Energy-Efficient Fixed-Coefficient FIR Filters for Millimeter-Wave Radios

Abstract: With the introduction of millimeter-wave antenna arrays in 5G base-stations and ever-increasing data volumes, the power consumption of the signal processing in digital radio systems has increased over the last decade. This, combined with increased cost and environmental awareness, has put focus on power optimization. This thesis investigates different aspects of optimizing one important component of digital radio systems, fixed-coefficient FIR filters. The thesis initially investigates subblocks of the FIR filters separately. The cost of fixed coefficient multiplication is found to be linearly dependent on the number of signed digits in the coefficient set while additions are linearly dependent on the number of summed bits. The second part of the thesis considers complete filters. First, two common filter architectures are compared and the direct form is found to be more efficient. Then, a mixed-integer linear program is formulated that minimizes the number of signed digits in the filter coefficients. This optimization is shown to be able to reduce the number of signed digits with 20–30% compared to the median which translates to a similar reduction in area and power. Furthermore, allowing the filter gain to be flexible is found to reduce the optimal number signed digits additionally with around 20%. For filter specifications where there are multiple solutions with the same optimal number of signed digits, a ranking score based on other coefficient features is suggested. In the final part of the thesis, the optimal solutions from the second part are optimized further. An algorithm for finding common sub-expressions in coefficient sets is developed. Using these sub-expressions when implementing the filter is shown to reduce both area and power consumption with 5–25%.