Module 2 · Lab 1: a GPU-ready EKS cluster¶

Format: hands-on lab Time: 50 min (incl. 5 min buffer) Cost: one GPU node, ≈ $0.35–1.00/hour depending on Spot

Goal: a GPU node running, the NVIDIA GPU Operator installed correctly for EKS, and your first CUDA container printing nvidia-smi output.

You start with a cluster that has never seen a GPU. You'll add each layer of the stack yourself, in the order the request flows through them:

flowchart LR
    A["Pod gpu-smoke<br/>nvidia.com/gpu: 1<br/><i>Pending</i>"] --> B[Karpenter]
    B -->|"matches"| C["NodePool gpu<br/>+ EC2NodeClass gpu"]
    C -->|"launches"| D["EC2 g5 / g6<br/>AL2023 NVIDIA AMI"]
    D -->|"node joins"| E["GPU Operator DaemonSets<br/>device plugin · GFD · DCGM"]
    E -->|"registers"| F["allocatable<br/>nvidia.com/gpu: 1"]
    F -->|"scheduler binds"| G["gpu-smoke <i>Running</i><br/>→ nvidia-smi"]

Steps¶

Step	What you do	Layer
2.1 Verify the base stack	Confirm Karpenter, Prometheus and KEDA are up and there are no GPU nodes	—
2.2 Create GPU capacity	Pin an AMI, write an `EC2NodeClass` and a `NodePool` for g5/g6 Spot with On-Demand fallback, taint the nodes	hardware, driver, toolkit
2.3 Install the GPU Operator	Helm install with `driver.enabled=false` — and understand why	device plugin, GFD, DCGM
2.4 Your first GPU pod	Deploy a CUDA pod, watch Karpenter launch the node, read `nvidia-smi`	scheduler, pod
2.5 Checkpoint	Validate, screenshot, take a break

Hard gate

Nobody goes to break without nvidia-smi output from their own cluster on screen. If you're stuck, raise your hand (or drop a message in the chat) early — the helpers are here for exactly this.

Everything you apply in this lab lives under modules/01-gpu-nodes/ and modules/02-gpu-operator/ in the repo.

Start with 2.1 →