A Static IR Drop Intelligent Prediction System Based on Heterogeneous Feature Fusion

Abstract

The application discloses a static IR Drop intelligent prediction system based on heterogeneous feature fusion, aiming at solving the problems of high calculation complexity of traditional IR Drop analysis methods and inability of existing deep learning methods to effectively capture global spatial features and unit instance topology information of a power delivery network (PDN). The framework combines the dual branches of a convolutional neural network (CNN) and a graph neural network (GNN), realizes effective extraction and feature fusion compensation of PDN multi-scale global to local spatial power features through a long-distance and local detail encoder (LLE) and a hierarchical and adjacent compensation group (HACG) module of the CNN branch, and adaptively aggregates PDN topology features of heterogeneous neighbors of different orders through a unit instance voltage perception (CVA) module of the GNN branch. Experimental results show that, compared with existing advanced methods, the framework has a significant advantage in prediction accuracy, and the prediction error is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit (IC) design and analysis technology, and specifically to a static IR Drop intelligent prediction system based on heterogeneous feature fusion. Background Technology

[0002] In ultra-deep submicron integrated circuit design, IR drop analysis of the power transport network (PDN) is a core step in ensuring chip performance and reliability. IR drop, caused by current flowing through the resistance of the PDN, can lead to voltage fluctuations, timing violations, and even functional failures. Traditional analysis methods, based on numerical simulation (such as Modified Node Analysis, MNA), require solving high-dimensional linear equations (GV = J), and the computational complexity increases exponentially with design size, especially inefficient in non-uniform power grids. While existing acceleration methods (such as multi-grid methods and boundary element methods) partially alleviate the computational burden, they still suffer from high memory consumption and long computation times for large-scale designs.

[0003] In recent years, deep learning methods based on convolutional neural networks (CNNs) (such as IREDGe and MAUnet) have achieved fast IR drop prediction through image-to-image mapping. However, these methods have the following limitations: due to the local perception characteristics of convolutional kernels, existing CNN models struggle to capture long-distance current superposition effects and the global spatial distribution patterns of PDNs, resulting in insufficient accuracy in predicting regional hotspots. Furthermore, the actual current path of a PDN is determined by its non-Euclidean topology, and when CNNs abstract the circuit as an image, a single pixel may contain multiple unit instances, leading to the loss of topological connectivity and unit-level current interference information. Existing methods rely on single-modal features (images or topology), making them insufficiently adaptable to complex PDN designs. Migrating to new processes or designs requires extensive retraining with labeled data, resulting in poor model generalization.

[0004] Meanwhile, while Graph Neural Networks (GNNs) have shown potential in circuit topology analysis (such as congestion prediction), their application in IR voltage drop prediction still faces challenges. Traditional GNNs often focus on first-order neighborhood aggregation, neglecting the influence of higher-order neighborhoods on voltage propagation, making it difficult to comprehensively characterize multi-order dependencies in complex current paths. Furthermore, existing methods have failed to effectively design differentiated feature fusion mechanisms for heterogeneous edge relationships (such as actual connections and virtual adjacencies), resulting in insufficient completeness and adaptability of topology feature extraction. These issues limit the application potential of existing technologies in efficiently and accurately predicting IR voltage drop distribution.

[0005] To address the aforementioned issues, this invention proposes an image-graph heterogeneous fusion framework (IGHF), which significantly improves the accuracy and generalization performance of IR voltage drop prediction by coordinating CNN and GNN branches and combining multi-scale spatial features and topology awareness capabilities, providing an efficient and reliable power integrity analysis tool for VLSI design. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to propose a static IR drop intelligent prediction system based on heterogeneous feature fusion. By combining CNN branches and graph neural network (GNN) branches, it can effectively capture global spatial features and topological information of PDN, thereby improving the accuracy of IR drop prediction.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: a static IRDrop intelligent prediction system based on heterogeneous feature fusion, the system comprising...

