Choosing a runner for GPU jobs
GPU runners are for testing ML models, CUDA builds, rendering and inference checks in CI. Only some providers offer them. Compare the GPU model and memory, the per-minute price, whether drivers and CUDA come preinstalled, and availability. Running GPU jobs in your own cloud account is the alternative when you need a specific instance type or have GPU quota already.
What GPU jobs need from a runner
- GPU model and VRAM: enough memory for the model or test you run.
- Drivers and toolkits: NVIDIA drivers, CUDA and the container toolkit preinstalled, or a way to bring your own image.
- Price and granularity: GPU minutes cost far more than CPU minutes, so per-second billing and fast start-up matter.
- Availability: GPU capacity is scarcer than CPU capacity. Check queue behaviour.
- Your own cloud: bring-your-own-cloud and open-source autoscalers can launch any GPU instance type your account has quota for.
Runners that fit, by kind
10 runners document what this workload needs. They are grouped by kind and listed alphabetically; the order is not a ranking. Shortlist two or three, then try them with your own workflows.
Managed runners for GitHub Actions
Third-party runners you switch to by changing the runs-on label in a workflow.
machine.dev
GPU, x64 and arm64 runners for GitHub Actions on dedicated AWS instances, with spot or on-demand tenancy chosen in the runs-on label.
- Dedicated AWS instance
- Ephemeral
- arm64
- GPU
Why it fits: Six NVIDIA GPU tiers and AWS Trainium and Inferentia2 runners are listed with CUDA and drivers preinstalled.
Runners from your CI platform or cloud
The first-party hosted runners of GitHub, GitLab, Buildkite and Azure, plus cloud-native runner services.
AWS CodeBuild-hosted runners
AWS CodeBuild projects that start a single-use runner in your AWS account for each GitHub Actions, GitLab or Buildkite job.
- Fresh build environment
- Ephemeral
- arm64
- Windows
- Docker layer cache
- GPU
- BYOC
Why it fits: Linux GPU compute types can be selected with the instance-size label.
Full profile →AWS CodeBuild-hosted runners alternativesWebsite ↗
GitHub-hosted runners
GitHub’s own Actions runners: a new VM per job on Linux, Windows and macOS, with larger, arm64 and GPU sizes billed per minute.
- Fresh VM
- Ephemeral
- arm64
- macOS
- Windows
- GPU
Why it fits: Larger runners include a 4-vCPU Tesla T4 machine on Linux and Windows.
GitLab-hosted runners
GitLab.com’s built-in runners: each job gets a single-use VM, metered in compute minutes with a cost factor per machine size.
- Ephemeral VM
- Ephemeral
- arm64
- macOS
- Windows
- GPU
Why it fits: A GPU-enabled Linux machine type with an NVIDIA Tesla T4 (or similar) is available on Premium and Ultimate.
Runners in your own cloud or on your hardware
Commercial products that launch runners inside your cloud account or on servers you own.
Actuated
Managed control plane that runs each CI job in a Firecracker microVM on servers you provide, for GitHub Actions and GitLab CI.
- Firecracker microVM
- Ephemeral
- arm64
- Docker layer cache
- GPU
- BYOC
Why it fits: GPU passthrough to jobs is documented for hosts switched to Cloud Hypervisor.
Aspect Workflows
Bazel-focused platform whose auto-scaling CI runners register with your existing CI and can run in your own AWS account or GCP project.
- Persistent VM
- arm64
- GPU
- BYOC
- Self-hostable
Why it fits: GPU runner groups are available for model training and inference tests.
Cirun
Hosted control plane that launches on-demand GitHub Actions runners in your own cloud accounts or on-prem hardware from a .cirun.yml file.
- Cloud VM
- Ephemeral
- arm64
- macOS
- Windows
- Docker layer cache
- GPU
- BYOC
Why it fits: GPU runners are included in every plan and the docs show GPU instances on Google Cloud and Vast.ai as a provider.
HyperEnv for GitHub Actions Runner
AWS Marketplace product that runs each GitHub Actions job on an ephemeral EC2 instance, deployed entirely inside your own AWS account.
