A Power Efficient Neuromorphic Computing System Based On the Memristor Crossbar Array
By: Jian Huang
Department: Computer Engineering
Faculty Advisor: Dr. Hao Jiang
The demand for high performance processors to boost computational speed and efficiency in matrix-vector and matrix-matrix operations have increased as machine learning and deep learning heavily relies on matrix operations in image, pattern, and speech recognition. Neuromorphic computing system utilizes the memristor crossbar array (MCA) to boost computational speed and efficiency by combining memory and matrix operation in hardware. In this work a 2-layer feed forward neural network based neuromorphic computing system is designed in IBM’s 130nm CMOS technology to process handwritten digits from the MNIST dataset. The proposed neuromorphic computing system implements a global-local circuit architecture to generate a pulse-width modulation (PWM) signal as its computational variable, compared to a traditional amplitude modulation (AM) signal. The designed neuromorphic computing system operates at 50MHz, classifies a 12x12 digit image in four clock cycles, achieves 96.66% classification accuracy on 10,000 MNIST images.