Apple A9X Geekbench

[Thala]

You're only arguing semantics at this point.

Pretty much! Would you like to argue without semantics?

In any case, an OS does not suddenly starts sucking only because it is compiled for another target architecture.

It might suck for you that you tried to run x86 apps and failed. I re-compiled apps for ARM before running them on Windows RT and that sucked, not surprisingly, much less.

 

[Arachnotronic]

Pretty much! Would you like to argue without semantics?

In any case, an OS does not suddenly starts sucking only because it is compiled for another target architecture.

It might suck for you that you tried to run x86 apps and failed. I re-compiled apps for ARM before running them on Windows RT and that sucked, not surprisingly, much less.

Most users aren't going to re-compile apps for ARM. I don't have the source code for many of my favorite applications, sadly.

 

[Thala]

Most users aren't going to re-compile apps for ARM. I don't have the source code for many of my favorite applications, sadly.

Exophase would like to have Linux on it for AArch64 development, Nothingness OS X and I would like to have Windows for the same reason.
You could argue that i could just develop for Windows Phone/Mobile. However the Windows Mobile Sandbox (WinRT runtime) is a piece of crap compared to what get offered by full Windows/Win32.
Regarding source code, there are lots of open source application readily compilable for the Win32 desktop.
Sure, if you are just a user, Windows RT might not be the right choice.

 

[cytg111]

Pretty much! Would you like to argue without semantics?

In any case, an OS does not suddenly starts sucking only because it is compiled for another target architecture.

It might suck for you that you tried to run x86 apps and failed. I re-compiled apps for ARM before running them on Windows RT and that sucked, not surprisingly, much less.

http://www.extremetech.com/computing/129137-microsoft-bans-third-party-browsers-from-windows-on-arm

"The issue seems to revolve around API access in Windows RT. Third-party developers (and apps) will only have access to the WinRT (Metro) API, while Microsoft’s own software will have access to the low-level (and old school) Win32 API. This means that Mozilla can build Firefox for Windows on ARM, but without access to Win32 it will be very hard to compete with Internet Explorer."

- Different. Kind. Of. Animal.

 

[paffinity]

A geometric mean change nothing since the bases are not weighted, the contribution of the score with the highest absolute value will keep being exagerated.

Using my previous exemple :

CPU A

Test 1 500 pts
Test 2 100pt pts

Score 600/2 = 300 pts


CPU B

Test 1 400 pts
Test 2 200 pts

Score 600/2 = 300 pts

Let s use the geometric mean :

CPU A score = sqrt(500^2 + 100^2) = 509.9

CPU B score = sqrt(400^2 + 200^2) = 447.2


That s even worse, the CPU B is now significantly slower, the GM just inflated the higher subscore contribution.

Think you need to double check what geomean is:

CPU A : 223.6
CPU B : 282.84

 

[Thala]

http://www.extremetech.com/computing/129137-microsoft-bans-third-party-browsers-from-windows-on-arm

"The issue seems to revolve around API access in Windows RT. Third-party developers (and apps) will only have access to the WinRT (Metro) API, while Microsoft’s own software will have access to the low-level (and old school) Win32 API. This means that Mozilla can build Firefox for Windows on ARM, but without access to Win32 it will be very hard to compete with Internet Explorer."

- Different. Kind. Of. Animal.

Unfortunately this is a wrong analysis of the situation. Technically you have full access to all Win32 low level APIs plus COM Interfaces like DirectX. It is just Microsoft enforced signing with a certain certificate in order to make the runtime load your application. Once the signing issue was resolved (by a hack, mind you?) your app starts and runs just fine. Of course big companies like Mozilla did not want to rely on a hack.

 

[cytg111]

Unfortunately this is a wrong analysis of the situation. Technically you have full access to all Win32 low level APIs plus COM Interfaces like DirectX. It is just Microsoft enforced signing with a certain certificate in order to make the runtime load your application. Once the signing issue was resolved (by a hack, mind you?) your app starts and runs just fine. Of course big companies like Mozilla did not want to rely on a hack.

Yea, worked for a while on 8.0, at 8.1 that door was closed again.. just did search, seems like it still is.

You would be right if not for the fact that microsoft is locking it down. Hard.

 

[Thala]

Yea, worked for a while on 8.0, at 8.1 that door was closed again.. just did search, seems like it still is.

You would be right if not for the fact that microsoft is locking it down. Hard.

The solution for 8.1 consists of using bcdedit (tool shipped with Windows since Vista) to change start configuration to test signing mode. Once in test signing mode Windows accepts any certificate including self-signed certificates.

