AWS Adds NVIDIA GPU Partitioning to SageMaker Endpoints

AWS says an H100 MIG partition delivered 53% to 89% more output tokens per second than a whole A100 in selected tests, while full-instance billing and workload-dependent break-even points limit savings.