[0008] The data preprocessing and input construction module is used to convert the original chip layout and netlist files into feature images and graph structures;

[0009] The CNN branch employs a power-multiscale fusion U-shaped network architecture, which includes long-range and local detail encoders (LLE) and hierarchical and neighbor-compensated group (HACG) modules to extract multi-scale global-to-local PDN spatial features.

[0010] The GNN branch includes a cell instance voltage sensing module (CVA) used to sense the non-Euclidean connectivity of the PDN and extract topological features.

[0011] The heterogeneous feature fusion module is used to align and fuse CNN features and GNN features to generate the final IR down-prediction result.

[0012] Specifically as follows:

[0013] The content includes the following steps:

[0014] This invention first converts chip layout data into feature images: based on spatial coordinate mapping, feature images of power supply, leakage power, and minimum path resistance are extracted. The importance permutation (PI) filter is used to filter the features of the input image, excluding non-critical features such as switching power and cell frequency, and retaining features such as total power, leakage power, and minimum path resistance to enhance the convergence of the analysis and reduce redundant features.

[0015] An undirected bidirectional heterogeneous graph is constructed based on the .dspf file to represent the actual non-Euclidean PDN topology. This graph contains all cell instances as vertices, and the connections between them define two types of edge relationships—actual physical connection edges (R...). l) and virtual space adjacent edge R d The virtual edge passes through the Euclidean distance threshold (||v) i ,v j ||2≤δ) is dynamically generated to construct a non-Euclidean heterogeneous graph data structure.

[0016] In the CNN branch based on multi-scale global-local spatial feature extraction, a Long Range and Local Detail Encoder (LLE) is first used to capture multi-scale global and local PDN spatial power features. The LLE combines a Long Range Module (LRM) and a Local Detail Module (LDM). The LRM is composed of Long Range Blocks (LRBs), which consist of a receptive network and a convolutional feedforward network (FFN). The former introduces 7×7 deep separable convolutions with large kernels to significantly improve the effective receptive field, containing a wider range of power and current distribution information; the latter uses fewer channels and more efficient, dense 1×1 layers to manage computational workload. The overall design uses an inverted bottleneck residual structure to reduce memory costs and incorporates skip connections to avoid gradient vanishing.

[0017] The Local Details Module (LDM) employs two layers of 3×3 small kernel convolutions and the Leaky ReLU activation function to extract fine-grained local current features; multiple LDMs are stacked to capture multi-scale details.

[0018] The features output by the Long Range Module (LRM) and Local Details Module (LDM) are fused using a 1×1 little kernel convolution operation to integrate global and local information and form a richer feature representation.

[0019] In the decoder section, a lightweight group aggregation bridge (GAB) is used to design hierarchical and adjacent compensation groups (HACG) to adaptively rescale and guide the fusion and compensation of encoder features at different scales. HACG includes two sub-modules: hierarchical compensation (HC) and adjacent compensation (AC). The HC sub-module first designs multiple sets of dilated convolutions with different dilation rates for the top layer of the decoder, superimposing feature maps from different encoder layers to more comprehensively model voltage fluctuation trends. In the specific implementation, the top layer of the decoder combines the outputs from all other encoder layers, and adaptively extracts information using multiple sets of dilated convolutions with different dilation rates. This process is implemented through the lightweight group aggregation bridge (GAB) module, which combines depthwise separable convolutions and bilinear interpolation to align the feature sizes of low and high layers, and adaptively extracts information through four sets of convolution groups with different dilation rates.

[0020] The AC submodule primarily addresses the unavoidable local information degradation problem in downsampling layers. It adaptively rescales and compensates for the corresponding decoder layers by performing GAB (Graphical Access Bypass) interactions between adjacent encoder feature maps. Specifically, GAB interactions between adjacent encoder feature maps guide IR (Inverse Regression) prediction from higher to lower layers. Unlike traditional direct jump connections, AC reduces feature loss during upsampling by interacting with features between adjacent layers.

