An ultra-low power in-memory computing cell for binarized neural networks

Abstract

Deep Neural Networks (DNN's) are widely used in many artificial intelligence applications such as image classification and image recognition. Data movement in DNN's results in increased power consumption. The primary reason behind the energy-expensive data movement in DNN's is due to the conventional Von Neuman architecture in which computing unit and memory are physically separated. To address the issue of energy-expensive data movement in DNN's in-memory computing schemes are proposed in the literature. The fundamental principle behind in-memory computing is to enable the vector computations closer to the memory. In-memory computing schemes based on CMOS technologies are of great importance nowadays due to the ease of massive production and commercialization. However, many of the proposed in-memory computing schemes suffer from power and performance degradation. Besides, some of them are capable of reducing power consumption only to a small extent and this requires sacrificing the overall signal to noise ratio (SNR). This thesis discusses an efficient In-Memory Computing (IMC) cell for Binarized Neural Networks (BNNs). Moreover, IMC cell was modelled using the simplest current computing method. In this thesis, the developed IMC cell is a practical solution to the energy-expensive data movement within the BNNs. A 4-bit Digital to Analog Converter (DAC) is designed and simulated using 130nm CMOS process. Using the 4-bit DAC the functionality of IMC scheme for BNNs is demonstrated. The optimised 4-bit DAC shows that it is a powerful IMC method for BNNs. The results presented in this thesis show this approach of IMC is capable of accurately performing dot operation between the input activations and the weights. Furthermore, 4-bit DAC provides a 4-bit weight precision, which provides an effective means to improve the overall accuracy.

Date of Award	2020
Original language	English