Apple A9X the new mobile SoC king

[catinthefurnace]

As opposed to Geekbench which can barely get 1 DP flop/cycle on a modern Intel core?

It's a tribute to just how powerful Intel cores are when it can score so high on a program that's optimized to make Apple cores look good.

Have you ever heard of Occam's razor? If your theory requires a conspiracy, it is almost always a bad theory.

And was it Apple that got caught cheating on mobile benchmarks, or was that other companies?

http://www.anandtech.com/show/7384/state-of-cheating-in-android-benchmarks

 

[Andrei.]

Let's wait for SPEC2000 score for A9, but I hope anandtech gives us SPEC06, LOL

SPEC 2006 requires about 10 times as long to run as SPEC 2000, might be a good throttling test too :biggrin:

SPEC2K already takes extremely long on mobile (Krait or A53 take about an hour), it's too early for 2006 to be a reasonable test.

 

[Thala]

No it isn't. As an example there's a single rounding mode.

Technically you are right, IEEE requires more than one rounding mode, however the support of Round-To-Nearest is in most cases not really a restriction.
How often did you change the default rounding mode in your live?

 

[Accord99]

Have you ever heard of Occam's razor? If your theory requires a conspiracy, it is almost always a bad theory.

Relatively unknown benchmark comes to prominence by being the poster boy for the 64-bit performance boost of the A7 cpu and despite clear problems with it in 2013, it's still somehow the primary benchmark used today.

And was it Apple that got caught cheating on mobile benchmarks, or was that other companies?

Of course, the only problem with that is despite Anand's insistence, how would you even know if iPhone does the same thing given Apples control of the entire system? There are many ways to recognize an application other than using its name. Or with Geekbench, where the "cheating" comes at the code and compiler level.

 

[Thala]

Or with Geekbench, where the "cheating" comes at the code and compiler level.

Give some evidence of you claims or just stop with your conspiracy theories.

 

[catinthefurnace]

Relatively unknown benchmark comes to prominence by being the poster boy for the 64-bit performance boost of the A7 cpu and despite clear problems with it in 2013, it's still somehow the primary benchmark used today.


Of course, the only problem with that is despite Anand's insistence, how would you even know if iPhone does the same thing given Apples control of the entire system? There are many ways to recognize an application other than using its name. Or with Geekbench, where the "cheating" comes at the code and compiler level.

And this big Geekbench, Google Octane, Kraken, Basemark, GFXBench, and WEBXPRT conspiracy is an easier theory for you to believe than the theory that Apple designed a really good SoC?

To each his own I suppose...

 

[Accord99]

Give some evidence of you claims or just stop with your conspiracy theories.

Geekbench is well recognized as being a collection of simple for benchmarks and over reliance of certain crypto benchmarks while not running the exact same workloads and using different compilers of various optimizing ability. It's really up to Geekbench to prove that it's a good cross-platform benchmark at all.

And this big Geekbench, Google Octane, Kraken, Basemark, GFXBench, and WEBXPRT conspiracy is an easier theory for you to believe than the theory that Apple designed a really good SoC?

I'm not saying that Apple didn't make a good SoC, it's just that they also have a great javascript engine that they've worked on far more than everyone else to get great scores on javascript benchmarks.

The real conspiracy is how after 2 years, we still have the same set of a terrible benchmarks being used.

 

[Nothingness]

Technically you are right, IEEE requires more than one rounding mode, however the support of Round-To-Nearest is in most cases not really a restriction.
How often did you change the default rounding mode in your live?

Never, but that's not the point, you claimed NEON is fully compliant Also subnormals are silently flushed to zero.

 

[Nothingness]

Geekbench is well recognized as being a collection of simple for benchmarks and over reliance of certain crypto benchmarks while not running the exact same workloads and using different compilers of various optimizing ability. It's really up to Geekbench to prove that it's a good cross-platform benchmark at all.

In the same way that Intel cheats with its compiler on SPEC 2006? Yeah, but no.

What about you prove it's not a good cross-platform benchmark? You're the one doing accusations after all

The real conspiracy is how after 2 years, we still have the same set of a terrible benchmarks being used.

