3DMark 2001SE and 3DMark03
FutureMark's old-but-good benchmarks round things out.3DMark 2001SE
All the mid-range cards converge to a point at the high-end setting, 3DMark 2001SE's DX8-leaning ranking all the boards pretty much as UT2003 did. The Radeon's experience the same slight CPU-limitation as the 6800 GT and Ultra initially, as they wait for the CPU to boost performance, rather than being held back by their own performance.
The 6600 GT driver is responsible for the performance difference, the interface having no effect on things at all.
3DMark03
At the upper setting, there's nothing to split 6600 GT and Radeon 9800 XT in terms of performance, as measured by 3DMark03. At the median setting, the AGP 6600 GT is ~15% faster than the XT with even more difference at the easier, lower setting.
However, as RTHDRIBL spells out, no matter what the basic tenets of performance each GPU is able to provide, the software architecture of the application using the accelerator is paramount. Doom3 is a prime example, running very well on recent NVIDIA hardware, exploiting nuances in the GPU's architecture. ATI's GPUs have games titles that show those GPUs off in the same way.
3DMark tests outline things nicely with investigation into other tests highlighting any marked differences in certain titles and applications.
