Abstract
Procedural layout generators are essential for automating analog IC design, yet their creation remains a labor-intensive manual task. This paper introduces a CNN-based framework that automates the layout-to-generator conversion process by identifying sub-cells within existing designs. Utilizing a multi-scale architecture and 21-channel matrix representations of GDSII files, the system achieves a 290x speedup in sub-cell assignment with 99.3% precision, significantly reducing the engineering effort required for analog layout automation.
Introduction
In analog integrated circuit (IC) design, layout creation is a significant bottleneck. Unlike digital design, which benefits from high levels of automation, analog layouts are typically handcrafted to meet stringent performance and matching requirements. To mitigate this, designers employ procedural layout generators—scripts that produce layouts based on variable input parameters such as transistor width or finger count.
However, developing these generators is time-consuming. A common practice is "layout-to-generator conversion," where a proven manual layout is transformed into a script. The primary hurdle in this workflow is "sub-cell assignment": the manual identification of thousands of components (vias, transistors, etc.) to match them with existing library scripts. For complex designs, this manual checking can take hours of intensive labor.
Proposed Methodology
The authors propose a Convolutional Neural Network (CNN) framework to automate sub-cell assignment. The methodology involves several key innovations:
- Layout-to-Matrix Preprocessing: GDSII layout files are converted into 21-channel matrices. Each channel represents a specific physical layer (e.g., Metal1, Poly, Doping), preserving the 3D spatial context of the chip.
- Multi-Scale Architecture: To handle features of varying sizes—from individual vias to large transistor arrays—the model uses a parallel architecture with three modified ResNet base models processing resolutions of 64x64, 128x128, and 256x256 simultaneously.
- Classification Strategy: Instead of a standard Softmax function, the model utilizes Sigmoid functions for each class. This allows for a "Not Generatable" classification when the probability for all known generators is low, preventing false positives on unfamiliar structures.
Experimental Results
The model was evaluated using synthetic data, negative samples, and real-world layouts from 28nm and 65nm CMOS processes.
- Efficiency and Accuracy: For a wireline receiver (RX) containing 4,885 sub-cell instances, the CNN reduced the examination time from 88 minutes (manual) to just 18 seconds—a 290x speedup. The model achieved 99.3% precision and 100% recall on this dataset.
- Generalization: The framework demonstrated impressive cross-technology robustness. A model trained on 28nm data successfully identified sub-cells in a 65nm design, as it learns geometric patterns rather than absolute dimensions.
- Robustness: On a transceiver (TRX) dataset with unseen layouts, the model maintained a Top-1 accuracy of 96.1% and a Top-3 accuracy of 98.2%.
- Code Reuse: For a simple inverter chain, the tool automatically referenced and reused approximately 33,000 characters of existing generator code, sparing designers from tedious manual scripting.
Conclusion
The CNN-based approach effectively automates the most tedious aspect of layout-to-generator conversion. By significantly reducing manual effort while maintaining high precision, this technique paves the way for more scalable analog design automation. Future work will focus on expanding the generator library and further automating the definition of parametric relationships between components.
Summary updated