On IVB, it appears MKL DGEMM decides to run on only eight cores when it is told to run on nine. Behaviour on SNB, HSW, BDW is fine. I tried different IVB chips, but to no avail. When instructed to run on ten cores, all ten cores are used (so it doesn't appear to be a thread pinning issue). I've seen irregularities on IVB chips in RAPL reported power consumption when going from eight to nine to ten cores. Is this related and expected behaviour (i.e. an optimization) because DGEMM knows something I don't? Is there a workaround to force it to use nine cores?
