For folks with GPUs, note that HDBscan is very optimized in cuML (https://docs.rapids.ai/api/cuml/stable/api/#clustering / https://developer.nvidia.com/blog/faster-hdbscan-soft-cluste...).