What this is

GPU Share is a small cluster: one coordinator on the internet, and Windows PCs with RTX 3060/3070 cards that run Docker jobs. Workers only connect outbound. You do not open ports on a home router.

Add a computer

1. Open Add this computer on the GPU PC (or use the join link from the cluster admin).
2. Paste the join code if it is not already in the page.
3. Copy the PowerShell command and run it on that PC. No Python or git install.
4. The worker installs to %LOCALAPPDATA%\GPUShare, starts now, and again when you log on.

If you run the installer as Administrator it also registers a Windows service. Docker Desktop still works best while someone is logged in.

NVIDIA and Docker

Joining only needs the NVIDIA driver. Running jobs needs Docker with GPU access:

1. Install current NVIDIA Game Ready or Studio drivers. Confirm with nvidia-smi.
2. Install Docker Desktop and use the WSL2 backend.
3. Enable NVIDIA GPU in WSL. In Ubuntu, nvidia-smi should work too.
4. Leave Docker Desktop running.

The worker runs docker run --gpus device=N .... Until that works, the PC can still appear online with a “docker gpu not ready” badge.

First test job

From the cluster dashboard (admin token), submit:

• Image: nvidia/cuda:12.4.1-base-ubuntu22.04
• Command: nvidia-smi

Click the job to watch logs. If the card is listed, the GPU path works.

VRAM on 3060 / 3070

One job per GPU. Set min VRAM on the submit form if a job must land on a 12 GB card. 70B models are out of scope unless you add multi-GPU later.

Other GPU work

Any Docker image that can use NVIDIA GPUs is a job: training scripts, ffmpeg, inference. Upload input files with the job; set output globs (for example out.pt, results/*) to pull artifacts back. LLMs are the same path — later you can run an Ollama or llama.cpp image as a normal job.

Troubleshooting

• PC never appears: rerun the installer; check it is online; Windows user should stay logged in.
• irm is blocked: Set-ExecutionPolicy -Scope Process Bypass then paste the command again.
• Job fails on docker: start Docker Desktop, confirm WSL2 GPU, retry nvidia-smi inside a CUDA image.
• Job sits in queued: no idle worker with enough VRAM, or all workers busy.