GPU runner

NVIDIA H100 80GB runners for GitHub Actions

Hopper-class GPU with FP8. Use when you need post-A100 throughput in CI. Cirun spins NVIDIA H100 80GB VMs up on your cloud account per workflow — pick whichever cloud has the SKU on your existing account, CUDA + drivers come from each cloud's own Deep Learning image.

Run a workflow on NVIDIA H100 80GB Docs

Why this fits

80 GB VRAM and ~989 TFLOPS FP16 — enough for 70B+ LLM evals and multi-batch inference.
Works with PyTorch, JAX, vLLM — CUDA + drivers come from your cloud's official Deep Learning AMI / image, not your workflow.
Ephemeral by default — each job gets a clean VM, so leaked weights or hung CUDA contexts can't poison the next run.

.cirun.yml

1runners:
2  - name: gpu-h100
3    cloud: aws
4    instance_type: p5.48xlarge
5    # Use AWS Deep Learning AMI GPU PyTorch on Ubuntu 22.04, or a
6    # Cirun-published H100-compatible AMI.
7    machine_image: ami-04823729c75214919
8    labels:
9      - cirun-gpu-h100

.cirun.yml

1runners:
2  - name: gpu-h100
3    cloud: gcp
4    instance_type: a3-highgpu-8g
5    # Use a GCP Deep Learning VM image family for pre-installed CUDA +
6    # drivers.
7    machine_image: projects/deeplearning-platform-release/global/images/family/pytorch-latest-cu121
8    labels:
9      - cirun-gpu-h100

Drop this in your repo root. The first workflow that requests the runner label spins this configuration up on your cloud account.

Where you can run it

AWS p5.48xlarge

192 vCPU · 2048 GiB host

Google Cloud a3-highgpu-8g

208 vCPU · 1872 GiB host

Ready to run your CI here?

Cirun is free for open source. For private repos, flat monthly plans by repo count — never per CI minute.

Run a workflow on NVIDIA H100 80GB Docs

Sources · verified 2026-05-25

NVIDIA — NVIDIA H100 80GB datasheet