156-run benchmark: RTX 3070 power limits vs local LLM efficiency, Gemma4 wins

bulletrhli · reddit · 2026-09-27

A data engineer ran 156 benchmarks (4 models × 6 power limits from 100-220W × 3 runs) on a Lenovo M920Q with an i5-8500T and an RTX 3070 8GB over OcuLink, using Proxmox LXC + Ollama + OpenWebUI, logging tokens/s, duration, and a tokens-per-watt metric.

Key findings:

Below a threshold, the power limit acts as the ceiling—useful guidance for anyone tuning local deployments on tight hardware.

Original post →

More from Infra

Infra channel →