Lower latency and higher throughput are better
Power consumption and energy per inference
Higher is better - GPU wins by 366×