In this paper we use the subdivision algorithm to approximate the box dimension of attractors of dynamical systems. Although in theory the subdivision algorithm provides a covering of the attractor with boxes of arbitrarily small diameter, in practice we have to overcome two obstructions: (1) ensure that the covering is (almost) minimal and (2) enhance the speed of convergence to the box dimension. We solve both problems and apply our results to the Hénon, Lorenz, Rössler and Chua attractors. The method suggested in this paper uses information from several subdivision steps and converges to the box dimension much faster than the expression in the definition of the box dimension which uses only one covering of the attractor with boxes of a prescribed diameter.