Undervolting a RTX 5090 Local LLM Rig Cut 150W and Actually Raised TPS 3%
MasterNomie · reddit · 2026-09-15
A local-LLM beginner shares a detailed tuning log running Qwen 3.8 27B NVFP4 on an i7-14700K + RTX 5090:
- GPU undervolt + VRAM OC: wrote LLM-generated benchmark scripts to sweep parameters and rank by TPS per watt. Power dropped 600W→<450W, temps 75°C→62°C, VRAM OC +2400MHz, TPS rose 172→178 (+3%), later 200 with Dflash2.
- CPU at 50% costs nothing: CPU only handles tokenization/scheduling; halving power changed TPS within noise, dropping CPU temps from 75-80°C to <55°C and silencing the AIO.
- Automation: a systemd service polls GPU usage every 30s and caps CPU at 50% when inference runs, plus a passwordless CLI to uncap for CPU-heavy jobs.
Result: from furnace to silent box with zero performance loss. Includes links to the 5090 undervolt guide and LACT tool.
More from Infra
- Your AI agent would cost 260% more without prompt caching, Arena data shows — arena · 2026-10-30
- Nebius runs auction-driven GPU pricing, lifting rates ~20% amid tight supply to 2028 — BenBajarin · 2026-10-30
- macOS update silently re-adds eMudhra root certs users had removed — steipete · 2026-10-30
- Bittensor's Terza Inference routes one API across miners' models on subnet 103 — const_reborn · 2026-10-30
- SGLang Summit 2026 sells out sponsor booths and Full Passes ahead of Nov 12–13 event — ying11231 · 2026-10-30
- OpenRouter rebuilds provider routing: 46% fewer capacity errors, 16.9% faster TTFT — steipete · 2026-10-30