GPU package on Windows

LAMMPS LAMMPS Mailing List Mirror
1

Dear Developers,
I tried to use my GPUs to run Lammps on Windows 10, using Pre-built installer package for windows
And I have some questions needed your advices:

  1. The GPUs didn’t run with full performance, but just around/less than 40% of capability. Do we have any way to force them to work with 100% capability?
  2. During simulations, I see that Lammps use almost RAM memory but didn’t use much physical memory of GPUs. Does we have anyway to balanced these memories or force Lammps to use GPU’s memory?
  3. The simulations using GPUs on Windows normally run well at the initial stage, but often get crash and stop when the run is going to end.
    I use version Lammps-11Aug2017 and MPICH2 for windows

Please give me some advices
Thank you so much

2

Dear Developers,
I tried to use my GPUs to run Lammps on Windows 10, using Pre-built
installer package for windows
And I have some questions needed your advices:
        1. The GPUs didn't run with full performance, but just around/less
than 40% of capability. Do we have any way to force them to work with 100%
capability?

​high GPU utilization does not automatically mean better performance. how
much the GPU is utilized depends a lot on your input and your GPU. since
the GPU package can run on the GPU and CPU concurrently, it is not always
best to run as much on the ​GPU as possible. for PPPM, it is often more
efficient to run it on the CPU and adjust the coulomb cutoff, so that Pair
style on the GPU and the rest on the CPU are optimally balanced.

         2. During simulations, I see that Lammps use almost RAM memory
but didn't use much physical memory of GPUs. Does we have anyway to
balanced these memories or force Lammps to use GPU's memory?

​the GPU package uses as much memory on the GPU as it needs. there is no
point to move more data to the GPU. it will just slow you down.​

         3. The simulations using GPUs on Windows normally run well at the
initial stage, but often get crash and stop when the run is going to end.
I use version Lammps-11Aug2017 and MPICH2 for windows

​this is likely an issue with your OpenCL GPU driver and/or your GPU/host
system. it may also be an issue with your input and choice of parameters,
that may not be suitable for mixed precision force evaluation. the LAMMPS
GPU source code​ has been well tested and used reliably on many machines
and by many people.

​axel.​