Databricks GPU clusters clusters support low-level CUDA development using Java libraries such as JCuda.
Databricks recommends using an init script to install JCuda to make it available on all cluster nodes.
Launch a GPU-enabled cluster with the preloaded Nvidia Libraries and use the notebook below to install an init script that attaches pre-built JCuda libraries to the GPU-enabled cluster.
Check out http://www.jcuda.org for details.
This tutorial is inspired by the classic Pi Estimation example at http://spark.apache.org/examples.html