Probabilistic Computing with p-bits: Optimization, Machine Learning and Quantum Simulation

Kerem Çamsarı, University of California

The slowing down of Moore's era has coincided with escalating computational demands from Machine Learning and Artificial Intelligence. An emerging trend in computing involves building physics-inspired computers that leverage the intrinsic properties of physical systems for specific domains of applications. Probabilistic computing with p-bits, or probabilistic bits, has emerged as a promising candidate in this area, offering an energy-efficient approach to probabilistic algorithms and applications.

Several implementations of p-bits, ranging from standard CMOS technology to nanodevices, have been demonstrated. Among these, the most promising p-bits appear to be based on stochastic magnetic tunnel junctions (sMTJ). sMTJs harness the natural randomness observed in low barrier nanomagnets to create energy-efficient and fast fluctuations, up to GHz frequencies. In this talk, I will discuss how magnetic p-bits can be combined with conventional CMOS to create hybrid probabilistic-classical computers for various applications. I will provide recent examples of how p-bits are naturally applicable to combinatorial optimization, such as solving the Boolean satisfiability problem, energy-based generative machine learning models like deep Boltzmann machines, and quantum simulation for investigating many-body quantum systems.

Through experimentally-informed projections for scaled p-computers using sMTJs, I will demonstrate how physics-inspired probabilistic computing can lead to GPU-like success stories for a sustainable future in computing.

Please sign up here in order to get the Zoom link and regular announcements of the upcoming talks.