Design of a Low Noise Amplifier with Programmable Gain for Neuromorphic Sensor Frontends

Always-on Keyword Spotting and Voice-Activity Detection (KWS/VAD) at the edge require microphone front-ends that combine sub-μW power with programmable gain and audioband noise performance. This thesis presents a two-stage Low-Noise Amplifer (LNA) for capacitive MEMS microphones, implemented in a 22nm FD-SOI technology and operating from a 0.8V supply. The LNA consists of a fixed-gain amplifier (FGA) followed by a Programmable-Gain Amplifier (PGA). The FGA is a fully differential inverter-input telescopic amplifier that sets the inputreferred noise and provides a nominal 32.8 dB closed-loop gain using capacitive feedback. Replica-bias branches reserve approximately 130mV at the ends of the telescopic stack, and a continuous-time DDA-based Common-Mode Feedback (CMFB) loop regulates the output around 0.4V. High-resistance DC paths at the amplifier inputs are realized with dedicated low-gm OTAs that replace ideal giga-ohm elements. The PGA employs a fully differentialNMOSinput amplifier with a binary-weighted capacitive feedback network, providing programmable gains of 1x-16x in 6 dB steps. The required enable signals are generated by a compact 3-bit digital decoder implemented using standard-cell logic from the same 22nm FD-SOI library. Cascading the fixed 44x FGA with the PGA yields an overall closed-loop gain range of 44x-704x (≈ 33 dB-57 dB) in uniform 6 dB steps, covering microphone levels from roughly 50 dB to 80 dB SPL within a fixed ≈ 160mVpp to 320mVpp differential output window. Typically, the maximum gain code achieves a -3 dB bandwidth of 9.6 kHz and an integrated input-referred noise of 54.4 μVrms over 50 Hz to 8000 Hz, satisfying the system-level noise target derived from a 3 dB input SNR requirement at 50 dB SPL. The typical total power consumption of the complete LNA, including CMFB, replica bias and low-gm OTAs, is 361.6nW at 0.8V. Across PVT corners the gain remains stable, while the bandwidth and noise vary moderately.