Originally posted by kyleb
actually it is ~8% more unfair in a strictly bit-to-bit analytical sence if you want to get into it Chalnoth. also the visual differences between the 3 levels of precision is something that deserves investigation when such claims are made. unfortunately none of us "regular" people has been given the chance to see then nv30 do fp32 and fp16 and we don't know how much int12 is going on either so then whole discussion is purely academic at this point.
Well, I can pretty much guarantee you that JC is not going to be using an integer format (framebuffer aside). The benefits of moving to floating-point are just too great. Not only that, but he's been going ga-ga over the possibility of using floating-point...
As an aside, I do think it's kind of funny that so many people seem to be ready to consider FP16 absolutely inferior to FP24. In particular, remember all of the 16-bit vs. 32-bit arguments of the past, back with the Voodoo3 vs. TNT2 Ultra?
Now, as then, there are always going to be differences, but I still feel that 16-bit FP is enough for most any color calculations, but the differences may become apparent for non-color data pretty quickly.
In particular, remember that it's the dark end of the spectrum that needs more accuracy, and 16-bit FP offers higher and higher accuracy (i.e. fewer visible errors) as the image gets darker.
One other thing to keep in mind is that 3DMark03 uses PS 1.4 for most of the shaders in their tests. PS 1.4 doesn't have the accuracy of FP16. So is there a reason to use FP32 in 3DMark03?