For platform engineers who already run Kubernetes and leave able to operate GPU clusters for training, batch and inference with fair sharing, gang scheduling and RDMA networking.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
Kubernetes was designed for stateless services that scale one pod at a time. AI work is different: a training job needs 64 GPUs to start together or not at all, an inference replica may span two nodes, GPUs are too expensive to leave reserved but idle, and the fast network is RDMA that the default pod network knows nothing about. This course teaches the components and patterns that close that gap.
You begin with the GPU node: the NVIDIA GPU Operator and the AMD GPU Operator, the device plugin, and Dynamic Resource Allocation, whose core APIs became generally available in Kubernetes 1.34. You then share GPUs with MIG and time-slicing, attach InfiniBand or RoCE interfaces with the Network Operator and Multus, and schedule batch work with Kueue quotas, JobSet, the Kubeflow Trainer and NVIDIA's open-source KAI Scheduler or Volcano. Later modules cover multi-node inference with LeaderWorkerSet, KServe and the Gateway API Inference Extension, Ray on KubeRay, storage through CSI drivers, GPU telemetry with the DCGM exporter, and how managed EKS, GKE and AKS differ. Facts are stated as of October 2026.
What you will be able to do
Install and upgrade the GPU Operator and Network Operator, and explain each component they deploy
Choose between the device plugin and Dynamic Resource Allocation for exposing GPUs, and write ResourceClaims for a given request
Configure MIG and time-slicing on Kubernetes and justify the choice for a stated mix of workloads
Attach RDMA-capable secondary networks to pods and verify GPUDirect RDMA from inside a pod
Implement team quotas, borrowing and gang scheduling with Kueue, JobSet and a GPU-aware scheduler
Deploy distributed training and multi-node inference with Kubeflow Trainer, LeaderWorkerSet and KServe
Monitor GPU utilization per namespace with the DCGM exporter and Prometheus and act on idle allocations
Syllabus
6 modules · 19 lessons · about 14 hours
1
Module 1: Making a Node GPU-Ready
The operators, drivers and device APIs that turn a Kubernetes node into a schedulable GPU node.3 lessons
Inside the NVIDIA GPU Operator: Driver, Toolkit, Device Plugin and Validators
Device Plugins Versus Dynamic Resource Allocation: ResourceClaims and DeviceClasses
AMD GPUs on Kubernetes: The AMD GPU Operator and Device Plugin
2
Module 2: Sharing GPUs and Wiring the Fast Network
Splitting GPUs among pods and giving pods direct access to the RDMA fabric.3 lessons
MIG and Time-Slicing Strategies on Kubernetes
The Network Operator, Multus and SR-IOV for InfiniBand and RoCE
Verifying GPUDirect RDMA and NCCL Performance from Inside a Pod
3
Module 3: Batch Scheduling and Fair Sharing
Queues, quotas and all-or-nothing scheduling for jobs that need many GPUs at once.4 lessons
Why the Default Scheduler Fails AI Jobs: Partial Starts and Deadlocks
Kueue: ClusterQueues, Cohorts, Borrowing and Preemption
Gang Scheduling with JobSet, KAI Scheduler and Volcano
Topology-Aware Placement for NVLink Domains and Network Locality
4
Module 4: Training on Kubernetes
Running distributed PyTorch and Ray workloads with checkpoints and failure recovery.3 lessons
Distributed PyTorch with the Kubeflow Trainer and torchrun
Ray on Kubernetes with KubeRay
Storage for Jobs: CSI Drivers, Parallel File Systems and Checkpoint Volumes
5
Module 5: Inference on Kubernetes
Serving models that span GPUs and nodes, and routing requests to them intelligently.3 lessons
Multi-Node Inference with LeaderWorkerSet
KServe and the Gateway API Inference Extension
Autoscaling GPU Inference: HPA, KEDA and Cluster Autoscaling Limits
6
Module 6: Operating a Shared GPU Platform
Telemetry, cost attribution and the differences between managed Kubernetes offerings.3 lessons
GPU Observability: DCGM Exporter, Prometheus and Per-Namespace Utilization
Managed GPU Kubernetes: EKS, GKE and AKS Compared
Upgrades, Node Health and Draining GPU Nodes Safely
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Kubernetes, NVIDIA, AMD, Kueue, Kubeflow, Volcano, Ray, KServe, Amazon EKS, Google Kubernetes Engine, Azure Kubernetes Service, or Prometheus. Product and company names are trademarks of their respective owners.
A plain-language tour of the stack that runs AI: what a GPU server is, why clusters need special networks and storage, how jobs are scheduled, what an AI data center needs in power and cooling, and how cloud and on-premises options differ. No prior infrastructure background required.
The operator's view of data-center GPUs: what the current NVIDIA and AMD systems are, how they are packaged and connected, how the driver, CUDA/ROCm and container stack fit together, how to monitor and triage failing GPUs, how to share and size GPU memory, and what power, cooling and sourcing decisions follow.
The Linux skills that differ on an AI cluster: choosing and imaging the OS, packaging GPU and RDMA drivers that survive kernel updates, NUMA and PCIe topology tuning, cgroup v2 isolation, shared identity and file systems, scientific software environments with Lmod, Spack and containers, and patching a fleet without breaking jobs.