Mixed-Signal Circuits and Architectures for Energy-Efficient In-Memory and In-Sensor Computation of Artificial Neural Networks
Bongjin Kim · 2019
Von Neumann architecture is recently facing a critical challenge with the high demands of energy-efficient computing hardware for a variety of machine learning tasks such as image classification. In particular, battery-operated mobile devices with limited power budget cannot process artificial neural networks (ANNs) with relatively low complexity by using traditional digital circuits and architectures. The key challenge with the traditional Von Neumann architecture is its energy-inefficient data access between memory and processor. Recently, in-memory computing architecture has gained significant attention as an alternative, especially for running mobile artificial intelligence applications. The memory access energy has been drastically reduced by using local in-memory processing elements which directly use the data stored in local memory. To further improve the efficiency, the processor based on mixed-signal circuits instead of conventional digital circuits have recently been actively researched. However, mixed-signal circuits have several critical drawbacks, including nonlinearity, PVT variation, and the overhead of ADC/DAC for interfacing with the external digital domain. In this work, we first review the recent mixed-signal circuits and architectures for in-memory computing using different embedded memories. In addition, we introduce the concept of in-sensor computation for integrating partial computing units in an image-sensor array using low-power mixed-signal circuit techniques.