Connect your GPU
This gets a node running on your machine so it can take AI training jobs. Most of it is one command — the rest is Windows prerequisites you only ever do once.
Heads up: this is a technical setup and you will be using a terminal. Nothing here is irreversible, and every step can be re-run safely if something goes wrong.
Why contribute a GPU?
- Put idle hardware to work. A gaming GPU sits unused most of the day. Pooled with others it adds up to real training capacity.
- You set the terms. The node only takes work while you have it switched on.
The point is to make serious compute reachable for independent developers and researchers who have the data but not the hardware, without going through a large cloud provider.
What you are actually running
You are about to install software from strangers on the internet, so here is exactly what it can and cannot do. Every line below points at the code that makes it true.
- A job contains no code. Somebody sending work picks from a fixed list — a network shape and some numbers, like width and depth. There is no field anywhere in a job that can carry a program, so there is nothing for your machine to be talked into running. jobSpec.py
- Their data cannot execute either. Datasets are read with pickle switched off. That matters because the normal way to load Python data does run whatever is inside it — which is exactly why it is not used here. artifacts.py
- Nothing connects in. Your node reaches out to ask for work; nothing on the internet can reach it. That is not a firewall setting you have to get right — the node's port is never published outside your own machine. docker-compose.node.yml
- It stops before your card cooks. The temperature is checked every few steps while training, and the job is abandoned rather than throttled through — the comment in that file reads "this is somebody else's hardware". trainer.py
- You can approve each job by hand. Off by default so the node keeps working while you are away; turn it on from your node page and nothing starts without you.
And what it does not promise. Training runs on your GPU, so it uses your electricity and your card's lifespan. The whole project is one person's work, not an audited product. The strongest thing that can honestly be said is that the code is short enough to read and it is all right here.
Before you start Windows, once per machine
Install the latest driver on Windows. Do not install a Linux driver inside WSL — the Windows one is what the container ends up using.
This runs Linux tooling on Windows without a dual boot. Everything after this point happens inside Ubuntu.
Then run one command inside Ubuntu
This installs Docker and the NVIDIA Container Toolkit, checks that a container can really see your GPUs, then fetches and starts the node. It is safe to re-run — every step checks before it acts.
Leave this as it is unless you are pointing the node at a different coordinator.
Loading…
That is the whole setup. When it finishes, head to the front door to register the node. The steps below are only for doing it by hand.
If you would rather do it by hand optional
Manual installation The script above does all of this for you
Open Ubuntu from the Start menu and run:
curl -fsSL https://get.docker.com | sh
sudo apt install -y wget
sudo systemctl restart docker
This is the step worth not skipping — it is where a broken driver or toolkit shows up.
docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi
You should see a table listing your card.
You can skip this. The node runs in a container built
on nvidia/cuda, which already carries the CUDA runtime. On
the host you only need the driver, Docker and the container toolkit.
Install the full toolkit only if you want to compile CUDA code
yourself — it is several GB.
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/12.8.1/local_installers/cuda-repo-wsl-ubuntu-12-8-local_12.8.1-1_amd64.deb
sudo dpkg -i cuda-repo-wsl-ubuntu-12-8-local_12.8.1-1_amd64.deb
sudo cp /var/cuda-repo-wsl-ubuntu-12-8-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-8
And the compiler, if you want to build CUDA programs:
sudo apt-get install -y cuda-compiler-12-8
The installer finishes by printing an address. Open it on the machine you just set up: that dashboard runs beside the agent, and it is where you register the node and watch it work.
It is http://localhost:3000 unless something was
already using that port, in which case the installer says which
one it picked.