AI Agents Streamline Chip Verification and RTL Debugging
A team from Columbia University, Google DeepMind, and Google has introduced a system called Back-to-the-Future that uses AI agents to assist with verifying chip designs. The approach organizes large simulation outputs into a structured database and lets multiple AI agents turn plain-language verification questions into database queries, while linking signal irregularities to RTL versions. On a test set of 150 queries, the system reached 95.33% execution accuracy, suggesting a route toward more automated post-simulation debugging.
Yang, Lobov, and Karpati describe Back-to-the-Future as an end-to-end agent-based framework for chip design verification. It addresses a gap: most prior LLM work in EDA has focused on generating static RTL code, while post-simulation verification and interactive waveform debugging remain largely manual.
BTTF converts large unstructured simulation dumps into a structured SQLite database. A multi-agent engine then turns natural-language verification questions into SQL queries that understand the database schema and links irregular signal behavior to RTL code versions. On a 150-query benchmark, it achieved 95.33% execution accuracy.
If such AI-agent verification tools mature, chip designers and verification engineers may spend less time manually tracing waveforms and more time reviewing higher-level results. Faster debugging could shorten development cycles for complex AI chips, potentially affecting hardware companies, their customers, and ultimately users of devices that depend on these chips. However, accuracy, reliability, and accountability remain important, since verification errors