Synaptic dynamics in ferroelectric devices and their application to deep neural networks

点击翻译按钮获取中文摘要

Nontechnical:One of the grand visions of the modern computing era has been to mimic the cognitive capabilities of the human-brain, and even to rival them. This vision is becoming possible due to recent advances in machine learning and artificial intelligence. Current computing technologies are still far from creating a digital entity that is as capable and as energy efficient as a biological brain. Digital learning systems are notoriously power hungry and can require a room-full of digital computer clusters. Compare that to the human brain which performs all its feats at a meager power budget of twenty watts and a weight of less than two kilograms. One reason for this inefficiency is that transistors, the basic building blocks of digital computers, do not function in the same way as synapses, the basis of biological computing. The proposed research aims at overcoming this barrier by making a relatively basic change to the structure of the transistor. An emerging material with ferroelectric properties, doped hafnium oxide, will be introduced into transistors. The new device is called a ferroelectric field effect transistor and can emulate the properties of biological synapses. In this project, the unique properties of the synaptic ferroelectric transistor will be used to design and optimize artificial intelligence cores such as deep neural networks that vastly exceed the performance and efficiency of the current state-of-the-art. The project will train participating students in an interdisciplinary setting that involves material science, circuit design, computer architecture, and neuro-science. The STEM outreach and education programs will help participating undergraduates, high school students and high school teachers to broaden their experience in computer science and novel semiconductor devices.Technical:The project will explore the rich domain dynamics in ferroelectric hafnia-zirconia alloy gated silicon transistors to build synaptic units for vector matrix multiplication crossbars. The architecture and system level work will entail the design and optimization of full-blown deep neural networks based on these ferroelectric crossbar kernels. Physics based compact models of ferroelectric transistors that account for the important details of domain dynamics will be developed which will tie the material-device level work and the architecture-system level work. A key feature of the project is its vertically integrated approach that involves different levels in the computing hierarchy from materials to systems. Innovations at all these different levels will ensure that the interesting properties of the emerging ferroelectric device technology can be fully leveraged to create an energy efficient and high-performance hardware platform for advanced machine learning and data intensive cognitive applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.