Home > Seminars > Modeling of Multilevel Ferroelectric Memory for Deep Learning Applications

Modeling of Multilevel Ferroelectric Memory for Deep Learning Applications

Start:

1/17/2019 at 2:00PM

End:

1/17/2019 at 3:00PM

Location:

116 O'Shaughnessy Hall

Host:

College of Engineering close button
headerbottom

Alan Seabaugh

Alan Seabaugh

VIEW FULL PROFILE Email: aseabaug@nd.edu
Phone: 574-631-4473
Website: http://www.nd.edu/~nano
Office: 230A Fitzpatrick Hall
Curriculum Vitae

Affiliations

Department of Electrical Engineering Frank M. Freimann Professor
College of Engineering Frank M. Freimann Professor
Research Interests: What limits density, speed, power, linearity, gain, noise, and efficiency in devices? What new device capabilities can boost electronic system performance?  Current research: tunnel field-effect transistors, atomically-thin transistors, ionic and ferroelectric memory, self ...
Click for more information about Alan
574-631-4473
Add to calendar:
iCal vCal

Deep neural networks (DNN) can perform cognitive tasks such as speech recognition and object detection with high accuracy, but with a large computational cost dominated by data movement. Hardware accelerators that rely on multilevel memory devices to perform in-memory computation can potentially improve the energy efficiency of DNNs by orders of magnitude. In this talk, we will present the characterization and modeling of polycrystalline ferroelectric films to implement multilevel memory devices and its application to DNN training. Ferroelectrics are a promising candidate for such devices because of their fast switching, ultrahigh storage density, and nonvolatile operation and CMOS compatibility. However, modeling the dynamics of polycrystalline ferroelectric films is challenging by the fact that they are composed of a multitude of grains having different switching thresholds. For this purpose, we developed a Monte Carlo simulation that can describe and predict the history-dependent polarization dynamics observed in ferroelectric films. Finally, we will discuss architectural considerations to mitigate device nonlinearity and limited resolution by leveraging the intrinsic redundancy of DNNs.

Seminar Speaker:

Cristobal Alessandri

Cristobal Alessandri

University of Notre Dame

Cristobal Alessandri received his Bachelor’s degree in Electrical Engineering at Pontificia Universidad Catholica de Chile. He is currently a Ph.D. Candidate in Electrical Engineering at the University of Notre Dame. His research interests include semiconductor devices and circuit design for machine learning applications.