AstroTune: AST-Assisted LLM Retrieval for Cross-Stage Design Flow Parameter Tuner
Modern VLSI design relies on EDA tools, which expose designers to high-dimensional and complex parameter spaces. Efficiently optimizing these parameters remains challenging, as manual tuning is time-consuming and heavily dependent on expert experience. Recent advances in automatic parameter tuning utilize conventional tuning algorithms and machine learning approaches, but most still rely on slow, sequential iterative adjustments and struggle to retrieve relevant prior design knowledge when semantic information is modified or obscured. We present AstroTune, a structure-aware framework for LLM-assisted parameter tuning in chip design flows. By integrating both RTL source code and its abstract syntax tree (AST), AstroTune can retrieve and transfer design knowledge based on both semantic and structural relationships, remaining effective even if semantic information is modified or obscured. During execution, AstroTune employs stage-by-stage multi-candidate generation, propagation, and pruning via tournament selection to accelerate iterative tuning. Experiments on public benchmark suites show that AstroTune achieves superior tuning quality and substantial runtime reduction compared to state-of-the-art works.
