lammps/lammps

[Feature Request] Mixed Precision GPU Reaxff/Kokkos Package

jobs-git opened this issue · 0 comments

Summary

The reaxff is already slower compared to classical MM given by its complexity. This is further compounded by the double precision only calculation which is about 2x to 64x slower than single precision. A mixed precision may result to speed up without too much lose in precision as fp32.

It was on the to do list of @stanmoore1 almost 4 years ago, I guess? https://www.osti.gov/servlets/purl/1806475