Personal AI Computer – DIY High‑Performance AI Workstation

What it is

A personal‑AI‑computer is an open‑source, DIY hardware guide for building a desk‑sized (or rack‑mounted) machine that can run large language models locally. The repository contains CAD files, bill‑of‑materials, assembly photos, BIOS‑tuning steps and wiring diagrams for several configurations ranging from a single‑user “2 × RTX 5090” box to an enterprise‑grade “8 × RTX 5090” rack.

Why it matters

• Local inference – By owning the hardware you keep model weights on‑prem, avoiding API costs, usage‑based billing and the risk of a provider shutting you out.
• Data privacy – All prompts stay on your own disk, so proprietary or confidential data never leaves the organization.
• Open stack – The hardware design is MIT‑licensed; the software side can be any open‑source inference engine (vLLM, Ollama, llama.cpp) or the provided Grid orchestrator that links multiple machines into a private AI network.

Core components (hardware)

Build	GPUs	CPU	Memory	Power	Size
2 × RTX 5090	2 × RTX 5090 (64 GB VRAM total)	Intel Xeon W5 (ASUS W790 ACE)	96 GB DDR5	1 600 W PSU	12.5×12.5×16 in
2 × RTX PRO 6000	2 × RTX PRO 6000 (192 GB VRAM total, ECC)	Intel Xeon W5	96 GB DDR5	1 600 W PSU	12.5×12.5×16 in
4 × RTX 5090	4 × RTX 5090 (128 GB VRAM)	AMD Threadripper Pro	96 GB DDR5	4 000 W PSU	20×20×24 in
4 × RTX PRO 6000	4 × RTX PRO 6000 (384 GB VRAM, ECC)	Intel Xeon w5‑3423	192 GB DDR5 ECC	4 000 W PSU	19.8×19.8×24.2 in
8 × RTX 5090	8 × RTX 5090 (256 GB VRAM)	2 × AMD EPYC 9004 (Genoa)	192 GB DDR5	8 000 W PSU (5U rack)	Rack‑mount

All builds use PCIe Gen 5 ×16 slots, high‑speed Ethernet (10 GbE / 2.5 GbE), and a modular open‑frame chassis that is fully documented with photos and CAD.

Software workflow

1. BIOS & GPU validation – setup.md walks you through enabling multi‑GPU support, installing NVIDIA drivers and stress‑testing each card.
2. Model serving – The recommended entry point is the Grid orchestrator (curl … | bash). Grid discovers the GPUs, launches inference servers (vLLM, Ollama, etc.) and can federate several Personal‑AI‑Computers into a private cluster.
3. Any local engine – If you prefer, you can run the models directly with vLLM, Ollama, or llama.cpp.

Who should use it

• Researchers & hobbyists who want a hands‑on platform for experimenting with quantized or full‑precision LLMs without cloud fees.
• Start‑ups / SMEs that need to keep proprietary data in‑house while still accessing the compute power of modern GPUs.
• Enterprises looking for a blueprint to assemble on‑prem AI servers that can be scaled from a single box to a rack of eight GPUs.

Getting started

1. Choose a build folder (e.g., 2x-5090/).
2. Follow the parts list and CAD files to source components and fabricate brackets.
3. Assemble the chassis, install the CPU/RAM, mount the GPUs on the PCIe risers.
4. Run the BIOS‑tuning script from setup.md.
5. Install Grid (or your preferred inference engine) and start serving models.

Community & contribution

The repo encourages users to submit their own modifications, alternative components or performance tweaks via pull requests or the “share‑your‑build” issue template. Notable community builds are highlighted in the README.

---

License: MIT – you can fork, modify and even sell your own version.