How to Increase FPS on Jetson Xavier NX for Real-Time AI & Video Processing?

[horiamommand]

Hello,

I am using a Jetson Xavier NX for real-time Object detection and tracking usually with Yolov8. My current setup includes:

YOLOv8n model converted to TensorRT

Torch, TorchVision, and ONNX installed with CUDA support

OpenCV compiled with CUDA support

Despite this, I am only achieving 12–15 FPS.

I would like to know:

How to properly increase FPS on the Jetson Xavier NX for real-time AI inference and video streaming.

Which settings, optimizations, or modes can maximize performance for TensorRT models on this platform.

Any best practices, recommended configurations, or system-level tweaks specific to Jetson Xavier NX.

waiting for your help.

[Y-T-G]

Did you convert the model to FP16?

[AastaLLL]

Hi,

Could you check GPU utilization with tegrastats?

$ sudo tegrastats

It’s common that the pre-processing or post-processing, which run on CPU, takes long time to finish.
To solve this, it’s recommended to run YOLO with Deepstream:

GitHub - marcoslucianops/DeepStream-Yolo: NVIDIA DeepStream SDK 8.0 / 7.1 / 7.0 / 6.4 / 6.3...

Thanks.

[horiamommand]

Thank you.
yes I did:
/usr/src/tensorrt/bin/trtexec --onnx=model.onnx --saveEngine=model.engine --fp16
now it is:

[horiamommand]

Thank you!

Okay, I’ll use DeepStream.

I used this command:

$ sudo tegrastats

Are there other ways to increase Jetson’s speed or FPS?

[AastaLLL]

Hi,

Do you run the inference at the same time?
The CPU and GPU loading shared above is less than 10% which is very low.

Thanks.