Dr. RTL
An agent-driven framework that uses Claude Code sub-agents to iteratively optimize RTL timing, verifying every proposed transformation against the original design with sequential equivalence checking before promoting it.
Dr. RTL is an agent-driven framework for RTL timing optimization presented at ICCAD 2026. Claude Code sub-agents analyze synthesized timing reports, propose RTL-level transformations, and verify each attempt against the original design using sequential equivalence checking, running through up to ten major versions with five minor attempts each.
For every minor attempt, a timing-analyzer agent selects the most critical timing paths and diagnoses their causes, an optimizer agent applies transformations that equivalence checking then validates for functional correctness, and a synthesis-evaluator agent runs synthesis and equivalence checking remotely and records a score combining timing and area. The lowest-scoring passing minor version is promoted to the next major version, and the process stops after ten major versions or when a full round produces no improvement. Separate skill-extractor agents distill reusable optimization patterns from completed runs.
Synthesis requires Synopsys Design Compiler, Synopsys Formality and Cadence JasperGold running on a remote EDA host, reached over paramiko and scp.