Deepstream 6.3+AGX Xavier, skeptical with astonishing performance

Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU) Jetson Xavier AGX
• DeepStream Version 6.3.0
• JetPack Version (valid for Jetson only) 5.1
• TensorRT Version 8.5.2.2

I am not trusting in what I am seeing. My equipment provider finally released an new BPS which allows me to use JetPack 5.1 and Deepstream 6.3. Previouslly, the limit was JetPack 4.6 and Deepstream 6.0. I see huge difference in my trivial tests between old and new libraries. The hardware is exactly the same: mic-730ai with AGX Xavier.

This is what I ran in my test: deepstream-test3-app (compiled) with the same RTSP stream being repeated over an over during the test, that is:

./deepstream-test3-app rtsp://blablabla:554
./deepstream-test3-app rtsp://blablabla:554  rtsp://blablabla:554
./deepstream-test3-app rtsp://blablabla:554  rtsp://blablabla:554  rtsp://blablabla:554
......

I collect for 5 minutes that terminal outputs, which look like this:
**PERF : FPS_0 (24.22) FPS_1 (24,26) FPS_2 (24,27)

Using DS6.3, JetPack5.1 the performance is almost at the possible limit, since my stream is 25fps. And it keeps almost the same even though I am increasing the number of RTSP streams. The previous configuration (DS6.0, JetPack4.6) decays quite fast, see the following chart.
benchmark

I am stalky on that. I am wondering if the more experiencing ones from this forum can give me a tip:

  1. DS is being tricky: it notices that I am providing repeated instances of the same input stream and it makes all the stuff of grabbing video, making inference and sending to the OSD with 1 source, and then, repeating the results to the n-other sources without processing. An so, the fps is quite the same all time.

  2. DS is being fair: it is really grabbing n sources, making inference in n frames, making the track in n frames, and providing n different outputs (although they look the same). The point is that I have not reached a number of sources that can make my plot start decaying.

I know that a good statistic demands a lot more samples, the value I obtain of FPS may not be too precise, but that huge difference stands tall. I wish I am in 2) but I am being skeptical of being in 1). What are your diagnosis?

The AGX Xavier supports 52x 1080p30 (H.265) videos decoding. Jetson Modules, Support, Ecosystem, and Lineup | NVIDIA Developer

There is no trick with DeepStream.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.