🚀 JetsonForge

---

🧠 Introduction

Building software for NVIDIA Jetson devices often presents challenges.
Typically, developers must have physical access to a Jetson board and manage varying JetPack, L4T, and system library versions.
This can make it hard to maintain a consistent, reproducible build environment—especially when supporting multiple Jetson models.

JetsonForge simplifies this by providing scripts and tooling to create cross-compilation Docker images on a standard x86_64 machine.
You can now build Jetson software without owning a Jetson device, enabling faster development and CI/CD workflows.

---

⚙️ Prerequisites

Before you begin, ensure you have the following installed:

• Ubuntu 22.04 LTS (desktop or server)
• NVIDIA Developer Account — required to access JetPack/L4T downloads
• NVIDIA SDK Manager — for downloading JetPack and preparing the sysroot
• Docker — installed and configured
• ~60 GB+ of free disk space — for images, SDKs, and sysroot data

Optional (but recommended):

• Basic knowledge of Docker, cross-compilation, and CMake

⚠️ Legal Notice:
You must download JetPack and L4T files only via your official NVIDIA Developer Account.
JetsonForge does not distribute proprietary binaries.

---

🚀 Getting Started

Follow these steps to set up your Jetson cross-compile environment.

---

1. Install Prerequisites

Set up Ubuntu 22.04 and install:

sudo apt update
sudo apt install -y docker.io nvidia-sdk-manager

---

2. Download JetPack and L4T Files

List available JetPack versions:

sdkmanager --query non-interactive --action downloadonly

To include archived versions:

sdkmanager --query non-interactive --action downloadonly --archived-versions

Example: Download JetPack 6.0 (r36.3.0) for Jetson Orin Nano 8 GB:

cd ~
mkdir jetpack-r36.3.0
sdkmanager --cli --action downloadonly   --download-folder ~/jetpack-r36.3.0   --login-type devzone   --exit-on-finish   --license accept   --product Jetson   --archived-versions   --version 6.0   --target-os Linux   --host   --target JETSON_ORIN_NANO_TARGETS

💡 This downloads all default JetPack 6.0 components.
To select only specific packages, launch sdkmanager in interactive mode.

---

3. Install Required Packages

After downloading, install dependencies:

sudo apt update
sudo apt install -y bzip2 zstd xz-utils dpkg-dev qemu-user-static pv

---

4. Clone JetsonForge Repository

mkdir ~/workspace
git clone https://github.com/hackash/JetsonForge.git
cd JetsonForge/scripts
chmod +x make-sys-root.sh

Set environment variables and generate the sysroot:

export JETSON_FORGE_DL=/home/ubuntu/jetpack-r36.3.0
export JETSON_FORGE_WORK=/home/ubuntu/workspace 
export JETSON_FORGE_OUT=/home/ubuntu/workspace/jetpack-6.0-aarch64.tar.zst
./make-sys-root.sh

Once complete, copy the generated sysroot archive into JetsonForge:

cp /home/ubuntu/workspace/jetpack-6.0-aarch64.tar.zst JetsonForge/sysroots

---

5. Build the Base Docker Image

From the root of the JetsonForge repo:

cd JetsonForge
docker build --network host   --build-arg TAR_ZST_NAME=jetpack-6.0-aarch64.tar.zst   -t l4t-cross-base:r36.3.0   -f docker/x86-cross/Dockerfile .

🏷️ Rename as desired, e.g. yourorg/l4t-cross:r36.3.0.

The resulting image l4t-cross-base:r36.3.0 replicates a clean JetPack 6.0 environment and can be used to build software for Jetson targets.

---

6. Build and Test an Example Application

Build the example CUDA project:

docker build --network host   --build-arg L4T_CROSS_BASE=l4t-cross-base:r36.3.0   --build-arg TARGET_JETPACK_TAG=r36.3.0   -t l4t-cross-example-app:r36.3.0-latest   -f examples/cmake-cuda/Dockerfile examples/cmake-cuda

You can now run this image on your Jetson device:

docker run -it --rm --runtime nvidia l4t-cross-example-app:r36.3.0-latest

Expected output:

Detected 1 CUDA device(s)

Device 0: Orin
  Compute capability: 8.7
  Total global memory: 7621 MB
  Multiprocessors: 8
  Max threads per block: 1024
  Clock rate: 624 MHz

🎉 Success!
You’ve cross-compiled a CUDA app on an x86_64 host (no GPU required) and executed it natively on a Jetson board.

---

📦 What’s Inside

Path	Description
docker/x86-cross/Dockerfile	x86_64 build image with AArch64 toolchain & JetPack sysroot
toolchains/aarch64-jetson.cmake	CMake toolchain configuration for cross-compiling
examples/cmake-cuda/	Minimal CUDA example and Dockerfile
scripts/verify-sysroot.sh	Verifies CUDA/cuDNN/TensorRT compatibility in the sysroot
.github/workflows/build-cross.yml	Example GitHub Actions CI pipeline

---

⚠️ Notes and Gotchas

• Ensure your sysroot matches the exact JetPack/L4T version of your deployment device.
• If a build step executes target binaries, compile those utilities for x86 or use QEMU emulation.
• Keep sysroots version-controlled and reproducible to prevent inconsistencies across environments.

---

🪪 License

This project is licensed under the MIT License.

---

❤️ Contributing

Pull requests and improvements are always welcome!
If you find bugs or want to suggest enhancements, open an issue on GitHub.

---

📫 Support

For setup questions or contribution discussions, reach out via GitHub Issues or start a Discussion thread.

---

❤️ Happy Vibe Coding!