Thermal Modeling of 3D Polylithic Integration and Implications on BEOL RRAM Performance

This paper presents thermal evaluation of a back-end-of-line (BEOL)-embedded chiplet integration scheme for dense 3D heterogeneous integration, where custom chiplets are embedded into the BEOL of an application processor tier and a monolithic memory device tier. Dielectric thermal conductivity study suggests that 3D ICs might require heat-sinking from multiple sides and, potentially, advanced backend heat spreading using higher thermal conductivity dielectric materials. The impact of 3D integration and cooling architectures on resistive random access memory (RRAM) reliability is quantified by estimating image recognition accuracy over time for a deep neural network hardware model. The worst-case accuracy drop at 10 years was estimated as 82% for air-cooling compared to 2% for liquid cooling.