A Memristive Neural Network Computing Engine using CMOS-Compatible Charge-Trap-Transistor (CTT)

摘要

A memristive neural network computing engine based on CMOS-compatiblecharge-trap transistor (CTT) is proposed in this paper. CTT devices are used asanalog multipliers. Compared to digital multipliers, CTT-based analogmultipliers show dramatic area and power reduction (>100x). The proposedmemristive computing engine is composed of a scalable CTT multiplier array andenergy efficient analog-digital interfaces. Through implementing the sequentialanalog fabric (SAF), the mixed-signal of the engine interfaces are simplifiedand hardware overhead remains constant regardless of the size of the array. Aproof-of-concept 784 by 784 CTT computing engine is implemented using TSMC 28nmCMOS technology and occupied 0.68mm2. It achieves 69.9 TOPS with 500 MHz clockfrequency and consumes 14.8 mW. As an example, we utilize this computing engineto address a classic pattern recognition problem-classifying handwritten digitson MNIST database - and obtained a performance comparable to state-of-the-artfully connected neural networks using 8-bit fixed-point resolution.