icurafu said:
I was objecting to the blanket statement that SGX530 is a better GPU. For the purpose of 3dmark, it is not "better" by the terms of the benchmarking tool which concluded that the SGX 535 was massively beaten by the GMA900 even with the SGX using a higher performing CPU, which is enough to defeat the sgx530 > gma950 assumption. We'll have a better idea when optismised Quake3 benchmarks are compared as they both use OpenGL.
I'm about to make a statement that's based off of experience, which is NOT to be confused with a blanket statement.
Fillrate-wise, yes, the GMA900/950 is going to be better at pumping texels to the screen.
But it's NOT fillrate that you're going to be concerning yourself with- for it's not what matters in the long run and hasn't for some years now. It's an important part, but it's not the big thing everyone makes it out to be.
In this day and age, fillrate doesn't mean as much if you're not able to process the texels because you're chundering on transform, clipping, and lighting operations on the polys you're submitting to the fragment pipeline. We won't even get into physics and geometry computation operations for which unified shaders bring to the table. This is a dead-on fact of how things work in the OpenGL and Direct3D pipeline- I know, I worked on drivers in the past, including the forefront of Linux OpenGL support.
This is the reason why GMA's are dog slow over IGPs with Vertex options. The main reason the unified shader IGPs with an otherwise identical performance profile are pasting the older generation is because you can now peel off threads to do Vertex shader operations. There's a reason that the X3500's actually able to play games that would otherwise be using a low-end Radeon or GeForce at minimum right now- that's because the X3xxx series have unified shaders and can allocate 1-3 of them for vertex portion operations.
For Quake3, it's not going to show much because Quake 3 doesn't USE vertex shading or fixed-function TCL operations to avoid dealing with the lack thereof that was at least partly present during that era. Heh... RagePRO's were able to play Quake3:Arena acceptably when it came out. I know, with John Carmack working on the Utah-GLX team ON the RagePRO and other chipsets, we were using betas of Q3:A for verification purposes on the drivers back then. As an aside, because of what I just stated, Quake 3 is NOT going to be a decent metric to compare the GPUs with. PERIOD.
If, for example, you were to try to run, say, Tribes 2 for Linux (if it could even run...) or something like X2, you'd see a differing story. Because these two games rely heavily on Vertex path operations, either as fixed functionality, or as vertex shaders, the GMA9XX series GPUs will be bashed in the head with them and drop to slideshow framerates. Why? Because the CPU is now being asked to do these operations and a lot of them and those aren't designed for that- that's what a shader processor does best.
And, this doesn't even get into the comparisons of how a properly executed tile engine renderer will also avoid doing fill operations where it doesn't need to be done. If you're not overlapping, the higher fillrate will matter. Unfortunately, this is NEVER the case- there's a LOT of overlapped rendering. For each overlap, you push data out to the pixels for each overlapping rendering. If you don't have a means to avoid this, you will need a high fill rate to put the scene's frame to the screen to get high framerates. What if you don't do the overlaps anywhere near as much or not at all? You don't need as high a fill rate to get high framerates. A tile engine is designed to minimize and/or remove most overlapping renders from the fragment path befort them happen. So, again, comparing fill rates isn't going to give you a good picture for which is best.
[ edit ]
As for 3DMark scores... I didn't think we had official ES drivers yet for the 3DMarkES benchmarks, which would be the proper comparison- not to mention it's my understanding they don't even have XP drivers for the SGX series GPUs... Where does one find the overall scores and what each valuation for each sub-test is (which is where you REALLY make comparisons...)? If you can't lead me there, I find it difficult to even begin to accept or discuss your position as it stands.
[ /edit ]
To be sure, there's extra pieces in the GMA9xx series that're not there in the PowerVR- but I think you're going to find that many of them are going to be doable in software on one of the shader threads or in the DSP engine.
