Using dynamic branch behavior for power-efficient instruction fetch

Abstract

Power consumption has become an increasing concern in high performance microprocessor design in terms of packaging and cooling cost. The fetch unit including instruction cache contributes a large portion of the total power consumption in the microprocessor The instruction cache itself suffers some hidden power consumption due to dynamic control flows. Although capturing the dynamic control flows to boost performance, conventional trace caches (CTC) may increase power consumption in the fetch unit due to its simultaneous access to both the trace cache and the instruction cache. By avoiding this simultaneous accesses, sequential trace caches (STC) achieve lower power consumption, but suffer a significant performance loss at the meantime. In this paper we propose dynamic direction prediction based trace cache (DPTC) which avoids simultaneous accesses to the trace cache and the instruction cache with the guide of fetch direction prediction. Experimental results show that dynamic prediction based trace cache can achieve 38.5% power reduction over conventional trace caches and an additional 7.2% reduction over STC, on average, while only trading a 1.8% performance loss compared to CTC.