Apple A9X the new mobile SoC king

[NTMBK]

Where does AnandTech have the Twister name from?

http://www.macrumors.com/2014/08/31/apple-hires-anand-lal-shimpi/

 

[jfpoole]

Why would the dataset be different? Makes comparing mostly useless.

The desktop dataset is too large for mobile devices -- it does not fit onto mobile devices with less than 1GB of RAM. Even if it did fit, the runtime would be unreasonably long for our users.

The mobile dataset it too small for desktop devices -- execution times would become small enough that timer precision could be an issue on some platforms.

The results are comparable -- we consistently measure less than a 1% difference between the desktop and the mobile dataset. See, e.g., the results for an iPhone 6s running the mobile dataset (left) and the desktop dataset (right):

http://browser.primatelabs.com/geekbench3/compare/3594692?baseline=3594646

 

[Eug]

The desktop dataset is too large for mobile devices -- it does not fit onto mobile devices with less than 1GB of RAM. Even if it did fit, the runtime would be unreasonably long for our users.

The mobile dataset it too small for desktop devices -- execution times would become small enough that timer precision could be an issue on some platforms.

The results are comparable -- we consistently measure less than a 1% difference between the desktop and the mobile dataset. See, e.g., the results for an iPhone 6s running the mobile dataset (left) and the desktop dataset (right):

http://browser.primatelabs.com/geekbench3/compare/3594692?baseline=3594646

So when does Geekbench 4 come out?

 

[witeken]

Why so?

Look at what Anand wrote

It wasn't until I wrote this piece that Apple's codenames started to make sense. Swift was quick, but Cyclone really does stir everything up.

 

[raghu78]

So are we 100% sure that ALL 6s units are Samsung and ALL 6s+ units are TSMC? Possible, but not guaranteed.

The logic is a chip name is related to a certain foundry and its associated FINFET process. so APL0898 is based on Samsung 14nm and APL1022 is based on TSMC 16nm. There is no logic in having iPhone 6S having two different chips. Any product needs consistency. Thats possible only if all chips used in a device are manufactured on the same process. So in conclusion APL0898 - iPhone 6S - Samsung 14nm.
APL1022 - iPhone 6S Plus - TSMC 16nm

 

[Eug]

The logic is a chip name is related to a certain foundry and its associated FINFET process. so APL0898 is based on Samsung 14nm and APL1022 is based on TSMC 16nm. There is no logic in having iPhone 6S having two different chips. Any product needs consistency. Thats possible only if all chips used in a device are manufactured on the same process. So in conclusion APL0898 - iPhone 6S - Samsung 14nm.
APL1022 - iPhone 6S Plus - TSMC 16nm

So in other words, you don't know for sure.

 

[raghu78]

So in other words, you don't know for sure.

I only know that chipworks thinks APL0898 is fabbed by samsung. anyway they will soon have die shots and metal layer shots.

 

[jfpoole]

Which workloads use NEON/SSE SIMD instructions? We do talk about SIMD instructions and not scalar instructions, right?

So Win version of GB4 will use VS2015?

SGEMM and DGEMM take advantage of SSE and NEON.

Geekbench 4 will be built with Visual Studio 2015 on Windows.

 

[jfpoole]

so when does geekbench 4 come out?

q1 2016.

 

[Systems analyst]

For the technology to show impact it must be either:
1) They use different designs (on RTL level)
2) They run at different clock-speeds

I consider both very unlikely.

ARM have standard RTL designs for their cores, which could be used on various production processes, but they also offer what they call "physical IP", including optimised versions of the RTL for a specific production process; for example, they finished work on developing their "physical IP" for TSMC's 16nm process last year; they then started working on optimising TSMC's 10nm process. They are working on optimising their RTL for Samsung and Global Foundries, and UMC also, I believe. They are even working on 7nm optimisation with an unknown partner.

Given this background information, I would think it likely that Apple would optimise their RTL per process, as the whole point of a custom design is optimisation.

 

[Arachnotronic]

The logic is a chip name is related to a certain foundry and its associated FINFET process. so APL0898 is based on Samsung 14nm and APL1022 is based on TSMC 16nm. There is no logic in having iPhone 6S having two different chips. Any product needs consistency. Thats possible only if all chips used in a device are manufactured on the same process. So in conclusion APL0898 - iPhone 6S - Samsung 14nm.
APL1022 - iPhone 6S Plus - TSMC 16nm

You do realize Apple has multi-sourced components for its iPhones for a long time, right? As long as the APL1022 and APL0898 are pin/electrically compatible and fit in the same package size, dual sourcing within a given model shouldn't be an issue.

 

[paffinity]

SGEMM and DGEMM take advantage of SSE and NEON.

Geekbench 4 will be built with Visual Studio 2015 on Windows.

If it used the Accelerate library you'd see quite a large increase in many of the tests, especially in the GEMMs.

 

[coercitiv]

I didn't know Typhoon either.

Pretentious names, BTW.

They probably got the names the same way Swift and Cyclone were discovered.

With Swift, I had the luxury of Apple committing LLVM changes that not only gave me the code name but also confirmed the size of the machine (3-wide OoO core, 2 ALUs, 1 load/store unit). With Cyclone however, Apple held off on any public commits. Figuring out the codename and its architecture required a lot of digging. Last week, the same reader who pointed me at the Swift details let me know that Apple revealed Cyclone microarchitectural details in LLVM commits made a few days ago (thanks again R!).

