You'll get a log_2-based scaling efficiency with nearly any batchsize increase, pending some limitations (memory, etc).
That should be enough at least to roughly sketch it out.