As a supplement to 3D WinBench 2000, we have also run MadOnion's 3D Mark CPU test, which focuses on the floating point-intensive geometry portion of the CPU. Rather than using hardware T&L, each CPU was set for its respective pipeline optimization (SSE and 3Dnow!).
Again we see the Duron 800 scale well in comparison to the 700 and 750MHz processors, and again the Celeron is left behind (only this time, by a smaller margin). This could be due to more SSE optimizations in 3D Mark or the 256-bit L2 bus of the Celeron versus the 64-bit bus of the Duron, though we would tend to suspect SSE optimizations.