Would be interesting to measure useful output vs compute time, as metric of how effective the models are.