← sources

Budget homelab built from hand-me-down PCs: OptiPlex on Proxmox + a Haswell box with a 1080 Ti for local LLMs

source: reddit (reddit:homelab)published 2026-10-11 11:31ingested 2026-10-12 02:38

retained text

Everything here is second-hand or was lying around. I'm a student running a small  
web dev side business, and the lab is where I learn and host my own tools.

\*\*Hardware\*\*

\*Main server "pve2": Dell OptiPlex 5050\*  
\- i3-7100, 32 GB DDR4, Crucial MX200 SATA SSD (system), Kingston A1000 256 GB NVMe (backups)  
\- Proxmox VE 9.2  
\- Yes, it's on Wi-Fi (MediaTek Wi-Fi 6 card, 5 GHz). No cable reaches that room.  
Wi-Fi can't be bridged, so all guests sit behind a NAT bridge with port forwards.

\*AI box "aipc": self-built\*  
\- ASRock H87 Pro4, i5-4590, 16 GB DDR3, GTX 1080 Ti 11 GB  
\- Ubuntu Server 26.04, Nvidia 580 driver, Ollama  
\- GPU power-limited to 200 W because the PSU is an old 530 W Thermaltake  
\- Ethernet over powerline plus USB Wi-Fi, with a small failover script that moves the IP

\*Retired: old tower (Gigabyte B85M-HD3G, i5-4570, 8 GB)\*, the previous Proxmox host,  
planned to become an OpenMediaVault NAS.

\*3D printer: Anycubic Kobra 2 Pro (bought used)\*  
\- Not fully working yet. It came second-hand with some issues I'm still sorting out  
\- Plan is to print parts for the lab once it runs reliably (fan shrouds, drive brackets, cable management)

\*\*What runs on it (12 LXCs + 1 VM)\*\*  
\- AdGuard Home (DNS for the whole network, HaGeZi lists)  
\- Home Assistant OS (VM) with Matter over Thread: Sonoff Dongle Plus MG24 as  
OpenThread border router, IKEA KAJPLATS bulbs  
\- Open WebUI + SearXNG (Docker), using the models on the AI box  
\- Samba file share  
\- Self-written tools: invoicing app, a read-only fleet dashboard (Node/Fastify/SQLite),  
a PCB documentation tool, backend for my own iOS food/training tracker (Whisper for voice input)  
\- A staging container for websites  
\- A container for AI coding agents  
\- An isolated "test lab" container on its own bridge (internet only, snapshot rollback)  
where local models get to run Linux commands before I trust them with anything  
\- Tailscale for remote access (HTTPS via Tailscale Services), no open ports at home

\*\*Local LLMs on the 1080 Ti\*\*  
\- Qwen 35B-A3B (MoE, IQ4\_XS), split GPU/CPU: \~21-27 tok/s  
\- Gemma 26B-A4B QAT: \~27-29 tok/s  
\- bge-m3 for embeddings  
\- Pascal is slow at FP16, so GGUF via Ollama/llama.cpp only

\*\*Backups and monitoring\*\*  
\- Nightly vzdump of all guests to the NVMe (7 daily / 4 weekly), failures go to Telegram  
\- restic for my two rented VPS, the home server pulls copies hourly  
\- Monitoring is plain bash + curl + systemd timers sending to Telegram: the machines  
watch each other, because a server can't report its own death  
\- Temperature watchdog that warns and shuts down cleanly on overheating

\*\*Lessons learned\*\*  
\- Measure disks before planning: my 15-year-old Seagate is healthy, a 2020 WD Blue  
with 505 hours is dead  
\- Unplug the PSU before swapping RAM. I killed a stick on standby power.  
\- The "GPU not detected" problem was a bent slot bracket

\*\*Next\*\*  
\- Tesla M40 24 GB is bought, with a used 850 W PSU and a blower shroud. Open question  
is whether the H87 board posts without Above 4G Decoding.  
\- Turn the old tower into a NAS and get the backups off the same machine  
\- Get a cable to the OptiPlex  
\- Get the used Kobra 2 Pro printing reliably

Happy to answer questions, and roast the Wi-Fi server, I know.

document fetched 2026-10-12 00:32 · canonical https://reddit.com/r/homelab/comments/1x35wmc/budget_homelab_built_from_handmedown_pcs_optiplex