Robust power-law computing for edge intelligence using spin-orbit torque magnetic tunnel junctions
Abstract
Power-law transformations are widely used in edge perception, where compact implementations must balance numerical precision, circuit resources, and device variability. Here, we demonstrate a stochastic power-law architecture based on spin–orbit-torque magnetic tunnel junctions (SOT-MTJs). An event-driven feedback loop adjusts the write voltage to regulate the probability of an all-1 event in an n-bit block. Under the fixed-point sampling model, this establishes a switching probability p = b1/n, from which m-bit event statistics yield the power-law target bm/n. Online probability locking removes the need for offline per-device probability-voltage characterization or stored calibration tables. Measurements on W-based SOT-MTJs, combined with digital analysis of the acquired bitstreams, demonstrate representative power-law operations with absolute deviations of 0.003 or less at the reported precision, as well as parallel decoding of multiple exponents from one source. A 4-bit gamma-correction example preserves dark-region image detail and reaches approximately 35 dB peak signal-to-noise ratio relative to the theoretical gamma-domain reference near 1000 raw samples. A 5 Oe device-level field perturbation and a joint ±90% simulation centered on the measured parameter means of 25 devices examine adaptation and the safe voltage operating boundary under the tested conditions. These results establish a nonlinear computing primitive with accuracy controlled by sampling depth and an operating point maintained through feedback.