AMD GPU systems
COSMA has six AMD MI50 GPUs hosted in the ga003 server. These are interlinked with a 4x and 2x InfinityFabric link between the GPUs, to enable direct data transfer.
There is also a MI100 GPU in ga004.
Relevant software is in /opt/rocm* and the AMD AOCC compiler is available as a module (module load aocc)
The rocm_smi.py command will provide information about the GPUs.
These systems are available for general use. If you use them, feedback would be welcome, in particular around performance, and the software environment.