The energy calculation results of -nb CPU and - nb GPU are inconsistent

GROMACS version: 2024-dev-20231024
GROMACS modification: Yes/No
Here post your question
When I calculate the same input using - nb CPU and - nb GPU, the output obtained using gmx energy - f ener.edr (using -nsteps 0) is inconsistent. And the results of running - nb gpu multiple times are also inconsistent.
The difference in Coul SR is significant.


This is my mdp

integrator = md
nsteps = 10000
nstlist = 10
nstfout = 0
nstxout = 0
nstvout = 0
nstxtcout = 0
nstcalcenergy = 100
nstlog = 0
dt = 0.002
constraints = h-bonds
nstenergy = 500
ns_type = grid

coulombtype = PME
rlist = 0.9
rvdw = 0.9
rcoulomb = 0.9

tcoupl = v-rescale
tc_grps = system
tau_t = 0.1
ref_t = 300
fourier_spacing = 0.1125
cutoff-scheme = verlet