Computing electrostatic interactions is a primary bottleneck in molecular dynamics (MD) simulations as a result of the slow decay of the Coulomb operator. To address this, we have recently developed a highly scalable method called the Gauss-Legendre-spherical-t (GLST) algorithm, which factors the long-range interactions into a finite sum of terms generated by utilizing the Gauss-Legendre quadrature and spherical t-designs to achieve O(N) scaling. This new algorithm also avoids the “all-to-all” communications required by the widely used particle-mesh Ewald method making it well-suited for modern distributed computing architectures. Here, we report new developments of the CUDA/C++ implementation of the method. In addition to optimizing the single-GPU implementation, we also introduce a multi-GPU scheme and investigate the scalability of the algorithm. We demonstrate precise control over the error of the calculation, as well as the performance of both the single- and multi-GPU libraries.
Gonzales et al. (Sun,) studied this question.