EDATracer outperforms Cursor and Claude Code on chip design analysis

Chip design generates a mountain of heterogeneous artifacts—source files, scripts, logs, netlists, and reports—and finding the evidence needed to debug or optimize a design often means digging across all of them. While LLM agents have shown promise for EDA assistance, they struggle to ground reasoning in tool-generated evidence and lack public benchmarks for large-scale cross-artifact analysis. A team from Texas A&M University, led by Phat Tieu, tackles this with EDATracer, an agentic framework described in a new arXiv paper (arXiv:2608.04032).

EDATracer organizes design artifacts into a knowledge graph paired with a semantic vector index, letting LLM agents retrieve relevant evidence across heterogeneous sources. The researchers built a large-scale evaluation set: an 18.9 GB dataset containing 2,787 synthesizable open-source chip designs, plus a 90-question benchmark covering factual, statistical, and reasoning tasks. Across evaluated agents, EDATracer achieved the best pass@1 accuracy, outperforming Cursor and Claude Code by 6.4 and 7.2 percentage points on average—while consuming 2.0-3.2x fewer tokens. That combination of higher accuracy and lower cost suggests EDATracer could make LLM-driven EDA analysis practical for real-world chip design workflows, where debugging and design-flow understanding are critical bottlenecks.