I don't see how "poorly parallelized" is relevant compared with, say, a decent OpenMP BLAS which is likely to benefit from any reduced power consumption that al

I don't see how "poorly parallelized" is relevant compared with, say, a decent OpenMP BLAS which is likely to benefit from any reduced power consumption that allows a faster clock. In general on compute nodes it's worth turning off cores dynamically which aren't being used by jobs, like on a half-full node.