Savings Plans Applicability to ML instances

Savings plans only apply to EC2 (compute and EC2 instance savings plans) and Fargate instances (compute savings plans). They do not apply to other services, such as SageMaker.

For cost optimization, SageMaker supports spot instances for training jobs, and for the hosting of real-time inference endpoints, look at:

• right-sizing and auto-scaling the endpoint
• Using elastic inference or inf1 instances (support for inf1 in SageMaker is on the roadmap)
• Compiling models with Neo to optimize model performance (for rightsizing)
• Using inference pipelines instead of multiple model endpoints
• Hosting in multi-model endpoints
• Hosting in self-managed compute, such as EC2 with RI, Lambda, or Fargate/spot.

Elastic inference, Neo, inference pipelines and multi-model endpoints work well in some, but not all use-cases, depending on use case, model framework, complexity, size, etc. so it will require some investigation and experimentation to figure out of they’re right for you, but they can help reduce inference cost significantly.