GTC 2020 S21976
Presenters: Evan Weinberg ,NVIDIA ; Stan Moore, Sandia National Laboratories
Cutting-edge investigations of material behavior via classical molecular dynamics simulation methods require application-specific, quantum-accurate interatomic potentials (IAPs). The SNAP machine-learning IAP, formulated in terms of general four-body geometric invariants, is trained against quantum electronic structure calculations. This enables the verifiably high-fidelity investigation of diverse material systems at length- and timescales unattainable by purely quantum calculations. Despite the high arithmetic complexity, achieving good SNAP performance with the increasing parallelism provided by modern GPU architectures is challenging. To address this, we have developed a novel parallelization over the geometric structure of the SNAP IAP, prompting memory layout optimizations which facilitate data reuse and reduce memory bandwidth requirements. The new SNAP algorithm will be deployed in the GPU-optimized LAMMPS implementation using the Kokkos templated C++ library.
Watch this session
Join in the conversation below.