[0021] In the GNN branch, a Cellular Voltage Sensing (CVA) module is designed: firstly, through a heterogeneous graph attention layer, key information from heterogeneous edge connections is adaptively rescaled and aggregated. For the actual connection edge R... l With virtual adjacent edge R d The key interactions between learning nodes are used to obtain attention factors, and their average value is also used as the attention factor for the connection weights of the two heterogeneous edges in the propagation layer. For each type of edge relation (r∈{R... I ,R d}), compute node v i With neighbors v j Attention weights:

[0022]

[0023] Attention weights are adjusted using Softmax. Normalization It is type r in v i and v j Attention score between, N i Indicates v i Neighborhood set, As a matrix One element is v in type r. i For v j Attention weights; after normalization, neighbor features are aggregated by edge type to obtain the output features.

[0024]

[0025] and These are the input and output of the l-th attention layer, respectively, where r represents the edge type in the relation set R. It is a hyperparameter matrix of type r, a matrix Note the weight matrix in type r.

[0026] Voltage representations are learned from neighbors through adjacent virtual and physical connections, incorporating voltage correlation features from neighbors at different distances. The MixHop method is employed to allow learning relationships between neighbors at different distances, and these relationships are mixed using different operators to extract meaningful signals, even in extremely sparse connections. The formula is:

[0027]

[0028] It is an adjacency matrix with self-connections. The mean of the attention weights for all edge types. These are learnable weights.

[0029] The image features extracted by the CNN branch and the graph features extracted by the GNN branch are aligned and fused. Specifically, the final graph sequence of the GNN branch is combined with the image output of the CNN branch through linear fusion to generate the final IRDrop prediction sequence, making full use of the advantages of both types of features to improve the accuracy and robustness of the prediction.

[0030] Beneficial Effects: The IGHF framework proposed in this invention effectively solves the problems of traditional CNN methods in IR drop prediction, namely, difficulty in capturing global features and lack of PDN topology awareness, by combining CNN and GNN branches. It achieves:

[0031] 1. Significantly Improved Prediction Accuracy: Compared with existing technologies, this framework can significantly improve the accuracy of IR voltage drop prediction, outperforming existing state-of-the-art methods on multiple evaluation metrics. Experiments show that on the public dataset CircuitNet-ICISC, IGHHF reduces the MAE (mean absolute error) by 24.6% (0.378mV vs. 0.501mV) compared to the state-of-the-art method MAUnet, and by 55.0% (0.378mV vs. 0.840mV) compared to IREDGe; the PCC (Pearson correlation coefficient) reaches 0.973, which is better than MAUnet (0.961) and IREDGe (0.927).

[0032] 2. Enhanced robustness of hotspot area prediction: The AMMAE (mean maximum MAE) index is 10.5% lower than that of MAUnet (24.790mV vs. 27.717mV), indicating that IGHF has higher prediction stability for worst-case scenarios (hotspot areas).

[0033] 3. Efficient Modeling of Non-Euclidean PDN Topologies: Traditional CNNs cannot directly handle irregularly connected PDN topologies. The Cellular Voltage Awareness (CVA) module in the GNN branch, however, overcomes the limitation of traditional GCNs, which rely solely on first-order neighbors, by fusing self-loops, first-order neighbors, and second-order neighbors through MixHop graph convolution. Experiments show that the MAE is 1.171mV when using only the GNN branch (without CNN), significantly outperforming the global performance of pure CNN (0.389mV), validating the necessity of topology modeling. Attached Figure Description

[0034] Figure 1 This is a diagram of the IGH framework structure.

[0035] Figure 2 This is a diagram of the LLE module structure.

