KR-20260063373-A - POWER DISTRIBUTION NETWORK OPTIMIZATION METHOD AND APPARATUS, INCREMENTAL IR-DROP ANALYSIS METHOD USING DEEP LEARNING

KR20260063373AKR 20260063373 AKR20260063373 AKR 20260063373AKR-20260063373-A

Abstract

A method and apparatus for optimizing a power distribution network of a semiconductor layout using deep learning are disclosed. The disclosed power distribution network optimization method comprises: a step of determining a candidate partition to remove a partial strap of the power distribution network among the sub-partitions of the semiconductor layout using a first graph for each partition of the semiconductor layout and a pre-trained first graph convolution network; and a step of determining a final partition to remove the partial strap among the candidate partitions using a maximum voltage drop value and a number of routing overflows for the candidate partitions, wherein the semiconductor layout is divided into the partitions and the partitions are divided into the sub-partitions.

Inventors

정영광
현대준
신영수
최소윤

Assignees

세종대학교산학협력단
한국과학기술원

Dates

Publication Date: 20260507
Application Date: 20241030

Claims (20)

A step of determining a candidate partition to remove a partial strap of a power distribution network among the sub-partitions of the semiconductor layout using a first graph for each partition of the semiconductor layout and a pre-trained first graph convolution network; and The method includes the step of determining the final partition to remove the partial strap among the candidate partitions by using the maximum voltage drop value and the number of routing overflows for the candidate partitions. The above semiconductor layout is divided into the above partitions, and the above partitions are divided into the above sub-partitions, Power distribution network optimization method.
In Article 1, The above first graph is A first node fully connected and corresponding to the above sub-partition, and an undirected edge connecting the first node. Power distribution network optimization method.
In Paragraph 2, The above first graph convolution network is An encoder that generates a first feature vector from a feature map for the above sub-partition; A first layout recognition graph network defined in a coordinate space, receiving a first graph including the first feature vector, and outputting a feature vector for the first graph; A second layout recognition graph network that receives a fourth graph defined in a latent space generated from the first graph and outputs a feature vector for the fourth graph; and A first artificial neural network that receives feature vectors for the first and fourth graphs and determines the candidate partition A power distribution network optimization method including
In Paragraph 3, The above first graph convolution network is A second artificial neural network that receives feature vectors for the first and fourth graphs and predicts the number of routing overflows. A power distribution network optimization method that further includes
In Paragraph 3, The above first feature vector is It is assigned to the first node mentioned above, and The weights assigned to the above undirected edges are Determined according to the spatial distance between the sub-partitions corresponding to the first node connected by the aforementioned undirected edge and the similarity between the first feature vector assigned to the first node connected by the aforementioned undirected edge Power distribution network optimization method.
In Paragraph 3, The above feature map is A cell density map showing the area occupied by standard cells and macro cells in the above sub-partition; A pin density map indicating the number of input and output pins in the above sub-partition; A routing demand map indicating the number of routings for signals in the above sub-partition; A power pad map indicating the presence or absence of a power pad in the above sub-partition; A PDN structure map showing the metal layer and via layer structure in the above sub-partition; A current map indicating the current consumption of a logic cell in the above sub-partition; and Voltage drop map representing the maximum voltage drop value in the above sub-partition A power distribution network optimization method comprising at least one of the following.
In Article 1, The step of determining the final partition above For each of the above candidate partitions, a step of calculating a first score corresponding to a ratio value of the number of routing overflows to the maximum voltage drop value; and Step of determining the sub-partition with the maximum first score as the final partition A power distribution network optimization method including
In Article 1, A step of generating a second graph for a candidate power distribution network of the final partition, wherein different partial straps that are candidates for removal in the final partition are displayed; A step of predicting a third graph including a node in the second graph where a voltage drop will occur, using the second graph and a pre-trained second graph convolution network; and A step of determining the partial strap to be finally removed among the removal candidates by using the change in the number of routing overflows for the candidate power distribution network corresponding to the third graph and the change in the maximum voltage drop value. A power distribution network optimization method that further includes
In Paragraph 8, The step of generating the second graph above Converting the above candidate power distribution network into a resistance network, and generating the above second graph from the above resistance network, The above resistance network is It includes resistors corresponding to partial straps and vias of the above-mentioned candidate power distribution network, voltage sources corresponding to power pads, and current sources corresponding to logic cells, and The second graph above is, It includes a second node of the resistance network, a third node corresponding to the voltage source and current source, and a directional edge connecting the third node. The above directional edge is Bidirectional edges connecting third nodes located in the same layer and unidirectional edges connecting third nodes located in different layers A power distribution network optimization method including
