Hardware
What your machine needs to lend GPU time.
Supported hardware
| GPU | Detection | Stepping aside | Notes |
|---|---|---|---|
| NVIDIA (GTX 16 series and newer) | Automatic via nvidia-smi | Yes | Windows and Linux. Keep the driver current. |
| Apple Silicon (M1 and newer) | Automatic | Not yet | The worker counts two thirds of unified memory as VRAM, the default Metal limit. |
| AMD Radeon | Automatic on Linux via rocm-smi | Not yet | On Windows, pass --vram <GB> because the worker cannot read VRAM there. |
The worker also needs:
- Windows 10 or newer, macOS 11 or newer, or a 64-bit Linux. Nothing to install beforehand; the install command brings what it needs.
- A stable connection. Bandwidth matters far less than staying connected.
What you earn
Pay follows the market rate for the GPU you lend. That rate fluctuates and is updated in real time. You keep 70% of what labs pay. Lending a larger share of the card earns more, because more of it is in use. See Allocation and Earnings.
The VRAM you lend is your card's total VRAM times your allocation. Labs ask for a minimum amount of memory per card, so more lent VRAM means your GPU fits more of their jobs.