COSMA has a number of GPU systems, which are available for use. These are:

• gn001: 10x NVIDIA V100 GPUs
• ga003: 6x AMD MI50 GPUs
• ga004: 1x AMD MI100 GPU
• login8b: 0-3x NVIDIA A100 GPUs
• mad04: 0-3x NVIDIA A100 GPUs
• mad05: 0-3x NVIDIA A100 GPUs
• ga005: 2x AMD MI200 GPUs
• ga006: 2x AMD MI200 GPUs

We have 3 NVIDIA A100 (40GB) GPUs, which can be moved (in seconds) between login8b, mad04 and mad05, hence the variable number above. If you have a particular requirement, please contact cosma-support. The default configuration is one GPU per node. These GPUs are part of a composible PCIe fabric using a Liqid infrastructure funded as part of ExCALIBUR.

To use mad04 and mad05, please submit to the cosma8-shm queue. To use the others, please ssh directly to them, from a login node, and be aware that they are a shared resource (i.e. other people might be using them)