NVIDIA GPU Operator

false for version 25.3.x.

true for version 25.10.x and later.

false for version 25.3.x.

Deprecated in version 25.10.x and later.

Controls the NVIDIA device plugin that exposes GPUs to Kubernetes.

The device plugin runs on every GPU node and is responsible for:

• Advertising nvidia.com/gpu resources
• Enabling GPU scheduling

The configuration directly affects scheduler decisions and GPU workload placement.

If this argument is misconfigured:

• GPUs might not be visible to Kubernetes.
• Pods requesting GPUs might remain in the Pending state.
• GPU availability might be inconsistent across nodes.
• Scheduling failures might occur at scale.

• Set devicePlugin.enabled to true .
• Allocate sufficient resources by using devicePlugin.resources .
• Use a consistent configuration across all GPU nodes.

Controls the NVIDIA Container Toolkit, which provides GPU access inside containers.

The toolkit runs on every GPU node and is required for GPU workloads to access GPUs inside containers.

If this argument is misconfigured, pods might be scheduled successfully but:

• GPU workloads might fail to start.
• Containers might not be able to access GPUs.
• Runtime failures might be difficult to diagnose.

• Set toolkit.enabled to true .
• Ensure that the toolkit is compatible with the container runtime.
• Allocate sufficient resources by using toolkit.resources .

Controls whether GPU Operator components can run on GPU nodes.

GPU nodes are often tainted. All GPU Operator DaemonSets must tolerate those taints to run on the GPU nodes.

If the required tolerations are missing:

• GPU Operator components cannot run on GPU nodes.
• GPUs might become unusable throughout the cluster.

• Configure tolerations that match the taints applied to GPU nodes.
• Validate the tolerations for all GPU node pools.

Controls how many nodes are updated simultaneously.

The value affects the number of GPU nodes that are temporarily unavailable during an update.

In a large cluster, an update might affect many nodes.

If the value is too high:

• Too many GPU nodes might be unavailable at the same time.
• GPU workloads might be disrupted.

If the value is too low, rolling out the update across thousands of nodes might take a long time.

Use a balanced value that is not overly aggressive. Set the value based on the cluster size and the amount of disruption that workloads can tolerate.

Defines the CPU and memory resources for the GPU Operator controller.

The operator manages components including:

• The device plugin
• The NVIDIA Container Toolkit
• Multi-Instance GPU
• NVIDIA Data Center GPU Manager

As the number of nodes increases, the operator manages more objects and requires more resources.

If the resources are undersized:

• Reconciliation might be slow.
• The operator might become unstable.
• Configuration changes and component rollouts might be delayed.

• Increase CPU and memory resources for large clusters.
• Monitor operator logs and memory usage.

Enables the Container Device Interface.

• Simplifies injecting GPUs into containers.
• Improves runtime consistency across nodes.

If the Container Device Interface is disabled, the older runtime integration might:

• Cause inconsistencies.
• Make troubleshooting more complex.

• Set cdi.enabled to true .
• Use the default settings unless you have a specific requirement.

Controls Multi-Instance GPU partitioning.

Multi-Instance GPU partitioning enables workloads to share GPUs and affects scheduler behavior.

If this argument is misconfigured:

• GPU resources might become fragmented.
• Pods might not find a matching GPU partition.
• Behavior might be inconsistent across nodes.

• Use a consistent Multi-Instance GPU strategy throughout the cluster.
• Avoid mixed configurations unless they are required.

Controls GPU metrics collection.

The exporter generates GPU metrics for each node, increasing Prometheus load and network traffic as the number of GPU nodes increases.

If this argument is misconfigured:

• Scraping metrics too frequently might cause high overhead.
• Scraping metrics too infrequently might provide insufficient visibility.

• Tune serviceMonitor.interval .
• Allocate sufficient resources by using dcgmExporter.resources .

Controls the NVIDIA Data Center GPU Manager host engine.

The optional component tracks GPU health across the cluster.

• If the component is disabled, visibility into GPU health is limited.
• If the component is enabled with insufficient resources, it might become unstable.

• Enable the component only when it is required.
• Allocate sufficient resources.

Controls the dependency check for Node Feature Discovery.

The dependency check affects node labeling and GPU detection.

If Node Feature Discovery is missing:

• GPU nodes might not be labeled.
• The operator might fail to detect GPU nodes.

• Use this argument only when Node Feature Discovery is already installed.
• Ensure that the required node labels are present and correct.

Disables the standalone NVIDIA GPU device plugin.

Disabling the standalone plugin prevents duplicate GPU device plugins from running in the cluster.

If this argument is configured incorrectly:

• Multiple GPU device plugins might run.
• The plugins might conflict.
• GPU resources might be reported incorrectly.

Disable the standalone NVIDIA GPU device plugin when using the NVIDIA GPU Operator.