AWS adds NVIDIA MIG GPU partitioning to SageMaker Inference

AWS announced on its official blog that SageMaker Inference now supports NVIDIA Multi-Instance GPU partitioning. The feature lets a physical GPU be divided into isolated partitions with dedicated compute and memory, so more models can share one endpoint and teams can choose a better-sized slice instead of a whole accelerator. AWS said the capability works with inference components, while single-model, serverless, and heterogeneous instance pool endpoints are not supported. Salesforce has helped define requirements and validate the initial implementation.