Optimization framework for energy-efficient and uniform jet impingement cooling for heterogeneous integration packaging

Abstract

The imbalance in heat power generated by various types of chips poses an obstacle to the reliability and performance of heterogeneous integration (HI) packaging technology, leading to excessive cooling that reduces the system's energy efficiency. We propose a framework to optimize the impinging nozzle arrangement for energy-efficient uniform jet cooling of HI packages. This framework utilizes a convolutional neural network (CNN)-based surrogate model that learns nozzle arrangements and heating scenarios to predict the temperature non-uniformity of the package. The potential optimal designs predicted by the CNN are used for re-training through an experimentally validated numerical analysis model. Combined with this active learning approach, the proposed hierarchical exploration algorithm accelerates optimization by gradually scaling the design options. The optimization results showed an increase in cooling uniformity by up to 39.5 %, while the cooling COP improved by up to 200 % across the investigated flow rate range (3–8 L/min). The optimized designs were experimentally validated with a maximum error of 4.34 % in average thermal resistance. Our framework achieved up to 45.7 % data savings compared to the random sampling-based approach. Along with a discussion on applying the CNN model to untrained conditions to further enhance optimization efficiency, our work represents a novel approach to broadly address the rapidly evolving diverse heating scenarios of HI, contributing to improved cooling energy efficiency in data centers and enhanced reliability of high-performance processors.