from what i understand, the voxels more or less dynamically adjust their resolution according to their distance from the modelview origin (where the player is). if the engine generated dynamic shadow maps per light source, would'nt the voxels need to be high resolution with respect to each light source, so that the shadow map isn't a bunch of huge blocks? i imagine that alone would be a pretty big performance hit
The dynamic adjustment is traversing the voxel representation when you render the view. This kind of per-pixel shadow map requires rerendering the scene for each light source - but in more or less the exact same way that the player view is rendered. Then some work rectifying that information. So it would be a big performance hit, but not obviously any worse than normal rendering. Although did someone say it was only running at 20fps as it was?