AT's iPhone 6S review is up

[ChronoReverse]

Not that I particularly like Geekbench, but in a world of imperfect benchmarks, why aren't there any of those in the review? It would at least be more useful than javascript benchmarks.

 

[swilli89]

If only I could get my hands on that chip outside of an iOS environment. It looks very interesting.

I really do wonder if Apple will take that step and use this chip in an "ultrabook" or something like it (or any of their future chips in the near-ish future)...they certainly seem to approach acceptable performance levels

Haha yes. If I could have android on a iPhone 6s I would do it in a heartbeat.

 

[Khato]

However unlike some posters here, I sincerely doubt that this design could clock much higher- leaving slack in the design for higher clocks means worse efficiency at lower clocks. (Intel demonstrated this years ago with their excellent Pentium M line.)

While there's certainly some margin in timing paths to account for yield variation, anyone who thinks that it could be overclocked a meaningful amount if it were possible is indeed delusional. Even increasing voltage dramatically wouldn't necessarily help much given the characteristics of the process. And even if you assume that it could increase frequency by 20% versus the ~10% average on Skylake that'd still only put it at around 2.2 GHz. And how much power would it be using at that point?

Here's an interesting question - how much larger would Apple's CPU cores be if they were designed to operate at frequencies comparable to Skylake? 2x? 3x? It's not at all trivial.

 

[cytg111]

Haha yes. If I could have android on a iPhone 6s I would do it in a heartbeat.

I dont know, Android is getting on my nerves, I might be up for a switch sooner rather than later.

 

[frozentundra123456]

Yea, I hated Android on the tablet that I had. Part of it was just that the hardware was slow (early Acer tablet) but I really did not like android either. Lots of people criticize it, but I much prefer my Win 8 tablet. I havent tried an apple tablet, but everyone I know that has one loves it.

 

[lefty2]

Not sure how relevant that link is...but 3dmark physics shows linear increases with thread counts in all platforms except for Apple (because of hard coding of thread counts).

Check this post:
http://www.futuremark.com/pressrele...results-from-the-apple-iphone-5s-and-ipad-air
The performance problem on iOS has nothing to do with threads.

 

[krumme]

While there's certainly some margin in timing paths to account for yield variation, anyone who thinks that it could be overclocked a meaningful amount if it were possible is indeed delusional. Even increasing voltage dramatically wouldn't necessarily help much given the characteristics of the process. And even if you assume that it could increase frequency by 20% versus the ~10% average on Skylake that'd still only put it at around 2.2 GHz. And how much power would it be using at that point?

Here's an interesting question - how much larger would Apple's CPU cores be if they were designed to operate at frequencies comparable to Skylake? 2x? 3x? It's not at all trivial.

I agree but it also shows that apple chose the right way with lower freq design from a business perspective. Cheaper process - dual sourcing. Far better perf/w. And as said. It targets the market where there is still growth.
Skylake is a fp tractor next to a a9 from a mobile perspective.

- yes different markets but trying to get Intel big cores to compete with future high perf arm and apple is futile.

Remember atom was for that?
Now look where we are.

 

[thunng8]

Check this post:
http://www.futuremark.com/pressrele...results-from-the-apple-iphone-5s-and-ipad-air
The performance problem on iOS has nothing to do with threads.

Yes, that is why one reason why it is slow.. but 3dmark never bothered to update their application when the Air 2 came out.

The air 2 has 3 cores and the 3rd core was not getting used. That just shows how bad their support for iOS is.

 

[lefty2]

Yes, that is why one reason why it is slow.. but 3dmark never bothered to update their application when the Air 2 came out.

The air 2 has 3 cores and the 3rd core was not getting used. That just shows how bad their support for iOS is.

The other link I posted clearly shows that Bullet physics is not multi-threaded. There's an experimental build that runs on windows only, but the main branch is definately not.

 

[Nothingness]

Spec2000 likes great caches.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.27.5974&rep=rep1&type=pdf

So something that likes caches isn't CPU bound? Caches don't belong to CPU? Anyway I agree SPEC 2k is too old (cf a previous post I made on this very thread).

 

[Arachnotronic]

So something that likes caches isn't CPU bound? Caches don't belong to CPU? Anyway I agree SPEC 2k is too old (cf a previous post I made on this very thread).

SPEC2k CPU is still better than any JavaScript benchmark.

 

[thunng8]

The other link I posted clearly shows that Bullet physics is not multi-threaded. There's an experimental build that runs on windows only, but the main branch is definately not.

That is false with 3dmark physics results.

eg.
http://www.futuremark.com/hardware/mobile/LG+G4/review
http://www.futuremark.com/hardware/mobile/Samsung+Galaxy+S6+Edge_/review

quad core a57 is more than 50% faster than dual a57.

 

[Vesku]

While there's certainly some margin in timing paths to account for yield variation, anyone who thinks that it could be overclocked a meaningful amount if it were possible is indeed delusional. Even increasing voltage dramatically wouldn't necessarily help much given the characteristics of the process. And even if you assume that it could increase frequency by 20% versus the ~10% average on Skylake that'd still only put it at around 2.2 GHz. And how much power would it be using at that point?

Here's an interesting question - how much larger would Apple's CPU cores be if they were designed to operate at frequencies comparable to Skylake? 2x? 3x? It's not at all trivial.

I'd guess somewhere at or under 2.5GHz would be where it would top out, i.e. hit a power wall. iPad version of the SoC should give us a better indication. Will be interesting to compare it to similarly clocked Broadwell and Skylake, the mobile i3s.

