Reliability and Scalability of p-Bits Implemented With Low Energy Barrier Nanomagnets

Probabilistic bits (p-bits) implemented with low energy barrier nanomagnets (LBMs) have recently gained attention because they can be leveraged to perform some computational tasks very efficiently. Although more error resilient than Boolean computing, p-bit-based computing employing LBMs is, however, not completely immune to defects and device-to-device variations. In some tasks (e.g., binary stochastic neurons for machine learning and p-bits for population coding), extended defects, such as variation of the LBM thickness over a significant fraction of the surface, can impair functionality. We examine if unavoidable geometric device-to-device variations can have a significant effect on one of the most critical requirements for probabilistic computing: the ability to “program” probability with an external agent, such as a spin-polarized current injected into the LBM. We found that the programming ability is fortunately not lost with reasonable device-to-device variations. The little variation in the probability-versus-current characteristic caused by typical device variability can be suppressed further by increasing the spin polarization of the current. This shows that the probabilistic computing with LBMs is resistant against small geometric variations, and hence will be scalable to a large number of p-bits.