HyperPlace: Harnessing a Large Language Model for Efficient Hyperparameter Optimization in GPU-Accelerated VLSI Placement
Magi Chen, Ting-Chi Wang · ACM Transactions on Design Automation of Electronic Systems · 2025
While GPU-based placers have demonstrated significant speed advantages over their CPU-based counterparts, hyperparameter tuning remains a bottleneck, often requiring substantial human intervention and expert knowledge. This challenge is particularly critical given the urgent need for rapid time-to-market solutions. Recently, Large Language Models (LLMs) have exhibited remarkable capabilities in zero-shot learning, context understanding, logical reasoning, and answer generation. In this work, we introduce HyperPlace, an innovative paradigm that leverages an off-the-shelf LLM to automate hyperparameter optimization using in-context learning techniques. Our approach transcends single-output black-box optimization methods by incorporating a batch optimization mechanism that evaluates multiple hyperparameter configurations simultaneously across several GPU computing platforms. We validated the effectiveness of our approach in placement quality, measured by Half-Perimeter Wire Length (HPWL), using DREAMPlace 2.0. To further demonstrate the capability of integrating our framework with other placers, we conducted additional experiments using Xplace 2.0. By employing the ISPD2005 benchmarks for our evaluation, HyperPlace enhances the placement tools with up to a 1.66% reduction in HPWL compared to their published results. Additionally, we evaluated HyperPlace on the ISPD2015 benchmarks, which incorporate fence region constraints not present in ISPD2005 benchmarks. Under these more complex constraints, HyperPlace achieves up to a 22.24% reduction in HPWL compared to the default settings of the placement tools, further demonstrating its adaptability across diverse placement scenarios and benchmark suites.