Running training jobs on a heterogeneous cluster

Using the heterogeneous cluster feature of SageMaker Training, you can run a training job with multiple types of ML instances for a better resource scaling and utilization for different ML training tasks and purposes. For example, if your training job on a cluster with GPU instances suffers low GPU utilization and CPU bottleneck problems due to CPU-intensive tasks, using a heterogeneous cluster can help offload CPU-intensive tasks by adding more cost-efficient CPU instance groups, resolve such bottleneck problems, and achieve a better GPU utilization.

See also the blog Improve price performance of your model training using Amazon SageMaker AI heterogeneous clusters.