CFIRSTNET: Comprehensive Features for Static IR Drop Estimation with Neural Network
Yu-Tung Liu, Yu-Hao Cheng, Shao-Yu Wu, Hung-Ming Chen · 2025
Read CFIRSTNET: Comprehensive Features for Static IR Drop Estimation with Neural Network by Yu-Tung Liu, Yu-Hao Cheng, Shao-Yu Wu, Hung-Ming Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12168.txt.
Introduction
Yu-Tung Liu, Yu-Hao Cheng, Shao-Yu Wu, and Hung-Ming Chen wrote this computer science paper, published in 2025. It speeds up a key chip design check, estimating voltage drop across a chip's power grid, by teaching a neural network on both picture-like layout features and circuit-list features.
Voltage drop forecasts guard a chip's reliability and speed, but the classic simulation is slow. Image-only models miss wiring detail, and list-only models miss spatial patterns. The authors' network, CFIRSTNET, fuses both feature families in one frame. A custom convolutional network reads the power grid's features and predicts the drop map. On the open benchmark from a 2023 design contest, the method took the best quality score. The authors present fast, accurate drop estimation as a practical need as chips keep shrinking.
Catalog record
| CREATOR | Yu-Tung Liu, Yu-Hao Cheng, Shao-Yu Wu, Hung-Ming Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-12168
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CFIRSTNET: Comprehensive Features for Static IR Drop Estimation with Neural Network Yu-Tung Liu, Yu-Hao Cheng, Shao-Yu Wu, Hung-Ming Chen Abstract . IR drop estimation is now considered a first-order metric due to the concern about reliability and performance in modern electronic products. Since traditional solution involves lengthy iteration and simulation flow, how to achieve fast yet accurate estimation has become an essential demand. In this work, with the help of modern AI acceleration techniques, we propose a comprehensive solution to combine both the advantages of image-based and netlist-based features in neural network framework and obtain high-quality IR drop prediction very effectively in modern designs. A customized convolutional neural network (CNN) is developed to extract PDN features and make static IR drop estimations. Trained and evaluated with the open-source dataset, experiment results show that we have obtained the best quality in the benchmark on the problem of IR drop estimation in ICCAD CAD Contest 2023, proving the effectiveness of this important design topic. Conference: IEEE/ACM International Conference on Computer-Aided Design; October 27–31, 2024; New York, NY, USA IEEE/ACM International Conference on Computer-Aided Design (ICCAD ’24), October 27–31, 2024, New York, NY, USA DOI: 10.1145/3676536.3676756 ISBN: 979-8-4007-1077-3/24/10 1. Introduction Analyzing the on-chip power delivery network (PDN) is critical to the modern integrated circuit (IC) design flow. Figure 1 is an illustration of a typical PDN. PDNs could be modeled as a network of resistance connecting the power pads (C4 bumps) and the standard cells. Static IR drop verification is an imperative step in PDN analysis. With the advance of technology, the IR drop issue occurs at the lower nodes leading to performance degradation. Moreover, excessive IR drop might also cause functional failure. Chip designers are supposed to have IR drop sign-off before the tape-out process and ensure to meet the IR drop constraint. If there is any IR drop violation, designers might need to perform the optimization step such as Engineer Change Order (ECO) to reorganize the circuit and attempt to resolve the problem. Figure 1. Illustration of a typical PDN structure: a resistive network between the power sources (C4 bumps) and the standard cells. However, during every iteration of ECO, EDA tools will assess the circuit repeatedly, i.e. the worst-case voltage drop of the circuit will be calculated. The conventional method to retrieve the voltage of every node was solving the linear equation ( 1 ) where G was a conductance matrix, V the target voltage vector to be solved, and J the vector of independent current sources. The number of nodes surged as the amount of cells used in modern design increased tremendously. The considerable computational time of the conventional nodal analysis method was so expensive that it became a costly overhead. Therefore, several methods were proposed to trade accuracy for speed and replace the traditional solution with a faster and more accurate IR drop estimation. (1) GV=J While previous works have proposed various