Next-Gen iPhone & iPhone Nano Speculation

[liolio]

That is quiet an amazing CPU, I'm amazed by what they pulled out. What impresses me further is that I was sort of expecting that to where they are now (in perfs) they would have implement some form of dynamic clocking/ turbo. they did not which means to me that they still have some room to significantly improve their perfs in future iteration.
Even without turbo the next ipad could quiet something as far as perfs are concerned.

 

[loekf]

That's the most interesting question, I think. Apart from x86, if Oscar can scale up to 2.0GHz or more, there's not much binding Apple to Intel for laptops.

Actually Oscar is the CPU in the M7 and Cyclone is the Swift MKII...

Interesting story also on the 64-bits in the A7. I didn't really do some reading on the ARMv8 ISA so far, but appearently it's a new (cleaner ??) ISA and not like in the PC world about accessing more memory.

 

[Pressure]

That's the most interesting question, I think. Apart from x86, if Oscar can scale up to 2.0GHz or more, there's not much binding Apple to Intel for laptops.

Oscar is the M7 co-processor

Cyclone is the replacement for Swift.

Update: Already mentioned by loekf above ... whoops.

 

[liolio]

Actually Oscar is the CPU in the M7 and Cyclone is the Swift MKII...

Interesting story also on the 64-bits in the A7. I didn't really do some reading on the ARMv8 ISA so far, but appearently it's a new (cleaner ??) ISA and not like in the PC world about accessing more memory.

I think there is an interesting paper, from D. Kanter, on Realworldtech.com website:
http://www.realworldtech.com/arm64/

 

[Turbotab]

Take an A7, fab it @ 20nm, add more RAM, cores, boost the clockspeed, and Apple would have an excellent low-power chip for an ARM port of OS X Server, if they so choose. Complete control of the entire stack from server to smartphone, that would be interesting.

 

[dZeus]

Take an A7, fab it @ 20nm, add more RAM, cores, boost the clockspeed, and Apple would have an excellent low-power chip for an ARM port of OS X Server, if they so choose. Complete control of the entire stack from server to smartphone, that would be interesting.

or a nice MacBook AIR CPU... (once Mac OS X / IOS are tightly enough integrated where a switchover to ARM for desktop OS does not impact functionality/compatibility too much).

how would its price compare to buying the CPU from Intel, as currently done?

 

[ltcommander.data]

https://twitter.com/anandshimpi/status/380327582786277376

@GermanJon @gruber as far as I know, the Intel deal fell through last year and hasn't been resurrected

Apparently the A7 or later chips could have potentially been manufactured at Intel but the deal fell through.

 

[Entropy]

Take an A7, fab it @ 20nm, add more RAM, cores, boost the clockspeed, and Apple would have an excellent low-power chip for an ARM port of OS X Server, if they so choose. Complete control of the entire stack from server to smartphone, that would be interesting.

That's one of several openings. The IPC of this core is so high, and the power draw (and die size) so low, that from a technical point of view, the field is open. What makes strategic sense, or sense for their customers/prospective markets is another matter, of course.

It wasn't a given that these mobile cores would evolve to this level of IPC, at least not at this point in time. This means that if Apple so chooses, it could utilize its mobile volumes to drive core development that could yield very competitive parts in other segments as well. (As opposed to for instance using their 20 million/year mac volume as CPU technology driver alone.)
From a financial point of view, given that Apple is using Intels more expensive offerings, I can certainly see the attraction in keeping that money for your own R&D and production. (For comparison, AMDs total annual revenue is $5billion. ARMs is less than 1$billion.)

 

[Pressure]

or a nice MacBook AIR CPU... (once Mac OS X / IOS are tightly enough integrated where a switchover to ARM for desktop OS does not impact functionality/compatibility too much).

how would its price compare to buying the CPU from Intel, as currently done?

If they have enough wafers they could make it much cheaper than Intels prices for Haswell ULT, which goes for something like $300 to $600.

 

[RDGoodla]

incredible increase in performance over x3.5 improvement in GL2.7 over the iphone5, in fact it's 20% better than the ipad4 !

Galaxy S4 octa uses SGX 544 MP3, which is 1.64x performance of iPhone 5. However Galaxy S4 leads iPhone 5 about 2.63 x in GL 2.7 rex offscreen 1080p, while it leads about 1.4x in GL 2.5 egypt offscreen 1080p. If we use GL 2.5 then iPhone 5s will lead iPhone 5 by about 2x, which is very close to official claim.

Apparently iPhone 5 scores too low in GL 2.7 according to its GPU spec. May I ask why? Thx!

 

[Turbotab]

That's one of several openings. The IPC of this core is so high, and the power draw (and die size) so low, that from a technical point of view, the field is open. What makes strategic sense, or sense for their customers/prospective markets is another matter, of course.

It wasn't a given that these mobile cores would evolve to this level of IPC, at least not at this point in time. This means that if Apple so chooses, it could utilize its mobile volumes to drive core development that could yield very competitive parts in other segments as well. (As opposed to for instance using their 20 million/year mac volume as CPU technology driver alone.)
From a financial point of view, given that Apple is using Intels more expensive offerings, I can certainly see the attraction in keeping that money for your own R&D and production. (For comparison, AMDs total annual revenue is $5billion. ARMs is less than 1$billion.)

