The research group from the Institute of Microelectronics at the Chinese Academy of Sciences has made a notable discovery: the noise properties inherent in ferroelectric diodes (Fe-diodes) are well-suited to fulfill the stringent requirements of edge artificial intelligence (AI) systems. These systems necessitate high-quality random entropy sources, especially when operating under high-frequency conditions and experiencing significant temperature fluctuations. By meticulously adjusting the resistive states and readout voltages, the team successfully generated and stably output a high-density shot noise. This noise is distinctively dual-regulated by both frequency and temperature, offering independent control over each parameter. Impressively, the noise density achieved is over two orders of magnitude greater than that of the conventional 1/f noise. Moreover, it remains robust and shows no signs of attenuation across a wide temperature spectrum, ranging from -40°C to 125°C.

Leveraging this advancement, the team constructed a Bayesian neural network chip. This innovative chip is based on a 3D 16-layer Fe-diode array and boasts the capability of in-situ training. Remarkably, it operates with an ultra-low energy consumption of just 25 fJ per program. In performance evaluations, the chip demonstrated a commendable MNIST recognition accuracy of 92.4%. Furthermore, it scored a minimum entropy of 0.9997 in the NIST randomness test, underscoring its reliability as a source of randomness. Notably, the chip achieves all these feats while occupying a minimal physical footprint, with an area of only 0.06 F² per state. This breakthrough paves the way for a scalable and efficient new paradigm in providing random entropy sources specifically tailored for edge AI inference applications.