How to Enable AMD's Next-Gen AIE4 NPU on Linux with the AMDXDNA Driver

Introduction

AMD's upcoming AIE4 NPU (Neural Processing Unit) marks a significant leap in on-device AI acceleration, and the Linux community is already preparing for its arrival. The AMDXDNA driver, which currently supports first-generation NPUs, is being extended to handle this new hardware. This guide walks you through the process of enabling AIE4 NPU support on your Linux system, from verifying kernel compatibility to verifying the driver installation. Whether you're a developer testing early patches or an enthusiast wanting to be ready, these steps will help you set up the driver environment.

What You Need

• A compatible AMD Ryzen AI system – the exact product with AIE4 NPU is not yet released; this guide assumes you have a development board or an early sample.
• Linux kernel 6.12 or newer – the AMDXDNA driver patches require recent kernel features.
• Git to clone the AMDXDNA driver repository.
• Build tools – gcc, make, flex, bison, elfutils-libelf-devel, and openssl-devel (or equivalent for your distribution).
• Root access for loading kernel modules.
• Optional: a supported AI framework (e.g., PyTorch, TensorFlow) to test NPU acceleration.

Step-by-Step Guide

Step 1: Check Hardware and Kernel Version

Before diving in, verify that your system actually contains an AIE4 NPU. For now, only pre-release hardware exists. Run lspci | grep -i 'np\|accel' to look for AMD NPU devices. Also confirm your kernel version with uname -r. If it's below 6.12, you'll need to upgrade (see Step 2).

Step 2: Update the Linux Kernel

The AMDXDNA driver patches for AIE4 rely on kernel interfaces introduced in version 6.12. Use your distribution's method to update, or compile a mainline kernel. For example, on Ubuntu:

1. Add the mainline PPA (if using Ubuntu): sudo add-apt-repository ppa:cappelikan/ppa
2. Install the latest kernel: sudo apt update && sudo apt install linux-generic
3. Reboot and select the new kernel from GRUB.

For manual compilation, download the kernel source from kernel.org, configure with make olddefconfig, build, and install. Ensure CONFIG_AMD_XDNA is enabled.

Step 3: Obtain the AMDXDNA Driver Source

AMD maintains the driver on GitHub. Clone the repository with AIE4 patches:

git clone --branch aie4-next https://github.com/amd/amdxdna.git
cd amdxdna

The aie4-next branch contains the latest enablement patches as of early 2025. If you're reading this later, check for stable branches.

Step 4: Build and Install the Driver

Compile the driver as a kernel module:

1. Prepare the environment: make kernel_prepare (or manually copy kernel headers).
2. Build the module: make
3. Install: sudo make install
4. Update module dependencies: sudo depmod -a

If you encounter errors, ensure you have the build tools listed in What You Need.

Step 5: Load the Module and Test

Load the driver module with:

sudo modprobe amdxdna

Check the kernel log (dmesg | tail -20) for AIE4 NPU recognition. You should see lines containing AIE4 or NPU. Then list the device:

ls /dev/amdxdna*

If a device appears, the driver is working. To make it load automatically at boot, add amdxdna to /etc/modules.

Step 6: Verify with a Sample AI Workload

Use the provided test utilities in the driver source:

cd amdxdna/utests
make
./run_all_tests.sh

Or run a simple inference from a supported framework. For PyTorch, ensure the torch.xdna backend is installed. A minimal test:

import torch
model = torch.nn.Linear(10,5)
x = torch.randn(1,10)
with torch.xdna.device(0):
    y = model(x)
print(y)

If no errors, your AIE4 NPU is fully enabled.

Tips and Troubleshooting

• Kernel version matters: If you use a distribution kernel older than 6.12, backporting patches is possible but advanced. Stick to the mainline kernel for simplicity.
• Check firmware: Some AIE4 NPUs require loading a firmware blob. The AMDXDNA driver may include a firmware loader; ensure your system has the correct firmware files from the /lib/firmware/amdxdna/ directory.
• Power management: The AIE4 NPU may be power-gated by default. You might need to enable it via echo on > /sys/bus/pci/devices/0000:00:08.0/power/control (replace the PCI address with your device's).
• Monitor driver updates: AMD continues to push improvements to the aie4-next branch. Regularly pull the latest commits with git pull and rebuild the module.
• Don't forget to unload older drivers: If you have a previous version loaded, run sudo rmmod amdxdna before loading the new one.
• Community help: Join the AMD Linux mailing list or the #amdgpu IRC channel for real-time assistance.

By following these steps, you'll be prepared to leverage the AIE4 NPU's performance for on-device AI inference as soon as the hardware becomes commercially available. The Linux enablement is progressing steadily, and your early adoption helps refine the driver for everyone.