GPU CI for GitHub Actions
Run tests and benchmarks on A100s, H100s, and B200s with a one-line change. No GPU infrastructure to manage.
From ubuntu-latest to H100 in one commit
Gauss plugs into the CI you already have, with nothing to install or operate.
- Change one line
- Point runs-on at a Gauss label. Your workflows, secrets, matrices, and required checks stay as they are.
- We spin up the GPU
- An isolated runner boots with the driver and CUDA toolchain in place. Gauss handles provisioning and capacity so your team doesn't have to.
- Results land in your PR
- Tests, benchmarks, and regression checks report back to GitHub like any other job. When the job finishes, Gauss destroys the runner.
.github/workflows/benchmark.ymldiff
jobs: benchmark: runs-on: ubuntu-latest runs-on: gauss-h100 steps: - uses: actions/checkout@v4 - run: make benchCI for software that runs on GPUs
Test CUDA kernels, inference engines, ML frameworks, and GPU libraries on the hardware your users actually run.
- On-demand accelerators
- Run CI on A100s, H100s, and B200s without provisioning or managing GPU infrastructure yourself.
- H100 SXM80GB HBM3·82%A10080GB HBM2e·64%B200192GB HBM3e·41%
- Ephemeral by design
- Every job gets a fresh, isolated runner that Gauss destroys on completion. Nothing carries over between runs.
- Provision18sRun job4m 12sDestroyNothing persists
- Native GitHub Actions
- Gauss runners behave like any other label. Your checks, matrices, and secrets work unchanged.
- benchmarkgauss-h1004m 12scompatgauss-a1002m 58skernelsgauss-h1006m 03s
- Regression detection
- Benchmark on identical hardware every run, so you catch a performance regression in the PR before it ships.
- attn_fwd latency+38%
Flagged at 7c3f2a1, before it reached main.
Get early access
Gauss is currently in private beta. Tell us what you're running and we'll help get your first GPU workflow online.
No spam. If your CI runs on GitHub Actions today, you're ready.