ASPEN: LLM-Guided E-Graph Rewriting for RTL Datapath Optimization

Datapath RTL optimization is a challenging multi-objective task that balances power, performance, and area (PPA). Traditionally, this process is done manually due to the complexity of balancing conflicting objectives. Recent approaches have explored rule-based rewriting using equality saturation with e-graphs, which solves the phase ordering problem but relies on manually designed rewrite rules and proxy PPA cost models. Meanwhile, large language models (LLMs) have been applied to RTL code generation and optimization due to their reasoning and programming capabilities. However, existing LLM-based methods lack a structured approach to represent and explore Pareto-optimal design points while ensuring equivalence in transformed RTL. We propose ASPEN, a system that leverages LLMs to guide e-graph rewriting while incorporating accurate, detailed feedback from EDA tools. Our approach employs an agentic system to propose and select rewrite rules, using real PPA feedback for extraction rather than proxy models. This reduces the need for handcrafted rewrite rules and eliminates the need for designing proxy-based cost models. ASPEN achieves average improvements of 16.51% in area and 6.65% in delay over existing e-graph-based RTL optimization approaches across a diverse set of benchmarks. Our results demonstrate that combining LLM-driven e-graph rewrite with real PPA feedback enables scalable and effective datapath RTL optimization.