Universal Random Number Generation with Field-Programmable Analog Array and Magnetic Tunneling Junction (MTJ) Devices

This paper presents a probabilistic-based design methodology for a random number generator according to any given probability density function. Our basic idea is to perform massively parallel analog computation through Gaussian mixture modelling. There are two main contributions in this work. First, we developed a new approach to performing Gaussian mixture decomposition with simple analog circuits. Second, we have presented a novel methodology to exploit the stochastic switching behavior of Magnetic Tunneling Junction (MJT) as a hardware-efficient Gaussian noise generator. Our resulting universal random number generator not only achieves extremely low energy consumption and ultra-high computing performance, but also is highly reconfigurable. Consequently, it can be widely applicable in many hardware-based signal processing applications, especially quite useful in the newly emerging stochastic based computing systems. Finally, to validate our design, we used field-programmable analog array to implement all required components.