Reducing microprocessor power consumption can alleviate expensive thermal management, improve reliability, allow increased performance, and reduce carbon emissions. Microarchitecture techniques that reduce instruction occupancy in the out-of-order engine to ultimately create more empty segments that can be disabled provide an attractive approach to reducing power consumption. This thesis presents two guiding principles to reduce occupancy for implementing such an approach. One involves downsizing the out-of-order engine as it becomes less effective in issuing instructions out of order. The second reduces instruction traffic by dampening traffic bursts. This thesis developed two techniques that implement these principles. One downsizes the out-of-order engine based on a data dependence index (DDI), and the second moderates the instruction traffic based on the sustained dispatch rate (SDR). The SDR, in addition to reducing occupancy, reduces the i-cache's power consumption. Total processor power savings are reported. Component power savings are also reported to show their relative effects on total savings. Simulations find that occupancy was reduced from 27% to 33% in the out-of-order engine and 40% in the fetch buffer. These occupancy reductions are, to our knowledge, the largest reported to date, with a performance loss of only 2.6%. Reduced occupancy results in fewer active segments, which directly contributes to power savings. Power models estimate that the savings for a single threaded, high performance processor can total to 27% of overall dynamic power consumption, including level-1 caches.
Read
