Energy and Power Aware Computing Through Management of Computational Entropy

摘要

With the increasing importance of computing in portable and other embedded settings, the power and energy consumed by computation has become an important consideration driving research in computer engineering and science. Historically, the study of power and energy have roots in the field of thermodynamics. The innovations that we propose in this work aim to develop a mechanism for designing algorithms that are power (energy) aware by directly working with their thermodynamic properties, as opposed to the traditional approaches wherein algorithm design is concerned with measures such as their running time, and space consumed. Overall, our research scope was to innovate energy aware algorithms, novel semiconductor circuits, methodology and models for energy aware algorithm design by the application of thermodynamics and randomization toward algorithm and semiconductor design and analysis. As a result of this research work, we innovated probabilistic CMOS or PCMOS technology, as a promising approach to addressing the CMOS device scaling challenges and as a dramatic shift from previous work. Devices based on this technology, where noise is harnessed as a resource to implement CMOS devices exhibiting probabilistic behavior, are guaranteed to compute correctly with a probability p. Additionally, this work also characterized an explicit relationship between the probability p with which the CMOS switch computes correctly, and its associated physical attributes such as the energy consumed by each switching step across technology generations. We extended the above characterizations into PCMOS laws governing the behavior of such devices and switching. There laws were also validated across various technology generations via simulations as well as physical measurements of PCMOS inverters.