Nutanix NCP-AI: Nutanix Kubernetes Platform and Containerized AI Workloads

Nutanix Kubernetes Platform and Containerized AI Workloads

NKP, formerly Nutanix Kubernetes Engine, provides lifecycle management for CNCF-conformant Kubernetes clusters from Prism Central. The exam focuses on GPU node pools, storage integration, and workload isolation.

Core rules

• Create clusters with control-plane and worker node pools; add GPU node pools on AHV hosts using passthrough or vGPU.
• Install the NVIDIA GPU Operator so the device plugin advertises nvidia.com/gpu.
• Request GPUs explicitly in pod specs; without a resource request, the scheduler will not reserve a GPU.
• Use taints/tolerations and node affinity to keep GPU nodes for GPU workloads.
• Integrate NDK CSI storage classes for persistent volumes; use namespaces, RBAC, and quotas for multi-team sharing.
• Expose inference through Service/Ingress; keep model artifacts in Objects or Files, not images.

Worked example

Deploy an inference pod needing one GPU: set resources.limits: nvidia.com/gpu: 1, add a toleration for the GPU node pool taint, mount an NDK persistent volume for the model cache, and set a readiness probe. The device plugin advertises whole GPUs, GPU limits equal requests, and scheduling needs both the resource request and the toleration.

Traps

• Requesting nvidia.com/gpu without the GPU Operator installed.
• Forgetting that GPU requests are integers and not overcommittable.
• Using the default storage class when NDK classes are required.
• Mixing teams in one namespace without quotas or RBAC.