TL;DR
PVTSizing 2.0 is a high-speed, "agent-ready" analog circuit sizing tool that bridges the gap between large language model (LLM) reasoning and precise numerical optimization. By combining loss-guided training with an optional LLM "circuit adviser," it achieves a 7.3× speedup over previous methods while ensuring designs remain robust across all process, voltage, and temperature (PVT) corners.
What Problem Does This Paper Address?
Analog circuit sizing—the process of picking device dimensions to meet performance specs—is a classic "whack-a-mole" problem. A set of parameters might work perfectly at room temperature but fail miserably in extreme cold or at low supply voltages. To ensure reliability, engineers must size circuits to satisfy constraints across dozens of PVT corners.
Traditional automated tools often struggle with this "multi-PVT" optimization because an adjustment that helps one corner often hurts another. Furthermore, while LLMs are great at high-level reasoning, they are notoriously bad at the precise numerical "fine-tuning" required for circuit design. Existing Electronic Design Automation (EDA) workflows also require significant manual effort to coordinate between simulators, scripts, and optimizers.
Methods
PVTSizing 2.0 introduces a two-layered approach: an Agentic Workflow for task coordination and a Fast Numerical Optimizer for the heavy lifting.
- Agentic Workflow: Instead of a human designer manually editing config files and checking logs, an LLM-based agent (like "Claude Code" or "Codex") orchestrates the process. It translates natural language instructions into tool inputs, launches simulations, and summarizes results.
- The Optimizer Core: The numerical engine is upgraded with two key features:
- Loss-Guided Training: It avoids wasting time on unnecessary model updates. If the internal "critic" model is already fitting the data well, it skips training cycles to save time.
- Critic-Guided Local Refinement: It uses Bayesian Optimization (specifically TuRBO) to explore local candidates. The "critic" screens these candidates across all corners before they are even sent to the simulator, preventing "regressions" where a local improvement breaks a different PVT corner.
- LLM Knowledge Injection: This is the "secret sauce." The optimizer can optionally ask an LLM adviser for help. Instead of the LLM picking exact numbers, it suggests search directions based on circuit topology (e.g., "try adjusting these two capacitors together to balance noise and peaking"). The optimizer then determines the exact magnitudes through simulation.
Datasets
The authors evaluated PVTSizing 2.0 on four real-world analog circuit benchmarks:
- FC-OTA: A folded-cascode operational transconductance amplifier.
- BGR: A bandgap reference circuit (highly sensitive to PVT).
- CMP: A StrongARM comparator.
- MS-OTA: A complex multi-stage OTA with 42 design variables.
These were tested across multiple process nodes (SMIC 180nm down to 40nm) and up to 45 PVT corners per circuit.
Key Insights and Ideas
- Separation of Concerns: The paper argues that LLMs shouldn't be the "solver." Instead, they should be the "coordinator" (agent) and the "strategist" (adviser), leaving the "tactical" numerical search to dedicated algorithms.
- Directions over Dimensions: The most effective way to use LLM circuit knowledge is to have it suggest relationships between parameters (directions in search space) rather than absolute values.
- Efficiency isn't just about Sample Count: While reducing the number of simulations is important, PVTSizing 2.0 shows that reducing the computational overhead between simulations (via loss-guided training) is a major source of speedup.
Key Results
- Reliability: PVTSizing 2.0 achieved a 100% success rate (40/40 runs) across all benchmarks, whereas traditional optimizers like BFGS or Hooke-Jeeves often failed on the more complex BGR and MS-OTA circuits.
- Speedup: The standalone optimizer achieved a 7.3× speedup in mean runtime compared to the original PVTSizing.
- LLM Boost: When paired with the "GPT-6 Astra" adviser, the system saw an additional 39.9% reduction in runtime and used 35.4% fewer simulation samples compared to the standalone version.
Limitations and Future Work
- Model Dependency: The benefit of "knowledge injection" depends heavily on the LLM's reasoning capability. As shown in the paper, older or smaller models (like GPT-5.5) provided negligible benefits or even slightly increased runtime due to overhead.
- Confidentiality: In industrial settings, sending circuit netlists to cloud-based LLMs is often a non-starter. Future work will likely focus on using local, specialized models that can run within a company's secure environment.
- Inference Costs: While the LLM saves simulation time, the cost of LLM API calls must be balanced against the cost of compute for simulations.
Summary updated