For anyone building ducts, cages, or fan mods in this thread: before you spend more time on hardware, read the software‑only approach here:
**Cooler GB10 Temps – (almost) no performance lost
**
Key points:
- Capping the GPU clocks (e.g.
sudo nvidia-smi -lgc 0,2000) gives essentially the same real‑world performance for many LLM/image workloads, with much lower power and temps. - A simple closed‑loop governor that holds your hottest thermal zone at a target temperature turns any cooling headroom directly into extra clocks, so better cooling shows up as MHz, not just °C.
If you’re evaluating a duct or cage, use that governor + hottest‑zone logging and report “this duct is worth X MHz at a fixed thermal target” instead of just “temps are lower”.