Increasing adder efficiency by exploiting input statistics

Abstract

Current techniques for characterizing the power consumption of adders rely on assuming that the inputs are completely random. However, the inputs generated by realistic applications are not random, and in fact include a great deal of structure. Input bits are more likely to remain in the same logical states from addition to addition than would be expected by chance and bits, especially the most significant bits, are very likely to be in the same state as their neighbors. Taking this data, I look at ways that it can be used to improve the design of adders. The first method I look at involves looking at how different adder architectures respond to the different characteristics of input data from the more significant and less significant bits of the adder, and trying to use these responses to create a hybrid adder. Unfortunately the differences are not sufficient for this approach to be effective. I next look at the implications of the data I collected for the optimization of Kogge- Stone adder trees, and find that in certain circumstances the use of experimentally derived activity maps rather than ones based on simple assumptions can increase adder performance by as much as 30%.