Optimisation for 3D Integrated Circuits using Continuously Improving Evolutionary Strategy Machine Learning Techniques

Three-Dimensional Integrated Circuits (3D ICs) are a revolutionary advancement in semiconductor technology, allowing vertical stacking of multiple active device layers to overcome the scaling challenges faced by conventional Two-Dimensional Integrated Circuits (2D ICs). While 3D ICs harness incredible advantages in terms of increased integration density, shorter interconnect length, and increased performance, they also present daunting challenges—primarily, the strong coupling between thermal and timing constraints. These challenges demand strong, multi-objective optimization techniques that can manage 3D ICs’ highly non-convex and multi-modal design space. In this paper, we introduce a novel optimization framework based on the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) integrated with a physics-based electro-thermal-timing simulation framework. By comparing CMA-ES with Bayesian optimization, we prove that CMA-ES produces superior convergence, robustness, and solution quality, especially under noisy and high-dimensional objective conditions. Our results evidence spectacular improvements in key figures of merit such as peak temperature, thermal gradients, clock skew, and overall simulation efficiency.