| Data storage | Amazon S3 | - Practically unlimited, highly durable storage for datasets and model files
- SageMaker, Athena and Glue read straight from S3, so you train without copying data
- Versioning keeps old dataset versions; lifecycle rules move cold data to cheaper tiers
|
| Notebooks | SageMaker Studio notebooks | - Managed JupyterLab with nothing to install
- Switch the instance from CPU to GPU when you need more power
- Built-in access to S3 data, experiments and the model registry
|
| GPU training | SageMaker Training on ml.g5.xlarge (A10G) / ml.p4d for large jobs | - Billed only while the training job runs; machines shut down automatically
- Managed Spot Training can cut training costs dramatically
- Scales from one GPU to distributed multi-GPU training
|
| ML platform & registry | Amazon SageMaker (experiments, model registry) | - One place for experiments, the model registry, approvals and deployment
- Lineage shows which data and code produced each model
- Access controlled with IAM roles
|
| Serverless serving | AWS Lambda (container image) or SageMaker Serverless Inference | - Scales to zero, so you pay nothing when nobody is using the model
- Handles spiky traffic automatically
- No servers to patch or manage
|
| GPU serving | SageMaker real-time endpoint on ml.g5 / Inferentia2 | - Low-latency real-time predictions with autoscaling
- Inferentia2 chips can lower the cost per prediction
- Production variants let you A/B test two models
|
| Batch predictions | SageMaker Batch Transform, scheduled by EventBridge | - Scores millions of rows without keeping a server running
- Reads input from S3 and writes results back to S3
- You pay only for the job's duration
|
| Pipelines | SageMaker Pipelines or Step Functions | - Repeatable train → evaluate → register workflows
- Conditional steps, e.g. only deploy if accuracy beats a threshold
- Retraining on a schedule or when new data arrives
|
| Vector database | Amazon OpenSearch Serverless or RDS PostgreSQL + pgvector | - Managed similarity search for RAG chatbots
- Hybrid keyword + vector search in OpenSearch
- pgvector keeps embeddings next to your relational data
|
| LLM access | Amazon Bedrock (Claude and others) | - Claude and other models through one API, with no GPUs to manage
- Prompts and data stay in your AWS account and region
- Built-in guardrails, knowledge bases and agents
|
| Monitoring | CloudWatch + SageMaker Model Monitor | - Alerts on latency, errors and traffic
- Model Monitor detects data drift against a training baseline
- Central logs make debugging predictions easier
|
| Privacy controls | VPC endpoints, KMS encryption, HIPAA-eligible services, keep data in one region | - VPC endpoints keep traffic off the public internet
- Encrypt data and models with your own KMS keys
- Many HIPAA-eligible services and compliance certifications
|