TL;DR
PANDA is an LLM-enhanced framework that automates analog circuit design from high-level intent all the way to final physical layout. By utilizing a Large Language Model (LLM) as a central coordinator to manage dependencies across topology, sizing, and layout stages, PANDA cuts design turnaround time from weeks to hours while maintaining performance through automated post-layout feedback loops.
What Problem Does This Paper Address?
Analog circuit design remains a stubbornly manual and experience-driven field compared to its highly automated digital counterpart. Traditional analog Electronic Design Automation (EDA) tools operate in isolated silos: topology synthesis, transistor sizing, placement, and routing are treated as independent steps.
This fragmentation creates three major "breakpoints":
- Lack of Intent: Conventional optimization algorithms explore design spaces blindly without understanding the high-level engineering intent.
- Stage Isolation: Schematic sizing is optimized without any awareness of physical layout realities, breaking downstream optimization loops.
- Open-Loop Execution: Existing flows cannot autonomously iterate based on post-layout simulation results. Human engineers must manually extract parasitic effects, re-evaluate performance, and tweak parameters.
PANDA aims to bridge these gaps by creating an end-to-end, intent-centric, self-improving design flow.
Methods
PANDA introduces an "agentic" workflow where an LLM acts as a semantic planner and constraint generator, while specialized, deterministic back-end EDA tools handle physical execution. The framework breaks the design process down into four tightly coupled stages using reusable task units called Skills:
- Topology Synthesis: PANDA leverages AnalogXpert to translate high-level user specifications into a subcircuit-level SPICE netlist (
topology.json). - Analog Sizing: Using MOSTAR (a multi-stage hierarchical Bayesian optimization engine that utilizes Graph Neural Networks), the framework determines optimal transistor parameters like width, length, and finger numbers (
sizing.json). - Constraint-Driven Placement: The system translates schematic structures into geometric layouts, optimizing for area and wirelength while strictly respecting symmetry and grouping constraints (
pl.json). - Routing and Layout Synchronization: Detailed routing is executed via SAGERoute. Crucially, PANDA automatically runs Parasitic Extraction (PEX) and post-layout simulations, feeding the resulting electrical metrics back to the LLM to iteratively refine design constraints until the target specifications are met.
[Design Intent] ➔ (LLM Coordinator) ➔ [Topology] ➔ [Sizing] ➔ [Placement] ➔ [Routing] ➔ [PEX / Simulation] ──(Feedback Loop)──↺
Datasets
As this is an EDA framework paper rather than a traditional machine learning dataset paper, the authors do not train on standard open-source datasets. Instead, they evaluate PANDA's performance using two practical circuit design case studies:
- OTA: A three-stage Operational Transconductance Amplifier with a differential input and single output.
- COMP: A StrongARM comparator circuit.
Key Insights and Ideas
- LLMs as Orchestrators, Not Creators: Instead of forcing the LLM to generate complex layout geometries directly (a task LLMs struggle with), PANDA uses the LLM to interpret human intent and generate structured constraints for robust, domain-specific EDA tools.
- Closing the Parasitic Loop: By automating the feedback loop between physical layout simulation (PEX) and early-stage optimization, the framework mimics a human engineer's iterative refinement process.
- Traceable Artifacts: Every stage of the pipeline outputs explicit, machine-readable JSON files. This eliminates opaque prompt chaining and ensures the AI's design decisions are entirely auditable and reproducible.
Key Results
- Massive Turnaround Speedup: PANDA reduces the analog design cycle from days or weeks of manual engineering iteration down to several hours.
- OTA Case Study: The framework successfully delivered a physical layout from a text prompt. The post-layout circuit retained a high gain of 76.11 dB (target: >75 dB) and a robust phase margin of 65.3° (target: >60°).
- Realistic Parasitic Tracking: The post-layout simulation successfully captured performance degradation—such as the Unity Gain Bandwidth (UGB) dropping from 10.84 MHz pre-layout to 4.192 MHz post-layout—proving that the framework accurately accounts for real-world physical overhead.
Limitations and Future Work
- Frequency Target Degradation: Although the main amplification behavior was preserved in the OTA case study, the post-layout frequency target (UGB) was not fully maintained due to severe parasitic effects. This indicates a need for better early-stage parasitic estimation.
- Future Directions: The authors plan to strengthen the framework's learning capabilities and portability. This includes integrating more powerful planning models, tightening the coupling between performance estimation and physical optimization, and developing cleaner service interfaces for back-end EDA tools.
Summary updated