We describe different implementations of the 3D Heisenberg spin glass model for Graphics Processing Units (GPU). The results show that the {\em fast} shared memory gives better performance with respect to the {\em slow} global memory only if a multi-hit technique is used.