As more and more transistors fit in a single chip, consumers of the electronics industry continue to expect decline in cost-per-function. Advancements in process technology offer steady improvements in system performance. The improvements manifest themselves as shrinking area, faster circuits and improved battery life. However, this migration toward sub-micro/nano-meter technologies presents a new set of challenges as the system becomes extremely sensitive to any voltage, temperature or process variations. One approach to immunize the system from the adverse effects of these variations is to add sufficient safety margins to the operating clock frequency of the system. Clearly, this approach is overly conservative because these worst case scenarios rarely occur. But, process technology in nanoscale era has already hit the power and frequency walls. Regardless of any of these challenges, the present processors not only need to run faster, but also cooler and use lesser energy. At a juncture where there is no further improvement in clock frequency is possible, data dependent latching through Timing Speculation (TS) provides a silver lining. Timing speculation is a widely known method for realizing better-than-worst-case systems.

TS is aggressive in nature, where the mechanism is to dynamically tune the system frequency beyond the worst-case limits obtained from application characteristics to enhance the performance of system-on-chips (SoCs). However, such aggressive tuning has adverse consequences that need to be overcome. Power dissipation, on-chip temperature and reliability are key issues that cannot be ignored. A carefully designed power management technique combined with a reliable, controlled, aggressive clocking not only attempts to constrain power dissipation within a limit, but also improves performance whenever possible.

In this dissertation, we present a novel power level switching mechanism by redefining the existing voltage-frequency pairs. We introduce an aggressive yet reliable framework for energy efficient thermal control. We were able to achieve up to 40% speed-up compared to a base scheme without overclocking. We compare our method against different schemes. We observe that up to 75% Energy-Delay squared product (ED2) savings relative to base architecture is possible. We showcase the loss of efficiency in present chip multiprocessor systems due to excess power supplied, and propose Utilization-aware Task Scheduling (UTS) - a power management scheme that increases energy efficiency of chip multiprocessors. Our experiments demonstrate that UTS along with aggressive timing speculation squeezes out maximum performance from the system without loss of efficiency, and breaching power & thermal constraints. From our evaluation we infer that UTS improves performance by up to 12% due to aggressive power level switching and over 50% in ED2 savings compared to traditional power management techniques.

Aggressive clocking systems having TS as their central theme operate at a clock frequency range beyond specified safe limits, exploiting the data dependence on circuit critical paths. However, the margin for performance enhancement is restricted due to extreme difference between short paths and critical paths. In this thesis, we show that increasing the lengths of short paths of the circuit increases the margin of TS, leading to performance improvement in aggressively designed systems. We develop Min-arc algorithm to efficiently add delay buffers to selected short paths while keeping down the area penalty. We show that by using our algorithm, it is possible to increase the circuit contamination delay by up to 30% without affecting the propagation delay, with moderate area overhead. We also explore the possibility of increasing short path delays further by relaxing the constraint on propagation delay, and achieve even higher performance.

Overall, we bring out the inter-relationship between power, temperature and reliability of aggressively clocked systems. Our main objective is to achieve maximal performance benefits and improved energy efficiency within thermal constraints by effectively combining dynamic frequency scaling, dynamic voltage scaling and reliable overclocking. We provide solutions to improve the existing power management in chip multiprocessors to dynamically maximize system utilization and satisfy the power constraints within safe thermal limits.