MTJ-Based p-Bit Designs for Enhanced Tunability

Abstract

Probabilistic bits (p-bits) can be viewed as tunable random number generators (RNGs), whose stochasticity coupled with their tunability makes them enablers for an emerging class of applications, including probabilistic computing. Their tunability is the feature that makes them unique to conventional RNGs. This paper studies the tunability range of existing p-bit designs reported in the literature, highlighting that existing designs have a limited input voltage range within which the p-bit’s stochastic response can be tuned, on the order of sub 0.5 V. This may greatly limit their scalability in large p-bit networks. Accordingly, this work proposes several variant p-bit designs that enable a wider input voltage tunability range and a more continuous response for p-bits. The designs employ both bipolar and continuous stochastic MTJs, and demonstrate an enhancement in the tunability range beyond 4 V (close to half of the supply voltage range).