Temperature in NEMD method continuously rising between fix nve while executing on GPU server

Dear LAMMPS-USERS:

I use 14May2016 lammps on my CPU while lammps-stable on my GPU. The newest lammps-stable version is very convenient to compile. My compiled procedure is as blew.

Compile fftw
Compile openmpi
Modify .bashrc
Compile lammps gpu : …/lammps/src execute " make mpi " ; …/lammps/lib/gpu modify Makefile.linux.single & execute " make -f Makefile.linux.single " ; …/lammps/src make mpi

Also here is my first part of the screen output :

LAMMPS (17 Nov 2016)
Lattice spacing in x,y,z = 4.35 2.51147 43.5
Created orthogonal box = (0 0 0) to (10005 30.1377 43.5)
3 by 1 by 1 MPI processor grid
Created 110400 atoms
48 atoms in group hot
48 atoms in group cold
Reading potential file BNC.tersoff with DATE: 2013-03-21
WARNING: Resetting reneighboring criteria during minimization (…/min.cpp:168)

*发件人:* "王淑婷" <[email protected]>
*发送时间:* 2017年3月24日 星期五
*收件人:* "Axel Kohlmeyer" <[email protected]>
*抄送:*
*主题:* Re: Re: [lammps-users] Temperature in NEMD method continuously
rising between fix nve while executing on GPU server

Dear LAMMPS-USERS:

I use 14May2016 lammps on my CPU while lammps-stable on my GPU. The newest
lammps-stable version is very convenient to compile. My compiled procedure
is as blew.

Compile fftw
Compile openmpi
Modify .bashrc
Compile lammps gpu : ../lammps/src execute " make mpi " ;
../lammps/lib/gpu modify Makefile.linux.single & execute " make -f
Makefile.linux.single " ; ../lammps/src make mpi

Also here is my first part of the screen output :

LAMMPS (17 Nov 2016)
Lattice spacing in x,y,z = 4.35 2.51147 43.5
Created orthogonal box = (0 0 0) to (10005 30.1377 43.5)
  3 by 1 by 1 MPI processor grid
Created 110400 atoms
48 atoms in group hot
48 atoms in group cold
Reading potential file BNC.tersoff with DATE: 2013-03-21
WARNING: Resetting reneighboring criteria during minimization
(../min.cpp:168)

--------------------------------------------------------------------------
- Using acceleration for tersoff/gpu:
- with 1 proc(s) per device.
--------------------------------------------------------------------------
Device 0: TITAN X (Pascal), 28 CUs, 12/12 GB, 1.5 GHZ (Single Precision)
Device 1: TITAN X (Pascal), 28 CUs, 1.5 GHZ (Single Precision)
Device 2: TITAN X (Pascal), 28 CUs, 1.5 GHZ (Single Precision)

​energy conservation in single precision typically worse than with double
precision. there is much more noise added through the significantly
truncated precision, especially when summing up forces.

please recompile LAMMPS with double presicion GPU support and compare
again.​
​if the temperature is still rising in the same way as with single
precision, we need to take a closer look.

axel.​