 

[quikah]

Comparison with 2C4T 2.4GHz Skylake: http://browser.primatelabs.com/geekbench3/compare/4048472?baseline=4149721

Single thread performance is highly competitive.

The huge performance difference of AES and SHA is kind of curious.

 

[Exophase]

The huge performance difference of AES and SHA is kind of curious.

Those use accelerated instructions on both platforms and are therefore subject to very big differences that aren't representative of other applications. I guess Apple doesn't think it's as important as Intel does. Given their respective market coverage it's probably justified.

I just ignore those scores. And unfortunately, that also means ignoring the composite integer and overall score... or just looking at the individual subtests in isolation. It's just as well though, I prefer to do that in benchmarks anyway.

 

[mikk]

Comparison with 2C4T 2.4GHz Skylake: http://browser.primatelabs.com/geekbench3/compare/4048472?baseline=4149721

Single thread performance is highly competitive.

You couldn't find a slower SKL result? This is clearly a Singlechannel device. So typical that people try to use a subpar Geekbench result from Intel. This is so stupid.

 

[Exophase]

You couldn't find a slower SKL result? This is clearly a Singlechannel device. So typical that people try to use a subpar Geekbench result from Intel. This is so stupid.

I did a search for Skylake and picked the first thing that looked close in clockspeed. I wasn't trying to cherry pick the results. If you have a fairer comparison by all means post that. Insinuating I have an anti-Intel bias and saying I'm being stupid doesn't really do anything.

I'm curious to see how much memory bandwidth impacts the single threaded integer subtests.

 

[mikk]

I did a search for Skylake and picked the first thing that looked close in clockspeed.

Do better next time with your search.

If you have a fairer comparison by all means post that. Insinuating I have an anti-Intel bias and saying I'm being stupid doesn't really do anything.

This is your job. At least use a dualchannel device. But damage is already done anyways as you can see.

 

[kimmel]

The the interface to sift through results manages to be even worse than the geekbench benchmark itself.

 

[kimmel]

http://browser.primatelabs.com/geekbench3/compare/4093592?baseline=3833358

ipad pro looks terrible compared to skylake! ()

edit: whoops, meant to edit the previous post.

 

[Exophase]

Do better next time with your search.

This is your job. At least use a dualchannel device. But damage is already done anyways as you can see.

Well I guess that's just too bad for you then. You don't get to complain about it then demand I find something that lives up to your standards better. You can do it yourself or you can keep whining about it because I'm not doing anything with that attitude.

 

[dahorns]

Or physics is an utterly non-representative PoS. When a single benchmark is an outlier, doubt it.

SPEC 2000 176.gcc has been shown to be competitive against an Haswell 4770K. And that is certainly way more interesting than what physics is measuring.

Uh, you should probably clarify that competitive means competitive with a down clocked processor. It is not actually competitive with a 4770k running stock clocks. Also, those numbers for the 4770k are bit out of whack with would you expect, so I'm not sure how much trust I'd put in them.

According to this guy (who may or may not post here) the A9 was matching a Core 2 Duo E6700 (2.6 ghz) on that particular benchmark (both ~3200).

Anyway, what's cool, we can pull Geekbench numbers for the E6700 and get the following:

Overclocked at 4.8 ghz
Single Core - 2516
Multi Core - 4720

Stock
Single Core - 1400
Multi Core - 2500

http://browser.primatelabs.com/geekbench3/search?dir=desc&q=E6700&sort=score

So, Spec2000 numbers (granted probably not optimally compiled for the A9) show the A9 and the E6700 as close, but with the E6700 winning the majority of tests. Yet, Geekbench shows the A9 crushing the E6700 at stock and matching the very best overclocked score listed. And you still see that result even if you ignore the encryption tests.

Point is, these cross-arch benches aren't exactly reliable or consistent.

 

[quikah]

You couldn't find a slower SKL result? This is clearly a Singlechannel device. So typical that people try to use a subpar Geekbench result from Intel. This is so stupid.

Something else that is interesting is the multichannel memory results show almost no scaling for the ipad. Does their architecture uses both cores no matter what for memory access?

 

[stingerman]

It's important to remember that Geekbench is not compiled in the latest version of Xcode.

 

[Exophase]

So, Spec2000 numbers (granted probably not optimally compiled for the A9) show the A9 and the E6700 as close, but with the E6700 winning the majority of tests. Yet, Geekbench shows the A9 crushing the E6700 at stock and matching the very best overclocked score listed. And you still see that result even if you ignore the encryption tests.

