A Neural Recording System with 16 Reconfigurable Front-end Channels and Memristive Processing/Memory Unit.

This paper proposes a neural recording system that supports recording of action potentials (APs), local field potentials (LFPs), or full-band neural signals (APs plus LFPs) by exploiting 16 reconfigurable front-end channels. To store and process the recorded signals, an off-chip 1 transistor 1 memristor (ITIR) crossbar array with optimized row/column selection logic is employed. The crossbar array is divided into four allocatable memory zones. The front-end channels are designed to filter the input signals to the bands of interest with a programmable gain of 100x or 1000x. Each front-end channel comprises an AC-coupled low-noise amplifier (LNA) which has an inherent high-pass filter (HPF), a passive first-order low-pass filter (LPF), and a variable gain amplifier (VGA). The VGA also acts as a butter to interface with the memristor array. A reconfigurable pseudo-resistor is used to set variable high-pass corner frequencies for different recording modes. The total input-referred integrated noise of the entire channel is less than 5.6$\mu \mathrm{V}_{rms}$ in the AP band (500- 10kHz) and less than 4.55$\mu \mathrm{V}_{rms}$ in the LFP band (1-500Hz). Implemented in a standard 180nm CMOS process, each channel occupies 0.035 m$\mathrm{m}^{2}$ chip area and consumes 20.4$\mu$W with a 1. 8V power supply. The memristor array is post-processed on the CMOS back-end-of-line (BEOL) in-house, which will be used to process and encode the channel outputs into memristive states.