Requirements
The machine
| Minimum | Recommended | |
|---|---|---|
| OS | 64-bit Linux (kernel 5.x+), or Windows 11 + WSL2 | Ubuntu 22.04 / 24.04 LTS |
| GPU | NVIDIA, 6 GB VRAM | 16–24 GB+ VRAM |
| NVIDIA driver | 575+ (CUDA 12.9) | Latest stable |
| CPU | 4 cores | 8+ cores |
| RAM | 8 GB | 32 GB+ |
| Free disk | 50 GB SSD | 100 GB+ NVMe SSD |
| Network | 25 Mbps down, stable | 100 Mbps+ symmetric |
Only two things are installed by you: Docker and the NVIDIA driver. Everything else ships in the image.
Which GPUs earn what
VRAM decides which models your card can hold, and that decides what it can be rented for. More VRAM is worth more than more speed.
| Tier | VRAM | Example cards |
|---|---|---|
| Entry | 6–8 GB | RTX 2060, 3050 / 3060, 4060 / 4060 Ti, 5060 (incl. Laptop) |
| Standard | 10–12 GB | RTX 3080, 2080 Ti, 4070 / 4070 Ti, 5070 |
| High | 16 GB | RTX 4070 Ti Super, 4080 / 4080 Super, 5070 Ti / 5080 |
| Premium | 24–32 GB | RTX 3090 / 3090 Ti, 4090 / 4090 D, 5090, RTX 4000–5000 Ada, A10 |
| Pro / Datacenter | 48 GB+ | A40, RTX 6000 Ada, RTX PRO 5000/6000 Blackwell, A100, H100, H200 |
NVIDIA only. AMD and Intel cards are not supported, and neither is Apple silicon. The inference engine the agent runs is CUDA-only.
Laptop cards count. A laptop 3060 with 6 GB is a real Entry-tier host.
Network
- Outbound to our backend, to Hugging Face for model weights, and to Cloudflare for the tunnel that carries inference traffic.
- No inbound ports and no port forwarding. The tunnel is outbound, so the agent works behind NAT and consumer routers without any configuration.
- Stability beats speed. The agent heartbeats continuously and is marked offline if it stops answering. A steady 25 Mbps line is a better host than a fast one that drops.
Docker fixes a container's DNS when it is created. Moving the machine between
networks silently breaks it until the container is restarted — the --dns
flags in the install commands prevent this. See
Troubleshooting.
Power and uptime
Your GPU only earns while it is online and idle enough to accept work. Two things quietly cost hosts money:
- Sleep. A machine that suspends stops serving, and a job in flight fails.
- Gaming or rendering on the same card. The agent measures free VRAM, so a busy card is offered smaller jobs or none at all.
A host does not need to be a dedicated machine, but the more predictable its free time, the more work it gets.
What the agent does to your machine
- Runs one container, plus one more per job while a job is running.
- Writes job workspaces to
/var/lib/neurogrid/workloads. - Stores its encrypted token in a Docker volume, not in your home directory.
- Needs the Docker socket, which is root-equivalent access to the host. The container is packaging, not a security boundary — run it on a machine you are comfortable giving that to.