Feb 24 (Tue) @ 1:00pm: "Stochastic Nanodevices for Scalable Probabilistic Computing," Kemal Selcuk, ECE PhD Defense

Location: Engineering Science Building (ESB), Room 2001
Zoom Link: https://ucsb.zoom.us/j/87044455795
Research Area: Computer Engineering
Research Keywords: Emerging Technologies for Computing, Stochastic Devices, Circuits & Systems, AI

Abstract

Probabilistic computing is a promising platform for energy-efficient machine learning and hard optimization tasks. Executing such probabilistic tasks on conventional, clocked digital hardware is highly inefficient due to the costs of emulating randomness in digital deterministic circuits. This dissertation advances probabilistic computing through hardware–algorithm co-design, from establishing new theoretical results for stochastic magnetic tunnel junctions to designing novel hardware-aware probabilistic algorithms. 

At the device level, this work theoretically establishes the noise amplifying features of stochastic magnetic tunnel junctions that produce voltage fluctuations far larger than ordinary RC noise at room temperature. This dissertation further proposes a new type of double-free-layer stochastic magnetic tunnel junction and theoretically characterizes it using the well-established 4-component spin-circuit formalism. These theoretical predictions were later confirmed experimentally. 

At the architecture level, the dissertation identifies the analog tunability and digital interfacing as the dominant costs of embedding stochastic devices and proposes digital-to-analog converter-free probabilistic architectures, enabling seamless integration with digital circuits. Further, techniques to combat device-to-device variations and probability bias are addressed using principles of stochastic logic, without relying on analog calibration loops.

At the algorithm level, the dissertation demonstrates that the natural physics of stochastic nanodevices exhibiting true randomness, random arrivals, and different fluctuation speeds can be leveraged to design hardware-aware algorithms such as on-chip simulated annealing and massively parallel Gibbs sampling. Together, these contributions establish a practical step towards scalable probabilistic computing with a focus on optimization and next-generation AI.

Bio

Kemal Selçuk is a Ph.D. candidate in the Electrical and Computer Engineering Department at UC Santa Barbara, advised by Prof. Kerem Çamsarı. He received his B.E. in Electrical Engineering from Sabancı University in 2021. His research focuses on hardware-algorithm co-design for next-generation AI systems, with an emphasis on probabilistic computing and stochastic devices that enable computing primitives for efficient hard optimization and artificial intelligence tasks.

Hosted By: ECE Professor Kerem Çamsarı

Submitted By: Kemal Selçuk <kemalselcuk@ucsb.edu>