KubeCon + CloudNativeCon North America 2024

Kind is a powerful tool for running local Kubernetes clusters using Docker. It is particularly useful for testing, development, and CI/CD workflows, offering features like multi-node cluster support, easy configuration, and cross-platform compatibility. However, providing access to GPUs in Kind is not a very straightforward process. There is no standard way to inject GPUs into a Kind worker node, and even with a series of "hacks" to make it possible, post-processing is still needed to isolate different sets of GPUs to different nodes. In this lightning talk, we introduce nvkind – a wrapper around Kind that encapsulates the steps necessary to make GPUs available to Kind worker nodes. Ideally, GPU support would have been added to Kind directly, but many challenges exist to make this possible. This talk discusses those challenges, how we've overcome them with nvkind, and the steps needed to eventually support GPUs directly within Kind itself.