[0036] Figure 3 This is a structural diagram of the HACG module. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0038] Example: See Figures 1-3 A static IR Drop intelligent prediction system based on heterogeneous feature fusion, the system comprising:

[0039] The data preprocessing and input construction module is used to convert the original chip layout and netlist files into feature images and graph structures;

[0040] The CNN branch employs a power-multiscale fusion U-shaped network architecture, which includes long-range and local detail encoders (LLE) and hierarchical and neighbor-compensated group (HACG) modules to extract multi-scale global-to-local PDN spatial features.

[0041] The GNN branch includes a cell instance voltage sensing module (CVA) used to sense the non-Euclidean connectivity of the PDN and extract topological features.

[0042] The heterogeneous feature fusion module is used to align and fuse CNN features and GNN features to generate the final IR down-prediction result.

[0043] Specifically, the data preprocessing and input construction module is used for:

[0044] Feature images are generated based on spatial unit coordinate mapping, including features such as total power, leakage power, and minimum path resistance;

[0045] An unweighted bidirectional heterogeneous graph is constructed based on the .dspf file. This graph contains all cell instances and their connections, as well as virtual edges based on the Euclidean distance between cells.

[0046] Among them, the Long-Range and Local Detail Encoders (LLEs) of the CNN branch include:

[0047] Long Range Module (LRM) increases the receptive field through large kernel depth convolutions to capture global PDN features;

[0048] The Local Details Module (LDM) extracts fine-grained local PDN features through small kernel convolution;

[0049] Branch fusion layer, used to fuse features extracted by LRM and LDM.

[0050] The Long Range Module (LRM) specifically includes:

[0051] The perceptual network, by introducing large-kernel deep convolution, significantly improves the effective receptive field and includes a wider range of power and current distribution information;

[0052] Convolutional feedforward networks (FFNs) manage computational workload, including feature dimensions and model nonlinearity, using more efficient and dense 1×1 layers.

[0053] The hierarchical and adjacent compensation group (HACG) module of the CNN branch includes:

[0054] The hierarchical compensation (HC) submodule is used to combine encoder feature maps of different semantic levels and achieve multi-scale information fusion and interaction through multiple sets of dilated convolutions.

[0055] The Adjacent Compensation (AC) submodule is used to adaptively rescale and compensate the corresponding decoder layers by interacting with adjacent encoder feature maps, thereby reducing information loss during upsampling.

[0056] The unit instance voltage sensing module (CVA) of the GNN branch includes:

[0057] Attention layer: Design a heterogeneous unit attention mechanism to adaptively rescale and aggregate important information from heterogeneous edge connections of different types;

[0058] The propagation layer learns voltage representations from neighboring cells through adjacent virtual and physical connections, and uses the MixHop method to combine voltage-related features from neighbors at different distances.

[0059] Example 2: See Figure 1 A static IR Drop intelligent prediction algorithm based on heterogeneous feature fusion is described below.

[0060] S1: The publicly available CircuitNet-ICISC benchmark dataset was used, which provides over 7000 samples from five open designs at the 28nm process node. Each design includes different layouts, composite circuit netlists, and different cell utilization rates. The training and test sets were split in an 8:2 ratio.

[0061] S2: Extract features from the chip's layout and netlist files, including the spatial distribution of total power consumption, leakage power consumption, and minimum path resistance features, to form a 320×320 pixel grayscale image. Use a permutation importance algorithm (PI) to filter key features. For each channel, perform 100 random permutations. If the change in the model's MAE exceeds a threshold after permutation, the feature is retained; otherwise, it is discarded (e.g., switching power consumption, effective path resistance).

[0062] S3: Based on the chip's .dspf file, each standard cell instance is a graph node, and node attributes include total power consumption, leakage power consumption, and minimum path resistance (consistent with the image input). Constructing actual physical connection edges (R... l If two units are directly connected by a metal wire, add a bidirectional edge. Construct a virtual space adjacency edge R. d Calculate the Euclidean distance between the center coordinates of the cell. If the distance is less than or equal to δ, add a bidirectional virtual edge.

[0063] S4: Use the `HeteroData` object from the PyTorch Geometric library to store nodes and edges, with a node feature dimension of 3.

