The Behavioral Circuit Implementation Method And AMS Simulation Of Memristive Neural Network

Memristor is a new type of devices that are programmable and nonvolatile.It is also a new kind of circuit element that enables in-memory computing.Memristors and CMOS transistors can be integrated on the same chip,forming the so called memristive neural network(MNN).A MNN has to be trained before it becomes functional.Among the two types of training methods,off-chip training and on-chip training,the latter is preferred due to its flexibility and convenience in application.Proposals in the existing literature mainly use digital circuitry to train MNNs.However,due to the analog nature of memristor,certain analog-todigital and digital-to-analog conversion circuitry is necessary for digital oriented implementations.That could cause a large amount of increase in circuit complexity,power and chip area.The main contribution of this thesis is a proposal of analog oriented circuit implementation method that avoids the use of data converter circuits.For this purpose,several novel circuit blocks have been introduced,including a memristor-array based circuit for gradient computation and storage,special circuitry for realizing negative synapse weights,currentmode programming circuit design,and special signal propagation control circuitry.All proposed circuit designs have been put together in a single MNN and verified by circuit-level behavioral simulation.The hardware description language Verilog-AMS was used to validate the whole design.Due to the fact that AMS simulation is a circuit level simulation that incorporate digital and analog circuit blocks in a cycle-accurate framework,simulation is very time and memory consuming.For this reason,the behavioral validation performed in this thesis was only focused on the pattern recognition problems of a small pattern set with low dimensions.Behavioral simulation has validated that all the proposed circuit blocks can function correctly and the training and forward computation circuits can produce the correct pattern recognition results.As a byproduct of this research,we also point out the time split in the behavioral simulation of MNN circuits,finding that the digital control circuits consume the most simulation resource in mixed-signal behavioral simulation.