Presentation
DeepOHeat: Operator Learning-based Ultra-fast Thermal Simulation in 3D-IC Design
DescriptionChip thermal optimization requires thousands of simulations. Neural network-based methods once trained can perform real-time inference for unseen new designs are therefore of primary interest. However, recent works either solve solely 2D temperature fields or don't generalize well to arbitrary unseen designs with multiple inputs. In this paper, for the first time, a physics-aware operator learning-based framework, named DeepOHeat, is proposed to solve a family of heat equations for ultra-fast multi-objective thermal optimization. We compared the proposed method against Celsius 3D, a SOTA commercial solver from Cadence Design Systems, to demonstrate its satisfactory accuracy on any unseen chip designs.
Event Type
Research Manuscript
TimeWednesday, July 12th1:40pm - 1:55pm PDT
Location3002, 3rd Floor
EDA
Timing and Low Power Design