Cooling is underrated: full-duty fans took a MI50 from 6 tok/s to 20 tok/s on Qwen
bigattichouse · reddit · 2026-11-07
A Reddit user found cooling dominates local inference speed: a MI50 running Qwen at Q6 normally gets 6-8 tok/s. Pinning all fans to 100% duty cycle immediately bumped it to 12 tok/s, and after letting the machine cool down, inference hit 17-20 tok/s for the first ten minutes — no other changes.
The author is now considering water cooling with a large radiator, and urges anyone running local models not to underestimate thermal headroom.
More from Infra
- Epoch AI: Huawei's largest China AI data center holds just 7% of Colossus 2's compute — Jsevillamol · 2026-11-07
- Musk says SpaceX will launch StarMind orbital supercomputers from Q3 2027, co-designed with NVIDIA — XFreeze · 2026-11-07
- Netflix builds ontology-driven observability: telemetry, incidents and services as one knowledge graph — JeremyCMorgan · 2026-11-07
- Google buys Crusoe's Wyoming data center site as hyperscaler takeover of neoclouds begins — SumitGup · 2026-11-07
- Oracle Down 66% From Peak, 301 Days After Michael Burry Went Short — kingslyj · 2026-11-07
- AI supplier multiples fall ~15% post-selloff, below analyst growth forecasts — BenBajarin · 2026-11-07