iPhone/Zune/iPod & More Prediction Thread

Yes it is true - proof

What am I missing in that link that doesn't tell me that it's a 535 at all?

Go ahead and run a comparison between a iPhone3GS/iPhone3G/Nokia N95 under low level 3D. Unless there's something seriously wrong with the driver the MBX with barely 1 TMU cannot under any circumstance yield higher fillrates than a SGX535 which has 2 TMUs (for which both the SGX in the iPhone3GS might even run at a higher frequency than the N95 MBX).

If you look at the fillrates between iPhone3G (MBX Lite) vs. N95 (MBX) the differences are easy to explain since the MBX has a slightly higher frequency and can execute one pixel/clock, whereby for MBX Lite it should be rather 1 pixel/2clocks.

It may very well not be SGX520, but then if it's a SGX53x variant there's a problem with the drivers. Those results point clearly in the 520 direction so far.
 
What am I missing in that link that doesn't tell me that it's a 535 at all?
There's a GL Environment tab...

Please note: Almost all of these results are vsync limited. They are NOT representative of actual peak performance.
 
There's a GL Environment tab...

Please note: Almost all of these results are vsync limited. They are NOT representative of actual peak performance.
Out of curiosity, I presume the benchmark is a Java application?
 
There's a GL Environment tab...

Please note: Almost all of these results are vsync limited. They are NOT representative of actual peak performance.

Argghh.....*runs against the next wall* I still don't get it why a 535 would give (even under a vsync limited scenario) lower fillrates than a MBX; in theory if both would have the same frequency the 535 would still have twice the fillrate (w or w/o overdraw).

(Oh and yes I'm awefully tired right now...)
 
GL_VENDOR Imagination Technologies

GL_VERSION OpenGL ES-CM 1.1 IMGSGX535-31.4

GL_RENDERER PowerVR SGX 535

whats the benefit of sticking an 535 inside the iPhone versus say the 530 in OMAP3?
 
re GLbenchmark - how low is the Samsung i8910 in the rankings. Nice to see PowerVR continuing to dominate though ;-)
 
re GLbenchmark - how low is the Samsung i8910 in the rankings. Nice to see PowerVR continuing to dominate though ;-)


Something badly BADLY wrong with that samsung i8910, its overall performance is slower that the ARM11/MBX 8510.
Some of its triangle & fill rates are 1/3 of the 8510, and 1/4 of the iphone




Look how much stronger the samsung CPU Float/Int figures are compared to the 3GS

Appl Samsung
Int 2378 2680
Float 8920 17182

The samsung has TWICE the Float performance.
 
I still don't get it why a 535 would give (even under a vsync limited scenario) lower fillrates than a MBX; in theory if both would have the same frequency the 535 would still have twice the fillrate (w or w/o overdraw).
Let me explain a few limits (for a vsync limited scenario):

GLBenchmark HD:
60 (fps) * 30 (seconds runtime) = 1800 frames

Swapbuffers:
60 (fps) * 10 (seconds runtime) = 600 frames

Geometry:
60 (fps) * 10216 (triangles per frame) = 612960 Tri/s

Fillrate:
60 (fps) * 320 * 480 (resolution) * 4 (layers) = 36864 kTex/s
40 (fps) * 320 * 480 (resolution) * 8 (layers) = 49152 kTex/s
40 fps is a 2:1 cadence thus one step below 60 fps. All this shows is that the fillrate in this particular test is below 73728 kTex/s. And it's entirely bandwidth limited.
 
Argghh.....*runs against the next wall* I still don't get it why a 535 would give (even under a vsync limited scenario) lower fillrates than a MBX; in theory if both would have the same frequency the 535 would still have twice the fillrate (w or w/o overdraw).

(Oh and yes I'm awefully tired right now...)

Only if the test draws twice as much per frame, which it may not be doing. They could easily get around the vsync, simply by flushing rather than swapping, if they cared to do so. Or, they could render to an offscreen framebuffer.
 
Last edited by a moderator:
Let me explain a few limits (for a vsync limited scenario):

GLBenchmark HD:
60 (fps) * 30 (seconds runtime) = 1800 frames

Swapbuffers:
60 (fps) * 10 (seconds runtime) = 600 frames

Geometry:
60 (fps) * 10216 (triangles per frame) = 612960 Tri/s

Fillrate:
60 (fps) * 320 * 480 (resolution) * 4 (layers) = 36864 kTex/s
40 (fps) * 320 * 480 (resolution) * 8 (layers) = 49152 kTex/s
40 fps is a 2:1 cadence thus one step below 60 fps. All this shows is that the fillrate in this particular test is below 73728 kTex/s. And it's entirely bandwidth limited.

Thanks Xmas now of course it's a LOT clearer. Why aren't you guys creating a techdemo along the Fablemark philosophy (witty and cute(tm)) that would also measure some performance aspects?
 
GL_VENDOR Imagination Technologies

GL_VERSION OpenGL ES-CM 1.1 IMGSGX535-31.4

GL_RENDERER PowerVR SGX 535

whats the benefit of sticking an 535 inside the iPhone versus say the 530 in OMAP3?

Twice the fillrate.
 
Yes you were on the right track. In retrospect it means that the 3GS graphics part by itself can turn out even above 4x times more efficient compared to the 3G.
 
Yes you were on the right track. In retrospect it means that the 3GS graphics part by itself can turn out even above 4x times more efficient compared to the 3G.

I am going to do a couple of fillrate tests this weekend just with 2d quads ... so far my game (3d) is maxing out at 60 fps (as compared to 35-40 on the old device) which doesn't tell me much.

Anyway, the real fun will start with ES 2.0 renderer specific code...
 
I am going to do a couple of fillrate tests this weekend just with 2d quads ... so far my game (3d) is maxing out at 60 fps (as compared to 35-40 on the old device) which doesn't tell me much.

Anyway, the real fun will start with ES 2.0 renderer specific code...

What kind of filtering are using for that game? Would it too bold to ask if something like anisotropic is possible at reasonable performance on the GS?
 
Mobile SoC designers had been more likely to select an SGX variant with increased pipelines before TMUs, seeming to indicate their belief of a better balance of performance versus power and/or silicon consumption with SGX5x0 cores.

Someone suggested that Apple chose the 535 for increased video functionality... I wonder if the 535 actually is more capable in that way and whether that would matter to a VXD3x0 licensee like Apple. Maybe for video encoding, but wouldn't Apple be using some other dedicated core/DSP for that?

With their selection of the 535, I wonder if they would've picked the 540 had it been available sooner.

If Apple has specified their VXD as well into this new SoC, they'd have a single mobile solution with the best CPU, GPU, and video core.
 
Back
Top