New and complex systems are being implemented using highly advanced Electronic Design Automation (EDA) tools. As the complexity increases day by day, the dissipation of power has emerged as one of the very important design constraints. Now low power designs are not only used in small size applications like cell phones and handheld devices but also in high-performance computing applications.
Embedded memories have been used extensively in modern SOC designs. In order to estimate the power consumption of the entire design correctly, an accurate memory power model is needed. However, the memory power model commonly used in commercial EDA tools is too simple to estimate the power consumption accurately.
For complex digital circuits, building their power models is a popular approach to estimate their power consumption without detailed circuit information. In the literature, most of power models are built with lookup tables. However, building the power models with lookup tables may become infeasible for large circuits because the table size would increase exponentially to meet the accuracy requirement.
This thesis involves two parts. In first part it uses the Synopsys power measurement tools together with the use of synthesis and extraction tools to determine power consumed by various macros at different levels of abstraction including the Register Transfer Level (RTL), the gate and the transistor level. In general, it can be concluded that as the level of abstraction goes down the accuracy of power measurement increases depending on the tool used. In second part a novel power modeling approach for complex circuits by using neural networks to learn the relationship between power dissipation and input/output characteristic vector during simulation has been developed.
Our neural power model has very low complexity such that this power model can be used for complex circuits. Using such a simple structure, the neural power models can still have high accuracy because they can automatically consider the non-linear power distributions. Unlike the power characterization process in traditional approaches, our characterization process is very simple and straightforward. More importantly, using the neural power model for power estimation does not require any transistor-level or gate-level description of the circuits. The experimental results have shown that the estimations are accurate and efficient for different test sequences with wide range of input distributions.