About gpu4.pics

A beautiful, open gallery of homelab & local-AI rigs. Every rig is a shareable card: the hardware, the model it runs, real tokens/sec, and — the part nobody else shows — perf-per-watt and the actual cost to run it.

Why it exists

The local-AI and homelab community produces an enormous stream of “show your rig” and benchmark content, but it’s trapped in Reddit threads, Discord scroll, and one-off blogs. There’s no canonical, beautiful, browsable home for it. gpu4.pics is that home — and every card is built to be posted back to the communities it came from.

How the numbers work

• Generation / prompt tok/s & TTFT come straight from your benchmark tool (llama-bench, ollama --verbose) or from LocalScore.
• Composite score follows LocalScore’s formula: 10 · ∛(prompt_tps · gen_tps · 1000/ttft_ms) (1000 ≈ excellent, 250 ≈ passable).
• Perf-per-watt = generation tok/s ÷ measured power draw.
• Electricity cost = (watts ÷ (tok/s · 3.6)) · $/kWh, at $0.17/kWh by default. This is electricity only — it excludes amortized hardware.
• VRAM fit is estimated from effective bits-per-weight per quant + a KV-cache approximation. It’s a guide, not a guarantee — treat the boundary as “try it.”

Where a value isn’t measured, we show “—” rather than guess. Transparency is the whole point.

Attribution

Seed benchmark data is drawn from LocalScore (an open, Apache-2.0 benchmark project by Mozilla Builders) — each card links back to its source. Community rigs are credited to their original posters with a link to the original thread; if a rig is yours, you can claim its card. We don’t re-host anyone’s photos.

Open source

gpu4.pics is open-source under Apache-2.0. The benchmark numbers are facts; the gallery is a gift to the community that produced them.