MLCommons consortium refines benchmark for low-power embedded neural motors