Packaging, Chiplet Design, and Multi-Physics Simulation
Session Chair: Changkai Yu, Jaguar Micro
Hierarchical Decomposition and Interconnection Method for Efficient Thermal Simulation of Chiplet-Based 2.5D Systems
Presenter: Shunxiang Lan, Shanghai Jiao Tong University
Abstract: Thermal simulation is a key step in validating the performance of chiplet-based 2.5D systems. However, when dealing with the temperature-dependent parameters, the global nonlinear iteration may result in heavy computational cost. In this paper, an efficient hierarchical decomposition and interconnection (HDI) method is proposed for efficient thermal simulation of chiplet-based 2.5D systems. Firstly, the HDI method resolves the 2.5D system into two parts: the linear region and the nonlinear region. In the transient simulation, the response of the linear region is divided into a zero-input (ZI) one and a zero-state (ZS) one. In order to characterize the linear time-invariant properties of the ZS response, the impact of the linear region is further decomposed into the contributions of the face elements at the interface. In this way, an equivalent thermal boundary condition in the form of the thermal resistance matrix is established to represent the linear region. Then, we combine it with the nonlinear region to calculate the complete response in the interconnection process. By this means, the tedious iteration is confined to the nonlinear region, thereby reducing the computational complexity and improving the efficiency. To validate the accuracy and efficiency of the proposed HDI method, thermal simulation of a typical chiplet-based 2.5D system is performed, and a speed-up of 48× is achieved compared to the conventional finite volume method.
Transformer-based S-parameter Extraction for Coupled Transmission Lines with Vias
Presenter: Qin Li, Southern University of Science and Technology
Abstract: S-parameters are a crucial tool for simulating and analyzing high-speed, high-frequency signal transmission, particularly in advanced packaging and PCB designs such as chiplets and 3DICs. However, traditional 2.5D/3D electromagnetic (EM) field solvers are computationally expensive and time-intensive, making them impractical for agile and automated design iterations. Existing neural network-based methods for S-parameter prediction are often limited to predefined 2D planar structures, restricting their applicability. In this work, we propose SFormer, a transformer-based deep learning framework that predicts the S-parameters of coupled transmission line pairs directly from their layout while supporting highly flexible configurations. Our method accounts for variations in multiple layout parameters, including line length, segmentation, bending direction, line spacing, line width, and line-to-ground distance. The layout is represented as a graph, processed by a graph convolutional network (GCN) to extract geometric and topological features, which are then refined by a transformer encoder. Finally, a multi-layer perceptron (MLP) predicts the corresponding S-parameter matrix elements. Additionally, we introduce a specialized scheme for efficiently generating S-parameters of multi-layer transmission lines connected by vias, enabling rapid extraction for truly 3D structures. Compared to commercial EM solvers, SFormer achieves over 500× speed improvement while maintaining a mean absolute error (MAE) below 1e-2 for 4-port network extraction up to 15 GHz in commercial design cases.
Graph Sequence-Based Prediction of Electrothermal Stress Evolution for Power Delivery Networks
Presenter: Yuwei Sun, Southeast University
Abstract: With the advancement of integrated circuit technology, electrothermal migration (EM-TM) in power delivery networks (PDNs) has become a key reliability challenge. Existing machine learning methods predict hydrostatic stress distributions in interconnects by avoiding computationally intensive Korhonen equation solutions. At the same time, they are unable to capture the trajectory of evolution of stress and actively seek the exact moment of failure occurrence. We propose a spatiotemporal graph-based prediction framework that integrates spatial structural features and captures early stress accumulation patterns from input sequences, enabling long-term forecasting of EM-TM coupling evolution. The experimental results show that our approach achieves higher prediction accuracy and stability on PDN datasets spanning up to 1,500 interconnect segments. Furthermore, it achieves a near 5× speedup in prediction compared to the single-timestep SOTA model and a 6896× acceleration over the commercial software COMSOL.
ATSim3.5D: A Multiscale Thermal Simulator for 3.5D-IC Systems based on Nonlinear Multigrid Method
Presenter: Qipan Wang, Peking University
Abstract: To resolve the rising temperatures in 3.5D-ICs, a thermal-aware design flow becomes increasingly crucial, necessitating an ac- curate and efficient thermal simulation tool. However, previous tools struggle to handle the unique heterogeneous multiscale structures in 3.5D-ICs and the nonlinear thermal effects caused by high temperatures. In this work, we present a multiscale thermal simulator for 3.5D-ICs. We propose a hybrid tree structure to generate multilevel grids and capture the multiscale features and employ the nonlinear multigrid method for quick solving. Compared to ANSYS Icepak, it exhibits high accuracy (mean absolute relative error < 1%, max error < 2 ℃), and efficiency (80× acceleration), delivering a powerful means to evaluate and refine thermal designs.
Hierarchical Decomposition and Interconnection Method for Efficient Thermal Simulation of Chiplet-Based 2.5D Systems
Presenter: Shunxiang Lan, Shanghai Jiao Tong University