Kubernetes GPU labs
Fifteen labs from your first GPU pod to a fleet you can schedule, watch, share and price.
Kubernetes is the container orchestrator that decides which node runs which pod (a pod being one or more containers scheduled as a unit), and on a GPU cluster it is where most operational questions get answered: why a training pod sits in Pending, which inference replica gets evicted first under memory pressure, whether a job lands on the A100 pool or the L40S pool, and why an expensive fleet runs at 15 percent utilization. The NVIDIA GPU Operator is the piece that makes a GPU visible to Kubernetes at all; it installs the driver, the container runtime hooks and the device plugin that publish nvidia.com/gpu as a resource a pod can request. Cluster orchestration and job scheduling, GPU monitoring with DCGM and GPU virtualization with MIG make up the AI Operations domain of NVIDIA's NCA-AIIO, and the same ground returns with more depth in the Kubernetes administration, workload management and troubleshooting domains of NCP-AIO.
- A GPU pod scheduled and verified with nvidia-smi on your own isolated Kubernetes cluster
- A single-node k3s cluster running the NVIDIA GPU Operator, with Helm values, RBAC and ResourceQuota manifests and a validation and troubleshooting runbook
- Manifests that place workloads by requests and limits, nodeSelector, nodeAffinity, tolerations and PriorityClass across a multi-pool GPU cluster
- Four broken GPU pods diagnosed and fixed from kubectl describe, plus rolling, rollback and blue-green rollouts of a model version behind readiness probes
- A Prometheus /metrics endpoint for GPU telemetry, a GPU health watchdog with auto-remediation, MPS and MIG sharing artifacts, and a GPU cost audit priced in dollars
9 labs in a Kubernetes cluster environment in the browser, 6 on dedicated NVIDIA GPU pods. Nothing to install.
- Basic Linux shell and the idea of a container image
- What a pod and a node are; the labs introduce every other Kubernetes object (Deployment, PVC, PriorityClass) as it appears
- No prior GPU Operator, Helm or kubectl experience; the welcome lab is a 5-minute smoke test
How the path is sequenced
The collection opens with a five-minute smoke test (schedule a pod that requests a GPU and read the nvidia-smi output), then has you bring up a single-node k3s cluster with the GPU Operator by hand and walk the chain from Helm install to a workload-ready nvidia.com/gpu, so you can attribute node labels, capacity and RuntimeClass to the component that produced them. The scheduling stage covers requests and limits and what they decide about throttling, eviction and survival, then nodeSelector, nodeAffinity and tolerations across a multi-pool cluster, then PriorityClass and preemption when GPU capacity is full. The workload stage picks the right controller (Deployment, StatefulSet or DaemonSet), keeps training checkpoints on PersistentVolumeClaims, builds a GPU container pipeline with a fail-fast smoke test, ships a new model version by rolling update, rollback and blue-green, and closes with a triage day that fixes four GPU pods each broken at a different lifecycle stage. The last stage operates the fleet: GPU telemetry exposed to Prometheus, a health watchdog with auto-remediation, four ways to share one GPU (CUDA streams, time-slicing, MPS, MIG), and a cost audit that turns NVML samples into dollars of waste.
Nine labs run in a Kubernetes cluster environment in the browser, with kubectl against your own isolated cluster, and six run on dedicated NVIDIA GPU pods. Every lab is split into checkable steps and takes 35 to 50 minutes, apart from the 5-minute welcome lab.
Get a GPU onto a pod
Smoke-test the cluster, install the GPU Operator on k3s yourself, then trace how a Helm install becomes nvidia.com/gpu.
Schedule with intent
Requests and limits, GPU-type targeting with nodeSelector, affinity and tolerations, and PriorityClass when capacity runs out.
Run, ship and triage workloads
Pick the right controller, persist checkpoints, build and ship a GPU container, roll out a new model version, then fix four stuck pods.
Operate the fleet
GPU telemetry into Prometheus, a self-healing watchdog, four ways to share one GPU, and a cost audit in dollars.
Run all 15 labs in this path, and every other lab and practice test on Preporato.
- All 15 labs in this path
- Dedicated GPU pods included
- Every certification practice test
- Cancel any time
Frequently asked questions
No. The nine Kubernetes labs give you an isolated cluster environment in the browser with kubectl ready, and the six GPU labs provision a dedicated NVIDIA GPU pod for your session. There is nothing to install and no cloud account to connect.
The AI Operations domain of NVIDIA's NCA-AIIO (cluster orchestration and job scheduling, GPU monitoring with DCGM, GPU virtualization with MIG) is the closest match. The Kubernetes administration, workload management and troubleshooting domains of NCP-AIO cover the same ground in more depth, and NCP-AII touches MIG configuration and DCGM health monitoring.
About nine hours in total: a 5-minute welcome lab, then fourteen labs of 35 to 50 minutes each. Progress is saved per step, and the labs stand on their own, so you can go straight to stuck-pod triage or GPU sharing if that is the gap you need to close.
All fifteen labs are included in Preporato Pro ($29.99 per month or $290 per year), together with every practice test and every other lab on the platform. Each lab page shows the full brief and step list before you subscribe.