One Box, Twelve Threads: My Homelab on a Single Mini PC

Most homelab stories I read start with a rack. Mine starts with one small box and two rules.

My day job is designing solutions for clients. At home I wanted the opposite: a machine I could change on a Sunday afternoon, break, and fix, without a change request. But it also had to be useful for the household, not just a toy.

So before buying anything, I wrote down two rules:

1. It’s a learning machine. It can fail, and I accept that.
2. When it fails, the house falls back to how it worked before. Internet still on, nothing important lost.

Almost every decision below comes from those two lines.

Five boards later

#

I went through more options than I want to admit: an HP ProDesk 400 G4, a Beelink SER8, a Minisforum MS-01, then a board with an i7-12650HX engineering sample. I ended up with a board sold as an i7-1255U, plus:

• 32 GB RAM (2 × 16 GB)
• One 1 TB SSD
• The integrated GPU (it matters later)

A mobile-class chip is a good fit for a box that runs 24/7: small, quiet, and easy on the power bill.

The CPU that isn’t what the box says

#

The first thing I ran was lscpu. Trimmed:

Model name:      Genuine Intel(R) 0000
CPU(s):          12
CPU max MHz:     4000.0000
L3:              18 MiB (1 instance)

Threads 0–7 sit on four P-cores and threads 8–11 on four E-cores. A retail i7-1255U doesn’t look like that.

Same thread count, different chip. “Genuine Intel(R) 0000” is the name you usually see on engineering samples, so my best guess is a pre-release part that ended up on a cheap board.

For a homelab, I’m fine with it. Four P-cores are nice for VMs, and I give up some single-core boost. For a production server, I would not take this risk: no official support, and no promise on stability.

Lesson: run lscpu on day one, before you trust the listing.

What I decided not to run

#

Some of my best decisions were the things I left out.

• No media server. I watch through debrid + Stremio, so I keep at most one or two films locally.
• No HDD array. One 1 TB SSD. The biggest user is about a week of camera footage, and that fits.
• No virtual router. I planned OPNsense, then dropped it. The ISP gateway stays the router. If my box dies, the internet doesn’t.

That last one is rule 2 in action. A virtual router is a great learning project, but it turns every Proxmox reboot into a household outage.

The stack

#

Proxmox VE runs on the metal, with the default storage split: about 100 GB local and about 800 GB local-lvm. On top of it:

Two VMs for the big things

• Coolify (4 vCPU, ~12 GB RAM) runs my Docker apps. The first app was Stirling-PDF; n8n with Postgres came next.
• Home Assistant runs in its own VM.

One LXC per small job

The split follows a simple rule of thumb. Apps I add and remove often go into Coolify. Infrastructure that other things depend on gets its own LXC. Monitoring especially: if the Coolify VM breaks, the tools watching it should not break with it. Uptime Kuma and Beszel each went in with the Proxmox community helper scripts.

The network: boring on purpose

#

The network is simple, and I want it that way:

• The ISP’s 10G PON gateway is the router.
• Two Tenda Nova mesh nodes run in AP mode on the same subnet.
• One camera plugs into a mesh node.
• Everything lives in 192.168.1.0/24. AdGuard Home sits at .53, because of course it does.

Tailscale runs in its own LXC, not on the Proxmox host. It advertises 192.168.1.0/24 as a subnet route and works as an exit node. From anywhere, my phone sees the home network as if I were on the sofa. On café Wi-Fi, I can send all my traffic out through home.

Public names, private addresses

#

This is my favorite trick in the whole setup. My domain htknguyen.dev uses Cloudflare for DNS:

So pdf.htknguyen.dev resolves for anyone in the world, but the answer is a private address. On my LAN or my tailnet, it opens Stirling-PDF. From anywhere else, it goes nowhere.

What I get:

• Clean names for every app, with no port forwarding and nothing exposed to the internet.
• A new app in Coolify gets a new subdomain with no DNS change, thanks to the wildcard.

The trade-off: the names are not secret. But a name is not a way in.

Where the homelab meets the house

#

Home Assistant is where the homelab touches real life.

The house runs on solar: 12 × 570 W panels (about 6.84 kWp), all facing west at a 15° tilt, on a LuxPower SNA6000 hybrid inverter with a battery. We don’t sell power back to the grid. EVN bills homes on a tiered tariff, so every kWh we make and use ourselves is a kWh that never reaches the expensive upper tiers. West-facing panels produce most in the afternoon, which is also when the house is hottest.

What’s connected so far:

• A Broadlink IR/RF hub in the bedroom controls the fan and the AC. The fan has no physical remote, so the hub is the remote.
• Frigate handles the cameras, sized for about a week of footage on the SSD.

Next on the list: inverter data (it has a Wi-Fi dongle), the Samsung AC, and monitoring the Tenda mesh, all inside Home Assistant.

War stories

#

IPv6 that looks alive

#

While setting up cloudflared in an LXC, the connection to Cloudflare just hung.

The cause: my LXCs get an IPv6 prefix from VNPT through router advertisements, so IPv6 looks ready. But outbound IPv6 connections hang, while IPv4 works fine. The fix was one flag:

--edge-ip-version 4

Lesson: when something hangs instead of failing, suspect IPv6.

n8n settings that don’t stick

#

n8n runs on Coolify from the n8n-with-postgresql template: n8n, task runners and Postgres 16, with the version pinned in the compose file. N8N_HOST, WEBHOOK_URL and N8N_EDITOR_BASE_URL are defined in that compose file, so the Environment tab is the wrong place to change them. Edit them in the compose editor.

The tunnel I built and removed

#

I wanted a stable webhook URL for POCs, to stop juggling ngrok links. I set up Cloudflare Tunnel with hooks.htknguyen.dev in front of n8n. Then I decided to tear the whole thing down and put n8n back to its original config.

Not every experiment has to stay. The IPv6 lesson stayed; the tunnel didn’t.

What’s next

#

• Game streaming. I can’t run a cable from the server to the TV, so the plan is Wolf (Games on Whales) with Moonlight, using the iGPU: PS2 games, NES-era tank games, maybe Project Diablo 2.
• A call-center demo that calls real phones, running on this box.
• More Home Assistant: inverter data, the Samsung AC, the mesh.
• Backends for side projects, like sync servers for apps I’m building.

What I’d tell myself on day one

#

• Write down what you will not run. It saves more time than any tool.
• Keep the router out of the lab.
• Big things in VMs, small infrastructure in LXCs, apps in Coolify.
• Private DNS plus Tailscale beats port forwarding.
• If it hangs instead of failing, suspect IPv6.
• Run lscpu before you trust the listing.

It isn’t a rack. It’s one small box, twelve threads, and a list of things I learned by breaking them. That’s exactly what I wanted.