approaches to trade accuracy with speed, they all faced limitations. Non-machine learning methods were swamped by complex circuits with hefty amounts of nodes. On the other hand, image-based or netlist-based machine learning (ML) methods focused on certain circuit features solely. We propose CFIRSTNET, a CNN approach along with comprehensive feature extraction to speed up the static IR drop estimation problem with minimal accuracy trade-off. In addition to the provided image-based features, netlist-based inputs are analyzed with KVL, KCL, and superposition in the feature extraction stage to model the electrical features of the PDN. The custom CNN model takes image-based features and netlist-based features for input and generates an IR drop prediction map with reasonable scaling, providing more accurate and more efficient estimations. Tested by the ICCAD CAD Contest 2023 open-source benchmark, CFIRSTNET mitigates the overhead of expensive computational resources and provides lower estimation errors in comparison to prior arts. This work could be further implemented in modern design flow and relieve the pain of heavy computational expenses. The main contributions of this paper are listed as follows: • CFIRSTNET harnesses the benefit of both image-based and netlist-based inputs and introduces a customized CNN model to perform static IR drop estimation. To the best of our knowledge, this is the first work regarding this combination as CNN model inputs that preserve circuit features while ensuring efficiency. • We propose the Hypothetical IR Drop Distillation that analyzes the power grid from the SPICE-based data in linear time and provides more comprehensive features for the customized CNN simultaneously. Our approach is independent of fabrication technologies, circuit designs (with different current source and voltage source distributions), and PDN structures. • The proposed model is scalable due to the properties of CNNs. CFIRSTNET has been tested with various chip sizes and has exhibited strong estimation results. The performance of CFIRSTNET is not affected by macros or irregular PDN arrangement. Compared to 1st place winner of ICCAD CAD Contest 2023, CFIRSTNET reduces average error by 60%, max error by 66%, increases F1 score by 59%, and achieves a 21.8x speedup. The remainder of this paper is organized as follows. Preliminaries to this problem are discussed in section 2 . The details of CFIRSTNET will be elaborated in section 3 and 4 . Section 5 and 6 present the experimental setup and results. Finally, Section 7 concludes the paper and discusses the possibilities of exploring future works. 2. Preliminaries Previously proposed methods introduced non-ML methods by modeling the whole power grid and finding the solution to Equation ( 1 ). Panda et al. ( Panda et al., 2000 ) provided a modeling way of resistance-inductance-capacitance (RLC) circuits. Moreover, Panda et al. ( Panda et al., 2000 ) and Zhao et al. ( Zhao et al., 2002 ) ignored the influence of the inductance in lower layers due to the small number compared to resistance and capacitance. This simplicity made some matrices symmetric and positive definite, which enabled Cholesky factorization to speed up the works. When it comes to tens or even hundreds of millions of nodes in the circuit, flattened circuit models require tremendous memory capacity and computational resources. Zhao et al. ( Zhao et al., 2002 ) proposed a hierarchy method to lower the memory cost, and speed up the circuit simulation. They generated micromodels for local grids and simulated the whole network. Kozhaya et al. ( koz, 2002 ) assumed that the sink resistance is larger than the grid resistance, which bridged the gap of the voltage drop between nearby grids. Hence, they used a more coarse structure to process the PDN and accelerate the framework by mapping out the solution with interoperation. The methods mentioned above traded off the accuracy to the speed. Other works such as Qian et al. ( Qian et al., 2005 ) took the hierarchical random-walk approach. However, the run time of these simulated methods was still prone to the increased node number. Recently, many ML methods have cast light on the EDA problems, few of which explored the static IR drop estimation problem. Compared with the traditional simulation-based methods, the runtime of ML methods is not affected by the number of circuit nodes. Large circuits could be processed and the predictions could be made in an unprecedentedly short amount of time. The ML models could