Training is close to traditional HPC in many ways. Inference is far simpler since it's a simple forward-going pipeline of a relatively small working set.