Earn from your unused GPU.
If your graphics card spends most of its time idle, Nodefield is building a way to put it to work — on your terms, with your control, and with no promise of earnings we cannot back up.
How providing works
- Register. Tell us your GPU, VRAM, operating system and availability on the waitlist. No payment details, no passwords for other services.
- Install the agent (beta). A lightweight desktop app detects your hardware, benchmarks it locally, and reports capacity to the coordinator.
- Set your limits. Choose when you are available and how much VRAM and power the agent may use. Pause any time.
- Complete work, get paid — once billing exists. Providers are paid only when their hardware successfully completes paid workloads. The proposed split is 80% provider / 20% platform, pending validation. No billing is active yet, so no earnings exist yet.
GPU compatibility
Compatibility is about VRAM, drivers and a supported inference runtime — not brand loyalty. The table below is our current guidance; the agent verifies your actual hardware before any work is assigned.
| Hardware | Status | Notes |
|---|---|---|
| NVIDIA RTX 3060 (12 GB), 4060 Ti (16 GB), 4070 and above | Compatible | Enough VRAM for 7B–8B quantised models; CUDA support. |
| NVIDIA RTX 3060 (8 GB), RTX 4060, GTX 16-series with 6 GB+ | Compatible | Suits 1B–4B models; 8B models only where VRAM allows. |
| AMD RX 6000 / 7000 series (8 GB+) | Experimental | Supported where the runtime supports it; performance varies by driver and OS. |
| GPUs with under 4 GB VRAM, integrated graphics only | Unsupported | Too little VRAM for the launch models. |
| macOS (Apple Silicon) | Experimental | Possible for small models via supported runtimes; Windows and Linux are the first targets. |
VRAM figures are for Q4-quantised launch models; see models.
Security — what the agent will never do
Never arbitrary code
The agent never executes arbitrary code supplied by marketplace customers. Workloads are sandboxed, validated inference jobs only — a model, a prompt, and strict resource limits.
Never hidden activity
Everything the agent does is visible in its dashboard: utilisation, VRAM, jobs completed, uptime. It never secretly uses your GPU.
Never your files
The agent does not access personal files. It reads hardware information and runs the inference runtime — nothing else.
Always your switch
Pause participation any time, set resource limits, or uninstall completely. Signed updates, encrypted coordinator traffic, and an emergency shutdown are part of the design.