Compute and containers: EC2, Lambda, ECR/ECS/EKS
Lesson 3 of 4 in The Supporting Cast: Data, Compute, and Pipes.
Most of the time SageMaker AI and Bedrock hide the compute, but the exam still expects you to know the raw materials. On EC2, learn the GPU families conceptually (chip specs are not tested): the P family carries training-grade GPUs — large memory, fast interconnects for multi-GPU and multi-node jobs; the G family leans toward inference and lighter training at friendlier cost; and AWS's own silicon splits by direction — Trn (Trainium) instances for training, Inf (Inferentia) instances for inference. When SageMaker runs a job or an endpoint, these same families are what it manages for you.
| Service | What it is | When ML touches it |
|---|---|---|
Amazon EC2 | Raw instances, including the GPU families (P = training, G = inference-leaning, Trn/Inf = AWS silicon) | Self-managed training or serving when you need total control — and the substrate SageMaker manages under its jobs and endpoints |
AWS Lambda | Serverless functions, scaling to zero | Event glue: light preprocessing, invoking endpoints, small CPU inference — bounded by its timeout and memory ceilings, and no GPUs |
AWS Batch | Managed batch job queues over containers | Scheduled, queue-driven jobs — feature builds, scoring sweeps, even GPU jobs — without owning a cluster |
Amazon ECR | The container registry | Where SageMaker pulls training and inference images; custom containers live here |
Amazon ECS | AWS-native container orchestration (including serverless Fargate capacity) | Serving lightweight model APIs and the microservices around a model |
Amazon EKS | Managed Kubernetes | Teams running their ML platform on Kubernetes (Kubeflow, Ray) or operating large self-managed GPU fleets |
The container thread ties it together: even fully managed SageMaker training runs a container — one of AWS's built-in images or your own pushed to ECR. That is why ECR appears in answers about custom frameworks: the "bring your own" path on SageMaker means bring your own image.
Interactive sorting exercise: Sort each scenario onto the compute service the exam wants.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.