Ferrimagnetic and non-collinear antiferromagnetic tunnel junctions for random number generation and probabilistic computing: A comparative study

The growing demand for higher throughput and energy efficiency in stochastic applications drives the search for true random number generators (TRNGs) with intrinsically fast dynamics. Magnetic tunnel junctions (MTJs) using ferromagnets (FMTJs) are attractive as TRNGs but are restricted to GHz switching speeds. Recent advances in ferrimagnetic (Fi-MTJs) and non-collinear antiferromagnetic (NC-AFMTJs) tunnel junctions offer a path toward faster magnetic-based TRNGs. This work presents a comprehensive comparative study of FMTJs, Fi-MTJs, and NC-AFMTJs as entropy sources for TRNG and probabilistic computing. Each device is modeled using the stochastic Landau–Lifshitz–Gilbert equation, adapted to reflect its unique physical properties, within the macrospin approximation. Three independent trials of 20 000 random bits are generated for each device by spin–orbit torque switching and evaluated using the NIST SP 800-22 statistical tests across varying sampling frequencies. The results show that the bitstreams from the FMTJ fail randomness tests beyond 0.1 GHz; however, those using the Fi-MTJ and the NC-AFMTJ remain statistically independent up to 2 and 20 GHz, respectively. The Fi-MTJ and the NC-AFMTJ are also employed as probabilistic p-bits in an Ising ground state search framework. Our analysis finds the Fi-MTJ and NC-AFMTJ-based p-bit networks using 16 interconnected devices achieve convergence speeds >30 × and >170 × faster, respectively, compared to the FMTJ. The enhanced performances can be attributed to faster switching dynamics enabled by exchange-coupled systems in Fi-MTJ and NC-AFMTJ devices and highlight their promise as ultra-fast, energy-efficient hardware for TRNGs and probabilistic computing systems.