6. Conclusions

as the level of inputs correlation changes, but also the values of total power consumption. For example, for duke2, the total power estimated under weakly correlated inputs was 3611.67 uW, while this value for strongly correlated inputs was 820.87 uW (there is a factor of 4 between the two). The same behavior has been observed for other circuits.To conclude, input pattern dependencies (in particular highly correlated inputs) is an extremely important issue in power estimation, despite other claims which assume independency and randomness on the primary inputs (or worse, throughout the circuit). From this perspective, power analysis needs analytical models to overcome this difficulty. The model we proposed here, based on conditional independence hypothesis while accounting for spatiotemporal correlations, is an efficient and robust analytical solution to this problem. We have proposed an efficient approach for power estimation in large combinational blocks fed by input streams which exhibit high levels of correlation. The work reported here addresses the relationship between logic and probabilistic domains and gives a sufficient condition for analyzing complex dependencies. From this perspective, the new concepts of conditional independence and isotropy of signals are used in a uniform manner to fulfill practical requirements for fast and accurate estimation. Under general assumptions, the conditional independence problem has been shown to be NP-complete; consequently, efficient heuristics have been provided for probabilities and coefficients calculation. A comparative analysis with the existing techniques and evaluations on benchmarks emphasize the effectiveness and universality of our approach.