Stochastic‐MTJ Sampler Arrays for In‐Array Monte–Carlo Estimation

Abstract

Stochastic magnetic tunnel junctions (sMTJs) with low energy barriers fluctuate randomly between two resistance states, supplying room‐temperature entropy without pseudo‐random‐number generators. This work develops a device‐faithful sMTJ probabilistic‐bit (p‐bit) model—telegraph switching, a current‐tunable sigmoid response, and Poisson‐distributed dwell times—using literature‐grounded perpendicular‐anisotropy parameters. The p‐bits are integrated into a programmable crossbar whose per‐column accumulation realizes a probability‐domain expectation estimator: the multiply–accumulate is retained as the mean output, while the estimator variance, which scales as over a ‐sample window, provides a calibrated and irreducible measure of uncertainty. A closed‐loop drift self‐calibration scheme re‐biases each cell through a shared digital‐to‐analog converter, turning device volatility from a liability into a managed quantity. A circuit‐level analysis in 22 nm CMOS documents an energy efficiency of up to 869 TOPS/W under a clearly defined read‐phase operation, and an in‐array Monte–Carlo and Bayesian‐uncertainty benchmark quantifies the energy saved by harvesting free thermal entropy instead of CMOS stochastic‐number generators. The architecture is positioned as a sampling accelerator, distinct from stochastic‐computing arithmetic and from coupled Ising machines.