I am using Lammps to study CVD of hydrocarbon employing ReaxFF potential. From the manual of Lammps, CUDA package can not accelerate a LAMMPS calculation employing ReaxFF potential. But some commands are supported by CUDA package,such as “fix npt/cuda”,”fix nvt/cuda’’. I use commands “fix npt/cuda” and “pair_style reax/c” together in my input script, whether CUDA package can accelerate the calculation employing ReaxFF(pair_style reax/c) or not?
The most expensive part of your calculation will be the force
evaluation in reax/c. It helps nothing at all if you move the
integration part to cuda. So, no, CUDA package cannot accelerate
reax/c.
I am using Lammps to study CVD of hydrocarbon employing ReaxFF potential.
From the manual of Lammps, CUDA package can not accelerate a LAMMPS
calculation employing ReaxFF potential. But some commands are supported by
CUDA package,such as “fix npt/cuda”,”fix nvt/cuda’’. I use commands “fix
npt/cuda” and “pair_style reax/c” together in my input script, whether CUDA
package can accelerate the calculation employing ReaxFF(pair_style reax/c)
or not?
you question is the perfect example for how dangerous it is to only
know a little and not spend any reading and explanations how GPU
acceleration works.and particularly how the USER-CUDA package is
implemented.
already, a simple consideration of amdahl's law and knowing how much
time is spend on computing the pairwise forces relative to everything
else (as shown in the output) makes your consideration totally
pointless.
in short, "a little bit of cuda" doesn't help. on the contrary.