Novel source-independent characterization methodology for embedded software energy estimation and optimization

In order to design a successful low-energy VLSI system, concurrent energy reduction at hardware and software levels is needed. The available techniques for embedded software energy estimation either provide unusable average-case results or require prohibitively complex hardware setups for cycle-accurate results. This paper introduces a new methodology for high-level software energy estimation for embedded systems. The methodology produces cycle-accurate results independent of the energy characterization process. The executed instructions as well as the transitions on the wires are taken into consideration for estimating the energy. This allows tradeoff between the accuracy and the complexity of the model. The methodology is generic and makes no assumptions about the measurement techniques or the architecture of the processor. The introduced methodology also allows successive improvements in the estimation accuracy with each step towards final silicon. The embedded ARM7TDMI RISC processor is modeled with this methodology and the errors are found to be less than 10%. For energy optimization, the model provides excellent relative accuracy too. Taking advantage of the relative accuracy, different code transformation techniques are discussed and employed to gain 32% energy savings.