That's a better explanation that the one I gave, maxx. Blame my lack of coffee. And that's it in a nutshell; a bump map can only say "this pixel is supposed to be higher than that pixel" and the software has to generate a slope from one to the other. A normals map can go right down to the texel level to specify not height but exactly which way that texel is pointing -- without having to slop an imposed slope over its neighbors. Say I had a bump map that was a tight checkerboard; one pixel white, one black. The result would be, at best, a kinda lumpy surface. Bump mapping can't derive a hundred little towers. In a normal map, though, I could make one of those diffraction grating patterns in which adjacent rows of pixels had crisp, distinct angles. That would be an interesting effect, by the way -- one of the reasons I look forward to a way of mathematically superimposing outside data on a normal map.