GPUs

Kubernetes has access to several virtual GPUs. Currently, the only GPU model we can offer is the NVIDIA A40 12Q.

GPUs are provided on a short-term basis only. We cannot predict how the GPU situation will develop or how long GPUs will remain in operation on our platform. We therefore reserve the right to remove GPUs from Kubernetes at any time. Please do not build services that depend on long-term GPU availability.

Currently, all GPUs are being utilized for JupyterHub, and all previous use cases have been well accommodated there.

If, however, there is a compelling reason why your service requires access to a GPU, please feel free to send a request with an explanation to cloud@uni-muenster.de, and we will evaluate the case together with you.

As a best practice, GPUs should not be held by long-running processes. Allocate them within jobs instead, even in the case of a service, so that the resources are freed as soon as the computations are completed.

Furthermore, it may take several minutes until the job is started with a GPU, as a separate virtual machine must be launched for each GPU pod. It is also possible that all GPUs are being used elsewhere, such as in JupyterHub, so one may need to wait until a GPU becomes available again.