Peoplenet performance on Jetson

• Hardware Platform : Jetson
• DeepStream Version : 5.0

Hi All,
I have implemented deepstream pipeline which uses peoplenet model for inference.
It gets the raw frame from appsrc and push it to deepstream pipeline for inference.
From the peoplenet ngc It says that It gives more than 200 fps inference speed for resnet34 model.
But I am getting max 80 fps.
I am providing 16 channels stream as input, and per channel I am only able to get max 5 fps.
appsrc0
asppsrc1 → deepstreampipeline → display
appsrcn
Can anyone suggest me what can be issue?
Do let me know if you need implementation or hardware configuration specifi issue.

Which platform are you using? Xavier or NX or Nano?

I am using Jetson xavier platform.

Ok, are you using int8 mode?

Yes I am using int8 mode.
I have tried both resnet34 and resnet18 model. But still got the performance same.

Ok, pls try following command before you run the test:

  1. Max power mode is enabled: $ sudo nvpmodel -m 0
  2. The GPU clocks are stepped to maximum: $ sudo jetson_clocks

Also what is your pipeline and can you refer DeepStream SDK FAQ - #11 by bcao to find which element is the bottleneck. And I think it is better to profile the model’s perf using trtexec .