TL;DR
Meta-RTL is a self-evolving framework that automatically designs the "harness"—the search logic and execution procedure—used by AI agents to optimize hardware designs. By using an LLM to analyze failures and iteratively rewrite its own optimization strategies based on real-world EDA feedback, Meta-RTL significantly outperforms hand-engineered agents, nearly doubling optimization utility on unseen hardware designs.
What Problem Does This Paper Address?
Optimizing Register-Transfer Level (RTL) designs for timing and area is a cornerstone of digital logic design. While Large Language Models (LLMs) have shown promise in rewriting RTL code to improve performance, they are usually governed by a "harness": a hand-written script that decides which design candidates to keep, how to prompt the model, and how to allocate the search budget.
The problem is that these human-designed harnesses often embed rigid assumptions—for example, only keeping the "best" candidate at each step. In reality, the best hardware optimizations often require a "stepping stone" approach, where a design might temporarily look worse (e.g., larger area) before a subsequent transformation makes it significantly better. Meta-RTL asks: instead of humans guessing the best search strategy, can we let the system evolve its own?
Methods
Meta-RTL employs a dual-loop evolutionary architecture:
- The Inner Loop: A fixed LLM (DeepSeek-V4.1-Flash) acts as the "worker," optimizing a specific RTL design under the rules of the current harness. It interacts with standard Electronic Design Automation (EDA) tools for synthesis and timing analysis.
- The Outer Loop: A "proposer" LLM (GPT-5.6 Sol) acts as the "engineer." It examines the logs and failures of the inner loop to hypothesize why the current strategy is failing. It then rewrites the harness code (Python) to implement new search mechanisms.
The evolution process follows four distinct stages:
- EDA Failure Attribution: Identifying why transformations failed (e.g., timing regressions or equivalence check failures).
- Harness Hypothesis: Proposing a targeted code change to the harness.
- Cost-Aware Probing: Testing the new harness on a small subset of designs to see if the hypothesis holds.
- Evidence-Based Revision: Finalizing the harness based on probe results.
Datasets
The authors evaluated Meta-RTL using two primary benchmarks:
- Development Set: 20 designs from the Dr. RTL benchmark, covering arithmetic, control logic, and communication modules. This set was used to evolve the harness.
- Unseen Evaluation Set: 10 more complex designs from the IC-RTL benchmark (e.g., image processing, hardware acceleration). These were held out entirely during the evolution phase to test how well the evolved strategies generalize.
Key Insights and Ideas
- Emergent Search Strategies: Through evolution, the system "discovered" sophisticated techniques that humans might overlook. These include Pareto Frontier Memory (keeping a pool of designs that trade off area vs. timing) and Executable Transformation Workspaces (maintaining stateful programs to refine complex rewrites).
- The "Stepping Stone" Effect: The evolved harness learned to preserve intermediate designs that were technically "worse" than the original but provided a better starting point for future optimizations.
- Generalization is Key: A major finding was that a harness evolved on simple arithmetic and control designs transferred exceptionally well to complex image processing and acceleration hardware, outperforming agents specifically designed for those tasks.
Key Results
Meta-RTL* (the best evolved harness) demonstrated significant gains over state-of-the-art baselines like Dr. RTL and RTLScout:
- Timing Improvements: On the 20 development designs, it reduced aggregate Worst Negative Slack (WNS) and Total Negative Slack (TNS) violations by 7.8% and 11.0%, respectively.
- Superior Generalization: On the 10 unseen designs, the improvements were even more dramatic, reducing WNS/TNS by 20.5% and 14.0%, with a 12.6% reduction in total area.
- Utility Score: Meta-RTL achieved a mean utility of 0.2660 on unseen designs, nearly doubling the 0.1388* achieved by the hand-engineered Dr. RTL.
Limitations and Future Work
- Computational Cost: The evolution process is resource-intensive, requiring multiple rounds of synthesis and LLM calls. While this cost is amortized if the harness is reused across many designs, it remains a barrier for small-scale projects.
- Stochasticity: LLMs are inherently probabilistic. The authors note that the robustness of these findings to LLM stochasticity warrants further evaluation through repeated trials.
- Future Directions: The authors suggest extending this "meta-evolution" approach to other parts of the chip design flow, such as logic synthesis heuristics or physical placement and routing.
Summary updated