Amazon ECS adds GPU sharing with NVIDIA MPS

AWS announced on October 27, 2026 that Amazon Elastic Container Service supports running multiple tasks concurrently on a single NVIDIA GPU through NVIDIA Multi-Process Service (MPS). AWS positions the feature for workloads that consume only part of a GPU, including small-model inference, embedding generation, multi-model serving and AV/ADAS simulations.

Per-container requirements guide placement

Containers declare their GPU memory requirements and can optionally specify a maximum share of GPU compute. ECS then places tasks on GPUs with available capacity as services scale and deploy. AWS says this enables denser placement without application-code changes, according to its launch notice.

The engineering opportunity is better use of otherwise idle capacity—not an automatic performance improvement. Teams should measure memory demand and request latency under concurrent load before increasing task density. A compute ceiling should not substitute for validating whether colocated workloads still meet their service targets. Evaluate cost savings against whether sharing actually reduces the capacity a deployment needs.

Managed Instances adds daemon recovery

On ECS Managed Instances, ECS monitors the MPS control daemon that enables sharing. If the daemon stops or becomes unhealthy, ECS automatically replaces the instance and moves tasks to healthy capacity, AWS says. The announcement attributes this monitoring and recovery behavior specifically to Managed Instances.

Teams should still validate request handling during instance replacement and task movement rather than assuming recovery will be invisible to clients.

Availability and setup

AWS lists GPU sharing as available on ECS Managed Instances and ECS on Amazon EC2 in all AWS commercial and AWS GovCloud (US) Regions.

The announcement outlines the setup: create a capacity provider with NVIDIA GPU instances, set the GPU-sharing strategy and memory in the task definition’s container definition, and create a service on that capacity provider. AWS lists the Management Console, CLI, SDKs, CloudFormation and other infrastructure-as-code tools as configuration options.