I think it depends on whether an algorithm is forced to traverse an array in the cache-efficient direction or not. Oftentimes you can't choose whether to make your outer loop rows or columns so the performance could go either way.