GPU nodes setup

You can add the GPU hardware to the existing or new ECS or OCP cluster as a worker node.

For information about GPU hardware requirements, see Additional resource requirements for CDE.

You must install nvidia-container-toolkit on the worker node. For more on nvidia-container-runtime migration to nvidia-container-toolkit, see Migration Notice. For information about the installation, NVIDIA Installation Guide. If using Red Hat Enterprise Linux (RHEL), use dnf to install the package. For an example with RHEL 8.8, see Installing the NVIDIA Container Toolkit.

You can use following options to advertise the GPUs in the Kubernetes cluster:

  • Nvidia device plugin: In ECS installation, if the Nvidia drivers are correctly installed, the Nvidia-device-plugin automatically advertises the GPU resource to the scheduler. Platform administrator need not deploy the Nvidia device plugin.

  • Node Feature Discovery Operator (NFD) and GPU Operator: OCP administrators must install NFD and GPU Operator for advertising the GPU resource to the Kubernetes scheduler.

If the Nvidia drivers are correctly installed, the above options should advertise the GPU resource to the scheduler. For more information, see NVIDIA Device Plugin documentation.