By leveraging idle computing time, researchers can double the speed of model training while preserving accuracy.