A High Accuracy and Hardware Efficient Adaptive Filter Design with Approximate Computing.

Least mean square (LMS) adaptive filters has been widely used in the fields of channel equalization, system identification and noise cancellation et al. Its cost function is essentially an approximation of the Winner filter, which consumes large amount of hardware resource with many multipliers and adders. However, due to the closed loop architecture, some errors can be eliminated after several iterations with careful design. In this paper, a high accuracy and low complexity approximate LMS adaptive filter is proposed by exploring approximate multiplier with specific error sign and system-level compensation strategy. The error generated by proposed approximate multiplier can be always compensated in the next iteration, which is accumulated in the traditional approximate design in some case. To generate the specific error sign, a novel 4–2 compressor with negative mean error and sign bit compensation scheme is proposed. Compared with the exact design, the proposed design achieves 15.10% of area reduction and 17.29% of power reduction, which also have significant accuracy advantage over previous approximate designs.