Originally posted by: BFG10K
The NV3x might be sligthly faster than the R3xx in Doom III but it'll be around 25%-50% slower in all DirectX 9.0 games. That's a tough trade and it's impossible to say that nVidia will be "pwning" anyone this generation.
Yes, it would probably be faster, however, that's when comparing it running a "special NV3x backend" utilizing FX12 w/a little FP16 for IQ because it runs ARB2 very slow (as per JC) vs ATI defaulting to the ARB2 path (no optimizations just a R2x00/R3x00 FP24 path), hence you'll have a slightly faster NV (special) path lower precision, a native higher precision ARB2 path with ATI at a "slightly" higher IQ (we'll see it's more than this).
So OGL is the same for NV3x as DX9, dumb down the API to be playable.
ATI DX9 = playable
NVidia DX9 = slow, API must be dumbed down to be playable.
ATI OGL (ARB2) = playable
NVidia (ARB2) =slow, API must be dumbed down to be playable.
As per Carmack, he HAS to make a special backend because NV3x utilizing the default ARB2 path is "very slow", why he mentioned this who knows, maybe he too is frustrated as valve was but this is his partner so a quick blurb maybe "punishment" enough, or the man just plain likes the challenge - either way you dice it up, NV is slow with the ARB2 path and precision + advanced features must be dumbed down.
