Amazon SageMaker Automatic Model Tuning now reuses SageMaker Training instances to reduce start-up overheads by 20x

Amazon SageMaker Automatic Model Tuning now reduces the start-up time of each training job launched to tune your models by 20x on average (from 2.5 minutes to 8 seconds). In scenarios where you have a large number of hyperparameter evaluations, the reuse of training instances can cumulatively save 2 hours for every 50 sequential evaluations.

SageMaker Automatic Model Tuning finds the best version of a model by running many training jobs on your dataset using specific ranges of  hyperparameters that you choose for your algorithm. SageMaker Automatic Model Tuning then chooses the most optimal hyperparameter values that result in a model that performs the best.

Before this launch, every training job launched as part of the tuning would incur on average 2.5 minutes of overhead to spin up and prepare a new cluster of SageMaker Training instances. This could become a bottleneck especially when training jobs would take only a few minutes to complete and overall slow down your tuning job. Starting today, SageMaker Automatic Model Tuning automatically re-uses a fixed cluster of training instances within each tuning job, thus reducing the average start-up time of each training job by 20x.

SageMaker Automatic Model Tuning reusable clusters is now available in all commercial AWS Regions. This new feature is turned on by default when launching your tuning jobs. To learn more about SageMaker Automated Model Tuning, please read the technical documentation.