A company wants to use SageMaker to deploy a model that requires GPU acceleration for inference but also needs to keep costs low when traffic is low. Which SageMaker feature should they use?