The issue of inference speed

[zhuo-limeade]

I have successfully reproduced dal-tiny, but when I used tools/analysis_tools/benchmark.py to test the inference speed on a single RTX3090, the inference speed of dal-tiny was much faster than the 16.55fps claimed in the paper, reaching 23.77fps. Is method of my inference speed test wrong?
My cuda version is 11.3, and the other configurations are consistent with the README.

translation：
我已经成功复现了dal-tiny，但是我在利用tools/analysis_tools/benchmark.py在单个RTX3090测试推理速度时，dal-tiny的推理速度比论文里宣称的16.55fps快了很多，达到23.77fps。请问我的推理速度测试方式有误吗？
我的cuda版本是11.3，其他的配置均与README中的一致。