GPU Runners

machine.dev offers 6 NVIDIA GPU types and 2 AWS AI accelerators for GitHub Actions. Each NVIDIA GPU is offered in multiple vCPU/RAM configurations so you can pair the right amount of CPU and memory with the GPU. All runners run Ubuntu 22.04 LTS with a configurable gp3 EBS root volume.

NVIDIA GPUs

Live prices and current spot interruption rates are at machine.dev/runners. The tables below show the best rates across all regions.

Configurations and pricing

T4G / T4 / L4 / A10G / L40S are available in three vCPU/RAM tiers each. RTX 6000 is available in six sizes including multi-GPU configurations. Pick a higher-vCPU configuration if your workflow does heavy CPU-side preprocessing before handing off to the GPU.

T4G — 16 GB VRAM, ARM64

T4 — 16 GB VRAM, X64

L4 — 24 GB VRAM, X64

A10G — 24 GB VRAM, X64

L40S — 48 GB VRAM, X64

The L40S has been used to fine-tune 70B-parameter models with QLoRA on a single GPU. See Crusoe’s writeup for a real-world example.

RTX 6000 — 96 GB VRAM, X64

RTX 6000 (NVIDIA RTX PRO 6000) gives you 96 GB of VRAM per GPU and scales up to 8 GPUs in a single runner — the largest VRAM tier in the public catalog. Well-suited for multi-GPU fine-tuning and large-context inference.

AWS AI accelerators

AI accelerator runners include the AWS Neuron SDK pre-installed.

AI accelerators are currently available in us-east-1, us-east-2, and us-west-2 only. See Regions for the GPU-by-region matrix.

Storage

Every runner gets a 100 GB gp3 EBS root volume by default with 6,000 IOPS and 250 MB/s throughput. You can scale up to 16 TB with custom IOPS and throughput using the disk_size, disk_iops, and disk_throughput labels:

runs-on:
  - machine
  - gpu=l4
  - disk_size=500          # 500 GB root volume
  - disk_iops=10000        # 10,000 IOPS
  - disk_throughput=750    # 750 MB/s throughput

Defaults are included at no additional charge. Increasing IOPS above 6,000 or throughput above 250 MB/s incurs prorated EBS charges. See Pricing for the EBS rate breakdown.

Instance metrics

machine.dev collects metrics by default for every job and renders them as sparkline charts on the dashboard. Collected metrics:

• GPU: utilization, memory usage, temperature, power draw
• System: CPU utilization, memory usage, disk I/O, network bytes in/out

Control collection per job with the metrics and metrics_interval labels:

runs-on:
  - machine
  - gpu=a10g
  - metrics=true           # Enable (default)
  - metrics_interval=10    # Collect every 10 seconds (default: 60)

To disable metrics for a job, set metrics=false.

Pre-installed software

GPU runners come with the NVIDIA driver, CUDA, cuDNN, the NVIDIA Container Toolkit, Docker, and Python pre-installed. AI accelerator runners also include the AWS Neuron SDK.

Versions are updated with each runner image build.

You can install any other version, build CUDA from source, or use your own Docker image — runners give you full root access.

Use it in a workflow

Minimal:

jobs:
  train:
    runs-on: [machine, gpu=l4]   # 4 vCPU, 16 GB RAM, 24 GB VRAM by default

Scale up the CPU/RAM:

runs-on: [machine, gpu=l4, cpu=16, ram=64]

Use spot pricing for cost:

runs-on: [machine, gpu=l4, tenancy=spot]

Pin a region:

runs-on: [machine, gpu=l4, regions=eu-south-2]

See Configuration options for the complete label reference.

Next steps

• CPU vs GPU decision guide — pick the right runner for your workload
• Cost Optimization — spot, checkpointing, right-sizing
• Pricing — full per-minute rates for every runner
• Workflow Setup — patterns for real workflows