amd/xdna-driver

AMD XDNA Driver (amdxdna.ko) for Linux and XRT SHIM library

What it solves

This project provides the necessary Linux kernel drivers and libraries to enable support for AMD XDNA devices (NPUs) on Linux systems. It allows users to run AI applications by bridging the gap between the hardware and the Xilinx Runtime (XRT).

How it works

The project consists of two primary components: the amdxdna.ko kernel driver (mirrored from the Linux kernel staging tree) and the XRT SHIM library. Together, these components allow the operating system to communicate with the NPU hardware. The driver is typically installed via DKMS to ensure it remains compatible across kernel updates, and the SHIM library provides the interface for XRT to interact with the XDNA device.

Who it’s for

Developers and users with RyzenAI processors who want to run AI inference applications on Linux (specifically Ubuntu, Arch Linux, and Fedora).

Highlights

  • Broad Distro Support: Provides build and installation paths for Ubuntu, Arch Linux, and Fedora.
  • XRT Integration: Includes a plugin for the Xilinx Runtime (XRT) to enable NPU functionality.
  • Firmware Management: Packages matching firmware binaries to prevent version mismatches between the driver and the hardware.
  • Inference Focus: Specifically designed to accelerate machine learning inference on AMD NPUs.

Related

  • Project
    intel/linux-npu-driver

    A Linux kernel driver for the Intel® NPU, enabling energy-efficient AI inference acceleration on Intel® Core™ Ultra CPUs.

  • Project
    Xilinx/XRT

    A runtime software interface that provides an abstracted layer for managing and interacting with AMD NPUs and FPGAs across different hardware platforms.

  • Project
    amd/RyzenAI-SW

    A suite of tools and runtime libraries for optimizing and deploying AI inference on AMD Ryzen AI-based PCs to maximize hardware acceleration.

  • Project
    aws-neuron/aws-neuron-sdk

    AWS Neuron is an SDK for high-performance deep learning and generative AI, enabling the deployment and profiling of workloads on AWS Inferentia and Trainium accelerators.

  • Project
    ROCm/aiter

    AITER (AI Tensor Engine for ROCm) is AMD’s open‑source library of high‑performance GPU kernels for AI inference and training on ROCm GPUs. It offers C++/Python APIs, multiple back‑ends (Triton, CK, ASM, FlyDSL), and is the default kernel backend for LLM serving frameworks like vLLM and SGLang. The library supports CDNA3/4 GPUs (MI300X, MI350, MI355X) and experimental RDNA3/4 GPUs, delivering up to 17× speed‑up on key operators such as MLA decoding and multi‑head attention. Installation is via a recursive clone and an editable pip install, with optional system Triton usage. Releases are bi‑weekly and include many‑linux wheels for ROCm 7.x and Python 3.10/3.12.