An evolutionary computational approach to configuring portable embedded system architectures

Portable embedded systems (e.g., medical equipment, cellular phones, pagers, and video game consoles) are being driven by consumer demands to be thermally efficient (produce less heat), perform faster, and have longer battery life. To design such a system, various hardware units (e.g., level one (L1) and level two (L2) caches, functional units, registers) are selected based on a set of specifications for a particular application. Currently, chip architects are using software tools to manually explore different configurations, so that tradeoffs for consumption, performance, and chip size may be understood.; However, when evaluating multiple design parameters simultaneously, the exploration space expands, design time increases and human errors become a concern. Genetic algorithms, which are effective in rapid global search of large and poorly understood spaces, have been modified for multiple objectives and applied to guide this process to an improved solution.; This dissertation presents a framework for an evolutionary approach to configuring an "ideal" embedded processor based on power consumption and performance. In addition, a database of simulation results that gives a more comprehensive evaluation of tradeoffs between power and performance, and of inter-dependence between parameter configurations is presented. Appropriate search techniques to reduce exploration space and decrease time-to-market are also discussed.