Elsa Gladiac 921DVI GeForce3 Ti500

Author : Wayne Date : 30th December 2001

3DVelocity would like to thank Elsa for their help and courtesy in providing this graphics card for review.

Technology Overview :

High Resolution Anti-Aliasing :

Because the image you see on your screen is constructed from individual pixels, there is simply now way to render a straight edge if it deviates from true horizontal or true vertical. Imagine trying to build a triangle from building blocks, no matter how many blocks you use the sloping lines will always take on a "staircase" appearance (aliasing).

One solution is to use smaller blocks, which while not solving the problem does make it less visible. The computer equivalent to using smaller blocks is increasing the resolution, though as we discussed earlier, this means more work for the graphics card and ultimately reduced framerates. The other solution is to change the colour of the pixels at the edge of a sloping face so they blend better with the object behind. This reduces the contrast of the edge and again makes the effect less visible.

So far only two accepted technologies exist for performing anti-aliasing, these being Supersampling and Multisampling.

Supersampling :

The idea behind Supersampling is that the scene is rendered at a much higher resolution than will be rendered in the final scene. Doing this creates extra pixels with extra colour data that can be used to more accurately define the eventual colour of the larger final pixels. This rather unrefined method is certainly effective, but because you are rendering more pixels than will be visible in the final scene. The performance hit can be significant. Using 2x supersampling renders twice as many pixels as would be rendered were anti-aliasing switched off, which immediately incurs a performance hit of one half. 4x supersampling render four times the pixels normally required leading to a performance of a quarter of what you would see with no anti-aliasing. In reality, the performance hit is higher than this because of the need to reconstruct the final scene based on the supersampled scene.

Multisampling :

Multisampling is an altogether more refined approach to anti-aliasing, though it needs more advanced hardware to pull it off. The idea behind multisampling is that rather than recreate an entire scene at a higher resolution, the GPU creates "virtual pixels" based on the data that would be used in creating that higher resolution scene at the very start of the process. Because only the raw data for a higher resolution render is used rather than the scene itself being created from start to finish, the final pixel can be created based on that data without needing to be filtered back down to its native resolution. This of course means less of a performance hit overall.

Another advance in this technology is something NVIDIA have christened Quincunx, named after the sampling pattern used. Below you can see a traditional 4x anti-alising sample pattern compared to the Quincunx sample pattern. In this example it appears that more samples are needed by the Quincunx method (5 samples) than are needed by the 4x sample pattern (4 samples), but when we place the pixels into a grid we see the truth of the matter.

Below you can the traditional 2x sample pattern on the left, while in the middle is a grid of pixels all using the 2x sample pattern. The example on the right shows how Quincunx is now able to sample from all five of its required areas using only the sample points created using the 2x pattern.

The upshot of this is that Quincunx is able to rival the quality of 4x supersampling while requiring the overhead normally associated with a 2x anti-aliasing method. This obviously makes for better framerates as the amount of data required in the frame buffer is decreases significantly. The chart below shows frame buffer storage requirements for each anti-aliasing method at a range of resolutions.

But the question remains, does it work? Take a look at the examples below and decide for yourself.

 

The biggest problem with Quincunx is that it tends to blur textures in the scene as well as blending edge pixels. If you look at the scene below, I've picked out an are then used enlarged section of that area taken using no anti-aliasing, 2x, Quincunx and 2x. Watch the clarity of the skulls and you'll notice they loose their sharpness when the Quincunx sampling pattern is used. Game resolution was 1024x768, 32bit, high quality.

Click to view image full size using 4xAA

Please allow time for this image to load

Page 5- Shaders