Next-Gen iPhone & iPhone Nano Speculation

http://www.chipworks.com/en/technical-competitive-analysis/resources/blog/inside-the-iphone-5s/

So the A7 is made at Samsung afterall and the M7 is a discrete chip with a Cortx-M3 by NXP.

 

There was a TSMC test chip out there though. Manufactured in 43rd week of last year.

 

First shot of the die from Chipworks:



Edit: oh and looks like Anand might have more info: http://www.anandtech.com/show/7355/chipworks-provides-first-apple-a7-die-shot

 

Looks like around 4 MB of SRAM over the GPU. That helps with transistor density and count.

 

These cores are quite massive - Anand puts them at 17% of die, hence 17mm².
Hence 4 cores (with 2x1MB L2) would take ~34mm², compare that to Kabini (Jaguar core size without L2 3.1mm²) where 4 cores including L2 of 2MB seem to be definitely smaller (below 30mm² I would guess).

 

Cool, Apple used a stacked sensor.

 

Managing top-end performance all around against devices well beyond its size class -- doing it without pushing toward the diminishing returns of high clock speed and core count nor necessitating a major foundry process advantage -- shows Apple's head is still in the right place with the A7's design. Once my performance expectations were readjusted to account for the limited thermal/power profie of the 5s's casing, I was really able to appreciate their accomplishment here.

None of the details uncovered so far in the analysis and investigation of their new hardware helps me to fully account for the oddities in their graphics benchmark performance: it has the pixel fill rate of a 433 MHz G64xx by my understanding and estimation but nothing resembling twice the performance of MediaTek's G62xx solution outside of that aspect of performance. To me, it almost feels like some kind of in-between variant of the two, like an SGX535 type situation with the ALU count of the lower configuration and the TMU count of the higher config.

Maybe other implementation details of the reference design of those cores, like clock rate or buffer size or memory interfaces, account for the discrepancies. Or, maybe, as was suggested to me in a bit of intriguing speculation, Apple didn't strictly implement one of the PowerVR reference designs in this situation. I have no doubt that the designs are Rogue based (GPU design takes an exeedingly high degree of specialization to produce competitive results... Apple won't be going solo on design for GPUs), but maybe they customized some stuff or used some kind of unannounced variant, like say a G6235. Really, there's just a lack of information to go on right now.

 

Comparing Anand's browser results to my 3GHz Phenom II, Cyclone is slightly faster per clock.

 

Did you normalize for clock speed or downclock the Phenom to 1.3GHz? I expect the latter would yield better performance per clock.

 

I just normalized for clockspeed. I will try re-clocking though and see what I get.

 

Ha sure if you downclock the CPU without downclocking RAM you'll get better perf/MHz

You have to DC memory too, but I don't know how fast iPhone RAM are.

 

Hmm, I think the Phenom II only had (official) support for DDR3-1333, at most. So the latest iPhone's memory might not be slower.

 

My memory is 1333Mhz DDR3, so it should be fairly comparable.

Here are the results:

Code:

                 Phenom II 1.3GHz                Cyclone 
SunSpider:       493.1                vs         416        (lower is better)

Kraken:          5465.8               vs         5904.9     (lower is better) 

Octane:          5201                 vs         5500       (higher is better)

Browser used: Chromium 28

 

Thanks! I really wonder how high Cyclone could clock.

In any case, that's a very impressive core.

 

Do you really think the A7 has two 64-bit channel running at that speed? I don't but I could be wrong

 

It's confirmed to be 64-bit (IE, single-channel DIMM equivalent), according to Anand's review. Could very well be dual 32-bit channels in A7 of course for better granularity/latency etc.

 

I looked at various Geekbench 3 stream scores, and it seems Phenom has lower BW than iPhone 5s. I'm surprised...

 

Apparently not, but I suspect latency is a lot more important than bandwidth, here.

 

The margin which separates the iPhone 5 from the iPad 3 in overall performance in 3DMark is tied a lot more to the difference between the Swift versus A9 CPUs than I expected. 3DMark does have sub-results which isolate graphics better, of course. Still, its stronger consideration for CPU and core count leads to a fairly different looking results page than some of the other benchmark pages out there, for certain device comparisons at least (since mobile SoCs tend to tie the highest performing GPUs with the highest performing CPUs at the high-end, many of the same devices will rank similarly on all suites, obviously.)

http://community.futuremark.com/hardware/mobile 3DMark

http://results.rightware.com Basemark X

http://gfxbench.com/result.jsp T-Rex GfxBench

 

Chipworks has published an annotated die shot:


http://www.chipworks.com/en/technic...ources/blog/inside-the-a7/?lang=en&Itemid=815

They figure the ~3MB above the GPU may be for the fingerprint sensor data. First, I wonder if that SRAM really isn't a power of 2. Second, does that make sense? Seems the fingerprint data stored would have to live in some non-volatile RAM, then possibly get expanded into that SRAM for quick analysis/computations. Still don't know if I buy that explanation though.