AMD GPU testbeds
COSMA contains multiple AMD GPU systems to allow code development, porting and performance benchmarking.
Specifications
Usage
If you do not already have an account on COSMA, please follow the instructions here, then request to join the project: do018.
MI100 node is accessible through the
cosma8-shm2queue. Use--nodelist=ga004to ensure allocation to the MI100 node.MI210 nodes are accessible through the
cosma8-shm2queue. Use--exclude=ga004to ensure allocation to the MI210 nodes.MI300X node is accessible through the
mi300xqueue.MI300A node is accessible through direct ssh. From a login node, use
ssh ga008.
Compiler modules
The hipcc family of modules include the ROCm software stack. You probably want to use the latest one (use module available hipcc to see options).
After loading this module, you may also want to load a standard compiler module, e.g. module load gnu_comp to provide a standard compiler (which may or may not offer better performance than the hipcc compiler for host CPU code).
Due to the way module dependencies are calculated, the order in which you load these modules is important. The last one loaded wil be the one that determines which versions of libraries you can then load (e.g. FFTW, HDF5, etc).
Notes
The AMD ROCm software stack is installed. ROCm 6.3.0 is available at /opt/rocm-6.3.0/bin/hipcc (2024)
ROCm 7.2.0 is also available (May 2026).
The latest version installed can be found under /opt/ - please check here.
ROCm is installed on the login8a login node, though this does not contain a GPU. This should facilitate compilation of codes where Internet access is requried.
CUDA code must be converted to HIP using the hipify script provided with ROCm.