Is FP64 support required for GPU accelerated runs?

User discussions
1

GROMACS version: 2022.3
GROMACS modification: No

Hello,

I currently run GROMACS on a cluster and my lab has received funding and is looking to potentially purchase a Mercury 2x NVIDIA L40S GPU Node. The GPU’s do not support FP64, only FP32. We use a default mixed precision GROMACS build, and reading the manual (https://manual.gromacs.org/current/reference-manual/definitions.html) as well as a couple of other blog posts seem to imply that only FP32 support would be okay. My concern lies in the fact that some values including aggregates like the virial require double precision. Would these GPU’s be compatible with GROMACS or is FP64 support required?

Thanks.
-Michael Cordes

2

Double precision (FP64) cannot be used with GPU accelerated builds/runs at all, only mixed precision works. So, the GPUs are compatible with GROMACS, but if you really need to run with double precision the GPUs cannot be used.

In most cases mixed precision (GROMACS uses double precision for some variables, such as the virial) works as well as, and faster than, double precision. Unless you’ve already read https://manual.gromacs.org/documentation/current/reference-manual/definitions.html#mixed-or-double-precision I would recommend it.