Notes on usage of Blackwell GPUs

Cluster 9 contains an NVIDIA RTX 5090 GPU, which is the first GPU in GPULab which was built on the Blackwell architecture.

In many cases, software will need to be recompiled in order to be able to run on this architecture. Please refer to NVIDIA’s Blackwell Compatibility Guide for more information on what changes need to be made to make your software compatible with this GPU.

Usage on JupyterHub

If you wish to launch a Jupyter instance via JupyterHub on this cluster, you must make sure to use an image in which the software has been compiled with support for the Blackwell architecture.

The default images that are currently (May 2025) available on the JupyterHub spawn page are not compatible with the Blackwell architecture. We plan to upgrade these images during the summer, as to minimize the risk of any impact on the deadlines for thesis and PhD students in the coming weeks.

If you want to use this cluster with JupyterHub before that, you can use the following images:

• PyTorch: quay.io/jupyter/pytorch-notebook:cuda12-latest
• TensorFlow: not available (there is currently no official TensorFlow release that supports the Blackwell architecture)