It is an unwritten law of development that you always want to get more stuff into your game than will actually fit.
3dfx FXT1 texture compression is just another tool for furthering this goal.
Compression algorithms are used for everything from .jpg images to sound files. In a nutshell, texture compression is
math's way of putting more information onto less storage space. If we were compressing nuts, a hammer would do nicely
but as we are trying to maintain image quality as close as possible to the original texture, it gets a bit more complex.
During the encoding (compression) phase, an assessment is made automatically for each 4x4 or 4x8 block of pixels in a
texture as to which of four compression algorithms (see page three for a breakdown of the four methods) is best suited
for that block. The file is marked with a 'compressed texture' header and automatically decoded (decompressed) when it gets to the hardware.
Encoding can be done before mastering and compressed textures included on the CD with uncompressed textures (for other hardware) or it can be done at runtime.
Theoretically, a developer could add the compression encoder to his load routine and, provided the encoding is done in a reasonable amount of time
(we are skeptical on this point), the textures would be compressed for the ride over to the graphics hardware and decompressed there.
Individual textures may have a specific compression method designated by the developer or the texture can be flagged for
no compression if it is felt that the compression process would degrade a critical image (perhaps one with small text or
difficult to compress content) beyond acceptable levels.
More detailed textures, either larger or with multiple layers (ie lighting & bump maps) in an environment or the same level
of texturing in an environment with freed-up bandwidth allocated to other things. Most likely some combination of the two.
The compression does create artifacts. How noticeable they are is still up for grabs. In any case, accuracy using the best
of the four methods for a particular block is closer than S3TC and the gains are pretty significant.
The images 3dfx chose to use for their presentation are (even in their own words) not the best.
"Uncompressed image (left) compared to the same image compressed using S3TC (center) compared to FXT1 (right). Note how much more accurate FXT1 is at resolving the finer details of the images." - 3dfx FXT1 Presentation
"High-resolution image - Uncompressed (Left) vs. same high-resolution image - Compressed (Right). While the compressed image on the right suffers no loss in visual quality, it is represented using 6 times less data than the one on the right." - 3dfx FXT1 Presentation