CUDA is for NVIDIA GPU. The best way to use Gaudi is to use a framework that is built for Gaudi. Currently, TensorFlow and PyTorch are integrated with SynapseAI and optimized for Gaudi.
In cases where a user would need to create custom kernels, Habana provides an SDK for creating kernels. Users can start on our Developer Site for custom kernels to access the tools, documentation and videos.
Additionally, In TensorFlow, you can write custom Ops. For more information, visit Custom OP example on our GitHub page.