OK let's get back on topic after that digression/episode...
Originally Posted by Jarred
it's also here
toward the middle.
"David Kirk: Anything that can be programmed in PS 2.0 just runs in PS 3.0 with no changes. It just runs faster, and looks better. Anything that can be programmed in PS 3.0 can be converted into multipass versions for PS 2.0. It just runs more slowly and looks less good."
AFAIK programmatically, a shader written in SM 2.0 should run in SM 3.0 with no changes. I agree with David Kirk in that regard. (If it didn't, something is very wrong, but I digress.) I disagree that it should run faster or it look better in all cases. I think it may look better if a shader was built that had artifacts or precision issues in FP24/DX 9.0b that could be corrected with FP32/DX 9.0c. Otherwise, they would probably look the same going from FP24 to FP32. I don't think an IQ increase would automatically or auto"magically" come from using a "pure" SM 2.0 shader with SM 3.0 for the majority of 2.0 shaders out there.
The speed increase may be dependent on nV's compiler more than DX 9.0c. If the shader is 2.0, it should run as a SM 2.0 shader under SM 3.0 from my understanding. I don't see how the compiler is making it faster just because the SM 3.0 cap is exposed. If the shader is the same source code when built under 2.0 and 3.0, I can only see the speed increase through a (valid) compiler optimization that may or may not come through the introduction of DX 9.0c. Maybe a programmer can elaborate on this...
All in all, some of that quote is market speak IMHO, but it has some grounding in reality.