Based on RHODS-3074:

Initial scope for RHODS-2387 indicated that taints should be added to individual nodes that had GPUs. However, in testing this, Luca found that node taints could not be changed after node creation (RHODS-3184) [possibly only in the default machine pool], so we switched to the current method of adding GPU nodes in a separate machine pool that had the nvidia.com/gpu taint applied.

We do not currently have instructions for users that have GPU and CPU nodes in the same pool.

According to RHODS-3074 comments, there should be two options for adding GPUs to your cluster:

1. Enabling GPU support for nodes in a new machine pool (Recommended, current process in RHODS-3235)

• appropriate when not all nodes have GPUs, and you want to schedule GPU workloads to GPU nodes, and CPU workloads to CPU nodes
• add nvidia.com/gpu NoSchedule taint to the whole GPU machine pool

-2. Enabling GPU support for nodes in an existing machine pool

• appropriate when all nodes have GPUs, and must be shared between both GPU and CPU workloads
• no taints applied because all nodes have GPUs, and applying a taint would block CPU-only workloads-
  Not applicable based on discussion with Landon, Erwan, Luca.

Some additional troubleshooting information in case of typos when setting the taint may also be useful