In the paper (https://dl.acm.org/doi/10.1145/3575693.3575702) there is also a TensorRT comparison. There, they are pretty much neck and neck. Sometimes faster, sometimes slower.