[0064] S5: In the CNN, a power multi-scale fusion Unet architecture is adopted. The Long-Range Local Encoder (LLE) consists of four levels of Long-Range Modules (LRMs) and Local Detail Modules (LDMs). In the LRM, the LRM is composed of Long-Range Blocks (LRBs), which consist of a receptive network and a convolutional feedforward network (FFN). The receptive network uses an inverted residual structure, first extracting global current distribution features through a 3×3 convolution followed by a large-kernel deep convolution (7×7 kernels) to reduce computational overhead, and finally normalizing through a layer. The convolutional feedforward network uses two 1×1 layers and a GeLU layer to reduce computation. In the LDM, two small-kernel convolutions (3×3 kernels) and a Leaky ReLU activation function are stacked to capture unit-level current details. The output features of the LRM and LDM are fused step-by-step through a branch fusion layer (3×3 convolution). The overall encoding structure is as follows: Figure 2 As shown.

[0065] S6: In the decoder hierarchy and adjacent compensation group HACG ( Figure 3The Hierarchical Compensation (HC) module integrates and fuses multi-level features from different scales using Group Aggregation Bridges (GABs). The Adjacent Compensation (AC) module uses GABs to compensate for information loss during upsampling. Finally, the feature maps are concatenated with those from cubic linear interpolation in the decoder and then aggregated via convolution.

[0066] S7: Convert the heterogeneous graph generated in step S2 and the CNN input into a sequence as the input to the GNN branch, and align the node features with the CNN branch image features using spatial coordinates.

[0067] S8: In the design of the Cell Instance Voltage Sensing Module (CVA), a Heterogeneous Graph Attention Network (HGAT) with Leaky ReLU activation is used in the attention layer to calculate the attention score for each edge type:

[0068]

[0069] S9: Calculate the attention weights for different edge classes using Softmax normalization, as shown in the following formula:

[0070]

[0071] Node features are aggregated by attention weights to generate a preliminary topological representation.

[0072] S10: Employs MixHop graph convolution to fuse self-loop, first-order, and second-order neighbor features.

[0073]

[0074] S10: Align and fuse the image features extracted by the CNN branch and the graph features extracted by the GNN branch, and generate the final IR reduction prediction result through a linear fusion module. The prediction result includes instance sequences and image output, and their sum is used as the total loss for supervised learning for model training.

[0075] S11: In the specific implementation, a loss function combining L1 / L2 loss is adopted. By combining these two losses, the accuracy and stability of the prediction can be comprehensively considered, improving the training effect of the model. The Adam optimizer is used, and five linear warm-ups and cosine annealings (1.5 × 10⁻⁶) are performed. -3 ) to (1×10 -5 ).

[0076] S12: Pre-train the CNN branch on 320×320 images for 100 epochs. During this process, a large amount of image data is used to train the CNN network, enabling it to learn the multi-scale spatial feature representations of the PDN. The purpose of pre-training is to ensure that the CNN branch converges quickly in subsequent joint training and provides good initial feature extraction capabilities.

[0077] S13: After pre-training, the parameters of the CNN branch are frozen, and then a GNN branch and a fusion block are added. The GNN branch is then jointly trained for 50 rounds. The purpose of this is to maintain the effective feature representations already learned by the CNN branch, while supplementing the topological information awareness capability of the PDN by introducing the GNN branch, thus laying the foundation for subsequent joint training. During the training phase, the model generates instance sequences and image outputs. The sum of these two outputs is used as the total loss for supervised learning to simplify the training process.

[0078] S14: The proposed IGH is compared with two excellent CNN-based methods, MAUnet and IREDGe. Accuracy and graph evaluation metrics, including Pearson correlation coefficient (PCC) and determination coefficient R, are adopted and comprehensively evaluated. 2 The metrics are: Mean Absolute Error (MAE) (mV) and Mean Maximum MAE (AMMAE) (mV). MAE is used to evaluate the accuracy of the forecast. AMMAE is used to evaluate the average error of the maximum IR drop, thus measuring our forecast performance in hotspot regions.

