TL;DR
This paper introduces Dr. RTL, a self-improving multi-agent framework that autonomously optimizes hardware designs (RTL) to achieve better performance and area than even top-tier commercial synthesis tools. By grounding Large Language Models (LLMs) in real-world Electronic Design Automation (EDA) feedback and a reusable "skill library," the authors demonstrate that AI can find significant optimization opportunities that traditional tools miss.
What Problem Does This Paper Address?
In digital circuit design, engineers write Register-Transfer Level (RTL) code, which synthesis tools then transform into a physical netlist of gates. There is a long-standing "synergy" problem: a clever RTL rewrite might improve a design's Power, Performance, and Area (PPA), but modern commercial synthesis tools are so advanced that they often perform these optimizations automatically.
This raises a critical question: Which RTL optimization opportunities actually remain after a high-end synthesis tool has done its job?
Most existing research evaluates AI-driven RTL optimization against "weak" open-source tools or manually degraded designs. This paper argues that for AI to be truly useful, it must be benchmarked against realistic, industrial-grade workflows and must be able to "ground" its decisions in actual timing and area reports from those tools.
Methods
The authors propose a three-pronged approach to autonomous hardware optimization:
- Realistic Benchmarking: They introduce two evaluation settings: RTL-OPT (reference-based), which compares AI results against human-optimized designs, and an Open-ended setting that challenges the AI to improve original, high-quality human-written code.
- Dr. RTL (The Agentic Framework): This is a multi-agent system that operates in a closed loop with commercial EDA tools.
- Timing Analysis Agent: Diagnoses bottlenecks by mapping critical paths from synthesis reports back to specific RTL structures.
- RTL Optimization Agent: Generates multiple candidate rewrites (e.g., sequential rebalancing, combinational simplification) in parallel.
- Evaluation Agent: Runs commercial synthesis and Sequential Equivalence Checking (SEC) to ensure the design still works correctly and measure PPA.
- Skill Learning: Instead of just guessing, Dr. RTL distills successful optimization trajectories into a Reusable Opt Skill Library. It learns "pattern-strategy" pairs (e.g., "if you see a long mux chain, try condition pre-computation") that it can apply to future, unseen designs.
- Synthesis Optimization: The authors extend the agent's reach beyond the code itself, exploring Script Tuning (optimizing the configuration of commercial tools) and Algorithm Discovery (rewriting the underlying logic synthesis algorithms in open-source tools like Yosys).
Datasets
The study utilizes two primary benchmarks:
- RTL-OPT: 36 handcrafted tasks covering arithmetic, control logic, and FSMs, paired with human-optimized "golden" references.
- Open-ended Dataset: 20 real-world designs (processors, communication interfaces, signal processing modules) averaging 812 lines of code.
Key Insights and Ideas
- LLMs need a "Ground Truth" Loop: Single-shot LLMs (like GPT-4 or DeepSeek) struggle to beat commercial synthesis. They only succeed when they can see the timing reports and iterate based on failures.
- The "Skill" Factor: By accumulating a library of strategies, the agent converges on optimal designs much faster. This allows the system to improve over time without needing to retrain the underlying LLM.
- Beyond the Code: Optimization isn't just about the RTL. Significant gains (up to 11% timing improvement) can be found just by having an agent tune the synthesis scripts or discover design-specific algorithms.
- Verification is the Boundary: The authors emphasize "Verification-Constrained Optimization." The agent is free to explore, but formal verification (SEC) acts as the guardrail to ensure the hardware's behavior never changes.
Key Results
- Dr. RTL Performance: On 20 real-world designs, Dr. RTL achieved average improvements of 21.3% in Worst Negative Slack (WNS) and 16.9% in Total Negative Slack (TNS), while also reducing area by 5.8% compared to commercial synthesis baselines.
- Script Tuning: Agent-driven tuning of commercial synthesis scripts provided an additional 7.4% to 11.1% timing improvement.
- Algorithm Discovery: In open-source flows, the agent discovered customized algorithms that reduced the Area-Delay Product (ADP) by 18.7%.
Limitations and Future Work
- Power Estimation: The current framework excludes power optimization because reliable power estimation requires workload-dependent switching activity, which is harder to simulate autonomously.
- Cross-Layer DSE: The authors envision a future where agents don't just look at RTL or synthesis scripts in isolation, but jointly explore the entire design space—from high-level architecture down to the specific EDA algorithms used to build the chips.
Summary updated