Introduction
Intel provides drivers, containers, and middleware to enable Intel GPUs on Linux. Intel® Open Middleware Xe (Intel® OMIX) has recently been added to simplify AI stack installation for Intel® Arc Pro B-Series Graphics.
Kernel mode drivers
The following kernel mode drivers provide support for Intel GPUs on Linux:
Xe KMD: for Intel GPUs based on Xe2 and newer architectures. Xe KMD is included in the upstream Linux kernel (6.11+) and provides native support for these GPUs.
i915 KMD: for GPUs based on Xe and earlier architectures. For Intel® Data Center GPU Max Series and Intel® Data Center GPU Flex Series, Intel provides out-of-tree i915 Dynamic Kernel Module Support (DKMS) drivers through the Intel GPU repository, with support for all hardware features of these GPUs.
Container images
To simplify deployment on supported host systems, Intel offers the following pre-configured container images as the fastest way to get started:
Intel® Open Middleware Xe (Intel® OMIX): AI stack container images with optimized deep learning frameworks and tools for Intel GPUs, enabling accelerated AI workloads.
PyTorch* on Intel® Hardware: Container images with PyTorch optimized for Intel GPU hardware, supporting both CPU and GPU acceleration.
See the release notes for each container for more information on supported GPUs and host operating system requirements. These containers are hosted at docker hub Intel repository.
Intel® Open Middleware Xe (Intel® OMIX)
Intel OMIX is an integrated compute stack for AI and general-purpose GPU workloads on Intel® Arc Pro B-Series Graphics. It includes the following compute runtime and oneAPI components. If you need to optimize or troubleshoot a specific component, refer to the documentation links provided for that component in the table.
Component |
Description |
Documentation |
|---|---|---|
Intel® oneAPI Level Zero |
A direct-to-metal programming interface for offloading accelerator devices. Its API can be tailored to device-specific needs and supports advanced language features such as function pointers, virtual functions, unified memory, and I/O capabilities. |
|
Intel® oneAPI DPC++/C++ Compiler |
A SYCL/DPC++ -based C++ compiler for developing high-performance applications targeting CPUs, GPUs, and other accelerators from a single codebase, built on LLVM. |
|
Intel® oneAPI DPC++ Library (oneDPL) |
A set of algorithms and utilities that complement the DPC++ compiler, providing parallel STL support and additional APIs for heterogeneous programming. |
|
Intel® Deep Neural Networks Library (oneDNN) |
An open-source, cross-platform library of optimized deep learning primitives for CPUs and GPUs, designed for use by AI frameworks and application developers. |
|
Intel® oneAPI Math Kernel Library (oneMKL) |
Optimized math primitives, such as BLAS, LAPACK, and FFT, accessible through a single portable API for CPUs and accelerators. |
|
Intel® Collective Communications Library (oneCCL) |
A library of optimized collective communication operations for distributed AI and HPC workloads, designed for easy integration into deep learning frameworks. |
|
Intel® Profiling Tools Interfaces for GPU |
A set of APIs and tools for profiling and tracing GPU workloads on Intel hardware, enabling performance analysis of applications using Level Zero and OpenCL. |
Intel OMIX is currently supported only on the following Intel® Arc Pro B-Series GPUs:
Intel® Arc™ Pro B50
Intel® Arc™ Pro B60
Intel® Arc™ Pro B65
Intel® Arc™ Pro B70