🖥️ EdgeBox AI¶

In Stock Edge AI Computer

AI inference where it matters — at the machine. The EdgeBox AI is powered by the NVIDIA Jetson Orin NX, delivering up to 100 TOPS of AI performance in a compact, fanless enclosure. Run real-time object detection, quality inspection, and anomaly detection without sending data to the cloud.

Quick Start¶

⚡ Running your first AI model in 20 minutes

1. Flash JetPack OS to the NVMe SSD 2. Connect power (12V DC), HDMI, and USB keyboard 3. Complete the NVIDIA initial setup wizard 4. Run the included ElephantAI demo (object detection on USB webcam)

Technical Specifications¶

AI & Compute
Module	NVIDIA Jetson Orin NX 16GB
AI Performance	100 TOPS
GPU	1024-core NVIDIA Ampere GPU
CPU	8-core ARM Cortex-A78AE @ 2.0 GHz
RAM	16 GB LPDDR5 (unified memory)
Storage
NVMe Slot	M.2 2280 PCIe Gen 4 (up to 2TB)
eMMC	64 GB (OS)
I/O
USB	4 × USB 3.2 Gen 2 (10 Gbps)
Ethernet	2 × 2.5GbE
Video In	2 × MIPI CSI-2 (4K camera)
Video Out	1 × HDMI 2.1, 1 × DP 1.4
Serial	1 × RS485, 1 × RS232
Power & Physical
Input Voltage	12V DC (5.5/2.5mm barrel)
Typical Power	15–35W (AI workload dependent)
Dimensions	160 × 110 × 55 mm
Cooling	Passive aluminium heatsink (fanless)
Operating Temperature	-10°C to +55°C

OS Setup¶

Flash JetPack¶

The EdgeBox AI uses NVIDIA JetPack 6.0 (Ubuntu 22.04 based):

# Use NVIDIA SDK Manager on a Linux host PC
# https://developer.nvidia.com/sdk-manager

# Or flash via USB with the included recovery cable
sudo ./flash.sh jetson-orin-nx-16 mmcblk0p1

JetPack version

Always use JetPack 6.0 or later for Orin NX support. Earlier versions do not support the Orin architecture.

AI Inference¶

Run the ElephantAI demo¶

# Clone the demo repo
git clone https://github.com/elephantronics/edgebox-ai-demo
cd edgebox-ai-demo

# Install dependencies
pip3 install -r requirements.txt

# Run real-time object detection (YOLOv8)
python3 detect.py --source /dev/video0 --model yolov8n.engine

Convert your own model to TensorRT¶

# Install torch2trt
pip3 install torch2trt

# Convert ONNX → TensorRT engine
trtexec --onnx=your_model.onnx \
        --saveEngine=your_model.engine \
        --fp16 \
        --workspace=4096

Benchmark inference speed¶

Model	Precision	FPS (EdgeBox AI)
YOLOv8n	FP16	~210 fps
YOLOv8m	FP16	~85 fps
YOLOv8x	FP16	~28 fps
ResNet-50	INT8	~480 fps

Camera Input¶

USB CameraMIPI CSI Camera

Any UVC-compatible USB camera works out of the box:

import cv2
cap = cv2.VideoCapture(0)  # /dev/video0
ret, frame = cap.read()

Connect a compatible MIPI camera (e.g. IMX477) to CSI port:

# GStreamer pipeline for CSI camera
pipeline = (
    "nvarguscamerasrc ! "
    "video/x-raw(memory:NVMM),width=1920,height=1080,framerate=30/1 ! "
    "nvvidconv ! video/x-raw,format=BGRx ! videoconvert ! appsink"
)
cap = cv2.VideoCapture(pipeline, cv2.CAP_GSTREAMER)

Troubleshooting¶

Issue	Fix
High temperature warning	Check ambient temperature; ensure heatsink fins are unobstructed
TensorRT conversion fails	Ensure JetPack and TensorRT versions match; clear `/tmp/` cache
Camera not detected	Check USB port (use USB 3.2 port); try `v4l2-ctl --list-devices`