Three scalable approaches to improving many-core throughput for a given peak power budget