Abstract
Monolithic 3D integration represents a major breakthrough in the quest for high-density, energy-efficient systems. Ferroelectric thin-film transistors (Fe-TFT) have garnered increasing attention due to their outstanding capability in realizing brain-inspired computing and compatibility with the back-end-of-the-line (BEOL) fabrication process. Nevertheless, monolithic 3D ICs inevitability suffer from excessive temperatures which degrade the device characteristics degrading the system performance. In this work, we are the first to demonstrate how existing Fe-TFT crossbar arrays can be employed to sense temperature and detect run-time thermal fluctuations. This enables the Fe-TFT array to self-adaptively adjust bias conditions and operate reliably for the entire temperature range. We demonstrate the proof-of-concept using meticulously calibrated device simulations and temperature measurements of fabricated BEOL Fe-TFT devices. Further, we perform an extensive device-to-system thermal modeling for Fe-TFT-based monolithic 3D ICs to (1) acquire accurate thermal maps, (2) assess the temperature's influence on the inference accuracy of deep neural networks, and (3) showcase the efficacy of our technique in defeating temperature effects.