How to build LAMMPS with gpu package on Maxwell GPU.

Dear all

I have a GPU node with Maxwell GPU card(nvidia geforce 980), and I want to build LAMMPS on it with gpu package. I can build LAMMPS(latest stable version) with cuda package at this node and it works perfectly. How to modify the makefile in lib/gpu folder? An example is best.

What’s more, if in a node there are two GPUs, one GPU is geforce 780Ti and one is geforce 980, is this feasible ?

Thanks

You can checkout README in lib/gpu to build libgpu.a.

lib/gpu/Makefile.linux can be a good example to start with. Generally, you may need to modify a couple of parameters such as CUDA_HOME to point to your CUDA install path, CUDA_ARCH to match the (common) compute capability of your card(s), etc.

The GPU package should work with the 2 GPUs, as you described, as long as the number of MPI tasks per GPU is greater than, or equal to, 1. The documentation explains this. (http://lammps.sandia.gov/doc/package.html and http://lammps.sandia.gov/doc/accelerate_gpu.html)

Hope it helps,
-Trung

Dear all

I have a GPU node with Maxwell GPU card(nvidia geforce 980), and I want to
build LAMMPS on it with gpu package. I can build LAMMPS(latest stable
version) with cuda package at this node and it works perfectly. How to
modify the makefile in lib/gpu folder? An example is best.

What's more, if in a node there are two GPUs, one GPU is geforce 780Ti and
one is geforce 980, is this feasible ?

please note that using two GPU models may lead to unexpected load
balancing problems, if the performance of the two GPUs is too
different.

Dear Dr. Axel Kohlmeyer:<br/><br/>Thanks very much for your advice. The problem has been solved.<br/><br/>Thanks

Thanks

The problem has been solved. After update cuda driver, GPU package works all right.

Thanks for your reply.