Cori GPU Nodes Software¶
The software stack optimized for Cori GPU nodes is maintained in a different module tree. You can access the stack using the cgpu module.
module load cgpu
Ideally, you should purge your modules through
module purge before loading the
cgpu module, this will remove the default Cori stack meant for the production nodes i.e. Haswell and KNL.
This page offers information about which compilers and MPI libraries are available for use on the Cori GPU nodes. It also describes methods of offloading code onto GPUs (CUDA, OpenMP, OpenACC, etc.) with the available system software.
Notes about using Intel MKL, Thrust, and other libraries are on this page.
On the Cori GPU nodes, we recommend that users build a custom conda environment for the Python GPU framework they would like to use; instructions are detailed on this page.
Instructions for using Shifter with CUDA on the Cori GPU nodes are provided.
The Cori GPU nodes provide a few tools for profiling GPU code; this page discusses how to use the tools, with examples given.
Several tools are available on Cori GPU nodeswhich can aid in debugging GPU code; this page offers examples and guidance.
There are a few existing known issues regarding the Cori GPU nodes; these are posted online at this page.