How do I deploy YOLO on an NVIDIA Jetson?

1. Flash JetPack

Start from a matching JetPack release, which bundles CUDA, cuDNN, and TensorRT for the board. Getting the versions right here saves a lot of pain later.

2. Install Ultralytics

Install the Ultralytics package on the device. Use the versions of PyTorch and torchvision built for your JetPack, not the generic desktop wheels.

3. Export to TensorRT on the device

Build the TensorRT engine on the same Jetson you will run it on, engines are hardware- and version-specific and are not portable between boards.

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="engine", half=True)   # builds a TensorRT engine (FP16)

trt = YOLO("yolo11n.engine")
trt.predict("video.mp4")

4. Squeeze out performance

• Use FP16 by default; try INT8 with a calibration set if you need more speed.
• Set the board to max power mode (e.g. nvpmodel + jetson_clocks).
• Prefer a small model (yolo11n/s) and the lowest imgsz that still works.
• For many camera streams, consider NVIDIA DeepStream for hardware-accelerated decode and inference.

5. Measure on the real feed

A Jetson Orin Nano and an AGX Orin are very different machines. Benchmark end-to-end FPS on your actual camera and resolution, including decode, so the number reflects production rather than a lab loop.

Hands-on tutorials