ParasGB: New Benchmark Suite for AMS Circuit Parasitic Estimation

A significant hurdle in the advancement of GNN-based parasitic modeling for analog and mixed-signal (AMS) circuits has been the absence of publicly available, high-fidelity RC benchmarks that support reproducible evaluation. To bridge this gap, researchers have introduced ParasGB, the first open-source benchmark suite specifically designed for pre-layout parasitic parameter prediction on circuit graphs. This suite aims to facilitate early-stage estimation of parasitic capacitance and resistance, which are increasingly critical for AMS circuit performance in deep submicron manufacturing processes, often leading to costly layout iterations. ParasGB offers a collection of large-scale, heterogeneous RC networks. These networks are extracted using commercial Electronic Design Automation (EDA) tools from actual tape-out-proven designs, ensuring their relevance and realism. The suite also includes a unified evaluation protocol that covers node-level ground capacitance, edge-level resistance, and edge-level coupling capacitance. Within this standardized framework, the developers have benchmarked various Graph Neural Network (GNN) architectures using a consistent training pipeline. This process has highlighted several challenges inherent in parasitic estimation, such as extreme label imbalance, long-tailed parasitic distributions, and strong structural heterogeneity. By providing a physically grounded and standardized benchmark, ParasGB creates an open platform for reproducible research in circuit graph learning and the development of parasitic-aware models, which is crucial for modern chip design.