Elements: TASChips: Thermal Analysis of Semiconductor Chips with High Efficiency, Accuracy and Resolution Enabled by Physics-Aware Reduced-Order Learning Techniques

This project develops TASChips, an open-source, high-performance simulation tool for thermal analysis of modern microprocessors such as CPUs, GPUs, and AI-accelerators. These microprocessors are central to the nation’s scientific, economic, and technological progress, but increasing computational demands lead to serious overheating challenges that can degrade performance, reliability, and energy efficiency of these microprocessors. TASChips addresses these challenges by enabling fast and accurate prediction of chip temperature distributions, allowing researchers and engineers to design more reliable, sustainable, energy-efficient systems. Its physics-based learning algorithms deliver accurate real-time thermal modeling capabilities at resolutions comparable to direct numerical simulations (DNS) with computational speeds even faster than dynamic thermal circuits. Such capabilities allow appropriate run-time assignments and redistributions of workloads based on dynamic hot spot distributions in the microprocessors. TASChips will be freely available to the broader research community, with extensive documentation and case studies, and integrated into educational activities. The project supports national interests by enabling better computing infrastructure, engaging STEM students in research through a REU program, and promoting innovation in thermal-aware design for next-generation computing systems. TASChips integrates physics-aware reduced-order learning models revised from Proper Orthogonal Decomposition and Galerkin Projection (POD-GP) to enable efficient and high-fidelity thermal simulation of semiconductor chips with tens to hundreds of thousands of cores. The tool includes multiple modeling approaches (GPOD-GP, EnPOD-GP, LEnPOD-GP, and MuPOD-GP) each designed for chips of varying complexity and scale. These models are trained on high-resolution simulation data and rigorously enforce physical principles of heat transfer throughout the computation, enabling both substantial speedups and fine resolution over existing thermal simulators, while maintaining a least square error ~1%. TASChips will be distributed via GitHub with both CPU- and GPU-based implementations. Collaborators across domains such as real-time scheduling, carbon modeling, and power systems simulation will guide use-case-driven development. Through this effort, TASChips will advance chip-level thermal simulation, support interdisciplinary research, and provide a foundation for scalable, real-time thermal management tools in high-performance computing. This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Division of Graduate Education within the Directorate for STEM Education. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.