It's workload-dependent. On-paper, ARM is more power-efficient at idle and simple ops, but slows down dramatically when trying to translate/compose SIMD instructions.