The "accepted truth" is that nVidia at least forces FP16 in 3DMark03.
They probably do even more than that, including precompiling the shaders to make them more efficient than what their drivers could do with a normal shader, and maybe they even use INT12 in some areas ( although that's really not sure at all, and it's in fact quite unlikely )
DX9 requires at least FP24 for most operations. But programmers can say a specific operation needs at least FP16 - that's why I say any serious programmer probably wouldn't do it the way Futuremark did it.
IMO, the best thing nVidia could do is manually setting FP16 for things that they think FP16 would be sufficent on. It might not give as much of a performance boost as forcing FP16 everywhere, but it would certainly be better than using FP32 everywhere!
And no one could complain about lower image quality too, if nVidia did it right.