ATI's SMOOTHVISION - The key to its "success"

Supersampling rises from the dead

Before we go any further it is very important that you understand a bit about the various forms of anti-aliasing. Without touching the issue of sampling patterns there are two major types of AA currently in use: supersampling and multisampling.

Supersampling was introduced with the NVIDIA GeForce2 and ATI's Radeon over a year ago and it achieves the AA effect by rendering the scene at a higher resolution (by some fixed factor in the x and y directions), taking color samples from all of the additional pixels generated at this higher resolution, blending them together and finally forming an anti-aliased image at the native screen resolution. For example, one method used to implement a 4X supersampled AA algorithm was to render at 2X the horizontal and 2X the vertical resolution meaning that at a native screen resolution of 1024 x 768 the supersampled buffer would have the same number of pixels as a 2048 x 1536 image.

This method of anti-aliasing produced excellent results from a visual standpoint however the penalty was an incredible performance hit. By forcing the GPU to render 4X as many pixels as the native scene required, these cards incurred a significant fill rate and memory bandwidth hit that made AA at resolutions greater than 800 x 600 and usually 640 x 480 useless. It was clear that today's and even tomorrow's video cards wouldn't have enough power to make supersampled AA a reality and thus the need for something more efficient was born.

With the introduction of the GeForce3, NVIDIA had produced a core that was specifically designed around a more efficient form of AA - multisampling. The true benefit of multisampling is the achievement of high quality edge anti-aliasing (removal of "jaggies") at the sacrifice of increased texture blurring. The way multisampling works is very similar to supersampling in that a certain number of samples are taken in the x and y directions to be used as subpixels for the final image. The difference is that the color of these subpixels is not determined by actually sending these additional pixels through the pixel pipelines, instead the color of the subpixels is determined by original pixels. This reuse of texture data results in two things that set it apart from supersampling:

1) a significant reduction in the amount of texture memory bandwidth required to enable AA
2) introduction of blurred textures because of the reuse of texture color data

Although the GeForce3 core supports both supersampling and multisampling, the currently available drivers only support multisampling which is by far the most efficient method of AA on the GeForce3. We incorrectly assumed that the 4X AA mode on the GeForce3 was using the supersampling capabilities of the core and thus labeled it as so in our last article.

With the realization of this and another clue in hand we began to question our conclusions as to what was causing the superior texture quality of SMOOTHVISION.