jetson-deploy

name	jetson-deploy
description	Deploying and benchmarking on Jetson Orin Nano. Use when deploying code to Jetson, running TensorRT inference, monitoring GPU usage, or benchmarking performance.

Jetson Orin Nano Deployment

Connecting to Jetson

ssh jetson-nano   # Via Tailscale
# or
ssh charlie@100.95.16.119

The Jetson hostname is charlie-jetson-orin-nano.

Device Specs

• Model: Jetson Orin Nano (8GB)
• JetPack: 6 (L4T R36.4.4)
• CUDA: 12.6
• TensorRT: 10.x
• Python: 3.10 (system)
• RAM: 8GB shared CPU/GPU

Deploying Code

# From Mac: deploy latest code
make deploy-jetson

# This runs: ssh jetson-nano 'cd ~/couch-vision && git pull'

The repo is cloned at ~/couch-vision/ on the Jetson.

Running the Perception Stack

ssh jetson-nano
cd ~/couch-vision

# Run with bag file
make full-stack BAG=bags/walk_around_university_all_data.mcap

# First run is slow — TRT engines auto-export (~10 min each)
# Subsequent runs start in seconds

TensorRT Engine Auto-Export

On first CUDA run, TensorRT engines are built automatically:

Model	Export Time	Engine Size
YOLOv8n	~10 min	~9 MB (INT8)
YOLOP	~8 min	~13 MB (FP16)

Engines are saved to perception/weights/ (volume-mounted). Delete *.engine to force re-export.

Monitoring Performance

tegrastats (GPU/Memory/Power)

# Real-time monitoring
tegrastats --interval 1000

# Log to file
tegrastats --interval 1000 > /tmp/tegrastats.log &

Output format:

RAM 2100/7620MB (lfb 1x4MB) SWAP 0/3810MB CPU [38%@1510,37%@1510,...]
GR3D_FREQ 99%@624 VIC_FREQ 0%@115 APE 174 CV0@47.3C CPU@48.8C SOC2@46.1C
SOC0@45.9C CV1@46.5C GPU@46.7C tj@48.8C SOC1@47.5C CV2@46.5C VDD_IN 5939mW
VDD_CPU_GPU_CV 2016mW VDD_SOC 1612mW

Key metrics:

• RAM — Memory usage
• GR3D_FREQ 99% — GPU utilization (target: >90% during inference)
• CPU [38%@1510] — CPU utilization per core
• tj@48.8C — Junction temperature (throttles at 97°C)
• VDD_IN 5939mW — Total power draw

jtop (interactive)

sudo jtop

Interactive TUI showing GPU, memory, power, temperature. Press q to quit.

Benchmarking

Quick benchmark

ssh jetson-nano
cd ~/couch-vision

# Start tegrastats in background
tegrastats --interval 1000 > /tmp/tegrastats.log &

# Run perception for 90 seconds
timeout 90 make full-stack BAG=bags/walk_around_university_all_data.mcap

# Check logs for FPS
docker compose -f perception/docker-compose.nav2.yml logs | grep "FPS"

# Check tegrastats
tail -20 /tmp/tegrastats.log

# Stop tegrastats
pkill tegrastats

Expected Performance

Metric	Value
Perception FPS	12-17
GPU utilization	~99%
RAM usage	~2.1 / 7.6 GB
Power draw	~6W total
Temperature	~47°C

Docker on Jetson

Jetson uses nvidia container runtime:

# Verify nvidia runtime
docker info | grep -i runtime

# Run with GPU access (automatic via docker-compose.yml)
docker run --runtime=nvidia ...

The docker-compose.nav2.yml automatically uses the nvidia runtime on Jetson (detected via platform).

Common Issues

Permission denied on Docker socket

sudo usermod -aG docker $USER
# Log out and back in

TensorRT engine incompatible

Delete engines and re-export:

rm perception/weights/*.engine
make full-stack BAG=...  # Will re-export

Out of memory

• Close other applications
• Reduce batch size (already 1)
• Use fast.yaml config (skips YOLOP)

Slow first run

Expected — TRT engine export takes ~10 min per model. Subsequent runs are fast.

Can't SSH to Jetson

Check Tailscale:

tailscale status
# Jetson should show as "charlie-jetson-orin-nano"

File Locations on Jetson

Path	Contents
~/couch-vision/	This repo
~/ros2_jazzy/	ROS2 Jazzy (built from source)
/tmp/tegrastats.log	Performance logs
perception/weights/	Model weights + TRT engines

Power Modes

Jetson Orin Nano supports different power modes:

# Check current mode
sudo nvpmodel -q

# Set to max performance (15W)
sudo nvpmodel -m 0
sudo jetson_clocks

# Set to power-saving (7W)
sudo nvpmodel -m 1

For benchmarking, use max performance mode.