Breaking the Conversion Wall in Mixed-Signal Systems Using Neuromorphic Data Converters

Abstract

Data converters are ubiquitous in mixed-signal systems, becoming the computational bottleneck in traditional data acquisition and emerging neuromorphic systems. Unfortunately, conventional Nyquist data converters trade off speed, power, and accuracy. Therefore, they are exhaustively customized for special purpose applications. Furthermore, intrinsic real-time and post-silicon variations dramatically degrade their performance along with the CMOS technology downscaling. Here, we review on our neuromorphic analog-to-digital (ADC) and digital-to-analog (DAC) converters that are trained using the online stochastic gradient descent algorithm to autonomously adapt to different design specifications, including multiple full-scale voltages, number of resolution bits, and sampling frequencies. We demonstrate the feasibility of our converters by simulations and preliminary experiments using memristive technologies. We show collective properties of our converters in application reconfiguration, logarithmic quantization, mismatches calibration, noise tolerance, and power optimization. The proposed data converters achieve a superior figure-of-merit (FoM) of 1 fJ/conv.