A high performance implementation for Molecular Dynamics simulations on a FPGA supercomputer

The design and implementation of an FPGA core that parallelises all the necessary operations to compute the non-bonded interactions in a MD simulation with the purpose of accelerating the LAMMPS MD software is presented in this paper. Our MD processor core comprised of 4 identical pipelines working independently in parallel to evaluate the non-bonded potentials, forces and virials was implemented on the nodes of a FPGA-based supercomputer, namely Maxwell. Implementing our FPGA core on multiple nodes of Maxwell allowed us to produce a special-purpose parallel machine for the hard-ware acceleration of MD simulations. The timing performance figures of this machine for the pairwise LJ and short-range Coulombic (via PPPM) interaction computations in the MD simulations of the solvated Rhodopsin protein systems show performance gains over the pure software implementation by factors of up to 14 on two or more nodes of the Maxwell machine.