Introduction
Designing analog circuits remains a predominantly manual and labor-intensive task. A primary challenge is ensuring robustness against PVT (Process, Voltage, and Temperature) variations and random silicon mismatches. These variations occur at two scales: global variations, which affect an entire wafer, and local mismatches, which occur between transistors on a single chip.
To guarantee manufacturing yield, engineers typically rely on exhaustive Monte Carlo simulations to verify "corner" cases. However, as design complexity grows, the high-dimensional search space makes traditional brute-force simulation prohibitively slow. GLOVA (Global and Local Variation-Aware design) addresses this by introducing an automated framework that optimizes for worst-case reliability with unprecedented sample efficiency.
The GLOVA Framework
GLOVA operates through a synergistic two-phase process: an optimization phase driven by risk-sensitive reinforcement learning (RL) and a verification phase utilizing hierarchical statistical filtering.
1. Optimization Phase: Risk-Sensitive RL
Unlike standard RL agents that aim to maximize average rewards, GLOVA employs Risk-Sensitive RL. This approach specifically penalizes designs that fail in worst-case scenarios, mimicking the "safety-first" intuition of expert human designers.
Key to this phase is an Ensemble-based Critic. By using multiple neural networks to predict performance, the system estimates a "reliability bound." This allows the agent to navigate uncertainty in the design space without requiring thousands of samples at every iteration. As shown in the performance tracking, the agent focuses on lifting the "Design Reliability Boundary" above the required constraints.
2. Verification Phase: Hierarchical Filtering
To avoid wasting compute on sub-optimal designs, GLOVA implements an "early-exit" strategy:
- μ-σ Evaluation: The system runs a small subset of simulations and uses the mean (μ) and standard deviation (σ) to statistically predict failure. If a design is likely to fail, it is rejected immediately.
- Simulation Reordering: Using Pearson correlation, GLOVA identifies which variations are most likely to cause failure and tests those first. This "worst-first" approach catches bugs significantly earlier than random testing.
Experimental Evaluation
The framework was tested on three 28nm CMOS circuits: a StrongARM Latch (SAL), a Floating Inverter Amplifier (FIA), and a complex DRAM Core (comprising an Offset Cancellation Sense Amplifier and subhole circuitry). The DRAM core is particularly challenging due to its sensitivity to local mismatches and massive parasitic capacitances.
Key Results
Compared to state-of-the-art tools like PVTSizing and RobustAnalog, GLOVA demonstrated:
- Extreme Efficiency: Up to an 80.5× improvement in sample efficiency and a 76.0× reduction in runtime for the FIA testcase.
- Robustness: In the DRAM core evaluation, GLOVA maintained a 100% success rate in finding valid designs, whereas previous tools saw success rates drop as low as 53%.
- Scalability: The framework converged on robust solutions for complex circuits with significantly fewer RL iterations than competing methods.
Conclusion and Future Outlook
GLOVA demonstrates that treating circuit verification as a statistical inference problem—rather than a brute-force task—can yield massive gains in design speed. By shifting the RL objective from average performance to risk avoidance, GLOVA aligns automated tools with the practical requirements of semiconductor manufacturing.
Future work aims to integrate layout-dependent effects (LDE) directly into the optimization loop. Moving beyond schematic-level intelligence to the physical layout phase represents the next major frontier in automated analog electronic design automation (EDA).
Summary updated