Next-Gen iPhone & iPhone Nano Speculation

[ltcommander.data]

Will they be able to move to sub 28nm before the next iPhone is due? If not they could be saving the big die for 28nm A8.

I just noticed from the keynote that Schillar said the A7 die is about the same size as the previous A6 but it has about twice as many transistors. Even if they hand laid-out everything to optimize for area could they actually double the transistor count in the same area if they only moved from 32nm to 28nm? If not, then the A7 could be using a new smaller process.

 

[Grall]

Unless I'm misunderstanding what they meant, that's going to be true for every iteration of a process that doubles at every iteration.

Yes of course, but how does that in any way diminish the accomplishment? So people go, "oh, they only DOUBLED performance compared to last gen, how unimpressive", but when you view it in graph form as in that keynote, it shows a clearer picture of just how big an improvement that actually is.

 

[Grall]

That comment would come over better if it actually was like $300 or something, not the $650 it currently is.

Bleeding edge portable devices have always been expensive as hell. It's nothing new. High-end feature phones often cost almost that much in the bad old days, with only a fraction of the effort spent in their design and manufacture compared to iP5. How many phones even had all-metal (almost) construction before iP5? Much less mirror-polished chamfered edges, laser-cut precision inlays and whatnot. Then all the work that has gone into the resident software, my god. There's millions of man-hours spent there.

 

[anexanhume]

I just noticed from the keynote that Schillar said the A7 die is about the same size as the previous A6 but it has about twice as many transistors. Even if they hand laid-out everything to optimize for area could they actually double the transistor count in the same area if they only moved from 32nm to 28nm? If not, then the A7 could be using a new smaller process.

Where did you get your transistor count for A6?

 

[Brilliantdeve]

Does anyone knows which PowerVR GPU used in iPhone 5s ?!

 

[DuckThor Evil]

5S seems pretty nice package overall, but the C with that pricing... Oh boy...

 

[anexanhume]

Does anyone knows which PowerVR GPU used in iPhone 5s ?!

It's probably either a G6430/6400 ~ 250 MHz or a G6230/G6200 ~ 500 MHz.

 

[Grall]

I'm more interested in how much, and how fast RAM is in that thing than which GPU it features, to be totally honest...

 

[ltcommander.data]

Where did you get your transistor count for A6?

I don't have the transistor count for the A6. In Schiller's verbal presentation of the A7 slide that lists "Over 1 billion transistors" as a bullet point he says:

It has over a billion transistors in it. And this fits in a die that's about the same size as the previous generation A6. It's about twice as many transistors.

The A6 was 97 mm2 and the A7 is 102 mm2, a 5% increase in die area. That means most of the claimed ~doubling of transistors is coming from increased density. Can going from 32 nm to 28 nm account for all that?

 

[Exophase]

Can going from 32 nm to 28 nm account for all that?

It's not simply going from 32nm to 28nm, but Samsung's 32nm to TSMC's 28nm. There's more to the density a design can achieve than the name given to the process node.

You can see for example that the Cortex-A9 in TI's OMAP4430 was much smaller on TSMC's 45nm than the A9 in Apple A5 or Exynos 4212 on Samsung's 45nm.

 

[anexanhume]

I don't have the transistor count for the A6. In Schiller's verbal presentation of the A7 slide that lists "Over 1 billion transistors" as a bullet point he says:

The A6 was 97 mm2 and the A7 is 102 mm2, a 5% increase in die area. That means most of the claimed ~doubling of transistors is coming from increased density. Can going from 32 nm to 28 nm account for all that?

I listened to the keynote since I figured that's where it came from.

32nm to 28nm is at best 20% denser. My guess is that the increased density comes from even more custom circuitry on the SoC. It seems very unlikely that it's 28nm TSMC, since the leaked A7 has a samsung product code. Problem is that we don't know any kind of transistor count for A57 (assuming A7 is at all similar to A57 in other than name only) or rogue to compare, but quad core is seeming more and more likely.

It's not simply going from 32nm to 28nm, but Samsung's 32nm to TSMC's 28nm. There's more to the density a design can achieve than the name given to the process node.

You can see for example that the Cortex-A9 in TI's OMAP4430 was much smaller on TSMC's 45nm than the A9 in Apple A5 or Exynos 4212 on Samsung's 45nm.

It's not fair to compare to the A9 in the A5. I've talked with someone in the industry familiar with people at Apple that suggested they used m-of-n encoding on their logic which is more transistor expensive but has benefits on speed versus power.

 

[Grall]

I'm more interested in how much, and how fast RAM is in that thing

Creepy Uncle Google claims the iP5S has 2GB RAM - twice the amount of the previous iP5. No idea the original source on this, as I believe the presentation did not mention this particular spec figure. Also no idea the speed, but could it use 1600MHz LPDDR3 perhaps or would that be too much...?

