Reduce Amazon SageMaker costs with these tips
Build, train, and deploy machine learning models at scale
Amazon SageMaker cost saving tips
Optimize your Amazon SageMaker costs with these tips
Use Lifecycle Configurations
SageMaker Notebooks can be expensive on an hour-by-hour basis. By using Lifecycle Configurations, you can configure your idle kernels to automatically shut down when they are not active. I often advice customers to shut down the kernels after 60 minutes. Shutdown script can be found on the AWS website.
Utilize ML Savings Plan
You can utilize machine learning saving plans for Amazon SageMaker. It is like the Compute Savings Plan for EC2, Fargate and Lambda. Often my customers opt for the no-upfront method to save up to 30% on their SageMaker instances. It is best to stack slightly smaller saving plans, so they are fully utilized.
Right-Size Notebooks
To get the most out of your SageMaker notebooks make sure you are using the right instance type. Monitor your SageMaker notebook CPU / memory usage and cut the instance size in half when a 4-week average is below 40%. You can use CloudWatch agent to collect memory and CPU usage.
Service information & pricing
About Amazon SageMaker
- About Amazon SageMaker
-
Amazon SageMaker is a full-featured machine learning service for building, training, and deploying ML models. It offers an array of tools for data scientists & analysts, handling of diverse data types, fast training infrastructure, MLOps automation. Amazon SageMaker is a great service with a lot of features but can expensive quickly. Optimizations depend on if you are using Studio (serverless) or Notebooks.
- Amazon SageMaker pricing
-
Amazon SageMaker is a flexible-priced ML platform with a free first two months trial. Post-trial, choose between On-Demand Pricing or SageMaker Savings Plans. Pricing for various features and capabilities is available. To optimize the cost further you need to split your cost saving in the training and notebooks part. For instance, to save on the training part of SageMaker you can use spot instances to save on training costs.