Open-source · MIT

Your own private
AI GPU. No cloud.

GhostGPU routes GPU compute from your laptop over Wi-Fi, through a Raspberry Pi 5, to an AMD RX 580. Run real local LLMs on hardware you already own — no subscriptions, no data leaving your network.

Get started → How it works
$0/ month
8 GBVRAM
~$205to build
100%local
👻
~35 tok/s
RX 580 · 8GB
How it works

A wireless bridge from your laptop to a real GPU

Your laptop runs llama.cpp in client mode and talks to a Raspberry Pi 5 over Wi-Fi. The Pi hosts an OpenAI-compatible server and offloads the heavy work to an RX 580 over a PCIe riser.

💻
Your laptop
llama.cpp · client mode
Sends prompts over the OpenAI-compatible API. No GPU required.
Wi-Fi → HTTP / WS
🍓
Raspberry Pi 5
the bridge
llama-server ROCm 6.0 GPU offload
PCIe x1 → riser
🎮
AMD RX 580
the muscle
VRAM8 GB GDDR5
Shaders2,304
The case for local

How it stacks up

Setup Monthly cost Latency Privacy GPU VRAM
👻 GhostGPU $0 ~50–150 ms ✓ 100% local 8 GB
OpenAI API (GPT-4o) $15–$60+ ~200–800 ms ✕ Cloud N/A
AWS EC2 g4dn.xlarge ~$380 ~100–300 ms ⚠ Cloud 16 GB T4
Google Colab (free) $0 ~300–1000 ms ⚠ Cloud 15 GB T4
Local PC (RTX 3070) $0 owned ~20–50 ms ✓ Local 8 GB
Measured on RX 580 · 5 GHz Wi-Fi

Performance benchmarks

Phi-3 Mini
Q4_K_M · 2.3 GB
28–35 t/s
Mistral 7B
Q4_K_M · 4.8 GB
18–25 t/s
CodeLlama 7B
Q4_K_M · 4.8 GB
17–23 t/s
Llama 3 8B
Q4_K_M · 5.0 GB
15–22 t/s
Mistral 7B
Q8_0 · 7.2 GB
12–16 t/s

Benchmarks vary by Wi-Fi conditions, system load, and layer offload configuration.

Where it's going

Roadmap

✓ SHIPPED
✓PCIe riser + RX 580 detection on Pi 5
✓ROCm install scripts for ARM64
✓llama.cpp HIP build scripts
✓llama-server systemd service config
✓OpenAI-compatible API over Wi-Fi
○ PLANNED
Web UI dashboard (Open WebUI)
Multi-model hot-swap via API
Docker container for easy deployment
Stable Diffusion image generation
Whisper speech-to-text support
mDNS autodiscovery (ghostgpu.local)
Power monitoring + thermal dashboards
Support for RX 6600 / 6700 XT
👻

Build your own GhostGPU

A spare RX 580, a Raspberry Pi, and an afternoon. Everything you need — scripts, configs, and docs — is in the repo.

★ Star on GitHub Read the docs →