Efficient Thermal-Aware Floor-Planning with Bayesian Optimization: A Simulation-Efficient Approach

Thermal-aware macro placement in floor-planning typically demands computationally intensive and time-consuming thermal simulations. In this work, we propose an approximation technique to mitigate this challenge. By integrating this approximation into the Bayesian Optimization (BO) framework, we achieve comparable placement results while significantly reducing the number of simulations, leading to substantial runtime improvements. Additionally, our method enables the strategic insertion of dead space to achieve specific temperature targets, allowing peak temperature (T) to be automatically constrained. This approach eliminates the need to balance multiple objectives, as wirelength (WL) can now be minimized as the primary objective, with thermal constraints inherently satisfied. Without this approximation, careful weighting of thermal and wirelength objectives is necessary for effective placement.