Comparative evaluation of deep learning workloads for leadership-class systems