- EC2 instance
- Ephemeral
- GPU
- BYOC
- Self-hostable
Why it fits: g4dn instance types with an NVIDIA T4 GPU can be selected per job.
Full profile →HyperEnv for GitHub Actions Runner alternativesWebsite ↗
RunsOn
Self-hosted GitHub Actions runner control plane installed in your own AWS account, launching one fresh EC2 instance per job.
- EC2 instance
- Ephemeral
- arm64
- Windows
- Docker layer cache
- GPU
- BYOC
- Self-hostable
Why it fits: GPU images with NVIDIA drivers and CUDA are published for x64 and Graviton GPU instances.
Open-source autoscalers for self-hosted runners
Software you deploy to run and scale self-hosted runners on Kubernetes, cloud VMs or bare metal.
GitLab Runner
The open-source agent that runs GitLab CI/CD jobs, with executors for Docker, Kubernetes, shell and autoscaled cloud instances.
- Depends on executor
- arm64
- macOS
- Windows
- GPU
- Open source
- BYOC
- Self-hostable
Why it fits: GPU configuration is documented for the Shell, Docker and Kubernetes executors.
Side by side
| Runner | CI systems | Runs on | Deployment | Isolation | Price | Largest runner | Docker layer cache | GPU | Open source | Company HQ |
|---|---|---|---|---|---|---|---|---|---|---|
| Actuated | GitHub Actions, GitLab CI | Linux, x64, arm64 | Your cloud or hardware | Firecracker microVM | $150/mo per server + your servers | Set by your infrastructure | Yes | Yes | No | United Kingdom |
| Aspect Workflows | GitHub Actions, Buildkite, GitLab CI, CircleCI | Linux, x64, arm64 | Your cloud or hardware | Persistent VM | Annual platform fee (quote) + cloud cost | Set by your infrastructure | Not documented | Yes | No | Not documented |
| AWS CodeBuild-hosted runners | GitHub Actions, GitLab CI, Buildkite | Linux, Windows, x64, arm64 | Your cloud account | Fresh build environment | $0.005/min | 72 vCPU / 144 GB | Yes | Yes | No | United States |
| Cirun | GitHub Actions | Linux, macOS, Windows, x64, arm64 | Your cloud or hardware | Cloud VM | $29/mo (3 private repos) + cloud cost | Set by your infrastructure | Yes | Yes | No | United Kingdom |
| GitHub-hosted runners | GitHub Actions | Linux, macOS, Windows, x64, arm64 | Platform-hosted | Fresh VM | $0.006/min | 96 vCPU / 384 GB | Not documented | Yes | No | United States |
| GitLab Runner | GitLab CI | Linux, macOS, Windows, x64, arm64 | Self-hosted | Depends on executor | Free software + your infrastructure | Set by your infrastructure | Not documented | Yes | Yes | Not documented |
| GitLab-hosted runners | GitLab CI | Linux, macOS, Windows, x64, arm64 | Platform-hosted | Ephemeral VM | $10 per 1,000 compute minutes | 32 vCPU / 128 GB | Not documented | Yes | No | Not documented |
| HyperEnv for GitHub Actions Runner | GitHub Actions | Linux, x64 | Your cloud or hardware | EC2 instance | $0.002/min fee + AWS cost | 192 vCPU / 768 GB | Not documented | Yes | No | Germany |
| machine.dev | GitHub Actions | Linux, x64, arm64 | Vendor-hosted | Dedicated AWS instance | $0.00297/min on-demand | 64 vCPU / 128 GB | Not documented | Yes | No | Australia |
| RunsOn | GitHub Actions | Linux, Windows, x64, arm64 | Your cloud or hardware | EC2 instance | €300/yr licence + AWS cost | 896 vCPU | Yes | Yes | No | France |
Frequently asked questions
Can GitHub Actions run jobs on a GPU?
Check before you rely on it. Products change quickly, and this page reflects public documentation as of November 22, 2026. Confirm anything that matters in the provider’s own documentation.
Cite this page
CI Runners. (2026). Choosing a runner for GPU jobs. cirunners.org. Last reviewed November 22, 2026. https://cirunners.org/use-cases/gpu/