The Hardware
This is what is running. I’m using cloud models while I learn, with local inference testing on the AI HAT+.
What Is Running This
Raspberry Pi 5 (8GB)
8GB RAM. ARM64. Low power, always on. The main computer.
Raspberry Pi 5 (16GB)
16GB RAM. ARM64. More memory for larger models.
Mini PC
16GB RAM. AMD Ryzen 4300U. For heavier workloads.
AI HAT+ (21 TOPS)
8GB RAM on the NPU. Testing local models.
OpenClaw
The agent framework. Skills-based architecture. Community-driven.
Ollama Cloud
GLM-5, Nematron via Ollama. Cloud models while I explore local inference.
Work in Progress
Real experiments. Real results. I share what I learn as I go.
Running Local AI
I’m experimenting with running small AI models directly on the AI HAT+. No internet needed. Still figuring out what works. If it does, I’ll have AI that runs even when the connection goes down.
Fast Storage – NVMe Experiments
What I Tried
Connected a fast NVMe SSD via USB to the Pi. Real-world result: about 32 MB/s. The USB connection was the bottleneck, not the drive.
What I Learned
Pi 5 USB 3.0 ports max out around 300-400 MB/s in ideal conditions. Cheap or older USB cables make it worse. The SSD itself is rated for over 1,000 MB/s, but the pipe was too narrow.
Next Plan
I have a 3.6TB NVMe drive. Going to connect it directly to the Pi’s PCIe slot instead of through USB. That should unlock the drive’s real speed: 3,500 MB/s.
Trade-off
The PCIe slot is also where the AI HAT+ goes. I can have fast storage OR local AI inference, not both at the same time. Figuring out which matters more right now.
Want to Build Something Similar?
This is my learning journey. If you want to build your own, or partner on OpenClaw, reach out.
YouGotThisAI LLC
© 2026 All Rights Reserved