Dr. RTL: Autonomous agentic framework for realistic RTL timing optimization
Dr. RTL is an agentic framework that autonomously optimizes Register Transfer Language (RTL) designs through multi-agent critical-path analysis and reusable optimization skills, achieving 21% WNS and 17% TNS improvements over commercial synthesis tools on real-world designs.
Score breakdown
Developers and hardware engineers optimizing RTL designs can now use an agentic framework that learns and reuses optimization strategies across designs, achieving better performance and area metrics than commercial tools without manual rule engineering.
- 01Dr. RTL uses a multi-agent framework for critical-path analysis, parallel RTL rewriting, and tool-based evaluation in closed-loop optimization
- 02The framework introduces group-relative skill learning to distill optimization experience into a reusable skill library of 47 pattern–strategy entries
- 03Evaluated on 20 real-world RTL designs, Dr. RTL achieves 21% WNS improvement and 17% TNS improvement over commercial synthesis tools
Dr. RTL addresses critical gaps in automated RTL optimization by moving beyond the unrealistic evaluation settings and limited methods of prior work. Existing approaches rely on manually degraded small-scale designs, weak open-source tools, and coarse design-level feedback with simple pre-defined rewriting rules. In contrast, Dr. RTL establishes a realistic evaluation environment using challenging real-world RTL designs and an industrial EDA workflow.
A novel contribution is group-relative skill learning, which compares parallel RTL rewrites and distills optimization experience into an interpretable skill library.
The framework employs a multi-agent architecture that performs closed-loop optimization through three key components: critical-path analysis, parallel RTL rewriting, and tool-based evaluation. A novel contribution is group-relative skill learning, which compares parallel RTL rewrites and distills optimization experience into an interpretable skill library. This library currently contains 47 pattern–strategy entries that enable cross-design reuse and accelerate convergence, with the capability to evolve over time. On 20 real-world RTL designs, Dr. RTL achieves average WNS (Worst Negative Slack) and TNS (Total Negative Slack) improvements of 21% and 17% respectively, along with a 6% area reduction, outperforming industry-leading commercial synthesis tools.
Key facts
- 01Dr. RTL uses a multi-agent framework for critical-path analysis, parallel RTL rewriting, and tool-based evaluation in closed-loop optimization
- 02The framework introduces group-relative skill learning to distill optimization experience into a reusable skill library of 47 pattern–strategy entries
- 03Evaluated on 20 real-world RTL designs, Dr. RTL achieves 21% WNS improvement and 17% TNS improvement over commercial synthesis tools
- 04Dr. RTL achieves 6% area reduction compared to industry-leading commercial synthesis tools
- 05The skill library can continue evolving over time to improve cross-design optimization and convergence speed