Aug 31 (Mon) @ 1:00pm: "Scalable and Trustworthy Operator Learning for Thermal Simulation and Optimization in 3D-IC Design," Xinling Yu, ECE PhD Defense
Abstract
Three-dimensional integrated circuits stack multiple dies to reach integration densities beyond the limits of planar scaling, but the same vertical integration concentrates power in a small volume and lengthens the path heat must travel to escape. The resulting temperature rise degrades timing, reliability and lifetime, so thermal-aware optimization has become a standard step of the 3D-IC design flow. It reduces to one computation repeated many times: predicting the temperature field of a candidate design, once for every design the optimizer visits. High-fidelity solvers are accurate but far too slow to call inside such a loop. DeepOHeat replaces them with a neural operator that maps a design configuration to its temperature field, trained from the governing heat equation rather than from solver-generated examples, so that a new design costs only a forward pass. Two obstacles have kept such a surrogate out of a real design flow. Training from physics alone is slow and memory-intensive, which limits the resolution at which a chip can be modeled. And the surrogate is trained on one distribution of designs but queried on the designs the optimizer visits, where its errors are not random: the search returns the placement with the lowest predicted temperature, so it favors exactly those the surrogate under-predicts.
This dissertation develops two successors. DeepOHeat-v1 makes physics-based training practical, cutting its time and memory cost by more than an order of magnitude, and scores every prediction during the search so that doubtful ones are refined by a conventional solver started from the prediction itself. DeepOHeat-v2 extends the approach to realistic multi-die stacks, where conductivity changes abruptly at material layers and at the walls of through-silicon vias, and where both mechanisms break down. It trains instead on a discretized energy form of the physics, which represents these discontinuities directly and lowers mean peak-temperature error from over 30 K to 0.55 K. It also replaces the global confidence check with a hotspot-aware criterion and reuses the resulting solver refinements as training data, so the surrogate improves on the designs the search visits and the loop calls the solver progressively less. In both cases the returned design matches what solving at every step produces, at a small fraction of the cost.
Bio
Xinling Yu is a Ph.D. candidate in Electrical and Computer Engineering at the University of California, Santa Barbara, advised by Professor Zheng Zhang, with research on physics-informed operator learning, neural PDE surrogates, and their use in thermal simulation and design optimization for 3D integrated circuits. This work has been published in TMLR, DAC, ASP-DAC, and IEEE TCPMT. Xinling received a B.S. in Mathematics from Beijing University of Posts and Telecommunications and an M.A. in Applied Mathematics and Computational Science from the University of Pennsylvania, and has interned at Hewlett Packard Labs and TSMC.
Hosted By: ECE Professor Zheng Zhang
Submitted By: Xinling Yu - Yu’s Email
