UROP Proceedings 2021-22

School of Engineering Department of Electronic and Computer Engineering 143 Compact Models for Circuit Design Supervisor: SHAO Qiming / ECE Student: XING Hexiang / ELEC Course: UROP1100, Fall As CMOS technology scales down, the power dissipation due to the leakage current becomes a serious concern, and spintronic devices are promising candidates for solving this issue. Spintronic devices have their own advantages such as high endurance, nonvolatile characteristic, low power dissipation, and good scalability. Among various spintronic devices, magnetic tunnel junction (MTJ), as the essential unit for magnetic random-access memory (MRAM), can be applied for various fields such as memory, in-memory computing, data security, and stochastic computing. Currently, spintransfer-torque MTJ (STT-MTJ) is widely applied. However, compared with STT-MTJ, voltage-control MTJ (VC-MTJ) can achieve higher density, lower power consumption, and faster speed. In this report, I summarize some basic applications of VC-MTJ in different devices such as magnetoelectric random-access memory, true random number generator, analogto-digital/stochastic converter, spintronic programmable logic, temperature sensor, in-memory computing, etc. Compact Models for Circuit Design Supervisor: SHAO Qiming / ECE Student: XU Zimo / CPEG Course: UROP1100, Fall DNN+Neurosim is a kind of integrated Framework that created by c++ language. It uses python wrapper to connect the neurosim and a machine-learning platform, Pytorch, to provide a mapping from algorithm to hardware that could be used to evaluate the chip training in area, energy efficiency and throughput. Deep Learning for Magnetic Domain Image Denoising and Super-Resolution Supervisor: SHAO Qiming / ECE Student: DUAN Qinkai / SENG Course: UROP1100, Summer I focused on learning during this project: autoencoder, t-sne, LSTM, Restricted Boltzmann Machine, and mumax3 tutorials. They are about graphic classification, time list prediction, and magnetic simulations. In many datasets, the dimension of the data is too big to visualize. So we need t-sne method to reduce the dimension to 2 or 3 while keeping the distribution. LSTM is an extension for RNN. It is used for predicting the next time step action in a time interval. It’s widely used in word and text recognition. Mumax3 can simulate how the magnetic material change in atom size based on equations in Micromagnetics.