Not many, since it's slow to begin with. You'll get a log_2-based scaling efficiency with nearly any batchsize increase, pending some limitations (memory, etc)

Not many, since it's slow to begin with. 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.