Measuring energy on a single machine running multiple concurrent GPU processes (Federated Learning simulation)

Dear authors,

Thanks for your nice work.

I have been using codecarbon for some time, and I was wondering what is the recommended way to use CodeCarbon to measure energy/emissions on a single machine where multiple GPU processes run in parallel?

Our use case is a federated learning simulation that launches many federated “clients” as separate processes on one GPU.

Right now, we run one EmissionsTracker per process (each simulated client), and this seems to be working reasonably. However, if we track per process, how does CodeCarbon derive the GPU power consumption? Is per-PID attribution supported via NVML, and could that double-count energy cost when multiple processes share a GPU? How should we think about CPU/RAM attribution when multiple processes are active? Are there docs/examples for multi-process, multi-GPU scenarios we should follow?

Thanks in advance! Any pointers to best practices or examples would be greatly appreciated.

Best regards,
Austin Tapp

No file or test is named in the issue. Start by reviewing the existing EmissionsTracker and NVML documentation, focusing on process-level GPU, CPU, and RAM attribution for concurrent clients. Done should be clear guidance or examples for single-machine multi-process and multi-GPU tracking without double-counting.