In Article 9, The feature value assigned to the above third node is It includes the voltage of the second node before removing the partial strap, a value indicating whether the power pad is connected, the current value of the second node connected to the current source, and the average value of the Euclidean distances from the second node to a plurality of power pads, The feature value assigned to the above-mentioned directional edge is Includes a resistance value between the second node and a value indicating whether the partial strap is a candidate for removal. Power distribution network optimization method.
In Article 10, The above second graph convolution network is An in-layer graph convolution network that generates a second feature vector for voltage change in the same layer from the feature values of the third node and bidirectional edge of the second graph corresponding to the same layer of the candidate power distribution network; An inter-layer graph convolution network that generates a third feature vector for voltage changes between layers of the candidate power distribution network from the second feature vector, the feature value of a unidirectional edge connecting an upper layer and a lower layer of the candidate power distribution network, and the feature value of a third node of the second graph connected to the unidirectional edge; and It includes a third artificial neural network that predicts whether there is a voltage drop at the third node from the second and third feature vectors, and The above-mentioned in-layer graph convolution network is If the above identical layer is not the top layer, the above third feature vector is received as input, and the above second feature vector is generated. Power distribution network optimization method.
In Article 1, The step of determining the partial strap to be removed among the above removal candidates is For each candidate power distribution network corresponding to the third graph above, a step of calculating a second score in which the change in the number of routing overflows with respect to the change in the maximum voltage drop value corresponds to a ratio value; and A step of determining the partial strap that is a candidate for removal in the candidate power distribution network corresponding to the third graph where the second score is maximum as the partial strap to be removed. A power distribution network optimization method including
A step of generating a first graph for a target power distribution network of a semiconductor layout; and A step of predicting a second graph including a node in the first graph where a voltage drop will occur, using the first graph and a pre-trained graph convolution network. A method for analyzing voltage drop in a power distribution network including
In Paragraph 13, The step of generating the first graph above The above target power distribution network is converted into a resistance network, and the first graph is generated from the resistance network, The above resistance network is It includes resistors corresponding to partial straps and vias of the target power distribution network, voltage sources corresponding to power pads, and current sources corresponding to logic cells, and The first graph above is, A first node of the resistance network, a second node corresponding to the voltage source and current source, and a directional edge connecting the second node. A method for analyzing voltage drop in a power distribution network including
In Paragraph 14, The above directional edge is Bidirectional edges connecting second nodes located in the same layer and unidirectional edges connecting second nodes located in different layers A method for analyzing voltage drop in a power distribution network including
In Paragraph 15, The feature value assigned to the above second node is It includes the voltage of the first node before removing the partial strap, a value indicating whether the power pad is connected, the current value of the first node connected to the current source, and the average value of the Euclidean distances from the first node to a plurality of power pads, The feature value assigned to the above-mentioned directional edge is A value including a resistance value between the first nodes and a value indicating whether the partial strap is a candidate for removal, Method for analyzing voltage drop in a power distribution network.
In Paragraph 16, The above target power distribution network is A power distribution network of a partition in which different partial straps that are candidates for removal are displayed in a preset partition of a semiconductor layout, and The part strap to be finally removed among the above removal candidates is Determined by using the change in the number of routing overflows and the change in the maximum voltage drop for the target power distribution network corresponding to the second graph above Method for analyzing voltage drop in a power distribution network.
In Paragraph 15, The above graph convolutional network is An in-layer graph convolution network that generates a first feature vector for voltage change in the same layer from the feature values of the second node and bidirectional edge of the first graph corresponding to the same layer of the target power distribution network; An inter-layer graph convolution network that generates a second feature vector for voltage changes between layers of the target power distribution network from the first feature vector, a unidirectional edge connecting an upper layer and a lower layer of the target power distribution network, and a feature value of a second node of the first graph connected to the unidirectional edge; and It includes an artificial neural network that predicts whether there is a voltage drop at the second node from the first and second feature vectors, and The above-mentioned in-layer graph convolution network is If the above identical layer is not the top layer, the above second feature vector is received as input to generate the above first feature vector. Method for analyzing voltage drop in a power distribution network.
In Paragraph 18, The above inter-layered graph convolution network is A GRU (Gated Recurrent Unit) that selectively reflects input data into the third feature vector Method for analyzing voltage drop in a power distribution network.
In Paragraph 13, The above voltage drop A voltage drop that exceeds preset constraints, Method for analyzing voltage drop in a power distribution network.

