The high stochasticity of collision cascades makes running large sets of molecular dynamics simulations mandatory to obtain meaningful statistics. The convergence of the number of defects and of clusters and of the Primary Knock-On Atom (PKA) penetration depth with respect to the number of simulations in the sets is investigated for PKAs of 1 keV and 5 keV in Si. Two methods for setting the initial directions of the PKAs are compared: one is entirely based on randomness and the other integrates symmetry considerations. The latter method eases the convergence of the results. Larger sets are needed to converge the PKA depth than the number of clusters and defects. We observe that the higher the energy of the PKA, the harder it gets to reach the convergence. We find that large sets of simulations enables to get rid of the influence of the initial position of the PKA.