Orin NX performance issue

I have a Jetson Xavier NX 8GB with Deepstream 6.0 running at 110fps with the YoloV3 example. For comparison, I have a new Jetson Orin NX 16GB with Deepstream 6.2. However, this only manages 52fps with exactly the same settings. The Xavier NX has already installed the patch Deepstream 6 YOLO performance issue - #9 by adventuredaisy . I think Deepstream 6.2 is also already patched. Why does the Orin NX perform so poorly? I would have expected a much higher performance from the device. DLA1 and DLA2 are not active.

• Hardware Platform (Jetson / GPU) Jetson Orin NX 16GB
• DeepStream Version Deepstream-6.2
• JetPack Version (valid for Jetson only) Jetpack 5.1
• TensorRT Version 5.1
• OpenCV without CUDA

• Hardware Platform (Jetson / GPU) Jetson Xavier NX 8GB
• DeepStream Version Deepstream-6.0
• JetPack Version (valid for Jetson only) Jetpack 4.6
• TensorRT Version
• OpenCV without CUDA

Have you disabled the power save mode of Orin before you tested the Orin performance? Jetson Orin NX Series and Jetson AGX Orin Series — Jetson Linux
Developer Guide 34.1 documentation (nvidia.com)

You can also use this method to enable the max clock: VPI - Vision Programming Interface: Performance Benchmark (nvidia.com)

What is the carrier board you are using for the Orin NX 16GB.
Are you using the deepstream “objectDetector_yolo” example
for your test?

Yes, the Jetson is running at maximum power. Jetson_clocks is active and NVP model 0 at MAXN. The Jetson consumes almost 25W when I run deepstream. So that shouldn’t be the problem.

@adventuredaisy I’ve been following your work for a while. You already have experience with the performance of Xavier and Orin. Can you maybe share some information? With your Xavier NX you get up to 55fps with Yolo. What performance can be expected with your Orin dev kit with the same configuration? Did you get your Yolo model up and running with DLA support?

I developed my own carrier board. However, I am currently using the Xavier NX dev kit to compare both devices. Yes I use the objectDetector_Yolo example with deepstream_app_config_yoloV3.txt

I tested with my Orin Nano 8GB board with DeepStream 6.2, the sample of /opt/nvidia/deepstream/deepstream-6.2/sources/objectDetector_Yolo/deepstream_app_config_yoloV3.txt performance is about 104 FPS with fakesink. Orin NX should be better than Orin Nano.

Please check your settings. You need to disable the clock gating with the script here : VPI - Vision Programming Interface: Performance Benchmark (nvidia.com)

I found the mistake. I moved objectDetector_Yolo example to another directory without recompiling it. Performance is now at 180fps with fakesink. Thanks

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