Quantum-Based Molecular Dynamics on Graphics Processing Units

Susan M. Mniszewski, CCS-3; Anders Niklasson, Ed Sanville, Marc J. Cawkwell, T-1 The performance of quantum-based molecular dynamics (MD) simulations using the LANL-developed code LATTE is limited by the time required to compute the density matrix at each time step. The density matrix has been computed historically via the diagonalization of the Hamiltonian matrix. We have investigated the performance and accuracy of the second-order spectral projection (SP2) algorithm that calculates the density matrix via a recursive series of generalized matrix-matrix multiplications. Owing to its simplicity, the SP2 algorithm is ideally suited to implementation on graphics processing units (GPU). We demonstrate that optimized implementations of this algorithm on single and multiple GPUs on a single compute node lead to significant speed-ups with respect to central processing unit-based implementations and diagonalization. Furthermore, the SP2 algorithm produced an error rate in the density matrix that are as good or better than traditional methods.