AWS Lambda adds GPU support for Event Functions on Managed Instances

·2026.11.19 04:39

Key point

The feature supports EC2 G-series instances with a 9.75 percent management fee and up to 90-minute timeouts for long-running inference tasks.

1 / 4

Details

  • AWS Lambda now supports GPU-accelerated Event Functions on Lambda Managed Instances (LMI), enabling serverless execution of ML inference, media processing, and scientific computing on EC2 G-series instances.
  • The feature eliminates the need to manage GPU drivers, CUDA versions, or clusters; functions are deployed as OCI container images and triggered by standard AWS events.
  • A demo using the Qwen/Qwen2.5-1.5B-Instruct model via the vLLM base image achieved a live, OpenAI-compatible endpoint in approximately 10 minutes, with four concurrent invocations finishing in the same time as a single one.
  • GPU-accelerated functions support container images up to 32 GB, 32 GB of ephemeral storage, and extended timeouts up to 90 minutes for asynchronous invocations.

Configuration and Limits

  • Supported instance types at launch include G4dn, G5, G6, G6f, Gr6, Gr6f, and G6e, with memory ranging from 2,048 to 32,768 MB and accelerator memory from 3 to 48 GB.
  • Per-execution environment concurrency can be set up to 48 requests, and capacity providers can scale up to 15,000 vCPUs.
  • Curated base images are available for CUDA, PyTorch, vLLM, and SGLang, allowing users to skip manual driver and library reconciliation.

Pricing and Availability

  • The feature is generally available and bills the underlying EC2 instance fee plus a 9.75 percent management fee for GPU instances and the standard Lambda request fee of $0.20 per million requests.
  • Existing Savings Plans and Reserved Instances apply to the EC2 costs, and users can create On-Demand Capacity Reservations to ensure predictable GPU availability.
  • GPU utilization, memory, and power metrics are visible in Amazon CloudWatch and the Lambda console without additional setup.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.