NVIDIA's GeForce 8800 (G80): GPUs Re-architected for DirectX 10

Our DX9FSAAViewer won't show us the exact sample patterns for CSAA, but we can take a look at where ATI and NVIDIA are getting their color sample points:

ATI
G70
G80
G80*

*Gamma AA disabled

As we can see, NVIDIA's 8x color sample AA modes use a much better pseudo random sample pattern rather than a combination of two rotated grid 4xAA patterns as in G70's 8xSAA.

While it is interesting to talk about the internal differences between MSAA and CSAA, the real test is pitting NVIDIA's new highest quality mode against ATI's highest quality.

G70 4X G80 16XQ ATI 6X

Hold mouse over links to see Image Quality

G70 4X G80 16XQ ATI 6X

Hold mouse over links to see Image Quality

Stacking up the best shows the power of NVIDIA's CSAA with 16 sample points and 8 color/z values looking much smoother than ATI's 6xAA. Compared to G70, both ATI and G80 look much better. Now let's take a look at the performance impact of CSAA. This graph may require a little explanation to understand, but it is quite interesting and worth looking at.

As we move from lower to higher quality AA modes, performance generally goes down. The exception is with G80's 16x mode. Its performance is only slightly lower than 8x. This is due to the fact that both modes use 4 color samples alongside more coverage samples. We can see the performance impact of having more coverage samples than color samples by looking at the performance drop from 4x to 8x on G80. There is another slight drop in performance when increasing the number of coverage samples from 8x to 16x, but it is almost nil. With the higher number of multisamples in 8xQ, algorithms that require z/stencil data per sub-pixel may look better, but 16x definitely does great job with the common edge case with much less performance impact. Enabling 16xQ shows us the performance impact of enabling more coverage samples with 8x multisamples.

It is conceivable that a CSAA mode using 32 sample points and 8 color points could be enabled to further improve coverage data at nearly the same performance impact of 16xQ (similar to the performance difference we see with 8x and 16x). Whatever the reason this wasn't done in G80, the potential is there for future revisions of the hardware to offer a 32x mode with the performance impact of 8x. Whether the quality improvement is there or not is another issue entirely.