A reduced-memory FIR filter using approximate coefficients for ultra-low power SoCs

This paper presents an ultra-low power (ULP) finite-impulse response (FIR) filter using a method that approximates filter coefficients on-chip without reliance on dedicated memory such as SRAM. In a system-on-chip (SoC) context, this method allows for full power gating of the coefficient unit without coefficient state loss, and runtime modifications of filtering specifications, such as filter order and cutoff frequency. Using trigonometric approximation methods for the sinc and resource sharing of computational units, a single coefficient is generated in five clock cycles. The approximation unit is compared against standard-cell-based memories, such as register and latch files, for energy and area, and the design is synthesized in 130nm CMOS consuming 6.9nW at 300mV and 6.5kHz.