Fig. 5: Schematic figure of heterogeneous parallelization in the NVNMD system.
From: Accurate and efficient molecular dynamics based on machine learning and non von Neumann architecture
The system consists of master processing unit (MPU), slave processing unit (SPU), and high-speed transmission interface (HTI). Calculation of each timestep of MD in the system consists of seven consecutive steps: building neighbor list (S1); encoding atomic information into compact data (S2); receiving and decoding atomic compact data (S3); evaluating potential energy surface (PES) (S4); encoding and sending data of PES (S5), decoding data of PES (S6), and other calculations, e.g., numerical integration (S7). Here, E and V represent the energy and virial of atomic system, respectively; Zi, Ri, Ei, Fi, and Nc(i) represent the chemical specie, coordinate, energy, force, and neighbor list of atom i, respectively; Rji, Fji, and Vji represent the relative coordinate, force component, and virial component between atom i and atom j.
