April 6, 2026
Description
I had a GPU lying around and wanted to get my feet wet with local LLMs. The design goal was to build a small form factor, power efficient AI rig with what I mostly had lying around at home.
The result: a cyberpunk-inspired mini PC that runs 7B parameter language models with GPU acceleration, generating around 48 tokens per second.
The parts are designed to require as little support as possible. You will need to print:
1x PSU Front
1x PSU Back
1x RPI5 Bracket
1x Label (Arasaka) - optional, you can also use "Label (Blank)" and put whatever design you want on it
I recommend printing in a material that won't warp at higher temperatures - I went with PETG. All parts are oriented the way they should be printed. Only "PSU Back" needs support (I used manual tree support in Orca Slicer, painted only on the 90 degree overhang).
For the label: print in two colors, no MMU required. Just swap filament at the layer where the letters start to extrude.
Assembly is straightforward:
PSU enclosure - Screw front and back PSU parts to the HDPlex. Screw the power outlet to the back part. There's enough room inside for wiring - just make sure the plug orientation matches your preference. Matching screws come with the HDPlex.
Pi mounting - Mount your RPi 5 to the bracket and attach the PCIe HAT. I used M2.5x10mm screws through the bracket. The standoffs that come with the PCIe HAT work fine, but the included screws had oversized heads - I recommend sourcing some with smaller heads.
GPU - Slide the GPU into the PCIe slot and mount the PCIe bracket in the back PSU part. Some wiggling may be required. Let me know if you can't get it to fit - I can easily adjust that part.
Label - Slide into place. You can secure it with glue, but I didn't find it necessary. Leaving it loose lets you swap designs.
Wiring - Connect the 8-pin cable to the GPU. One 4-pin 12V CPU cable connects to the PCIe HAT. For the Pi's 5V power, I tapped the 5V line from the SATA power cable and connected it to the Pi's 5V GPIO pins. You could also power it via USB-C if you prefer (less clean, but works). If you're doing custom sleeving, you know the drill.
Software - See below.
Raspberry Pi 5 (8GB RAM recommended)
Active Cooler (better safe than sorry)
MicroSD Card (256GB+ recommended, Class 10 / A1)
HDPlex 250W GaN PSU
AMD GPU of your choice - I use an Asus RX 5600 XT. Any card supported by the open source amdgpu driver works. Must be AMD - NVIDIA cards do not have full driver support on RPi 5 at the time of writing.
4x M3x8mm screws
4x M2.5x10mm screws
Cables, sleeving and plugs (optional - the HDPlex comes with all necessary wiring)
For the software stack, I built R.O.A.S.T. (Radeon On ARM, Serving Tokens) - a setup script and model manager that handles everything automatically:
Installs a kernel with amdgpu support
Builds llama.cpp with Vulkan backend
Manages models as systemd services
Optionally sets up Open WebUI for a chat interface
One-liner install on a fresh RPi OS Trixie Lite:
wget -qO- https://raw.githubusercontent.com/stylesuxx/roast/master/install.sh | sudo bash
See the R.O.A.S.T. repo for full documentation, GPU selection guide, and integration with coding tools like aider and Continue.
Obviously I don't hold any rights on Cyberpunk property. CDPR - please don't sue me. This is just fan work, I love Cyberpunk...
License:
Creative Commons — Attribution — Noncommercial — Share Alike