First Demonstration of a Bayesian Machine based on Unified Memory and Random Source Achieved by 16-layer Stacking 3D Fe-Diode with High Noise Density and High Area Efficiency

In this work, a Bayesian machine based on unified memory and random source using 16-layer stacking 3D Fe-diode is experimentally demonstrated for the first time. A 3D integration of Fe-diode can provide high area efficiency for storage and ultra-high memory density per unit area (0.26F²/bit) can be realized. Meanwhile, shot noise in Fe-diode is selected as the random source owing to the frequency-independent high noise density compared with other types of noise. A low cost TRNG circuit based on shot noise is then proposed to generate stochastic bit-streams (SBS) with eliminated correlation, validated by ACF and NIST tests. The Bayesian machine is finally implemented by unified memory and random source based on 16-layer 3D Fe-diode array and excellent properties (95.31% MN1ST recognition accuracy, lowest RMSE of 0.026 and 10% noise tolerance) can be successfully achieved.