Amazon ECS Now Supports Sharing GPUs for Container Tasks

The new capability allows multiple containers to run on a single NVIDIA GPU, increasing resource density on EC2 instances.

Updated on Oct. 28, 2026 in Data Centers

Amazon ECS Now Supports Sharing GPUs for Container Tasks

Live Poll

Do you believe shared GPU resources will improve your own cloud application performance and cost efficiency?

Amazon has introduced GPU sharing for Amazon ECS on EC2, allowing multiple tasks to execute concurrently on a single NVIDIA GPU. This update is now available across all AWS commercial and GovCloud regions.

Why it matters

Many inference models and embedding tasks occupy only a small fraction of a GPU's total capacity. By enabling task density on a single chip, AWS allows users to optimize infrastructure costs by improving utilization rates for under-taxed hardware.

The feature supports 1 to 100 percent compute allocation across a wide range of NVIDIA-powered instance families, including G4dn through P6-B200. Operators require the ECS container agent version 1.107.0 or later and the 20261019 optimized AMI.

The players

Amazon Elastic Container Service

A fully managed container orchestration service that simplifies running and scaling containerized applications on AWS.

NVIDIA

The primary designer of GPUs and specialized compute architectures used in high-performance cloud data centers.

The details

This capability leverages the NVIDIA Multi-Process Service (MPS) — a runtime architecture that allows multiple processes to submit work to a GPU simultaneously. By configuring a sharingStrategy with an nvidiaMps object in resource requirements, containers share a single CUDA context — the environment used by the GPU to execute parallel code. ECS manages placement by calculating the cumulative memory requests of all containers assigned to the GPU to ensure it does not exceed the hardware limit.

October 27, 2026: AWS announced the GPU sharing feature for ECS.
October 19, 2026: The minimum required ECS AL2023 GPU-optimized AMI version was released.

The Tech Race

Cloud providers are increasingly competing to maximize hardware utilization as GPU supply remains constrained and inference workloads become more prevalent. This integration follows industry efforts to move beyond static, single-task GPU assignments toward granular, shared-resource models.

Engineers using EC2 instances can now stack smaller inference models on the same GPU to increase throughput. Note that this feature is not supported on Fargate, ECS Anywhere, or for tasks running in ECS Express Mode.

The takeaway

AWS is prioritizing higher density for inference workloads to address common GPU underutilization. Monitor your resource requirements and agent versions to ensure compatibility with shared-GPU configurations in your production clusters.