A photo of the actual cluster is coming — the one shot no AI-generated site can fake.

The hardware

Four HP EliteDesk 705 G4 mini PCs, each with an AMD Ryzen 5 Pro 2400G (4 cores / 8 threads). They're palm-sized, quiet enough to sit in an office, and were bought used off corporate leases — the exact strategy from the hardware guide. No dedicated GPUs anywhere in the cluster: everything, including AI inference, runs on CPU.

NodeMachineCPURAMTypical role
Node 1HP EliteDesk 705 G4Ryzen 5 Pro 2400G16 GBCluster services, monitoring stack
Node 2HP EliteDesk 705 G4Ryzen 5 Pro 2400G8 GBOllama inference worker
Node 3HP EliteDesk 705 G4Ryzen 5 Pro 2400G8 GBOllama inference worker
Node 4HP EliteDesk 705 G4Ryzen 5 Pro 2400G16 GBOllama inference, heavier services

Running Proxmox VE 9.2 as a four-node cluster. Measured power-draw and monthly-cost numbers are getting their own post — real kill-a-watt readings, not spec-sheet estimates.

What runs where

Almost everything lives in LXC containers rather than VMs — the reasoning is inVM vs LXC. Each service below links to the guide that shows you how to build the same thing.

Why it's built this way

Three principles, learned the slow way. Used mini PCs beat new towers — the performance-per-dollar of off-lease business machines is unbeatable, and four small nodes teach you clustering in a way one big box never will. CPU-only is a feature — it keeps power draw and noise down, and proves how much of a homelab (including local AI) genuinely doesn't need a GPU. Everything gets documented — the cluster feeds asecond-brain knowledge vault, and the posts here are that vault's public face.

Want to build something similar? The Start Here roadmap walks the whole path in order, from first hardware to a monitored, backed-up cluster.

The name over the bench

This cluster is the working floor of Peira Labs (say PAY-rah) — from the Greek πεῖρα, "the trial, the attempt," the root of empirical. I chose the name because it describes what this shelf is for. A lab isn't a display case: it's where claims get tested. When a guide on this site says a command works, it's because that command ran here — on one of these four machines — and the note you're reading is what survived the attempt. The wrong turns got written down too; those are usually the parts that save you at eleven at night.

So when you see the Peira Labs name on this site, the newsletter, or the tools that come out of this cluster, it all points back to this bench: try it for real, measure it, keep the receipts.