Session

DescriptionSpecifications remain the foundation of RTL design and verification, yet they are often ambiguous, inconsistent, incomplete, or quickly outdated. This session explores how AI can systematically improve specification quality and usability across the silicon lifecycle. We focus on using AI to detect inconsistencies, contradictions, missing or incorrect information, and violations of technical writing rules, while providing actionable recommendations. We also discuss generating specifications from meeting recordings, transcripts, and email threads, and deriving downstream documentation such as verification plans, coverage models, formal strategies, sign-off documents, microarchitecture descriptions, and low-level design specs. Beyond generation, we address practical challenges: handling large specifications beyond LLM context limits, incorporating diagrams and schematics via multimodal models, supporting automated diagram creation, and making specifications "AI-friendly." Finally, we examine incremental formalization, spec drift detection, legacy spec reconstruction, and AI-validated cross-company interface documentation to reduce integration and contractual risk.