Point is, these cross-arch benches aren't exactly reliable or consistent.

Geekbench does have a lot of problems, and really it's even worse going cross platform because they don't actually use the same datasets between the Windows and iOS builds. They scale the scores to calibrate it out but this only works if the difference is linear and repeatably so between uarchs I doubt it's that close. Hopefully this is improved with GB4 (along with taking the accelerated encryption tests out of the main scores...) I probably wouldn't have brought up a GB comparison if not for the thread opening with it.

But SPEC2k has its own problems. They're more subtle but in a way that makes it more insidious. The big one is how much ICC is gamed for it. Now I don't know how much this applies to a current GCC or Clang/LLVM score vs whatever ICC was used for these numbers but back when VIA wrote about it in 2011 the difference was about 20% improvement for their processor and 25% for Atom. That was with both optimized with PGO.

http://www.centtech.com/wp-content/uploads/2014/09/WP1-NanoX2-whitepaper-1-3.pdf

There's a score breakdown and the biggest advantage is seen in 255.vortex and 300.twolf, which were over 50% faster using ICC on Nano. Those are also the two that show the biggest leads in the linked article, very similar ~50% leads actually. That makes it seem pretty evident to me that ICC is optimizing (maybe breaking...) these benchmarks in a way that GCC isn't.

Less obvious but still notable are 181.mcf and 197.parser which both exhibited ~33% leads using ICC. 181.mcf again shows one of the biggest differences.

For all of GB's faults at least it uses similar (same?) compiler versions for the different builds; that still won't ensure optimizations are applied equitably to all platforms but it'll make the difference a lot closer. Compilers like GCC are also a lot less likely to employ benchmark breaking optimizations because they don't have the marketing incentive ICC does. The AEC scores are still really bad and should be ignored, but otherwise I'd rather take a benchmark that has questionable relevance or representation but is fair than one that is relevant and unfair.

 

[Nothingness]

Uh, you should probably clarify that competitive means competitive with a down clocked processor. It is not actually competitive with a 4770k running stock clocks.

Yep sorry, it's so obvious to me that an A9 can't compete with a 4770K running at full speed (which would be a useless comparison given the TDP anyway) that I didn't feel the need to say at equal clock. I edited the message, thanks :thumbsup:

Also, those numbers for the 4770k are bit out of whack with would you expect, so I'm not sure how much trust I'd put in them.

According to this guy (who may or may not post here) the A9 was matching a Core 2 Duo E6700 (2.6 ghz) on that particular benchmark (both ~3200).

What makes you think the RWT guy number are wrong? If we scale the result, we get about 3800 for a 2.6 GHz.

Now if you look at the entry in SPEC db, you'll notice two things:

1. this is using icc which is known to be heavily optimzed for SPEC
2. this is running as 32-bit code and many of the SPEC tests are slower in 64-bit mode which the RWT guy used.

So the Intel E6700 score is inflated and I see nothing obviously wrong in the RWT result.

 

[Exophase]

Something else that is interesting is the multichannel memory results show almost no scaling for the ipad. Does their architecture uses both cores no matter what for memory access?

What do you mean by multichannel memory results? You mean the multicore memory tests? That means that a single core is capable of saturating the memory bus on its own. Which is what you'd expect from a good design. The system should be able to access nearly full bandwidth with the memory level parallelism generated through with a single core. This is possible through out of order memory requests creating multiple outstanding cache misses and prefetching.

 

[cytg111]

The solution for 8.1 consists of using bcdedit (tool shipped with Windows since Vista) to change start configuration to test signing mode. Once in test signing mode Windows accepts any certificate including self-signed certificates.

partial-jailbreak discovered bout a month ago .. 2 years too late and still a big hassle to execute right. I would not, from this fact, deduce that windows on arm == windows on x86.

 

[Exophase]

partial-jailbreak discovered bout a month ago .. 2 years too late and still a big hassle to execute right. I would not, from this fact, deduce that windows on arm == windows on x86.

Is that really what he's saying?

It sounds to me like he wants an ARM platform for some reason and wants a Windows-ish platform for some reason and that such a combination on a hypothetical Apple machine would suit his needs. Not that this should be a thing everyone should start using.

It does all seem pretty rhetorical though, because there's no way anyone's going to cobble together enough Windows drivers to make this thing work at all on an Apple SoC, at least not without Apple's help. Nevermind doing it with DirectX support.

 

[cytg111]

Is that really what he's saying?

- I believe it is, however it is of little consequence, furthering this dispute would gain very little. Whatever the point im pretty certain its laid out flat by now.