Probabilistic Computing with Binary Stochastic Neurons

Abstract

Deterministic bits that represent "0" or "1" form the basis of all digital computing. At the other end of the spectrum, quantum bits represent a superposition of "0" and "1" and form the basis of the emerging field of quantum computing. Recently, it has been shown that probabilistic bits or "p-bits" which fluctuate between "0" and "1" can be correlated using techniques borrowed from neural networks and can be used to address a broad class of problems relevant to machine learning and quantum computing. It has been shown that scaled p-bit implementations can be realized in hardware by making slight modifications to existing MRAM technology that are nearing production in Gb level densities. Even though there are no fundamental obstacles, implementing MTJ-based p-bits in gigabit scales can be challenging as they require a careful design of low- barrier nanomagnets which, unlike high-barrier nanomagnets, remain relatively unexplored. In this paper, we propose and experimentally demonstrate the operation of a compact, low-level p-bit emulator that retains much of the physics of the MRAM- based p-bit. The emulator uses most of the components of the original mixed-signal design and simply replaces the low-barrier nanomagnet with a fluctuating resistor circuit. Our emulation allows us to build asynchronously operating rudimentary p- computers to explore potential difficulties that could arise in scaled nanodevice based p-computers.