Jetson Xavier NX can not support multi video stream well by using deepstream

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hello,

I got a problem when using deepstream to do multi-camera video analysis on Jetson Xavier NX.

The problem is the GPU usage will up to 99%, and the multi-camera debugging video will be very stuck and lagging, when I use deepstream to do multi-camera video analysis.

multi-camera means 2 or 3 RTSP cameras.

The GPU usage will be about average 40%-60% when doing single-camera (1 RTSP camera).

Official specification said Xavier NX can support 16 cameras video streams analysis, but why The GPU usage will be so high when only use 2 or 3 RTSP cameras?

Some details of my program are listed below:

  1. The nvinfer part include pgie and tracker. The pgie is peoplenet34 of TLT pretrained models, and the tracker is libnvds_mot_klt.

  2. Some screenshots after the program run are uploaded

many thanks!

01
011920×1442 168 KB

02
021920×1442 312 KB

03
031920×1442 369 KB

04
041920×1442 174 KB

3

Hi,
By default we run ResNet10 to demonstrate 16-source inputs. Please run deepstream-app with this config file for a try:

/opt/nvidia/deepstream/deepstream-5.1/samples/configs/deepstream-app/source30_1080p_dec_infer-resnet_tiled_display_int8.txt

In your case, it looks like the model is too heavy and GPU is overloaded for multiple sources. We suggest adjust interval in [primary-gie]. Please refer to the document:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_ref_app_deepstream.html#primary-gie-and-secondary-gie-group
Frequently Asked Questions — DeepStream 6.1.1 Release documentation

4

Regarding the reproducibility of the claimed DS performance, please refer:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_Performance.html#jetson

To optimize the DS performance, please refer the following best pratice:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_ref_app_deepstream.html#performance-optimization

5

Hi DaneLLL, thanks a lot for your advices.

I’ve replaced the nvinfer model from Resnet34 (TLT-pretrained peoplenet model) to Resnet10 of my program, and the GPU usage was down to 30% - 50% by using 2 cameras, and the latency of 2 cameras were both lower than before and be acceptable.

BTW, I hope the next gen Jetson Xavier will have more powerful GPU (10 times than now).

6

We do have JETSON AGX XAVIER which can deliver 32 TOPS AI Performance, you might find more information here: AI-Powered Autonomous Machines at Scale | NVIDIA Jetson AGX Xavier

If you need more computation power, EGX platform can fulfill your requirement (10x AI performance), more information: The EGX Platform

8

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.