ResNet-18 very slow on Jetson Nano

[Myrgued]

Hi there,
I have been running a PyTorch-ResNet-18 on the Jetson Nano 4GB (JetBot) with MAXN and jetson_clocks active. I am using Jetpack 4.4.1. I measured the model's performance with the method you provided in your benchmark test.py. Latency is ~65 ms and throughput ~15.4 fps. Input is (1, 3, 224, 224) and I use 16FP. When measuring the TRT model, the results are quite exactly as the ones in your benchmark. What could cause this gap in performance of the PyTorch model?
Thanks for any hints!