Abstract
To resolve the rising temperatures in 3.5D-ICs, a thermal-aware design flow becomes increasingly crucial, necessitating an accurate 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°C), and efficiency (80× acceleration), delivering a powerful means to evaluate and refine thermal designs.
The Challenge of 3.5D-IC Thermal Management
The semiconductor industry is rapidly transitioning toward 3.5D Integrated Circuits (3.5D-ICs), which integrate 2.5D packaging (chiplets on an interposer) with 3D stacking. While this architecture offers significant performance gains, it introduces a "thermal nightmare." These systems are inherently multiscale, featuring massive components like heat sinks alongside microscopic features such as Through-Silicon Vias (TSVs).
Existing simulation tools generally fall into two categories, neither of which is ideal for 3.5D-ICs:
- High-fidelity industrial tools (e.g., ANSYS Icepak, COMSOL) utilize dense meshes to capture fine details, resulting in prohibitive computational costs and slow simulation times.
- Fast academic tools (e.g., HotSpot) often oversimplify geometry or ignore critical nonlinear effects—the fact that silicon's thermal conductivity drops and leakage power rises exponentially as temperatures increase.
The authors aim to bridge this gap with ATSim3.5D, a tool that is both "structure-aware" (handling multiscale geometry) and "physics-aware" (handling nonlinearity) without sacrificing speed.
The ATSim3.5D Framework
The ATSim3.5D framework is built upon three core innovations designed to handle the complexity of modern packaging.
Hybrid Tree Grid Generation
To avoid the memory waste of uniform meshing, ATSim3.5D employs a hybrid quadtree approach. It begins with a global partition of the entire package and performs local refinement only in critical areas—around chiplets, hotspots, and complex internal structures. This allows the simulator to handle the dimension mismatch between large cooling components and tiny heating elements efficiently.
Nonlinear Multigrid Solver (FAS-MG)
Standard solvers often struggle with the nonlinear equations that arise when material properties change with temperature. ATSim3.5D implements the Full Approximation Scheme (FAS) multigrid method. Unlike traditional multigrid methods that only transfer errors between grid levels, FAS transfers the full solution. This enables the solver to handle temperature-dependent conductivity and leakage power more robustly and efficiently than standard iterative techniques.
Voigt-Reuss Bound Modeling
To accurately model "heterogeneous" cells containing a mix of materials (e.g., epoxy and silicon), the simulator uses the Voigt-Reuss bound to calculate effective thermal conductivity. This approach provides a more physically grounded average than simple arithmetic means, improving overall accuracy.
Experimental Evaluation
The researchers evaluated ATSim3.5D using two primary benchmarks: a model of the Nvidia V100 GPU (featuring a GPU core and four HBM modules) and a 3.5D Near-Data-Processor (3.5D-NDP) system (a 2×2 array of chiplets with DRAM stacked on logic). Results were compared against ANSYS Icepak for linear simulations and COMSOL Multiphysics for nonlinear ground truth.
Key Results
- Efficiency: ATSim3.5D achieved an average 80× acceleration over ANSYS Icepak.
- Accuracy: The simulator maintained high precision, with a Mean Absolute Relative Error (MARE) of less than 1% and a Maximum Error (MaxE) of less than 2°C.
- Robustness: In nonlinear scenarios factoring in temperature-dependent leakage, ATSim3.5D maintained high accuracy (0.83% MARE for the V100 system), whereas some competing academic tools like ATSim3D failed to converge or could not handle the 3.5D geometry.
- Scalability: The tool demonstrates near-linear scaling as the number of TSVs increases, overcoming the "memory wall" that plagues traditional simulators.
Conclusion and Future Outlook
ATSim3.5D provides a powerful, efficient means to evaluate and refine thermal designs for next-generation 3.5D-ICs. While highly effective, the authors note areas for future improvement, such as implementing finer vertical subdivision for materials like epoxy and refining global grid resolution to further reduce coarse-grained errors. Additionally, future work will focus on generalizing these methods to non-rectilinear structures and transient (time-varying) thermal simulations.
Summary updated