Has anyone done, or seen, any analysis on energy used for ML training rather than clock time? I’ve done some of my own for comp heavy workloads, but it wasn’t M

Has anyone done, or seen, any analysis on energy used for ML training rather than clock time? I’ve done some of my own for comp heavy workloads, but it wasn’t ML on a GPU.