GPU support belongs to Managed Instances

AWS announced GPU-accelerated Event Functions on Lambda Managed Instances on November 18. The feature runs functions on supported EC2 G-series instances, with Lambda handling the instance lifecycle, GPU drivers, routing and scaling.

AWS lists inference, media processing, rendering and numerical computing as uses. Availability is limited to commercial regions where both Lambda Managed Instances and the relevant EC2 G-series instances are offered.

Check the quota before preparing a deployment

The developer guide says the EC2 Running On-Demand G and VT instances quota defaults to zero vCPUs. An account needs a sufficient quota before it can deploy these functions; an announcement of regional availability does not grant capacity automatically.

GPU functions use container images, rather than ZIP packages. The guide lists a 32 GB image limit and one GPU or GPU fraction per execution environment. AWS supplies base images with supported frameworks, while allowing custom containers.

The fee is additional to EC2 compute

AWS's pricing page lists a GPU G-series management fee of 9.75% of the EC2 On-Demand instance price, plus standard EC2 instance charges and $0.20 per million requests. Applicable EC2 pricing options can reduce the compute charge; they do not turn this into free GPU execution.

The Managed Instances model does not separately bill each request's execution duration like the default Lambda compute model. Provisioned event-poller usage can have its own charges. Check the selected region, instance, capacity and event-source configuration before estimating a total bill.