Now that we've seen the performance numbers, let's now take a closer look at exactly how the Matrox, NVIDIA and ATI cards perform anti aliasing. ATI's SMOOTHVISION is actually a form of supersampling, where pixel data is up-sampled, re-sampled, processed and then sent to the buffer. The result is a high-quality anti-aliased image, but the large data requirements can put a serious strain on the memory bus. The preceding ATI anti aliasing benchmarks give a good indication of the performance implications, which also scale up exponentially as the resolution is increased.
NVIDIA uses a better anti aliasing method called multisampling. The actual process is exactly how it sounds, and it takes multiple samples of the same pixel data which is then processed into the final pixel data. This reuse of pixel data puts much less strain on the memory bus, but the actual multi-sampling does hit the GPU with some additional processing overhead. NVIDIA has alleviated this by advocating the hard-coded Quincunx AA pattern for the GeForce3 and through the dedicated Accuview Anti Aliasing found on the GeForce4 Ti and MX cards. The GeForce4 Ti in particular, provides an excellent combination of image quality and performance, especially at higher resolutions.
Matrox uses a new technique called 16x Fragment Antialising with the Parhelia, and this is different than either multisampling or supersampling. Instead of the entire scene being subject to anti aliasing, FAA-16x takes the angle that only edge data really needs to be processed, and that this data only represent only 5-10% of the pixels in a given scene. The Parhelia GPU utilizes FAA-16x to identify fragment and non-fragment data using 16x subpixel accuracy, and then fragment data is written to the specific buffer and the rest goes to the standard frame buffer. Once the fragment data has been supersampled and processed, it too is sent to the framebuffer and displayed.
This process is akin to Hidden Surface Removal or any of the myriad occlusion removal techniques that 3D cards use to save on bandwidth. In HSR, data is polled and if found to be occluded by some other object, then there is really no reason to process/store it. The same goes for FAA-16x, which uses basic supersampling for its AA technique, but only processes the data its algorithm determines as representing edge data. The potential benefits are enormous and FAA-16x should only have a minimal impact on overall performance. Best of all, as resolutions increase, overall edge data decreases and a FAA-16x engine will scale quite well to higher resolution anti aliasing.
All of the above factors come into play for the Matrox Parhelia, and its FAA-16x anti aliasing performance is top-notch. But as with other forms of data culling (like HSR) no process is truly perfect. If a card uses supersampling or multisampling, there is a higher overhead involved, but it is still working with a full scene of data. With FAA-16x, the algorithm predetermines the data that is anti aliased and leads to a few potential issues. One is image artifacts, and the other is what we like to call skipped edges.
To their credit Matrox has fully disclosed the potential for image artifacts, and even included large screen shots in the reviewer documentation. The official explanation is that some games are incompatible with the FAA-16x algorithm and can lead to image quality issues. We're still a bit unsure as to what this means, because if the Parhelia is actually inspecting full scene pixel data using 16x accuracy, it is strange that any game could be incompatible. Regardless, there are certain games that display image artifacts using FAA-16x, but the exact number and actual titles affected is as yet unknown.
On the flip side, we found no mention of skipped edges in the documentation, or situations where anti aliasing was not used to smooth some very blatant jaggies. Whether this is another incompatibility or simply a function of the FAA-16x algorithm is unknown, but just as we examined the incredible performance benefits of FAA-16x, so too should the potential issues be identified as well.