Description

Power distribution network optimization method and apparatus using deep learning, incremental voltage drop analysis method The present invention relates to a method and apparatus for optimizing a power distribution network using deep learning and a method for analyzing voltage drop, and more specifically, to a method and apparatus for optimizing a power distribution network of a semiconductor layout using deep learning and a method for analyzing voltage drop. During semiconductor layout design, the design of the power distribution network that supplies power to the logic circuits is carried out alongside the design of the logic circuits. Since logic circuits, or logic cells, are placed beneath the power distribution network, it is important that the network be designed to minimize signal routing congestion, as the network affects signal routing within the logic cells. However, if a power distribution network is designed considering only signal routing congestion, power supply problems to logic cells may occur due to voltage drop in the network; therefore, it is necessary to optimize the power distribution network to minimize voltage drop in the network while reducing signal routing congestion. Previously, power distribution networks were optimized primarily by using commercially available analysis tools to analyze the number of routing overflows, which indicate routing congestion, and voltage drops within the network. However, since these optimization methods consume significant time and cost, there is a demand for more efficient power distribution network optimization methods. Relevant prior art includes patent documents such as Korean Registered Patents No. 10-0593803 and No. 10-2441442, and Korean Published Patent No. 2022-0157383, as well as non-patent documents including "Hierarchical Layout-Aware Graph Convolutional Network for Unified Aesthetics Assessment, Dongyu She; Yu-Kun Lai; Gaoxiong Yi; Kun Xu, IEEE, 2021", "Integrated power distribution network synthesis for mixed macro blocks and standard cells, D. Hyun, W. Lee, J. Park, and Y. Shin, IEEE, 2023", "Machine-learning-based dynamic IR drop prediction for ECO, Y.-C. Fang et al., in Proc. Int. Conf. on Computer-Aided Design, Nov. 2018", and "Stable Power Plan Technique for SoC Implementation, Seo Young-ho, Kim Dong-wook, Journal of the Korea Institute of Information and Communication Engineering." There is "Volume 16, Issue 12". FIG. 1 is a diagram illustrating a power distribution network optimization method using deep learning according to an embodiment of the present invention. FIG. 2 is a diagram illustrating a semiconductor layout and a power distribution network according to an embodiment of the present invention. FIG. 3 is a diagram illustrating a first graph convolution network according to an embodiment of the present invention. FIGS. 4 to 6 are drawings for illustrating graphs of a candidate power distribution network according to an embodiment of the present invention. FIG. 7 is a diagram illustrating a second graph convolution network according to an embodiment of the present invention. The present invention is susceptible to various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the invention to specific embodiments, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. Similar reference numerals have been used for similar components in the description of each drawing. The present invention proposes a method for optimizing a power distribution network of a semiconductor layout using deep learning. One embodiment of the present invention performs optimization by determining a partial strap to be removed from a power distribution network, wherein the partial strap to be removed is determined to be one that can reduce routing congestion while satisfying voltage drop constraints even if removed. A power distribution network is generally composed of multiple layers and may consist of a metal layer and a via layer. A metal strap is placed in the metal layer, and vias connected to the metal strap are placed in the via layer that electrically connects the metal layer. At this time, a portion of the metal strap located between the connection point of the via and the metal strap is defined as a partial strap. In addition, since semiconductor layouts or power distribution networks can be represented as graphs, one embodiment of the present invention utilizes a graph convolution network as a deep learning model for optimizing the power distribution network. The graph convolution network outputs feature vectors of input tensors like a general convolution network, but outputs feature vectors for graphs rather than images. One embodiment of the present invention uses a first graph convolution network to determine which part