Warp Engine Explained

[Himself]

I have the tech specs, must admit I have trouble gleaming information out of them as to the actual capabilities of the various parts. I wouldn't mind a reaoned run though of the warp engine myself, from someone who knows what they are talking about, if that is possible to find here. Never mind the silly banter from the peanut gallery when anything on topic comes up.

[SwAmPlAdY]

Warp engine specs are a touchy subject in companies. I doubt that no more than 5 people in each company actual now the real insides of their warp engines.

Some wise ass can allways say that it draws triangles real fast

I would be surprised if matrox did somekind of hardware T&L with the G400 because it will cause a big hit in its fill-rate and triangle rate due to "overdraw."

What's the solution? Increase fill-rate or switch to a different architecture.

AFAIK, only tile based rendering is capable of ignoring the "overdraw" problem by rendering only what you see. Basically, the scene is mapped into separate tiles and each of these tiles are rendered. Instead of rendering everything within the scene, only the objects that you see are rendered.

Currently, only the PowerVR based chips offer an affordable solution but I can be wrong. The only other way is to just increase the fill-rate.

Hardware based T&L cards also have a hard time with the vertex buffers calculation.

Then again, I can be totally wrong.

SwAmPy

[Himself]

I agree, once you bring the T&L operations over the the card side of the equation you have to change the way you do things dramatically. You no longer have the results of the view changes to work with, so culling is an issue to be handled differently for instance.

Overdraw is a byproduct of surfaces, a circle for instance wouldn't have any more overdraw than a square would have, but would involve more polygons, so it's not necessarily a strict overdraw problem. Q3 for instance uses extra polygons for curved arches, which also isn't an overdraw problem. You get overdraw when you try to do things like leaves on trees or have a lot of objects in general.

GigaPixel is another company interested in tiled architecture, unlike BitBoys, they actually have working silicon.

As for Matrox, who knows? I think they are getting the most out of the ram they are using right now, having to use bandwidth on vertex buffers can't speed things up, and extra steps in calculating mean lost clock cycles, so you need more than just the capability to do the math you need an architecture to match.