Re: why do fx cards perform so badly in 3dmark05
1st: Looking at pure shader fillrate numbers, ATi GPUs are roughly 2x faster than equivalent NVidia GPUs. High-end NV30,35 are 4 pipelines architectures, while mainstream NV31,34,36 are only 2 pipelines architectures (forget the ability to pull out 4 pixels in certain situations): that is half of their ATi counterparts. No surprise if in shader intensive apps their performance sink.
2nd: There's a noticeable speed hit when using more than 2 (4 in FP16) temp registers. I wonder why NVidia overlooked this point in the NV3x architecture, considering that in shader model 1.4/2.0 texture registers cannot be used to hold temp information like in PS 1.1
Although the whole picture for NV3x doesn't look good, it seems that shader model 2.a gives some nice speed boosts over plain 2.0 in 3DMark05 and, hopefully, in games too. NV3x users should be grateful to ATi for making their new R4xx line of cards PS 2.b only. This gives the developers some incentive to include in their games support for PS 2.a/b too.
Pentium Dual E2180 (2GHz, 200MHz FSB)
Abit L95C (The one and only mATX with dual PCIe slots!) + 2x1 Geil Black Dragon PC6400
XFX 9600GSO 384MB
Little freak wonder