SHF: Small: Native Stochastic Computing Based on Memristors

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

The objective of this research is to create a new stochastic computing paradigm to reap the full benefits of emerging memristor devices for the next generation of energy-limited and high-performance applications. The computing paradigm takes advantage of the nondeterministic memristors for native stochastic computing, where the randomness required by stochastic computing will be intrinsic to the system without resorting to expensive stochastic number generation. The native stochastic computing system consists of memristor memory and stochastic computing elements that are seamlessly integrated. To evaluate the approach, test vehicles including a stochastic calculator and a clustering processor are created by the hybrid integration of memristor memory and stochastic computing circuits. The native stochastic computing based on memristors demonstrates key advantages in energy and speed over a competitive CMOS baseline, and is best positioned for compute-intensive, data-intensive and probabilistic applications. The interdisciplinary nature of this research encourages a tight integration of device and computer hardware research areas and provides excellent training to students in cutting-edge nanoelectronics, integrated circuit design, and computer architecture. The graduates of this research program possess skill sets of critical need in both industry and academia. The research contributes to the undergraduate and graduate curriculum, in the forms of updated undergraduate courses and special topics offered in graduate courses. The research outcomes, along with two test vehicles, are disseminated through professional seminars, workshops, and as part of high school outreach programs. An emphasis on undergraduate and minority mentoring helps draw a broad and diverse participation over the course of this research project.