Originally posted by: Ackmed
Originally posted by: BenSkywalker
The FX's had to have their own path written.
Not quite. nVidia gave Carmack all the explicit features and extensions he wanted for DooM3 and he used them. Carmack has been talking about the DooM3 engine since back in the GeForce1 days and you can go check his quotes from then yourself. He wanted 64bit color(AKA FP16) he wanted really fast register combiner ops and he wanted loads of stencil op power- the FX architecture with nV's proprietary extensions.
When playing in the standard ARB path, the Radeons are much faster.
When playing the ARB2 path the big difference for the nV parts in terms of performance is that they are running in FP32 vs FP16(talking about the NV35 parts here). ATi's boards are running a 'lower quality' setting by using FP24. Carmack says that there is no discernable quality difference between FP16, FP24 and FP32 for DooM3. You have to say Carmack is wrong or say that ATi is faster running medium quality then nVidia is running high quality to validate the above statement.
The fact still is, doing a default path, ATi is faster. In HL2 and Doom3, the FX's had to have extra code written for them to get playable frames. To me, thats sad.
"Lower quality" as you say is what DX9 calls for. FX's use FP16 all the time instead of FP32, just to get better frames. Yeah, lets tout FP32, then not use it.
If you cards have to keep getting major help from the devs, then thats a sign that it wasnt a very well designed product to me. Look at Farcry, the NV3x doing just PS 1.1, while the R3x does 1.1 and 2.0. And when you force 2.0 via a 3rd party software, the FX's CRAWL.