Originally posted by Ruined
aapo addressed that claim nicely here.
How so? First, the DX9 minimum for calculating 32-bit data types is FP24. The NV30-NV34 calculate everything at FP16 or less, because FP32 hasn't even been allowed in the drivers up to now. Only the NV35 currently supports FP32, and it's still slow overall compared to the R350 because of the register usage problem. It remains to be seen how much Nvidia can optimize for HL2, but the developers themselves have said ATI's cards are much faster. It depends on the degree that HL2 uses pixel shading, and the precision required. The NV35 will take a huge performance hit much sooner if there are lots of full-precision shaders, so Nvidia will likely have to force FP16 for most or all of the shaders to try stay to under the penalty threshold.
If proper FSAA relies on the pixel shaders as well then the situation is even worse.
I think HL2 might be able to run as fast on NV35 as R350, but only after Nvidia rewrites shaders for lower precision, which may or may not result in a noticeable loss in IQ.