Then do the world a favor: develop a good benchmark. Or point us to what is a great benchmark, and why it is great.

 

[R0H1T]

I'm not saying that Apple didn't make a good SoC, it's just that they also have a great javascript engine that they've worked on far more than everyone else to get great scores on javascript benchmarks.

The real conspiracy is how after 2 years, we still have the same set of a terrible benchmarks being used.

No what you, & some others, are saying is that Geekbench is terrible cause it favors Apple right? Also JS benchmarks mean nothing, when it comes to AMD though single core & single thread benches are the be all & end all of computing conveniently discarding the lead AMD has in graphics.

Yeah, if anything you should know that since iOS doesn't run the fastest browser(s) with their native JS engines the benchmark scores are even more impressive.

Checkout kraken with x64 nightly builds of firefox on windows, that's the fastest browser for the particular benchmark. Now for Octane it's chromium & dartium, latest builds, while sunspider runs best on chrome & some custom builds of FF as well!

Browsers are by far the most popular apps, across all platforms, & to term JS benches as meaningless just reeks of ignorance.

 

[ShintaiDK]

The problem is people mixes overall performance and specific performance. Specially when specific performance depends massively in this case on the software ecosystem.

People look at Geekbench and JS browser benches and concludes X CPU can compete or beat Y CPU.

I think a lot of people got a big lesson on memory bottlenecks just with the Broadwell-C benchmarks for example. Nobody would claim Broadwell(-C) had a 30% IPC increase over Haswell.

 

[Nothingness]

The problem is people mixes overall performance and specific performance. Specially when specific performance depends massively in this case on the software ecosystem.

People look at Geekbench and JS browser benches and concludes X CPU can compete or beat Y CPU.

I definitely agree. You need many benchmarks to compare two different CPUs in particular when they don't use the same instruction set.

 

[R0H1T]

The problem is people mixes overall performance and specific performance. Specially when specific performance depends massively in this case on the software ecosystem.

People look at Geekbench and JS browser benches and concludes X CPU can compete or beat Y CPU.

I think a lot of people got a big lesson on memory bottlenecks just with the Broadwell-C benchmarks for example. Nobody would claim Broadwell(-C) had a 30% IPC increase over Haswell.

The other problem that people, especially on this board, fail to see is that absolute performance is not necessarily the pinnacle or more appropriately the defining factor in judging a given platform. The AT frontpage displays the same content on (iOS) safari as it does on chrome & win10, this makes the JS benches even more impressive despite the fact that chrome & FF's native JS rendering engines on windows are miles ahead of anything on iOS.

This is why the Apple ecosystem is thriving, not necessarily because it offers the "best in class" performance but most certainly due to the fact that it has the best software (mobile) ecosystem coupled with a highly optimized hardware powering it.

 

[ShintaiDK]

The other problem that people, especially on this board, fail to see is that absolute performance is not necessarily the pinnacle or more appropriately the defining factor in judging a given platform. The AT frontpage displays the same content on (iOS) safari as it does on chrome & win10, this makes the JS benches even more impressive despite the fact that chrome & FF's native JS rendering engines on windows are miles ahead of anything on iOS.

This is why the Apple ecosystem is thriving, not necessarily because it offers the "best in class" performance but most certainly due to the fact that it has the best software (mobile) ecosystem coupled with a highly optimized hardware powering it.

That's the huge benefit of a controlled platform. Same goes for consoles for that matter. Legacy is both a burden and benefit. The Apple approach would never work on a grand IT scale. Even if they had world monopoly. But it does give a much better perception experience.

Just take my own Windows 10 with Edge and IE11.
Kraken gives:
IE11 1532ms.
Edge 728ms.

Same system and 6700K in both.

 

[stingerman]

Geekbench is well recognized as being a collection of simple for benchmarks and over reliance of certain crypto benchmarks while not running the exact same workloads and using different compilers of various optimizing ability. It's really up to Geekbench to prove that it's a good cross-platform benchmark at all.



