Alternative Least Mean Square adaptive filter architectures for implementation on Field Programmable Gate Arrays

Abstract

The Least Mean Square (LMS) adaptive filter is a simple well behaved algorithm which is commonly used in applications where a system has to adapt to its environment. Architectures including the direct, transposed and hybrid forms are examined in terms of the following criteria: speed, power consumption and FPGA resource usage. Both the transposed and hybrid forms, which are derived from the delayed LMS, allow for higher speeds without significant increases in power or area. Results for both these adaptations are independent of filter length with the maximum speed of the 16 tap transposed form being over 4 times greater than the speed of a 16 tap direct form implementation. For FPGA implementation, the transposed form is optimal, as power and area are not significantly greater than values found for the direct form, despite the higher maximum frequency. Even at greater numbers of taps, the maximum frequency of the transposed form is not degraded, despite the input data bus driving an increased number of multipliers.