Anyway, pretentious indeed: they are on collision course for Category 6. :sneaky:

 

[EightySix Four]

Looks like they're mixing and matching even in the same type of phone (not just split for + and non + models):

It was a surprise to find two different application processors in two otherwise identical phones. As pictured below, there is a difference in the die size for the APL0898 (Samsung) and the APL1022 (TSMC).

http://www.chipworks.com/about-chipworks/overview/blog/a9-is-tsmc-16nm-finfet-and-samsung-fabbed

 

[jfpoole]

If it used the Accelerate library you'd see quite a large increase in many of the tests, especially in the GEMMs.

Only on Apple platforms, though. Other platforms don't have an equivalent to Accelerate.

 

[Arachnotronic]

Looks like they're mixing and matching even in the same type of phone (not just split for + and non + models):




http://www.chipworks.com/about-chipworks/overview/blog/a9-is-tsmc-16nm-finfet-and-samsung-fabbed

I guess some people owe Idontcare an apology. He was spot on, as he was in saying TSMC won the entirety of the Apple 20nm.

 

[Eug]

Looks like they're mixing and matching even in the same type of phone (not just split for + and non + models):

http://www.chipworks.com/about-chipworks/overview/blog/a9-is-tsmc-16nm-finfet-and-samsung-fabbed

And Boom! There ya go.

I guess with a relatively new process and 13 million phones needed in 3 days, sourcing can be a problem.

Oh and I thought this was the ultimate geek statement.

"With the same exact SoC design implemented in two different technologies, it just doesn’t get any better than this for us at Chipworks to benchmark the technologies!"

q1 2016.

Looking forward to it.

 

[paffinity]

Only on Apple platforms, though. Other platforms don't have an equivalent to Accelerate.

On x86 you could use MKL. Other's would be out of luck. These libraries show the true performance of the system though...

 

[NTMBK]

I guess some people owe Idontcare an apology. He was spot on, as he was in saying TSMC won the entirety of the Apple 20nm.

This.

 

[Thanatosis]

I guess some people owe Idontcare an apology. He was spot on, as he was in saying TSMC won the entirety of the Apple 20nm.

http://www.chipworks.com/about-chipworks/overview/blog/a9-is-tsmc-16nm-finfet-and-samsung-fabbed

This says nothing about them finding APL0898 and APL1022 *both in the 6s*, they are clearly referring to the 6s plus which they already stated had only APL1022. I am pretty sure by "identical phone" they were referring to the otherwise identical components in the 6s and 6s plus. It is highly unlikely that Apple would source two different foundries for the exact same chip when they already have a simpler way to use the TSMC chips while maintaining consistency by using them in the 6s Plus.

Nobody should apologize for pointing out when a popular poster is wrong. There is nothing wrong with being wrong, but you shouldn't try to silence other forum members just because you like a poster. He acted like he knew exactly where the chips came from and when the rubber met the road it is clear that's not the case at all.

By the way, I would point out that even if they are sourcing 6s A9 from both Samsung and TSMC, IDC was still wrong. He claimed they were all TSMC which is patently false. He gave zero information and then pointed at other posters being "fanboys". If anything he should be apologizing to US.

 

[Arachnotronic]

http://www.chipworks.com/about-chipworks/overview/blog/a9-is-tsmc-16nm-finfet-and-samsung-fabbed

This says nothing about them finding APL0898 and APL1022 *both in the 6s*

That's what "otherwise identical" implies.

Nobody should apologize for pointing out when a popular poster is wrong. There is nothing wrong with being wrong, but you shouldn't try to silence other forum members just because you like a poster. He acted like he knew exactly where the chips came from and when the rubber met the road it is clear that's not the case at all.

It's not clear to me.

 

[Thanatosis]

That's what "otherwise identical" implies.



It's not clear to me.

It's not clear to me whether they are referring to the 6s or 6 plus and 6s... but obviously you have chosen to interpret the vague statements on the quoted chipworks page to IDC's benefit. I am not as confident as you in your mentor's infallibility, sadly, so in my mind it remains to be seen whether they are sourcing APL0898 and APL1022 both for the 6s. I doubt it.


Regardless, there is a trend emerging.
https://medium.com/@InertialLemon/apple-s-bitcode-telegraphs-future-cpu-plans-a7b90d326228

 

[Fjodor2001]

It's quite cool if we now have the exact same chip design on two different process techs. For once it ought to be possible to make a fair comparison between TSMC 16 nm and Samsung 14 nm process tech with regards to characteristics and metrics. No blaming on uArches on different processes differing in design, affecting clock frequency and transistor density, etc thus mot making a fair comparison possible.

It would be even more fun if the same chip design was on Intel 14 nm too, so all of them could be compared on equal terms.

 

[Arachnotronic]

It's quite cool if we now have the exact same chip design on two different process techs. For once it ought to be possible to make a fair comparison between TSMC 16 nm and Samsung 14 nm process tech with regards to characteristics and metrics. No blaming on uArches on different processes differing in design, affecting clock frequency and transistor density, etc thus mot making a fair comparison possible.

It would be even more fun if the same chip design was on Intel 14 nm too, so all of them could be compared on equal terms.

It's funny that people are acting surprised that a chip vendor was able to dual source a chip. The Qualcomm MDM9615 baseband inside of the iPhone 5/5s was manufactured by both Samsung and TSMC but nobody really made a big deal of it.

 

[EightySix Four]

It's funny that people are acting surprised that a chip vendor was able to dual source a chip. The Qualcomm MDM9615 baseband inside of the iPhone 5/5s was manufactured by both Samsung and TSMC but nobody really made a big deal of it.

It's pretty uncommon (never?) for a current gen high performance SOC to be dual sourced on two different processes. No one is benchmarking two basebands from two different sources against each other because there is nothing to be learned. This is the current leading edge fabs (excluding Intel) with the exact same processor and there is an opportunity to learn a ton about those processes because of it.

It also explains the geekbench scores being all over the place.