Usually you will be using different geometry for collision detection -- so there is no copy; there is just different data. For example, one could use axis-aligned bounding boxes or object-oriented bounding boxes. Once computed, they can be used separate from the vertex/face geomtry of the mesh. When using complex per-face collision detection, the cd data comes closest to the actual mesh data. But even then, you can usually gain a lot since you don't need to have duplicate vertices (otherwise required for multi-texturing or materials) and you don't need to have extra triangles in flat surfaces. In such a case it is still best to store it separately.

Greetz,

Illco