Re:One more thought on this... (Score 1)

Actually, Quake 1's renderer did this. When a non-static polygonal object reached a certain distance from the camera, it would jump into a different renderer that just plotted the object's vertices. If a particular triangle in the model happened to be more than 3 pixels "large" at that distance (meaning there would be gaps), it would recursively subdivide the triangle, plotting the vertices along the way, until the subdivision yielded triangles 3 pixels in size. Crude, but fast.

To stay on topic, this accelerator IMHO represents a fairly significant advance in graphics hardware (not that it's new, but that it represents an intent to bring the hardware closer to the consumer market). As good as textured/lit/AA/bumpmapped/envmapped polys look, people need to remember that they're still just approximations. Take any polygonal/curved object, and keep increasing the resolution of detail. Eventually you're going to end up with just vertices. So while those approximations are the hip and in thing now, it's important to remember that eventually they will no longer be sufficient.

It should also be noted that when they say "voxels" they are talking about actual volume data, meaning a 3d array of samples. Delta Force/Commanche/Bladerunner/Tiberian Sun are all 2d simplifications.