iPhone benchmarked with GLBenchmark

[kishonti]

Hi All,

OpenGL ES performance numbers for iPhone are available at http://www.glbenchmark.com/phonedetails.jsp?D=Apple%20iPhone

iPhone is very close to the top, but Nokia N95/N93 are still somewhat faster in high level OpenGL ES 1.1 tests.

Although iPhone's CPU is faster than Nokias', it lacks matrix-palette support (GPU accelerated skinning) which drags its OpenGL ES 1.1 performance down.

Current 1.1.x iPhone also has 60 fps screen refresh limit.

[Shifty Geezer]

Gosh, the float performance of the Apple CPU is way beyond anything else in that database. How is this being used in the system. Performance wise other phones are managing similar on a sixth the power - I'm curious over Apple's choice of FP strong CPU.

[darkblu]

about time.

I'm curious over Apple's choice of FP strong CPU.

my speculation would be something along the line of an absent VGP.


on an unrelated note, can somebody enlighten me as to why this test measures some typically fillrate/bandwidth-bound tasks (like tex res and filtering modes) in triangles?

[Fafalada]

Although iPhone's CPU is faster than Nokias', it lacks matrix-palette support (GPU accelerated skinning) which drags its OpenGL ES 1.1 performance down.
Off topic - but what is the size of supported MatrixPalettes on this class of GPUs?
If this is along the lines of 2000era accelerators(or certain modern handhelds for that matter) it might as well be ignored as a relevant benchmark.

on an unrelated note, can somebody enlighten me as to why this test measures some typically fillrate/bandwidth-bound tasks (like tex res and filtering modes) in triangles?
Yea I wondered about that too.

[Lazy8s]

The iPhone's consistency in performance across the different tests implies a solid implementation of its drivers, API, and related software environment.

Interestingly, the API reports that a VGP Lite is on board.

[kishonti]

on an unrelated note, can somebody enlighten me as to why this test measures some typically fillrate/bandwidth-bound tasks (like tex res and filtering modes) in triangles?

the same geometry is rotated with different lighting, texturing, filtering, etc settings in all lower level tests. using triangles/s as a unit makes easier to compare

[Lazy8s]

Is a VGP Lite really included? Why no matrix-palette support?

The clock speed of the graphics processor truly must be around 100 MHz for the iPhone to rank so highly.

[Simon F]

Off topic - but what is the size of supported MatrixPalettes on this class of GPUs?
If this is along the lines of 2000era accelerators(or certain modern handhelds for that matter) it might as well be ignored as a relevant benchmark.
Basically, are you asking how big is the constant store for the vertex program?

[Fafalada]

Basically, are you asking how big is the constant store for the vertex program?
The way kishonti worded his comment it sounded like hw he was referring to is fixed function :???:
But yea, info about constant space would be interesting to know either way. :razz:

[kishonti]

The way kishonti worded his comment it sounded like hw he was referring to is fixed function :???:
But yea, info about constant space would be interesting to know either way. :razz:

Yes, iPhone is OpenGL ES 1.1 = fixed function.
The matrix palette size on OpenGL ES is usually 9, but there are other supported sizes like 16 (ARM/Falanx) or 31 (AMD, ES 1.1 Extension pack).

[Simon F]

The way kishonti worded his comment it sounded like hw he was referring to is fixed function :???:
Ahh. Whether an API which is fixed-function can expose a programmable vertex unit is a completely different matter.

[darkblu]

Ahh. Whether an API which is fixed-function can expose a programmable vertex unit is a completely different matter.

ooh.

gosh, i'm slow today. it took me the above _and_ lazy's post to trigger my reaction to go and check more closely the gl env report. good news, nevertheless.


the same geometry is rotated with different lighting, texturing, filtering, etc settings in all lower level tests. using triangles/s as a unit makes easier to compare

so basically you're doing a '3d mark game test' there. i see.

i, for one, would appreciate a more synthetic test designed to tell me the fillrate implications of those parameters (i.e. tex res and filtering modes) rather than measuring triangles for the sake of relating across the board. moreover, you can use a measure like pixels/s to the same effect.

[roninja]

Nice to see PowerVR dominating all of the top 15 or so devices in the GLBenchmark database.

[frogblast]

Is a VGP Lite really included? Why no matrix-palette support?


A more accurate description of the extensions the iPhone supports is available in Apple's developer documentation, which requires a free account+NDA to access.

