Nvv4l2decoder performance on jetpack 4.5.1 vs. 4.4


I observed bad performance of my application using gstreamer+nvv4l2decoder on jetpack 4.5.1. So I decided to investigate, TensorRT gives me the same result on inference, so discarded, but h264 decoding gives me differences.
If I do a comparison with a simple gst-launch of the same IP RTSP camera, same stream, I get this result:

On the left: r32.4.3
On the right: r32.5.1

The camera is RTSP, 720p, h264 at 25fps.
The cmd line is (more or less the same pipeline as in my application):
gst-launch-1.0 uridecodebin uri='rtsp://user:pwd$xxx:554/Streaming/Channels/1' ! nvvidconv ! fakesink

The pipeline generated is the same (I checked it with GST_DEBUG_DUMP_DOT_DIR converted to png).
This was tested with official ubuntu images provided by nvidia, both jetson nano boards are the devkit B01 with power mode MAXN.

Can anybody help me?

SOLVED: it was because I did not run jetson_clocks…

Thanks for the update.