Homelab · self-hosting · local AI
Build the box.
Measure the watts.
We run a small test bench of local-AI and always-on boxes and write up what they actually draw, how hot they get and what they are worth. Every part was bought with our own money, and every number was read off a meter, not a spec sheet.
See the builds →HOW EVERY ARTICLE IS MADE
Bought. Run. Measured.
Idle and load power from a wall meter. No vendor numbers.
Start with the question in front of you
- Two GPUs. One board. 48 GB.Build logBest budget GPU for local LLMs in 2026: 6 GB, 8 GB or 12 GB, measured
September 2026 prices and our bench numbers: 12 GB is the line that matters for local LLMs. A refurbished RTX 3060 12 GB beats a new 8 GB card for Ollama on a $250–$550 budget.
Latest from the lab
- comparisonsRTX 3060 Ti vs RTX 2060 for Ollama: we benchmarked each card alone
We benched an RTX 3060 Ti and an RTX 2060 alone in Ollama. The 2060 runs 7B models at 76% speed, drops to 61% on llama3.1:8b, and neither card fits a 14B model.
- guidesNVIDIA power limits for local LLMs: we swept 100 to 200 W so you can cap with confidence
We swept an RTX 3060 Ti from 100 to 200 W under Ollama. 150 W cost 3% speed for 25% less draw. Here's the cap we'd pick, the nvidia-smi commands, and how to reapply it at boot.
Own hardware, own meter, own money. Affiliate links are disclosed on every page and never decide a verdict. Read our editorial policy.