 

[Hi-Fi Man]

So something that likes caches isn't CPU bound? Caches don't belong to CPU? Anyway I agree SPEC 2k is too old (cf a previous post I made on this very thread).

The wonderful PowerPC 750 with external L2 cache . Appropriate considering the age of that benchmark. Don't forget cache used to be on the motherboard too.

 

[Nothingness]

SPEC2k CPU is still better than any JavaScript benchmark.

Yes and no. No because more benchmarks is always better. And yes because JS depends so much on the JS compiler that using it to compare different architectures is worse than using Geekbench alone.

An example of the obsolescence of SPEC 2k is the fact that 176.gcc scales linearly on 4770K and its 8MB L3 (this was posted on RWT some weeks ago).

 

[Gikaseixas]

Makes me wanna buy Apple again
This could be the foundation for bigger, faster and actively cooled solutions. I bet they have plans to eventually equip future MacBook's with ARM cpu's and stop paying 3rd party for processors.

 

[knutinh]

That would be great, but will never happen because there's no business case in providing such a benchmark. The amount of work associated with delivering of highly hand optimized code for each CPU family is just too much for any 3rd party. What should the benchmark test? Intel would love heavy optimized AVX2 (and AVX512) FP32/FP64 with random memory access patterns and it would annihilate anything from ARM. ARM would love general purpose code (preferable with fixed functions) with sequential memory access into cache.

I would be interested in:
1. Simple low-level tests. How many float32 adds or mults or multiply-accumulate or divs can be carried out per second when data is hot in the cache. Code should be hand-optimized for the architecture.
2. Representative cache/memory efficiency tests. I don't know what to ask for here, but getting some feel of what memory footprint/access patterns is possible before memory really starts to bog things down.
3. "hw + compiler tests". What kind of performance can be had for inner-loops written in (sensible) plain C using manufacturer recommended compiler and sensible compiler switches.
4. Simple library tests. What are the fastest FFT, array multiplication etc libraries available for each platform (people generally don't implement these things themselves).
5. A wide range of "real-world" applications within different domains using the latest code/compilers.

Synthetic tests can give us some idea about the fundamental limits of the machine. Hopefully those ideas can be extrapolated to a range of practical use cases. For concrete use cases (how fast will Quake run), it will always be safer to measure that application directly. But without specific knowledge about how that application is written, it is hard to extrapolate findings

I imagine that many real-world applications have highly suboptimal code due to things like delivering a working product on time. While this can be said to be representative of "how the world is" and something that a given cpu just have to cope with, the specific suboptimal-ness of one application could be very different from another in how it differentiation between cpu A and cpu B.

-k

 

[ShintaiDK]

So something that likes caches isn't CPU bound? Caches don't belong to CPU? Anyway I agree SPEC 2k is too old (cf a previous post I made on this very thread).

Not computational no

Its no different than the Broadwell-C contra Haswell/Skylake case, except in this case its just L4

Both are great products, but you just have to remember how they primary gained it. It is a shame tho with all the one sided benchmarking.

 

[ShintaiDK]

In terms of the graphics scores, if anyone tries to compare with the PC. They have to remember Windows uses FP32 and IOS/Android FP16. This has a huge performance impact.

 

[stingerman]

However unlike some posters here, I sincerely doubt that this design could clock much higher- leaving slack in the design for higher clocks means worse efficiency at lower clocks. (Intel demonstrated this years ago with their excellent Pentium M line.)

Snapdragon 805 can burst up to 2.7Ghz, if only for a few seconds. Unlike Qualcomm, Apple does't play that false marketing card. It's not a question of worse efficiency at lower clocks, the requirement is to clock high at the lowest possible voltage. It's the voltage that creates the exponential heat increase as well as battery drain. In benchmarks, Apple's A9 can sustain peak performance for the life of the battery. It gives me every indicator that under a wider cooling envelope, these things can hit sit significantly higher clocks.

Don't forget Apple has the disadvantage of needing to clock their processors for the best yield at the start of a given node. That means that the 1.85Ghz of the A9 provides them the best yields when the process node is nascent. It takes about a year to perfect yields, that is in the time of the A10. I wouldn't be surprised to see a 2.2 to 2.4Ghz A10, stacked with HBM.

 

[stingerman]

In terms of the graphics scores, if anyone tries to compare with the PC. They have to remember Windows uses FP32 and IOS/Android FP16. This has a huge performance impact.

How much precision do you need at HD or probably sub HD quality (Aren't most mobile games upscaled)?

 

[stingerman]

Snapdragon 805 can burst up to 2.7Ghz

Another detail to add here, is that the 805 goes back to 1.3Ghz after a few moments. So look at it this way, Snapdragon, doubled its clock rate with a voltage jump. So similarly, it's not unreasonable to expect that the A9 could leap forward at a higher voltage, and with a proper heatsink maintain that speed.

 

[witeken]

What does Arachsomething think of http://seekingalpha.com/article/3634056-apple-does-intel-inside-iphone-make-any-sense?

 

[StinkyPinky]

Gold rating for a phone that has worse battery life. What an epic fail. Sure it is faster and has a couple new features, but we're talking $800 dollars here. You shouldnt have to sacrifice battery life on a $800 "upgrade".

All phone makes seem to do this. The new iphones are already ridiculously thin, I would much prefer they make then 20% fatter and heavier and add in a decent battery. Seems I'm in the minority though.

The CPU performance of the iphone 6s is crazy fast however. Intel need to stop screwing around.

 

[stingerman]

I would much prefer they make then 20% fatter and heavier and add in a decent battery

They do make it, it's called the iPhone 6s Plus