Efficiency of classical molecular dynamics algorithms on supercomputing hardware

Development of new HPC architectures proceeds faster than the corresponding adjustment of the algorithms for such fundamental mathematical models as classical molecular dynamics. The wide variety of choices poses the requirement of clear guiding criteria for the computational efficiency of a particular model on a particular hardware. LINPACK benchmark can no longer serve this role. In this work we consider a practical metrics of the time-to-solution versus the computational peak performance. In this metrics we compare different hardware (both legacy and modern) on the example of the LAMMPS software packages widely used for atomistic modeling. We show that the metrics considered can serve as a universal unambiguous scale ranging different combinations of CPUs, accelerators and interconnects.