Power modeling of embedded memories

After Moore’s Law the number of transistors on an integrated circuit doubles every 18 months. New circuits are furthermore clocked with increasing frequencies. This development not only leads to an increase of the available functionality, but also to a rise of the electrical power consumption of these systems. The power consumption of integrated circuits is problematic in two respects: on one hand the power must be fed into the system, on the other hand the heat produced on the chip must be dissipated. An increased power consumption consequently leads to reduced battery lifetimes and increased energy costs. Today already 80% of the office related power consumption in the USA stems from computers. The heat development of integrated circuits reduces their reliability and lifetime. The required cooling measures (ceramic packages, heat fins, fans, etc.) furthermore increase the system cost. For the development of low power systems it is necessary to estimate and consider the power consumption already in the early stages of design. For such estimates models of the system blocks are required. This thesis describes a methodology for the generation of models of the power consumption of embedded memories. Memories have a special importance among the system blocks as it is forecast that more than 90% of the area of newly developed systems will be occupied by memory within ten years. In addition to the requirements on models, like accuracy, speed and mathematically closed form the presented approach specifically adresses the requirements on the modeling procedure. These are mainly the protection of intellectual property and low modeling costs. As a key point of the approach a method for the generation of nonlinear (signomial) models from empirical data is presented. This regression based method allows the generation of piecewise model functions, which can be used for optimisation through geometric programs. The presented models reduce the prediction error by up to 95% compared to existing approaches.