An on-chip framebuffer (32x16 pixels) enables efficient full-scene anti-aliasing (FSAA). Kyro supports ordered grid four-sample FSAA. This means that the source image (aliased) is four times the size of the down-sampled image (anti-aliased). Down-sampling does not impose additional overhead. The bandwidth and memory storage involved in FSAA are actually less than that needed for rendering at the source resolution. Therefore, Kyro is slightly faster with four sample FSAA enabled than without (Figure 8). For GeForce, downsampling is not part of the rendering pipeline. It is an additional step that consumes external bandwidth and external memory storage. Its four-sample FSAA performance trails that of Kyro (Figure 8).
Figure 8: Fillrate graph for 'Serious Sam Test 2'. 16-bit, four-sample anti-aliasing.
Common settings: 'Quality' rendering, lens flare disabled.
Kyro setting: compiled vertex array disabled.
Figure 9 shows the 16- and 32-bit fillrate graphs for Kyro and GeForce2 MX. 'Serious Sam Test 2' utilizes hardware transformation if it is available. GeForce2 MX leverages on this with a commanding lead in 16-bit renders. It falters at 1280x1024 and is overtaken by Kyro at 1600x1200. In 32-bit renders, bandwidth takes precedence over the benefits of hardware transformation at 1024x768 and higher resolutons.
Figure 9: Fillrate graph for 'Serious Sam Test 2'.
Common settings: 'Quality' rendering, lens flare disabled, bilinear filter.
Kyro setting: compiled vertex array disabled.
