Enabling quantum Monte Carlo simulation of solids with large and low symmetry cells

Computer simulations help us predict and explain the properties of functional materials for energy applications. With Exascale supercomputers, quantum Monte Carlo (QMC) simulations become affordable for studying materials with hundreds of atoms in a simulation cell. Single particle orbitals of the many body wavefunction in QMC play a crucial role in the accuracy and speed of the simulation. 3D cublic B-spline orbital presentation is a very common choice for solid thanks to its easy conversion from plane wave orbitals and fast evaluation in real space. With low symmetry structures, the memory footprint of B-spline orbitals grows quadratically with respect to the atom count easily beyond the memory capacity of a single GPU, a restriction from QMCPACK parallelization that assigns one GPU to each MPI rank for maximal efficiency. This work presents how we distribute the memory usage and computation of B-spline orbitals and how QMCPACK performance can be improved with more resident walkers on a GPU.