I'm not saying that Apple didn't make a good SoC, it's just that they also have a great javascript engine that they've worked on far more than everyone else to get great scores on javascript benchmarks.

The real conspiracy is how after 2 years, we still have the same set of a terrible benchmarks being used.

Think the benchmarks that include the whole system are more realistic of actual real world experience. And, that includes power utilization during benchmarking. Why? Because experience has shown us that real world apps rely on the compiler and specialized API's that also evoke additional high performance circuitry that frees the generalized cores.

And really, we are buying a whole system and not just the CPU cores. However, It's just really cool knowing how powerful the CPU in your pocket is.

 

[Thala]

Also subnormals are silently flushed to zero

Not really silent. It is well documented, that NEON flushes to zero and it is a configuration option of VFPv4.

 

[Nothingness]

Not really silent. It is well documented, that NEON flushes to zero and it is a configuration option of VFPv4.

By silent I meant no flag is set to let you know subnormals were seen.

I really don't see the point discussing this any further, at least not in this thread.

Let's get back to what matters: A9 is crushing Skylake.... Hmm wait :biggrin:

 

[Fjodor2001]

At this pace of progress, when will the Apple SoCs overtake the Intel laptop CPUs in performance, and desktop CPUs... ?

 

[ShintaiDK]

Here is another example of why Geekbench is terrible and favours Apple designs rather heavily. (Apple improves memory bandwidth drastically between generations.)

50% performance boost with 50% faster memory. Pretty much scales linear with memory speed.

While a more computational heavy physics bench shows much less.

This is also reflected with the iPhone 6 Plus. As soon as it comes to real computational work it tanks massively.

 

[Nothingness]

Here is another example of why Geekbench is terrible and favours Apple designs rather heavily. (Apple improves memory bandwidth drastically between generations.)

This just shows that people who only look at the global score don't know what they are doing. The same applies to SPEC BTW.

While a more computational heavy physics bench shows much less.

This is also reflected with the iPhone 6 Plus. As soon as it comes to real computational work it tanks massively.

Now that is ridiculous: you pick the single known outlier and make an unsustained claim that it is representative of "real computational work". Do you really think an iPhone 6 is twice slower than a Galaxy 3 at "real computational work"? What makes you think 3dmark benchmark is not complete sh*t?

 

[ShintaiDK]

This just shows that people who only look at the global score don't know what they are doing. The same applies to SPEC BTW.

Now that is ridiculous: you pick the single known outlier and make an unsustained claim that it is representative of "real computational work". Do you really think an iPhone 6 is twice slower than a Galaxy 3 at "real computational work"? What makes you think 3dmark benchmark is not complete sh*t?

If Geekbench scales completely linear with memory speed. Do you really think its any indicator of CPU performance?

 

[witeken]

Here is another example of why Geekbench is terrible and favours Apple designs rather heavily. (Apple improves memory bandwidth drastically between generations.)

50% performance boost with 50% faster memory. Pretty much scales linear with memory speed.


This is also reflected with the iPhone 6 Plus. As soon as it comes to real computational work it tanks massively.

How can you even react to him seeing all those quotes of you in his sig?

 

[mavere]

Here is another example of why Geekbench is terrible and favours Apple designs rather heavily. (Apple improves memory bandwidth drastically between generations.)

50% performance boost with 50% faster memory. Pretty much scales linear with memory speed.
.

That graph specifically shows Geekbench Memory scores. Are you complaining that increasing memory bandwidth offers corresponding increases in memory bandwidth tests?

 

[Rakehellion]

Have you ever heard of Occam's razor? If your theory requires a conspiracy, it is almost always a bad theory.

And was it Apple that got caught cheating on mobile benchmarks, or was that other companies?

http://www.anandtech.com/show/7384/state-of-cheating-in-android-benchmarks

Boom. Headshot. First Volkswagen, now this?

 

[mavere]

The multi-year mental gymnastics needed to elevate 3DMark Physics from an interesting & notable outlier to the last bastion of Real World(tm) performance is marvelous and impressive. Bravo.