You can, theoretically, have a model that is so biased with more polygons on one side that it would affect performance worse than double. Think of something like a diamond ring: the diamond has thousands of facets, but the back only has a few polygons.