From GPU Mining Rig to Local Inference Tower

ASIC hashboards do not become GPUs. An S19 does not run Llama. The reusable pieces are power, airflow, and the habit of measuring watts. Corporate version of that split: hashcenter migration.

What survives from a mining frame

• PSU headroom and breakout. Inference is bursty. Cheap consumer PSUs hate 200 W idle → 1,000 W in a hundred milliseconds. Mining-era server PSUs plus a breakout board still make sense. Credit the homelab and old Bitcointalk threads that documented this.
• Open-air or 4U chassis with real exhaust. Desk-silent it is not.
• 120 mm fan language you already know from shrouds.

What does not survive

• Risers that were only ever 1x for mining. Inference wants the GPU in a real x16 slot when you can give it one.
• Undervolt BIOS habits aimed at hash/watt, not stable VRAM.
• The fantasy that six leftover 8 GB mining cards equal a 48 GB unified pool. They do not.

Conversion path

1. Inventory: GPU model and VRAM, PSU wattage and rails, motherboard chipset, whether it is 120 V or 240 V.
2. Pick a model class that fits that VRAM. 3090 guide, fit tool.
3. Install a current NVIDIA or AMD driver stack, then Ollama or vLLM. Credit those projects.
4. If the frame is a fire hazard or the PSU is tired, we quote a new tower rather than “save” it.

Quote: custom GPU inference rig. Heat reuse: heating with inference. Bitcoin-layer hardware (Bitaxe, heaters) stays one click away on this miner-shaped page — not on a CFO Law 25 article.

Custom GPU rig, Canada · All three computers