Thinking ahead. At a SME level at least, because of myriad issues, including outsourcing, cost-cutting, low wages and general lack of understanding / appreciation of IT, the recruitment / retention of competent IT staff is a real issue.

The appeal of a one vendor solution for what remains of your IT infrastructure not in the cloud or already outsourced is very appealing, especially with Apple's reputation for ease of use, security and general mass appeal. I mean give your employee a Blackberry, and you're likely to get it thrown back in your face, but give them an iPhone etc.....

I wonder if Apple has any R&D plans to use the Cyclone core in an experimental HPC co-processor as a Xeon PHI rival, but back to reality..

 

[wco81]

Apple is not big in servers.

It sounds like Cyclone is at parity with Bay Trail according to the Anandtech review?

So can Apple produce the A7 less expensively than buying Bay Trail from Intel? Though Intel would have to integrate Rogue.

 

[tangey]

Galaxy S4 octa uses SGX 544 MP3, which is 1.64x performance of iPhone 5. However Galaxy S4 leads iPhone 5 about 2.63 x in GL 2.7 rex offscreen 1080p, while it leads about 1.4x in GL 2.5 egypt offscreen 1080p. If we use GL 2.5 then iPhone 5s will lead iPhone 5 by about 2x, which is very close to official claim.

Not sure what throwing in the galaxy S4 does to the discussion.
Apple have often used gpu compute as the basis of its graphics improvement. Assuming the fillrates are correct, Gpu compute must be up about x3. Gl2.7 is also way over x3.

Apparently iPhone 5 scores too low in GL 2.7 according to its GPU spec. May I ask why? Thx!

Not sure its "too low", but generally, SGX has relatively poor shader performance.

 

[silent_guy]

We still have to wait for a comparison of a real application compiled for both 32bits and 64bits on the 5s, but for Geekbench, the difference is a whopping 40%. This is probably an upper bound. Does anyone remember how much faster 64bits was vs. 32bits on the same CPU for x86 and for similar kind of benchmark?

(Imagine how much faster the 5s could have been if Apple didn't only care about thickness of the phone and started paying attention to performance! )

 

[Lazy8s]

RDGoodla's point was that the iPhone 5's GfxBench T-Rex performance was low relative to what would be expected extrapolating from the scores of the Galaxy S4 Exynos 5410, the iPad 4, etc.

It's something I've always wondered about, too. Improved drivers for the iPhone 5, like the iPad 4 recently saw, would probably resolve most of the imbalance, I presume.

 

[Gubbi]

We still have to wait for a comparison of a real application compiled for both 32bits and 64bits on the 5s, but for Geekbench, the difference is a whopping 40%. This is probably an upper bound. Does anyone remember how much faster 64bits was vs. 32bits on the same CPU for x86 and for similar kind of benchmark?

The Geekbench numbers are boosted by 800% performance gains in crypto-hashes and a 200% boost from vectorization of floating point benchmarks using doubles (SIMD can operate on doubles now).

For the remainder, the biggest improvements are in compression benchmarks which are likely helped by the larger D$. General speedup of general code spaghetti seems to vary from 5 to 30% That's still very impressive.

Cheers

 

[tangey]

RDGoodla's point was that the iPhone 5's GfxBench T-Rex performance was low relative to what would be expected extrapolating from the scores of the Galaxy S4 Exynos 5410, the iPad 4, etc.

It's something I've always wondered about, too. Improved drivers for the iPhone 5, like the iPad 4 recently saw, would probably resolve most of the imbalance, I presume.

Ah, I missed his point.

I think the big differential between GL2.5 and GL2.7, Samsung v iphone5" might be at least in part due to fillrate.

Fillrate seems more important in GL2.5 that GL2.7.

IMG recently did a blog posting that shows how close to theoretical maximum, SGX cores got to fill rates. However the stand out one was the non-named highly clocked one, which had to be the Samsung chip. It appears that higher clocked multi-core SGXs have problems hitting their fillrates. Or perhaps it was just the way Samsung implemented it.

http://withimagination.imgtec.com/i...vrs-market-leading-fillrate-efficiency-part-8

"only one product with a very high clock frequency sits at just below 70% efficiency".

So you might say that the Samsung was under performing in the GL2.5 versus iphone5 as much as overperforming in GL2.7.

 

[anexanhume]

According to the teardown, looks like M7 is actually just part of the A7. Not a separate chip.

http://www.ifixit.com/Teardown/iPhone+5s+Teardown/17383/2

 

[ninelven]

It sounds like Cyclone is at parity with Bay Trail according to the Anandtech review?

Considering the Z3770 in Anand's comparison is clocked at 1.46GHz with burst up to 2.4 GHz, I'd say Cyclone/A7 looks impressive indeed.

 

[tangey]

full iphone5s up on glbench

http://gfxbench.com/result.jsp