Originally Posted by gstanford
Just as an update to the 3dmark05/06 issue I mentioned earlier, I asked a friend for an ATi screenshot of the same general area.
You should be able to notice that the same sawtooth artifacting is present on the shadows with the ATi card, therefore this has to be a futuremark issue.
The problem is in their Cascaded Shadow Map implementation, it has very good shadow resolution 5x(2048x2048), but the shadowed area is very large so the texels can be quite big in screenspace.
In addition to that, their implementation doesn't keep same shadow texel locations in the worldspace, so shadow texels follow camera and causes very annoying temporal aliasing.
The small light bleeding under the shadowing object comes from the fact that they cast shadows from the backfaces. (from lights point of view)
This is very traditional trick to make sure that there is no shadow acne on surface, but side effect is that shadow actually starts slightly after the shadowing object ends. (additional space comes from slight depth bias when rendering the shadow map.)