Skip to content

GPU runners: track alternatives to the RunsOn custom AMI #183

Open

Description

@quantecon-services

Why

GPU lecture builds (lecture-python.myst) run on RunsOn with a custom AMI, which needs AMI maintenance and Packer rebuilds (docs/dev/GPU-AMI-SETUP.md). Two alternatives were recorded in docs/FUTURE-DEVELOPMENT.md, which #178 removed (source). This issue keeps them tracked.

Alternatives

GitHub-hosted GPU runners

  • No AMI to maintain; the runner is selected with runs-on:.
  • Availability and pricing need checking before this is worth evaluating.
  • Trade-off: likely more expensive per minute than RunsOn, and possibly less choice of instance and GPU type.

A GPU container on an nvidia/cuda base

  • CUDA and cuDNN pre-installed next to the scientific Python stack.
  • Matches the container pattern the CPU builds already use.
  • Still needs a GPU runner (RunsOn or GitHub-hosted) started with --gpus all.
  • Trade-off: an image of roughly 15–20 GB, so pull time could cancel out the setup savings unless layer caching works well.

Related

Checklist

  • Check whether GitHub-hosted GPU runners are generally available to the org, and at what price
  • If they are, compare a cold lecture-python.myst build against RunsOn on time and cost
  • Decide whether a CUDA container is worth prototyping, given the image size

Activity

  1. added
    infrastructureSubstantial CI / build / deploy / tooling / automation work, or behaviour-preserving restructuring
    on Sep 24, 2026
  2. added a commit that references this issue on Nov 15, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    infrastructureSubstantial CI / build / deploy / tooling / automation work, or behaviour-preserving restructuring

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions