Reddit User Shows Strong Cooling Significantly Boosts Local LLM Inference Speed

On 2026-11-07, Reddit user bigattichouse posted in the LocalLLaMA community describing a hands-on experiment that suggests aggressive cooling can materially speed up local LLM inference. According to the post, the user had been running Qwen for local coding work on an MI50 at Q6 quantization, at a historical rate of roughly 6-8 tokens per second. The user then wrote a script that drives the graphics card fans to a 100% duty cycle, which is noticeably louder, and immediate inference throughput rose to about 12 tokens per second. After letting the machine sit and cool with the fans running, a further inference run reached 17-20 tokens per second for the first 10 minutes or so. The user says they are now considering adding water cooling and a larger radiator, and describes the gain as a huge difference. The story so far consists of this single community post. No vendor has published independent benchmarks, and no temperature readings, clock speeds, or the exact Qwen version were provided, so the size and mechanism of the gain remain to be verified by others.