With reference to figure 3, it can be seen that fillrates (y-axis) initially rise with resolution (x-axis) before reaching their respective peaks. At resolutions higher than these, the increased space needed to store framebuffer and Z-buffer in video board memory displace textures that would otherwise be stored in video board memory to AGP memory. Therefore, texture transfer takes place predominantly across the slower AGP bus. The result is a sharp dropoff in fillrate beyond the peak. The graph shows that Radeon attains a peak fillrate at a higher resolution (1280x1024 or 1.3 megapixels) compared to GeForce2 (1024x768 or 0.79 megapixels). Thus, Radeon's Z-buffer compression postpones the onset of AGP texturing in the MBTR demo.

Note: performance figures are denoted in millions of pixels rendered onscreen per second (framerate x resolution). "Fillrate" is used loosely as it is not quite the same as the actual accelerator fillrate, which is a few times higher if overdraw and multi-pass rendering are taken into consideration. Graphs are plotted showing fillrate (Y-axis) versus onscreen pixels (X-axis). The number of onscreen pixels denotes screen resolution.

As a side note, the fillrate graph of GeForce2 using an earlier set of drivers (ver 5.22) is also shown. Compared to the more current 6.18 drivers, fillrates are markedly lower at all resolutions. It is even more significant that fillrates obtained with 6.18 drivers scale linearly at lower resolutions in contrast to the earlier 5.22 drivers. A linear relationship between fillrate and resolution is possible in the absence of bandwidth bottlenecks. The 6.18 drivers remove the bottleneck that was present with the 5.22 drivers by relocating geometry data from video board memory to AGP memory. This prevents the geometry memory stream from competing for video board memory bandwidth.