With a new and intrinsically parallel algorithm for Tight Binding Molecular Dynamics we obtain a performance of 3.4 Gigaflops per million dollar on a cluster of 8 Hewlett Packard workstations in a simulation of 216 Silicon atoms. One time step with this new algorithm takes as much time for the 216 atom system as one time step with a conventional algorithm on a NEC-SX3 supercomputer. In addition, the linear scaling of new algorithm allows us to calculate systems of unprecendented size which are not any more accessible by the combination of standard algorithms and vector supercomputers.