[0079] S15: Compare the performance of the three prediction models (IGHF, IREDGe, MAUnet) on five RTL hardware designs (ZERO-RISCY, RISCY, RISCY-FPU, VORTEX, NVDLA) and the overall results. The results show that the IGHF model performs better across all metrics (PCC, R...). 2 The IREDGe model significantly outperformed the other two models in terms of overall MAE (0.378) and AMMAE (24.790), especially in reducing errors: its overall MAE (0.378) and AMMAE (24.790) were 55% and 40% lower than IREDGe, respectively, and 24% and 20% lower than MAUnet. The IREDGe model performed the worst, with an MAE as high as 5.750 on NVDLA hardware, far exceeding IGFH (1.311) and MAUnet (1.374), and an AMMAE of 392.454 on VORTEX hardware, revealing its insufficient generalization ability to complex hardware. While MAUnet was inferior to IGFH, its overall PCC (0.961) and R... 2(0.874) is close to the level of IGHF, indicating a certain competitiveness in correlation metrics. From a hardware perspective, the prediction errors of VORTEX and NVDLA are significantly higher than those of the RISCY series hardware (e.g., IGHF's ​​AMMAE on VORTEX is 267.136, 12 times that of RISCY-FPU), suggesting that these two types of hardware may have higher complexity or data noise. Overall, IGHF exhibits the best prediction stability and accuracy, and is recommended as the first choice for high-precision scenarios.

[0080] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A static IR Drop intelligent prediction system based on heterogeneous feature fusion, characterized in that, include: The data preprocessing and input construction module is used to convert the raw chip layout and netlist files into feature images and graph structures; The CNN branch employs a power-multiscale fusion U-shaped network architecture, which includes long-range and local detail encoders as well as hierarchical and neighbor compensation group modules, for extracting multi-scale global-to-local PDN spatial features. The GNN branch includes a unit instance voltage sensing module, which is used to sense the non-Euclidean connectivity of the PDN and extract topological features. The heterogeneous feature fusion module is used to align and fuse CNN features and GNN features to generate the final IR down-prediction result; The hierarchical and neighbor-compensation group modules of the CNN branches include: The hierarchical compensation submodule is used to combine encoder feature maps of different semantic levels and achieve multi-scale information fusion and interaction through multiple sets of dilated convolutions. The neighboring compensation submodule is used to adaptively rescale and compensate the corresponding decoder layers by interacting with the feature maps of neighboring encoders, thereby reducing information loss during the upsampling process.

2. The static IR Drop intelligent prediction system based on heterogeneous feature fusion according to claim 1, characterized in that, The data preprocessing and input construction module is specifically used for: Feature images are generated based on spatial cell coordinate mapping, including total power, leakage power, and minimum path resistance features; An unweighted bidirectional heterogeneous graph is constructed based on the .dspf file. This graph contains all cell instances and their connections, as well as virtual edges based on the Euclidean distance between cells.

3. The static IR Drop intelligent prediction system based on heterogeneous feature fusion according to claim 1, characterized in that, The long-range and local detail encoders in CNN branches include: The long-range module increases the receptive field through large-kernel deep convolutions to capture global PDN features; The local detail module extracts fine-grained local PDN features through small kernel convolution; The branch fusion layer is used to fuse features extracted from long-range modules and local detail modules.

4. The static IR Drop intelligent prediction system based on heterogeneous feature fusion according to claim 1, characterized in that, The unit instance voltage sensing module of the GNN branch includes: Attention layer: Design a heterogeneous unit attention mechanism to adaptively rescale and aggregate important information from heterogeneous edge connections of different types; The propagation layer learns voltage representations from neighboring cells through adjacent virtual and physical connections, and uses the MixHop method to combine voltage-related features from neighbors at different distances.

Images