Clarifications

Don't we already have multi-sampling?

Notice that we have made distinctions between this new multi-sampling anti-aliasing and the multi-sampling that is found on the Voodoo5. This issue has been a hotly discussed, online topic.

The Voodoo5 renders multiple samples (hence it is sometimes also referred to as multi-sampling) to achieve a form of super-sampling. Super-sampling, in its simplest of definitions, is combining multiple samples to achieve each pixel. The Voodoo5 does this. On a Voodoo5, you will render your scene multiple times and buffer these into four (or however many are supported in hardware) separate buffers.

As the scene is sent to the screen there is a combine unit that averages each of the four buffers at a sub-pixel level to produce the final, displayed image. As we can see, this is considerably different than the form of multi-sampling that we are discussing here. The terms themselves can certainly be confusing, but everything makes a lot more sense once we understand exactly what we kind of multi-sampling we're talking about.

The future

What of the future of anti-aliasing, particularly now that 3dfx can no longer promote the feature? Well, we see certainly see hardware moving towards multi-sampling. And over the next year we expect many chip makers to adopt this implementation.

What of the long term, though? Well, as with most anything, there are a variety of directions that can be taken. What we would like to see, however, is stochastic multi-sampling. As discussed, multi-sampling can use jittering to help increase its quality; yet even with jittering you have a pattern, and as we know the human eye is extremely good at noticing or discerning patterns.

Random noise, however, is very difficult for the eye to detect. By using stochastic sampling,we no longer have a set pattern, instead, the pattern is completely random and random patterns produce noise. With this noise, we are capable of delivering a much higher quality anti-aliasing, while requiring fewer samples. It is often felt that 4x4 super/multi-sampling is sufficient with stochastic sampling (this is commonly used within the CGI industry). While there are more advanced forms of stochastic sampling, such as Poisson Disc distribution, these can be extremely expensive to implement.

While one might argue that stochastic super-sampling is truly the best route for total quality, we would argue that it is not the best path for at least the next several years. A balance must be met between overhead and quality, and we feel that stochastic multi-sampling is better at achieving this balance at this point in time. Presently, super-sampling is simply too demanding on hardware to be done in real-time at high sampling rates, while multi-sampling isn't. By delivering multi-sampling we achieve a very nice balance.