As an engineer I am pretty skeptical about this claim. I am not tempted to believe that it is possible to recreate 22 bits from 16 bits input, without some kind of catch. Actually, if I want I can create 22-bit output from any data you send me. Its really simple: just add noise. Yes, I can create 22-bit color output, just by taking 16-bit output and adding some random bits to make it 22-bit... but is this really 22-bit color? Are the 22-bit colors shown by the 3dfx technique really closer to the 24bit color that it "should" be? Again questions, but no answers.

What 3dfx probably does is this: they use dithering and in some way, they are trying to reverse this dithering. Its not completely impossible to do that, BUT dithering has one huge problem: it reduces the spatial resolution (the tradeoff for gaining color resolution). So basically, the result of all this is a reduction in the sharpness of the image. Do we want that kind of trade-off? Is this why Voodoo1's images where so blurry? Naturally, we are kept in the dark about how it really works. I can't judge how good all of this works since I have never seen a Voodoo3 in action. (hint ;-) )

Anyway, I wrote a small article about Dithering. I wasn't planning to release it right now but I will VERY soon, as it seems to fit into this little discussion. Stay tuned for it to be here soon.

So should we believe him or not? I don't know. Maybe 3dfx should just get used to the fact that 3D sites and 3D coders do not like it that 3dfx doesn't support Real 32-bit rendering. They will be criticized for it, just as they will be criticized for not supporting higher resolution textures. I think Tarolli should either just leave this whole discussion and get used to it, or he should explain how it works. This kind of "believe me it works" column sounded too much like marketing crap. In the end, we know that at some stage of the rendering 3dfx is stuck with 16-bit quality, and no matter what they do this fact will stay... and it will cause some kind of image quality reduction.

Just quick note about the floating point Z-Buffer. He is right about that, but this is nothing new. This technique has been used by others too, and they don't make a big fuss about it. For example VideoLogic's PowerVR uses a 32-bit floating point Z-buffer, but you don't hear them claim to have an equivalent 44-bit linear Z-Buffer.

This whole column looks a lot like 3dfx is defending their lack of features by a whole lot of techno talk. Perhaps they should start worrying more about the competition.