Statistical and deterministic power optimization techniques using multiple supply and threshold voltages

Technology scaling has been the driving force behind the growth of the semiconductor industry over the past several decades. However, with process dimensions shrinking to the nanometer regime, a number of effects such as increased power dissipation and process variability, now limit performance. Leakage power, which has been observed to grow exponentially with process scaling, now contributes a significant fraction of the total power dissipation even in active modes of operation. In this dissertation, we address a range of crucial issues involved in power dissipation in current technologies. We develop a deterministic power optimization technique that is based on the use of multiple supply voltages in conjunction with threshold voltages, and is targeted at total power minimization. This is in contrast to most of the prior work in power optimization that focuses on either dynamic power or leakage power minimization independently, and is able to reduce power dissipation by 40% over benchmark circuits at normal activity. We also develop a framework to guide process designers to choose optimal supply and threshold voltages in their process to enable maximum benefits from such power optimization techniques.; The growing importance of leakage power is exacerbated by increased process variability. Leakage power is known to be highly susceptible to process variations and results in a wide distribution of leakage power, with high-leakage samples having up to 20X higher leakage compared to low-leakage samples. It therefore becomes imperative that low-power design techniques consider the impact of variations while performing optimization. However, most of the current work in statistical analysis and modeling has been targeted towards timing analysis. In this dissertation, we propose the first analytical technique to predict the distribution of leakage power---this is then used to develop a statistical leakage power optimization technique for the first time. This technique shows that statistical methods can be used to reduce leakage in deterministically optimized fast-corner parts by 50%. The final portions of this dissertation are aimed at performing true parametric yield analysis considering the impact of process variation on both power and performance, along with their correlation. This analysis technique is accurate and computationally efficient and therefore represents an ideal candidate to form the core of future parametric yield optimization techniques.