While I have limited resource usually to train my machine learning models, I often find that my hyperparameter optimization procedure is not necessary using all my GPU and CPU, and that is because the results also depend on the batch size in my experience.
If you find in your project that a low batch size is necessary, how do you scale your project? In a multi-GPU scenario, I could imagine running different hyperparameter settings on different GPUs, but what other options are out there?