RTHDRIBL and UT2003
RTHDRIBL
With 66.93 on the PCI Express board, the score (46.882) shows that the interface is having no effect. The extra 100MHz of memory clock helps the AGP board a fair bit (7% of new performance for 11% increase in memory clock), showing the basic clocks of the AGP board, at least in RTHDRIBL, are well balanced.
Versus the Radeon's, this hand-written shader test is an ass-kicker for the NVIDIA boards, losing out to ATI's year-old GPU despite more shader power and higher clock speeds. I like using this test to show how software optimisations are responsible for a massive chunk of potential performance. You see it in action heavily here. It's not that NV43 has worse fragment shader and render to texture performance than R360, it's just the shaders are hard for NVIDIA's shader compiler to deal with, impacting performance.
UT2003
The new driver affords more UT2003 speed and the memory clock increase to 1000MHz gets you a bit too, especially at the median setting. Compared to the Radeon's, the 6600 GT's raw clock speed accounts for the difference in mostly multitexture-heavy, shader-light performance in UT2003, offset by the R380's excellent DX8-level performance. R360 is great for UT2003's engine, NV43's raw core speed and memory bandwidth advantage offset a clock-for-clock performance deficit.
