Probabilistic computing is a computing paradigm that can solve certain problems more efficiently than traditional digital computing. While digital computing deals with deterministic binary bits that are either 0 or 1, probabilistic computing deals with probabilistic bits (p- bits) that are sometimes 0 and sometimes 1. This is distinct from quantum computing that deals with quantum bits (q-bits) which are a superposition of 0 and 1 (and hence a mixture of both 0 and 1 all the time). Quantum computing usually requires the most hardware resources and digital computing the least, with probabilistic computing between the two. Most of the hardware resources in probabilistic computing are devoted to generating specific correlations between two or more p-bit streams. This project will study and demonstrate a system that will greatly reduce the hardware burden associated with generating correlations. The results will make probabilistic computing much more efficient than it currently is. The project will educate K-12, undergraduate, and graduate students in this field to increase the pool of skilled scientists and engineers while advancing the field of computing.
One of the major challenges in probabilistic computing is the complex hardware needed to generate required correlations between probabilistic bit streams. This hardware usually consists of microcontrollers, analog-to-digital converters, shift registers, etc., that consume significant power and vastly expand the system?s footprint on a chip. In this project, an ultra-compact and extremely energy-efficient correlator or anti-correlator will be studied and demonstrated that can generate tunable degrees of anti-correlation between two p-bit streams. The approach is implemented with two magnetic tunnel junctions (MTJs) whose soft layers are in close proximity and hence dipole-coupled. Bit states are encoded in the resistance states (high or low) of the MTJs. One MTJ generates a p-bit when driven by a spin-polarized current delivering a spin-transfer-torque. The current sets the resistance state high or low with a probability determined by its magnitude, while the bit state of the other MTJ is determined by dipole coupling with the first. Very strong dipole coupling will result in perfect anti-correlation, while very weak dipole coupling will result in no correlation. The effect of dipole coupling will be controlled by applying (electrically generated) local strain to the second MTJ, which modulates its internal energy barrier, thereby modulating the degree of anti-correlation between the p-bits from 0% to 100%. This project will result in new understanding of devices that use emerging nanomagnetic physics for the next generation of computing.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
