Figure 5 shows a graph that compares the fillrates of the Voodoo5 5500 and the GeForce2 GTS (with and without the benefit of hardware transformation) in Test Drive 6, a Direct3D racing game.
A combination of factors including narrow AGP bandwidth, the absence of AGP texturing and shared frame buffer access, account for the Voodoo5 5500's fillrate decline when rendering at high resolutions, especially at 1600x1200. In contrast, the Voodoo5 5500's low-resolution fillrates are comparable to the GeForce2 GTS with hardware transformation enabled and ahead of the GeForce2 GTS without the benefit of hardware transformation. This indicates the efficiency of the Voodoo5 5500's Direct3D drivers.
It is notable that the high-resolution fillrates of the GeForce2 GTS, without the benefit of hardware transformation, overtake the fillrates with hardware transformation enabled (Fig. 5). This is because Test Drive 6, with hardware transformation enabled, strains local memory bandwidth with geometry processing. On the other hand, software transformation utilizes system resources for geometry processing and does not encroach upon local memory bandwidth. This does not impact performance significantly at the lower resolutions (640x480, 800x600 and 1024x768) where bandwidth is not an issue. At higher resolutions (1280x1024, 1600x1200) where bandwidth is a bottleneck, the additional geometry bandwidth causes performance to drop below software transformation. Applications that have been designed for hardware transformation make use of AGP bandwidth rather than local memory bandwidth for geometry processing.