[kishonti]

A more accurate description of the extensions the iPhone supports is available in Apple's developer documentation, which requires a free account+NDA to access.

That should be about iPhone firmware 2.0. GLBenchmark results are about 1.1.x.

[Fafalada]

Ahh. Whether an API which is fixed-function can expose a programmable vertex unit is a completely different matter.
Not budging huh? :razz:
Anyway thanks for reminding me why I dislike working with these generalized APIs - not to mention I'm about to start work where I'll no longer be able to avoid them :sad:

i, for one, would appreciate a more synthetic test designed to tell me the fillrate implications of those parameters (i.e. tex res and filtering modes) rather than measuring triangles for the sake of relating across the board.
Especially since measuring triangles just for the sake of it can very well hide the actual performance implications of varying those parameters.

[Simon F]

Not budging huh? :razz:.
Sorry - I don't actually know. I've only "driven" the VGP directly, not via OpenGL ES. <shrug>

[Lazy8s]

Between MS VS 1.1 and 2.0 lies the level of functionality of the VGP, I believe.

[Scott_Arm]

Quake3 on iPod Touch with multiplayer?

http://hardocp.com/news.html?news=MzE2MzcsLCxoZW50aHVzaWFzdCwsLDE=

[Simon F]

Quake3 on iPod Touch with multiplayer?

http://hardocp.com/news.html?news=MzE2MzcsLCxoZW50aHVzaWFzdCwsLDE=

It was interesting reading the forum post associated with that link - with people asking why it doesn't run doom when cell phones do.

I suppose if there is an OpenGL ES port of doom available then it would; but if it's just the old software renderer then why would you want to?

[Scott_Arm]

I finally saw the video today. The control looks a little weird. Guess that part hasn't been sorted out yet. Still, this bodes very well for iPhone, iPod touch gaming.

[rezuth]

It was interesting reading the forum post associated with that link - with people asking why it doesn't run doom when cell phones do.

I suppose if there is an OpenGL ES port of doom available then it would; but if it's just the old software renderer then why would you want to?
But... It does...

http://www.youtube.com/watch?v=1oKcLRTzcbU

[kishonti]

A more accurate description of the extensions the iPhone supports is available in Apple's developer documentation, which requires a free account+NDA to access.

The limited numer of iPhone OpenGL ES extensions was a bug in GLBenchmark website, sorry we didn't recognized that (only the first line was printed).

The full list is now available: http://www.glbenchmark.com/phonedetails.jsp?benchmark=pro&D=Apple%20iPhone&testgroup=gl

[Sc4freak]

Gosh, the float performance of the Apple CPU is way beyond anything else in that database. How is this being used in the system. Performance wise other phones are managing similar on a sixth the power - I'm curious over Apple's choice of FP strong CPU.
That's probably because the CPU in the iPhone actually has a hardware accelerated FPU.

As far as I can tell, it's one of the only ones that do. My Dell Axim x51v, for example, has to simulate floating-point math in software (which, as you can imagine, is incredibly slow).

[muratmat]

That's probably because the CPU in the iPhone actually has a hardware accelerated FPU.

As far as I can tell, it's one of the only ones that do. My Dell Axim x51v, for example, has to simulate floating-point math in software (which, as you can imagine, is incredibly slow).

OMAP2420 (ARM1136JF-S) has a VFP, and it has been mounted on N93, N95 and so on. In my opinion, the low FPU results on those phones are due to a compiler issue / missing option. VFP hard instructions are correctly emitted by the Apple toolkit.

[Fox5]

Gosh, the float performance of the Apple CPU is way beyond anything else in that database. How is this being used in the system. Performance wise other phones are managing similar on a sixth the power - I'm curious over Apple's choice of FP strong CPU.

They're using modified OSX libraries right? Maybe OSX was programmed with a heavy FP dependence.

[Lazy8s]

Some of the more recent changes to the GLBenchmark rankings are interesting.

All of the results of high triangle rates of around 2M/sec scored by various HTC and LG devices containing ATi/AMD/Qualcomm graphics have disappeared, though the scores of upwards of 6.5M/sec achieved by the iPhone 3G S would've left them far behind anyway.

PowerVR's dominance of the top twenty-six spots has finally come to an end with the appearance of a next-generation, Z430 powered device, the Acer F1. However, the F1 isn't actually a product on the market yet.