R9700 at 160W for llama.cpp: ~20% pp loss, near-silent fans
LippyBumblebutt · reddit · 2026-11-19
To quiet a loud R9700 Pro cooler, a Reddit user benchmarked llama.cpp inference (recent ROCm build, -40mV undervolt) across 100W-300W power limits. The default floor is 210W; they had Qwen3.8-27B code a livepatch kernel module to remove the lower limit.
Findings: pp (prompt processing) scales roughly linearly down to 180W, tg (token generation) only drops below 140W. Going 300W→210W costs 10% pp; another drop to 160W costs 13% more, while tg stays within 5% throughout. At 160W the fan runs at 45% — audible but not annoying. Voltage from -80mV to 0mV made no difference at 140W. A worthwhile trade for quiet desktop use.
More from Infra
- US blocks Chinese firms from renting Nvidia chips abroad while Microsoft nears $500M OpenAI deal for ByteDance — kevinnbass · 2026-11-19
- vikhyatk: we are overinvesting in GPU compilers, underinvesting in CPU compilers — mohbibi_ · 2026-11-19
- KVARK Flash: a proxy that caps DeepSeek's thinking per request, 30-50% fewer tokens — Ill_Contribution_717 · 2026-11-19
- Polymarket launches Anthropic IPO perps, ANT trades near $104 ahead of Nasdaq debut — Polymarket · 2026-11-19
- Baseten runs DeepSeek v4.1 Flash at ~2,500 tok/s on Vera Rubin, 1.7x faster than GB300 — baseten · 2026-11-19
- Cloudflare launches Spend Management: auto-pause Workers billing at limits — JeremyCMorgan · 2026-11-19