 45nm is supposedly 40,000 sheets according to the article, so Apple is claiming half of Samsung's capacity at that node if that is what Apple will actually be using for the A5. I don't know if Samsung's 32nm process is ready for volume production yet. Either way Apple will still need 45nm capacity to keep the A4 in production for another year for low-end models. Perhaps capacity constraints are also partially responsible for Samsung looking to use Tegra in future Galaxy Tabs. Certainly one way to constrain the competition is to lock up the capacity for flash memory, SoC, and displays.
The article also speculates that Samsung's Fab 2 Phase 2 expansion in Austin, Texas may be geared toward supplying Apple. Perhaps, this fab is one of the investments that Tim Cook was talking about? Intrinsity is in Austin Texas as well, so it would be sensible for Apple to pay to lock in guaranteed access to a local fab for prototyping and future production. Cook said deals with three suppliers which could then be Sharp and Toshiba for LCD production and Samsung for SoC production.
Last edited by a moderator:
Since they have quadrupled they fab capacity, I think it's safe to say that they estimate ~4x more sales of iOS devices.
They said they were paying to 3 suppliers. You generally want to stick to a single fab to avoid having to tape out multiple versions of your chips.
ltcommander.data
Regular
I didn't mean that Apple is looking for 3 SoC fab partners, just that Tim Cook said that Apple invested in 3 supplier relationships which I'm thinking could be Toshiba and Sharp as speculated before for dedicated LCD capacity and Samsung for dedicated SoC capacity.
http://www.appleinsider.com/article..._ipad_2_components_5_million_unit_supply.html
It looks like the specs for the iPad 2 are coalescing. No retina display, just the same resolution but thinner with lower reflection for outdoor viewing. They are also claiming the dual core A9 will run at 1.2GHz and will have 2 channels of LPDDR2 at 1066MHz for 512MB total. Up from a reported 800MHz RAM frequency in the iPhone 4. Front and back cameras will be common with the iPod Touch instead of the iPhone 4 in quality. And interestingly, despite all the talk of Apple moving to a combined GSM/CDMA chip from Qualcomm the iPad 2 looks to still retain separate WiFi, GSM, and CDMA models. Perhaps production timing meant that unification will have to wait for the iPhone 5. All in all, the specs look pretty reasonable and are consistent with previous claims so perhaps they are true.
Well, the opinions of one Taiwanese analyst doesn't really constitute coalescence in my book.
While 64 bits worth of LPDDR2 seems reasonable (I guess that's what 2 channels would refer to), it would be neat to have some kind of confirmation on that.
I'm still holding out hope on higher resolution, mainly because it would sell the device to me. Otherwise, I'll simply sit on my wallet, there is nothing in higher performance in and of itself that makes it more attractive to me, nor to my very elderly parents. Higher resolution to make it print like in terms of content appearance - that would bring value and interest, and I'd buy it for my parents immediately. As it stands however, it doesn't work all that much better as a reader than the other devices out there - they are all too low res.
The funny thing with the entire resolution speculation for Apple's next iPad is that it renders from one absurd extreme (2048*1536) to another (1024*768). Ok I'll admit that "absurd" is debatable, but the analyst himself claims higher GPU power and an according bandwidth increase to feed that added GPU power and resolution is meant to remain the same?
What exactly am I missing here?
What exactly am I missing here?
That does seem odd, doesn't it? Conceivably, Apple could move from SGX535 to dual SGX543 in order to improve graphics performance alone. The improvements in main memory performance would be needed to allow dataflow capabilities to somewhat keep up with the increases in CPU and GPU processing abilities.
After all, the processing capabilities per pixel took a hit going from 3GS to iPad/iPhone4.
However, the rumors specify a huge increase in processing capabilities, for devices whose responsiveness are arguably already defined mostly by internet access speed. It is difficult to believe Apple would go to those lengths without having something specific in mind to actually use that performance for. Resolution fits the bill. To any prospective buyer, it is an immediately apparent benefit, and would provide a unique selling point vs all others, (smaller form factor tablets in particular!).
A lot of great thoughts came up in this thread. One thing I would like to see discussed somewhere is where we are heading in terms of memory. 64-bit wide LPDDR2 would seem like an obvious incremental next step for Apples upcoming SoC. What's next then? 128-bit like in the A15 documents? SPMT LPDDR2, seeing as Samsung, ARM, Hynix and LG among others are members of that consortium? Rambus, of course, has an offering everyone will shy away from. DDR3 following the laptop example? Et cetera. Given how close we are in time to greater bandwidth needs, the lack of clear roadmap is remarkable.
D
Deleted member 13524
Guest
"No one" means "not a substantial enough amount of people to make Apple even think of spending gazillions of money to make a cheap-ass-to-make oversized iPod touch into something with an unprecedented pixel density for its size class".
Some people may wipe their asses with 100€ bills, just for the fun of it. Doesn't mean the European Union will ever redesign the bills to make them fluffier.
While frequencies per GPU increased in A4 I don't think they were enough to cover the resolution increase in iPhone4 vs. 3GS. Frequency for the iPad GPU could be even higher than in iPhone4.
In theory if we'd have a SGX535/2TMUs@200MHz vs. a SGX543 2MP/4TMUs@100MHz you'd have the same raw on paper 400MPixels/s fill-rate. At first glance it sounds like a weird decision but not if the iPhone's GPU block frequency should be in the 100-150MHz ballpark vs. =/>200MHz for the next iPad.
I used the above numbers more for an example than anything else, but assuming the MP2 in the iPad2 is truly clocked at somewhere 200MHz, that's 800MPixels/s and could mean up to 60% higher fill-rate than in the current iPad.
Good point about the clocks. They are tricky to make a good call of.
It seems reasonable to assume that if the same SoC is used in the iPhone/iPod touch, clocks will be adjusted downward to preserve battery life in the smaller form factor. But choosing the "dual-GPU and lower clocks" as the default for the iPad, rather than single GPU and normal clocks for the process, seems a bit wasteful of die area. (And possibly licensing costs - does having dual or quad 543s cost more per SoC than having one?) And the iPad has rather ample battery capacity so small savings in power draw doesn't seem compelling. It would seem reasonable to assume that the GPUs will be clocked at what would be "normal" for the process. Whatever the heck that would be.
The oft quoted clock for the SGX543 is 200MHz@65nm (presumably at TSMC LP), and how that translates to Samsungs 45nm process is anybodys guess. I think we can assume that the clocks for the GPU won't be below 200MHz, and might be higher. Additionally, Samsung demonstrated SoCs produced on 32nm LP HKMG back in June 2010 with promises of full production capability just about now. And if that is the process to be used, the additional capabilities wouldn't imply that much compromising need to be done. The SoC wouldn't need greater die size than the A4, and the additional capabilities wouldn't have to come at very restrained clocks for the process for power reasons. Apple has been pointed out as one of Samsungs "leading edge" customers.
Of course, knowing is difficult, but I find the jump from the A4 to the rumored part surprisingly ambitious for the same process node, and again, definitely so if there isn't any particular intended use for the added capabilities.
Sometimes, surprises can actually be nice.
