Step 2: Install GPU Kernel Driver and Intel® Graphics Compute Runtime for GPU Accelerator Support
To use Intel® oneAPI tools with Intel GPUs, you need to install the following components:
Kernel mode driver (KMD) consisting of the i915 kernel driver and the related firmware. The KMD component is a kernel driver of the Linux* kernel. To support recently released GPU products, an updated version of the KMD may have to be installed. This needs to be done on each node separately or embedded as a part of the compute node OS image.
If a Linux kernel provided by the Linux distribution does not have all the desired features enabled in the i915 driver yet, use more recent Linux kernel from kernel.org.
Intel Graphics Compute Runtime environment, you can install it using one of the following methods:
Local installation: install the runtime environment on each compute node or into the compute node OS image
Shared installation (recommended): install the runtime environment on a shared file system. This method is preferred on a cluster. It limits the changes needed to the OS node image, reduces the size of the compute node image, and enables multiple versions on the runtime environment to be made available in parallel.
Local Installation of the Intel Graphics Compute Runtime Environment
The process for installing the Intel Graphics Compute Runtime environment on a local file system is described in Intel GPU Drivers Installation Guides. Use this process for standalone nodes, as it installs the runtime components into directories that are used for OS components, for example, /etc and subdirectories of /usr.
Shared Installation of the Intel Graphics Compute Runtime Environment
To install the Intel Graphics Compute Runtime Environment components to a location different from the one used for OS components, for example, to a shared network file system, follow the steps below:
Get the installation packages and their dependencies either through direct download using wget or by downloading them from Intel online repositories via Linux* package managers. The download method depends on the tools and the Linux distribution used.
The following examples demonstrate how a given <package> can be downloaded from the online repositories to a <local_dir> without installing it. Before using the commands below, make sure that the corresponding repository environment is set up correctly.
RHEL:
dnf install --downloadonly --downloaddir=<local_dir> <package>
SLES:
zypper --pkg-cache-dir download <local_dir> <package>
Ubuntu:
cd <local_dir && apt-get download $(apt-cache depends \ --recurse --no-recommends --no-suggests --no-conflicts --no-breaks --no-replaces \ --no-enhances <package> | grep -E '^[a-zA-Z0-9]')
Unpack the packages and their dependencies. Unpacking is used to avoid installing packages to OS directories and to keep copies of multiple versions of the same software available on the system. The method of unpacking depends on the tools and the Linux distribution used. The following examples demonstrate how to unpack a given <package> into <install_dir> without installing it with the package manager:
RHEL/SLES (.rpm):
mkdir -p <install_dir> && cd <install_dir> && rpm2cpio \ path_to_package.rpm> | cpio -idm --no-absolute-filenames
Ubuntu (.deb):
mkdir -p <install_dir> && dpkg -x <path_to_package.deb> <install_dir>/