If the PS 2.0 results that we've seen coincide with the ARB2 results, then the performance is all over the board, and is definitely low compared to what the hardware should be capable of. I seriously doubt it's a 16-bit vs. 32-bit float issue (at least, not all of it).
And Hellbinder, the NV30 OpenGL path doesn't "just" support INT12 and half-floats. Otherwise there'd be no GL_NV_float_buffer extension.
Regardless, it still remains to be seen just how much quality is lost by going for FP16. I really don't think it will be noticeable for the vast majority of operations.
"Physics is like sex. Sure, it may give some practical results, but that's not why we do it." - Richard P. Feynman