3 ms·

Unneeded when using local models, as every workload produces a novel pattern of coil whine from the GPU.