 

[anexanhume]

Creepy Uncle Google claims the iP5S has 2GB RAM - twice the amount of the previous iP5. No idea the original source on this, as I believe the presentation did not mention this particular spec figure. Also no idea the speed, but could it use 1600MHz LPDDR3 perhaps or would that be too much...?

Is the product code legible in the package shot? You could figure it out from that.

edit: This leak (which suggests TSMC) also says it's still Elpida and still 1GB. Probably LPDDR3 now though. http://www.macrumors.com/2013/06/24/apples-prototype-iphone-5s-based-on-new-a7-chip/

 

[loekf]

Is the product code legible in the package shot? You could figure it out from that.

edit: This leak (which suggests TSMC) also says it's still Elpida and still 1GB. Probably LPDDR3 now though. http://www.macrumors.com/2013/06/24/apples-prototype-iphone-5s-based-on-new-a7-chip/

The why now 64-bits for the Swift CPU still puzzles me. It's not the RAM. Is it because lot of devs write/compile 64-bits apps these days ? iPads with > 4 GB RAM to enable use-cases, programs we didn't see yet (e.g. proper video editing) ?

 

[Deleted member 13524]

The why now 64-bits for the Swift CPU still puzzles me. It's not the RAM. Is it because lot of devs write/compile 64-bits apps these days ? iPads with > 4 GB RAM to enable use-cases, programs we didn't see yet (e.g. proper video editing) ?

The theory that it's a long-term strategy to unify all their products under their own SoCs seems to be the most plausible.

 

[anexanhume]

The theory that it's a long-term strategy to unify all their products under their own SoCs seems to be the most plausible.

They did say "Desktop class performance" multiple times.

 

[ltcommander.data]

The why now 64-bits for the Swift CPU still puzzles me. It's not the RAM. Is it because lot of devs write/compile 64-bits apps these days ? iPads with > 4 GB RAM to enable use-cases, programs we didn't see yet (e.g. proper video editing) ?

Apple kept stressing forward thinking, but maybe it's less specific future applications that benefit from 64-bit at which point the A7 may not have the performance to offer a good experience anyways, than a practical matter of support policies. At some point, iOS devices will need more than a 4GB address space and will need 64-bit support. Apple will then want to go 64-bit only for the OS (32-bit apps will be supported longer) as soon as possible in order to simplify development. Apple's long support policies could actually be working against them here. iPhone's now seem to be getting 4 years of OS updates. So if the iPhone 5S were to have been 32-bit, Apple wouldn't be able to transition to 64-bit only until iOS 11 in 2017. As it is, the A6 in the iPhone 5/5C should still be supported in iOS 9 in 2015 and iOS 10 can be 64-bit only in 2016.

 

[tangey]

32nm to 28nm is at best 20% denser. My guess is that the increased density comes from even more custom circuitry on the SoC. It seems very unlikely that it's 28nm TSMC, since the leaked A7 has a samsung product code. Problem is that we don't know any kind of transistor count for A57 (assuming A7 is at all similar to A57 in other than name only) or rogue to compare, but quad core is seeming more and more likely.

Two things.
1)
Do you not think it is very un-apple like to NOT announce this new chip as Quad-Core 64-bit. I would have thought marketing would have been all over the fact that they had gone quad core. Or in the opposite view, would want to ensure people knew they weren't still "only" using dual core.

2)
However, 1) given the increased overall transistor count, 2)the virtually same die size on the smaller node, 3) the relatively modest Rogue implementation the graphics performance increase would imply, I suggest they have prioritised cpu performance on this soc, which would be consistent with quad-core.

Only other thought it that they have remained dual-core and felt it necessary to include a significantly larger memory cache than previous implementations.

 

[Entropy]

Two things.
1)
Do you not think it is very un-apple like to NOT announce this new chip as Quad-Core 64-bit. I would have thought marketing would have been all over the fact that they had gone quad core. Or in the opposite view, would want to ensure people knew they weren't still "only" using dual core.

2)
However, 1) given the increased overall transistor count, 2)the virtually same die size on the smaller node, 3) the relatively modest Rogue implementation the graphics performance increase would imply, I suggest they have prioritised cpu performance on this soc, which would be consistent with quad-core.

Only other thought it that they have remained dual-core and felt it necessary to include a significantly larger memory cache than previous implementations.

Well, if they can provide "up to" 2x the performance per thread over the A6 Swift, on similar process nodes and within the same power envelope, it's bloody fantastic. Witchcraft and sorcery, in one chip!

Benchmarks will tell some of the story, of course, but it will be very interesting indeed to see the inevitable chipworks teardown as well.

 

[Miksu]

I think the fingerprint reader could work great with iPad, as that device is quite often shared with family members. Each user could have their own profile, which the iPad would automatically load based on the fingerprint. So when you press the Home button, you're taken to your own unique home screen with your own apps.