fp16 mode of onnxparser cannot speed up on Xavier

Sinse the onnxparser api has changed in TensorRT5, I found that when using builder.setFp16Mode(true) to enable fp16 mode, it cannot speed up and the fp16 plan file is larger than fp32 plan file.
Could you give me a example to show how to using fp16 mode with onnxparser?

My envirs: