A worst-case noise analysis method for power distribution networks based on deep learning

Abstract

The present invention discloses a worst-case noise analysis method for a power distribution network based on deep learning, which involves how to quickly calculate the worst-case noise of each sub-area of ​​a power distribution network under a given excitation based on a convolutional neural network. The result of the dynamic noise simulation of the power distribution network is used as a label, and the input current information and the distance information from the circuit sub-area to the voltage source are extracted as features to establish a data set; a noise analysis network consisting of a distance feature processing module and a dynamic noise prediction module is constructed, and the distance feature processing module reduces the dimension of the input distance feature, and then splices it with the current feature and inputs it into the dynamic noise prediction module to obtain the worst-case noise of each sub-area under the excitation. This method can quickly and accurately predict the worst noise of the power distribution network under a given excitation, and effectively improves the efficiency of the dynamic noise analysis of the power distribution network.

Description

Technical Field

[0001] The present invention relates to the fields of design circuit verification and power integrity analysis, and in particular to a worst-case noise analysis method for a power distribution network based on deep learning. Background Art

[0002] When current flows through a wire with resistance, it will cause an ohmic voltage drop, generate power supply noise, and affect the stability of the power supply. With the development of ultra-large-scale integrated circuit technology, the integration of chips is getting higher and higher, and the impact of power supply noise on circuits is becoming more and more obvious. The performance of integrated circuits is highly dependent on the stability of the power supply. Excessive power supply noise is equivalent to reducing the power supply voltage, which will lead to a decrease in circuit noise tolerance and an increase in logic gate delay, making the chip unable to complete normal functions. In order to ensure the performance of the chip, it is necessary to perform noise analysis on the power distribution network, find out the areas with excessive noise in the circuit, and modify and optimize them. Therefore, noise analysis is an important part of power network design.

[0003] Traditional technology mainly uses the improved node method to analyze the noise of the power distribution network. According to Kirchhoff's law, the equations of each node are listed, and then the noise of the node is obtained by solving the linear equations. As the chip scale increases sharply, the complexity of the linear equations that need to be solved also increases greatly, so traditional simulation tools need to spend a lot of time to complete the noise analysis of the power network. Since the circuit needs to be repeatedly modified and verified during the design process, long simulation time will lead to inefficient circuit design. In order to improve the speed of noise analysis, machine learning methods have begun to be applied to noise analysis of power networks.

[0004] Non-patent document 1 (Z. Xie, et.al. "PowerNet: Transferable Dynamic IR Drop Estimation via Maximum Convolutional Neural Network." ASP-DAC, 2020) proposes a method for finding areas with excessive noise in a power grid using a convolutional neural network. The worst noise of the circuit under a given stimulus is obtained by inputting the power consumption and flipping information of the logic gate into the trained network, which improves the speed compared to traditional simulation software. However, the network can only predict the noise of one sub-area at a time, so it still takes a long time when facing large-scale circuits, and it cannot effectively improve the efficiency of noise analysis.

[0005] In order to achieve fast power grid noise analysis and improve the efficiency of circuit design, the applicant proposed a worst-case noise analysis method for power distribution networks based on deep learning. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention proposes a worst-case noise analysis method for a power distribution network based on deep learning, which can quickly calculate the worst-case noise of the power distribution network under a given excitation.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A worst-case noise analysis method for a power distribution network based on deep learning, the method comprising the following steps:

[0009] S1 For a given power distribution network, divide it into M×N sub-areas, input a random triangle wave current into each sub-area as excitation, take the current feature matrix of the entire power distribution network and the distance feature from the center point of each sub-area to each VCC voltage source in the power distribution network as the features under the excitation, use simulation software to obtain the worst noise of each sub-area as the noise label under the excitation; each time the triangle wave current excitation is changed, a set of features and labels are obtained, and by randomly assigning different input currents, a data set for power distribution network noise analysis is constructed;

[0010] S2 builds a noise analysis network based on deep learning, which includes two parts: distance feature processing module and dynamic noise prediction module:

[0011] S2.1 distance feature processing module: the matrix constructed by the distance feature is input into the distance processing network composed of several convolutional layers, several deconvolutional layers and several activation layers, the distance feature matrix is ​​reduced in dimension, and the processed distance feature is output;

[0012] S2.2 Dynamic noise prediction module: The feature matrix composed of the current feature and the processed distance feature is input into a noise prediction network composed of several convolutional layers, several deconvolutional layers and several activation layers to obtain the worst noise of each sub-area within the simulation time predicted by the noise analysis network;

[0013] S3 extracts the current characteristics and distance characteristics of the given power distribution network and input current information to be analyzed and inputs them into the noise analysis network to obtain the worst noise within the target time.

[0014] Furthermore, in S1, the power distribution network used in the process of establishing the data set has multiple layers of wiring, and the circuit includes decoupling capacitors, and multiple sets of data are generated by changing the input triangular wave current.

[0015] 3. According to the method for worst-case noise analysis of a power distribution network based on deep learning in claim 1, the method for establishing a power distribution network dataset in S1 specifically comprises:

[0016] S1.1 Circuit simulation: For a given power distribution network, divide it into M×N sub-areas, input a random triangle wave current into each sub-area as excitation, and use simulation software to obtain the worst noise of each sub-area within the simulation time T, that is, a noise matrix of size 1×M×N, as the noise label under this excitation;

[0017] S1.2 Feature extraction: The dynamic noise simulation time step is t, and the starting point current of each simulation time step is extracted as the current feature. Then the current feature matrix size of the entire power distribution network is T / t×M×N; the number of VCC voltage sources in the power network is k, and the distance from the center point of sub-area i to each VCC voltage source is expressed as a vector (d i1 , d i2 , ..., d ik ) and use it as the distance feature, so the distance feature matrix size of the entire power distribution network is k×M×N; the current feature and the distance feature are used together as the features under the triangular wave current excitation in step S1.1, that is, the feature matrix size is (T / t+k)×M×N.

[0018] Furthermore, in S1.1, RedHawk software is used to perform vectorless dynamic noise simulation on a given power distribution network, a .pwl file is used to describe input current information in different states, and current is assigned to each sub-area in the form of BPA.

[0019] Furthermore, in S1.2, the distance d from the center point of sub-region i to the VCC voltage source j is ij is the Euclidean distance, and its calculation process is:

[0020]

[0021] where x i and i are the coordinates of the center point of sub-region i, x j and j are the coordinates of VCC voltage source j respectively.

[0022] Furthermore, in S2.1, the distance processing network is composed of 8 convolutional layers, 2 deconvolutional layers and 8 activation layers, wherein the convolution step size of 2 convolutional layers is 2, which plays the role of downsampling, and the step size of the deconvolution layer is 2, which plays the role of upsampling. Except for the first convolutional layer and the last convolutional layer, there is no activation layer behind them. The other convolutional layers and deconvolution layers are followed by an activation layer using the ReLU function.

[0023] Furthermore, in S2.2, the noise prediction network consists of 11 convolutional layers, 3 deconvolutional layers and 13 activation layers, wherein the convolution step size of 3 convolutional layers is 2, which plays a role of downsampling, and the step size of the deconvolution layer is 2, which plays a role of upsampling. Except for the last convolutional layer, there is no activation layer behind it. The other convolutional layers and deconvolution layers are followed by an activation layer using the ReLU function.

[0024] Furthermore, S2 also includes S2.3: The training of the noise analysis network based on deep learning is an end-to-end process, and the distance processing network and the noise prediction network are trained as a whole, and the L1 norm of the difference between the predicted worst noise and the actual noise label is used as the loss function, that is, the loss function is:

[0025]

[0026] in, represents the worst noise of sub-region i in simulation time predicted by the deep learning-based noise analysis network, V i represents the worst noise of sub-area i analyzed by the simulation software during the simulation time.

[0027] The beneficial effects of the present invention are as follows: the present invention provides a worst-case noise analysis method for a power distribution network based on deep learning, which uses a convolutional neural network to mine the relationship between the dynamic noise of the power distribution network and the input current and voltage source position, so that the trained noise analysis network can use the above characteristics to predict the worst-case noise of each sub-region of the power network under a given excitation. By extracting the input current of each sub-region of the power network and the distance characteristics from the center point of the sub-region to the voltage source, the noise analysis network proposed by the present invention can quickly predict the worst noise with high accuracy, and can effectively improve the efficiency of dynamic noise analysis of the power network compared with existing simulation software. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a noise analysis flow chart of the present invention;

[0029] Figure 2 It is a noise analysis network structure diagram of the present invention;

[0030] Figure 3 It is a distribution diagram of the noise analysis results of the present invention. DETAILED DESCRIPTION

[0031] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 As shown in FIG. 1 , a worst-case noise analysis method for a power distribution network based on deep learning is shown in FIG. 1 , and the specific implementation steps are as follows:

[0033] S1 establishes a data set for the power distribution network, including circuit simulation and feature extraction:

[0034] S1.1 Circuit simulation: For a given power distribution network with multi-layer wiring and decoupling capacitors, it is divided into M×N sub-areas, each of which is the same size. A random current is input to each sub-area, where the input current is divided into 5 states, each of which is a current segment with a duration of 0.02ns. The current value of state 0 remains unchanged at 0, state 1 is composed of a triangle wave with a width of 0.02ns and a peak value of 16mA, state 2 is composed of two triangle waves with a width of 0.01ns and a peak value of 16mA, state 3 is composed of a triangle wave with a width of 0.01ns, a peak value of 16mA and a zero input current, and state 4 is composed of a triangle wave with a width of 0.02ns and a peak value of 8mA. The input current information of each state is described by a .pw1 file, and the input current state of each sub-area is specified in the GSC file, and the current is assigned to each sub-area in the form of BPA. The simulation software RedHawk is used to perform vectorless dynamic noise simulation on the power grid, and the worst noise of each sub-area within the simulation time T is obtained, that is, a noise matrix of size 1×M×N, which is used as the noise label under this stimulus. In this example, the values ​​of M and N are both 50, and the value of T is 0.2ns.

[0035] S1.2 Feature extraction: The dynamic noise simulation time step is t, and the starting point current of each simulation time step is extracted as the current feature. Then the current feature matrix of the entire power distribution network is The size of is T / t×M×N. The number of VCC voltage sources in the power grid is k, and the distance from the center point of sub-area i to each VCC voltage source can be expressed as a vector (d i1 , d i2 , ..., d ik ), where the distance from the center point of sub-region i to the VCC voltage source j is the Euclidean distance, and its calculation process is:

[0036]

[0037] Among them, x i and yi are the coordinates of the center point of sub-region i, respectively, j and j are the coordinates of VCC voltage source j respectively, and the distance from the center point of the sub-region to the VCC voltage source is taken as the distance feature, so the distance feature matrix of the entire power distribution network is The size is k×M×N; the current feature and the distance feature are connected as the feature under this excitation, that is, the feature matrix The size of is (T / t+k)×M×N. In this example, the value of t is 1ps and the value of k is 35.

[0038] For a given power distribution network, each time the input current of a sub-area is changed, a set of feature matrices and labels can be obtained. By randomly assigning different input currents to each sub-area of ​​the circuit, a data set for power grid noise analysis is constructed. A total of 1,000 sets of data are generated and randomly divided, with 700 sets as training sets, 100 sets as validation sets, and 200 sets as test sets.

[0039] S2 builds a noise analysis network based on deep learning, which includes two parts: distance feature processing module and dynamic noise prediction module.

[0040] S2.1 The distance feature processing module consists of a convolution layer, a deconvolution layer, and an activation layer. The distance feature matrix of size k×M×N Processed into a feature matrix of size 1×M×N like Figure 2 As shown, the structure of the distance processing network is:

[0041] ① The distance feature matrix The convolution layer with a 3×3 input kernel, a stride of 1, a padding of 1, 35 input channels, and 8 output channels obtains a feature matrix of size 8×M×N.

[0042] ②The feature matrix The input convolution kernel is 3×3, the stride is 2, the padding is 1, the number of input channels is 8, the number of output channels is 8, and the convolution layer and the subsequent ReLU activation layer are used. Then the output feature matrix is ​​input into the convolution kernel 3×3, the stride is 1, the padding is 1, the number of input channels is 8, the number of output channels is 8, and the convolution layer and the subsequent ReLU activation layer are used to obtain a convolution layer with a size of The characteristic matrix

[0043] ③The feature matrix The input convolution kernel is 3×3, the stride is 2, the padding is 1, the number of input channels is 8, the number of output channels is 8, and the convolution layer and the subsequent ReLU activation layer are used. Then the output feature matrix is ​​input into the convolution kernel 3×3, the stride is 1, the padding is 1, the number of input channels is 8, the number of output channels is 8, and the convolution layer and the subsequent ReLU activation layer are used to obtain a convolution layer with a size of The characteristic matrix

[0044] ④The feature matrix The deconvolution layer with a 3×3 input kernel, a stride of 2, a padding of 1, 8 input channels, and 8 output channels, followed by a ReLU activation layer, yields a network of size The characteristic matrix

[0045] ⑤The feature matrix and Connected to get the size The feature matrix is ​​input into a convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 16, and an output channel number of 8, and a subsequent ReLU activation layer. The output feature matrix is ​​then input into a deconvolution layer with a convolution kernel of 4×4, a step size of 2, a padding of 1, an input channel number of 8, and an output channel number of 8, and a subsequent ReLU activation layer to obtain a feature matrix of size 8×M×N.

[0046] ⑥The feature matrix and The feature matrix of size 16×M×N is obtained by connecting them together, and the convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 16, and an output channel number of 8 is input into it, and the subsequent ReLU activation layer is input into the convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 8, and an output channel number of 1 to obtain the feature matrix

[0047] S2.2 Dynamic noise prediction module consists of convolution layer, deconvolution layer and activation layer. Distance features after processing Connect them together to get a feature matrix of size (T / t+1)×M×N Input the noise prediction network to obtain the worst noise within the predicted simulation time. Figure 2 As shown, the structure of the noise prediction network is:

[0048] ① The feature matrix The convolution layer with a 3×3 input kernel, a stride of 1, a padding of 1, a 201 input channel number, and a 16 output channel number is followed by a ReLU activation layer to obtain a feature matrix of size 16×M×N.

[0049] ②The feature matrix The input convolution kernel is 3×3, the step size is 2, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used. Then the output feature matrix is ​​input into the convolution kernel 3×3, the step size is 1, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used to obtain a convolution layer with a size of The characteristic matrix

[0050] ③The feature matrix The input convolution kernel is 3×3, the step size is 2, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used. Then the output feature matrix is ​​input into the convolution kernel 3×3, the step size is 1, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used to obtain a convolution layer with a size of The characteristic matrix

[0051] ④The feature matrix The input convolution kernel is 3×3, the step size is 2, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used. Then the output feature matrix is ​​input into the convolution kernel 3×3, the step size is 1, the padding is 1, the number of input channels is 16, the number of output channels is 16, and the convolution layer and the subsequent ReLU activation layer are used to obtain a convolution layer with a size of The characteristic matrix

[0052] ⑤The feature matrix The deconvolution layer with a 3×3 input kernel, a stride of 2, a padding of 1, 16 input channels, and 16 output channels and a subsequent ReLU activation layer is obtained. The characteristic matrix

[0053] ⑥The feature matrix and The size of the positive connection is The feature matrix is ​​input into a convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 32, and an output channel number of 16, and a subsequent ReLU activation layer. Then the output feature matrix is ​​input into a deconvolution layer with a convolution kernel of 3×3, a step size of 2, a padding of 1, an input channel number of 16, and an output channel number of 16, and a subsequent ReLU activation layer, and the size is obtained. The characteristic matrix

[0054] ⑦The feature matrix and Connected to get the size The feature matrix is ​​input into a convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 32, and an output channel number of 16, and a subsequent ReLU activation layer. The output feature matrix is ​​then input into a deconvolution layer with a convolution kernel of 4×4, a step size of 2, a padding of 1, an input channel number of 16, and an output channel number of 16, and a subsequent ReLU activation layer to obtain a feature matrix of size 16×M×N.

[0055] ⑧The feature matrix and The feature matrix of size 32×M×N is connected, and the convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 32, and an output channel number of 16 is input, and the ReLU activation layer is followed. Then the output feature matrix is ​​input into the convolution layer with a convolution kernel of 3×3, a step size of 1, a padding of 1, an input channel number of 16, and an output channel number of 1 to obtain a prediction noise matrix of size 1×M×N.

[0056] S2.3 The training of the noise analysis network is an end-to-end process. The distance processing network and the noise prediction network are trained as a whole. The L1 norm of the difference between the predicted worst noise and the actual noise label is used as the loss function, that is, the loss function is:

[0057]

[0058] in, represents the worst noise of sub-region i predicted by the noise analysis network during simulation time, V i represents the worst noise of sub-area i analyzed by the simulation software during the simulation time.

[0059] S3 extracts the current characteristics and distance characteristics of the given power distribution network and input current information to be analyzed and inputs them into the noise analysis network to obtain the worst noise within the target time. Figure 3 This is a noise distribution diagram predicted by the method of the present invention, and the noise unit is mV.

[0060] The trained noise analysis network was tested on the test set. The network took 2.45 seconds to calculate the worst noise of 200 test samples, while the traditional simulation software took 13 seconds to simulate the dynamic noise of one sample, which is more than 1000 times faster. The noise prediction results were compared with the simulation software. The average relative error of the noise prediction was 1.14%. The VCC voltage source of the power grid was 1.0V, and the average error of the noise prediction accounted for 0.11% of the power supply voltage.

Claims

1. A worst-case noise analysis method for power distribution networks based on deep learning, characterized in that: The method comprises the following steps: S1 For a given power distribution network, divide it into M×N sub-areas, input a random triangle wave current into each sub-area as excitation, take the current feature matrix of the entire power distribution network and the distance feature from the center point of each sub-area to each VCC voltage source in the power distribution network as the features under the excitation, use simulation software to obtain the worst noise of each sub-area as the noise label under the excitation; each time the triangle wave current excitation is changed, a set of features and labels are obtained, and by randomly assigning different input currents, a data set for power distribution network noise analysis is constructed; S2 builds a noise analysis network based on deep learning, which includes two parts: distance feature processing module and dynamic noise prediction module: S2.1 distance feature processing module: the matrix constructed by the distance feature is input into the distance processing network composed of several convolutional layers, several deconvolutional layers and several activation layers, the distance feature matrix is ​​reduced in dimension, and the processed distance feature is output; S2.2 Dynamic noise prediction module: The feature matrix composed of the current feature and the processed distance feature is input into a noise prediction network composed of several convolutional layers, several deconvolutional layers and several activation layers to obtain the worst noise of each sub-area within the simulation time predicted by the noise analysis network; S3 extracts the current characteristics and distance characteristics of the given power distribution network and input current information to be analyzed and inputs them into the noise analysis network to obtain the worst noise within the target time.

2. According to claim 1, a method for worst-case noise analysis of a power distribution network based on deep learning is characterized in that: In S1, the power distribution network used in the process of establishing the data set has multiple layers of wiring, and the circuit contains decoupling capacitors. By changing the input triangular wave current, multiple sets of data are generated.

3. The method for worst-case noise analysis of a power distribution network based on deep learning according to claim 1, characterized in that: The method for establishing a power distribution network data set in S1 specifically includes: S1.1 Circuit simulation: For a given power distribution network, divide it into M×N sub-areas, input a random triangle wave current into each sub-area as excitation, and use simulation software to obtain the worst noise of each sub-area within the simulation time T, that is, a noise matrix of size 1×M×N, as the noise label under this excitation; S1.2 Feature extraction: The dynamic noise simulation time step is t, and the starting point current of each simulation time step is extracted as the current feature. Then the current feature matrix size of the entire power distribution network is T / t×M×N; the number of VCC voltage sources in the power network is k, and the distance from the center point of sub-area i to each VCC voltage source is expressed as a vector (d i1 ,d i2 ,…,d ik ) and use it as the distance feature, so the distance feature matrix size of the entire power distribution network is k×M×N; the current feature and the distance feature are used together as the features under the triangular wave current excitation in step S1.1, that is, the feature matrix size is (T / t+k)×M×N.

4. The method for worst-case noise analysis of a power distribution network based on deep learning according to claim 3, characterized in that: In S1.1, RedHawk software is used to perform vectorless dynamic noise simulation on a given power distribution network, a .pwl file is used to describe input current information in different states, and the current is assigned to each sub-area in the form of BPA.

5. The method for worst-case noise analysis of a power distribution network based on deep learning according to claim 3, characterized in that: In S1.2, the distance d from the center point of sub-region i to VCC voltage source j ij is the Euclidean distance, and its calculation process is: where x i and i are the coordinates of the center point of sub-region i, x j and j are the coordinates of VCC voltage source j respectively.

6. A method for analyzing worst-case noise in a power distribution network based on deep learning according to any one of claims 1 to 5, characterized in that: In S2.1, the distance processing network consists of 8 convolutional layers, 2 deconvolutional layers and 8 activation layers, wherein the convolution step size of 2 convolutional layers is 2, which plays the role of downsampling, and the step size of the deconvolution layer is 2, which plays the role of upsampling. Except for the first convolutional layer and the last convolutional layer, there is no activation layer behind them. The other convolutional layers and deconvolution layers are followed by an activation layer using the ReLU function.

7. A method for analyzing worst-case noise in a power distribution network based on deep learning according to any one of claims 1 to 5, characterized in that: In S2.2, the noise prediction network consists of 11 convolutional layers, 3 deconvolutional layers and 13 activation layers, wherein the convolution step size of 3 convolutional layers is 2, which plays a role of downsampling, and the step size of the deconvolution layer is 2, which plays a role of upsampling. Except for the last convolutional layer, there is no activation layer behind it. The other convolutional layers and deconvolution layers are all followed by an activation layer using the ReLU function.

8. A method for analyzing worst-case noise in a power distribution network based on deep learning according to any one of claims 1 to 5, characterized in that: The S2 also includes S2.3: The training of the noise analysis network based on deep learning is an end-to-end process, in which the distance processing network and the noise prediction network are trained as a whole, and the L1 norm of the difference between the predicted worst noise and the actual noise label is used as the loss function, that is, the loss function is: in, represents the worst noise of sub-region i in simulation time predicted by the deep learning-based noise analysis network, V i represents the worst noise of sub-area i analyzed by the simulation software during the simulation time.

Images