At that point, how does this compare with simply running the model on the CPU?