Neural network based modeling method and prediction method for high-speed serial channel transmitter
By constructing an equivalent circuit for the transmitter and utilizing a neural network model, the problems of low efficiency and insufficient accuracy of transmitter simulation models in existing technologies are solved, enabling fast and accurate signal integrity analysis of high-speed serial channels and improving the efficiency and accuracy of simulation tools.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2023-05-12
- Publication Date
- 2026-07-21
AI Technical Summary
Existing transmitter simulation models suffer from inefficiency and insufficient accuracy in high-speed channels, especially when considering the impact of output load. The application scope of existing machine learning models is limited, and they cannot meet the signal integrity analysis requirements of high-speed channels.
A neural network-based approach is adopted to construct the equivalent circuit of the transmitter and model the transmitter using a neural network model. A dataset is constructed using the input signal, load capacitance value, matching resistor value and pull-up voltage, and the neural network is trained to predict the output signal, thereby achieving fast and accurate signal integrity analysis.
It improves simulation speed and model accuracy, enables rapid prediction of output waveforms under different load conditions, enhances simulation and design efficiency, and is suitable for complete high-speed serial channel simulation analysis.
Smart Images
Figure CN116757135B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of high-speed serial channel simulation analysis, and specifically to a high-speed serial channel transmitter modeling method based on neural networks. Background Technology
[0002] High-speed signals need to be transmitted between chips through a communication link consisting of a transmitter, a propagation medium, and a receiver. As integrated circuit process dimensions continue to shrink and operating frequencies continue to increase, parasitic effects and losses in the circuit become more severe. Crosstalk, reflections, and oscillations encountered during signal transmission in the channel become more pronounced, causing waveform and timing distortions in the received signal, i.e., signal integrity issues. To ensure that the receiver can sample the correct transmitted signal from the received waveform, we need to perform signal integrity analysis on the high-speed channel.
[0003] As a crucial component of high-speed channels, the transmitter, typically acting as a buffer, requires accurate modeling of its nonlinear characteristics for sound signal integrity analysis. Existing transmitter simulation models are primarily categorized into transistor-level models and behavioral-level models. Transistor-level models provide detailed circuit designs and semiconductor parameters for simulation tools, enabling precise simulation analysis; however, for complex circuits, simulation is time-consuming and inefficient. Behavioral-level models, on the other hand, do not provide specific circuit structures but obtain simulation results through table lookups, offering faster simulation speeds, but their accuracy is limited by the precision of the provided data tables. Analyzing the signal integrity of high-speed channels requires inputting large amounts of data for verification; therefore, how to model the transmitter to quickly and accurately obtain the output signal is a pressing issue.
[0004] In recent years, machine learning has been widely applied in various fields and has achieved good results. Some works have studied the use of machine learning to model the nonlinear characteristics of buffers. However, these works have not considered the influence of the output load on the nonlinear characteristics of the buffer, only considering the characteristics of the buffer itself, thus limiting the application scope of the model and making it unsuitable for simulation analysis of high-speed channel transmitters. To solve the above problems, a neural network-based modeling method for high-speed serial channel transmitters is proposed. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a modeling and prediction method for high-speed serial channel transmitters based on neural networks.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A method for modeling a high-speed serial channel transmitter based on a neural network, the method comprising the following steps:
[0008] 1) Construct the equivalent circuit of the transmitter: The transmitter is equivalent to the load, propagation medium and receiver connected to its output port. The equivalent circuit includes the transmitter, load capacitor, lossless transmission line and matching resistor connected in sequence.
[0009] 2) Constructing a dataset: Perform transient simulation on the equivalent circuit of the transmitter. The input signal is a square wave with an initial value of 0, a length of at least one cycle, and a certain rise / fall time. The output signal of the transmitter obtained from the simulation is used as a label. The input signal, load capacitance value, matching resistor value, and pull-up voltage together constitute the features. By changing the amplitude of the input signal, the rise / fall time, the load capacitance value, the matching resistor value, and the pull-up voltage, multiple simulations are performed to obtain multiple sets of data, which constitute the dataset. The dataset is divided into training set and test set according to the proportion.
[0010] 3) Training the neural network model: The transmitter is modeled using a neural network model. The network input consists of an input signal of length p {x(t-p+1),x(t-p+2),…,x(t-1),x(t)}, the load capacitance value, the matching resistor value, and the pull-up voltage. The network output is the output signal y(t) at time t. Zeros are padded in front of the input signal so that the network can predict the output signal sequentially starting from t=0. The error between the network's predicted output signal and the simulated output signal is used as the loss function. The network is then trained on the training set to obtain the transmitter model.
[0011] Further, in step 1), one end of the load capacitor is connected to the output port of the transmitting end, and the other end is grounded; one end of the lossless transmission line is connected to the output port of the transmitting end, and the other end is connected in series with the matching resistor; the other end of the matching resistor is connected to the pull-up voltage source; wherein the characteristic impedance of the lossless transmission line is equal to the characteristic impedance of the propagation medium in the high-speed serial channel; the magnitude of the matching resistor is equal to the characteristic impedance of the propagation medium; and the magnitude of the pull-up voltage source is the equivalent open-circuit voltage of the receiving end of the high-speed serial channel.
[0012] Furthermore, the transmitting end is a nonlinear device, and the output port of the transmitting end is also connected to a load capacitor, which is grounded, while the receiving end is a linear circuit.
[0013] Furthermore, in step 2), the transient simulation is implemented using the simulation tool HSPICE, and the simulation time is the same as the length of the input signal.
[0014] Furthermore, in step 2), the test set is used to test the model after modeling.
[0015] Furthermore, the neural network structure in step 3) includes an input layer, several hidden layers, and an output layer, wherein the number of neurons in the input layer is p+3, the number of neurons in the hidden layer is 18, and the number of neurons in the output layer is 1. The output of each hidden layer is activated by the tanh activation function.
[0016] Furthermore, in step 3), the length p of the input signal depends on the electrical characteristics of the transmitting device itself.
[0017] Furthermore, in step 3), the mean square error between the output signal predicted by the network and the output signal obtained from the simulation is used as the loss function. The formula for calculating the loss function is:
[0018]
[0019] Where m represents the length of the input signal, and y(i) represents the value of the i-th output signal obtained from the simulation. This represents the value of the i-th output signal of the neural network.
[0020] The present invention also provides a method for predicting the output signal of a high-speed serial channel, comprising: for a high-speed serial channel, obtaining the load capacitance value, matching resistor value, and pull-up voltage by performing equivalent circuit modeling on its transmitting end, and predicting the corresponding transmitting end output signal of the high-speed serial channel under a given input signal using the trained transmitting end model as described above.
[0021] The beneficial effects of this invention are as follows: This invention provides a modeling and prediction method for the transmitter end of a high-speed serial channel based on neural networks. The proposed equivalent circuit for the transmitter end can effectively describe the impact of the transmitter load. Simultaneously, the neural network is used to model the nonlinear characteristics of the transmitter end, thereby enabling the prediction of the output waveform based on the input signal and achieving signal integrity analysis of the transmitter end. Compared with the simulation tool HSPICE, the high-speed serial channel transmitter end model proposed in this invention has high accuracy and efficiency, and can quickly and accurately predict the output waveform corresponding to different input signals under different loads. As part of a high-speed serial channel, the neural network-based transmitter end model proposed in this invention can be combined with the subsequent propagation medium and receiver end to achieve simulation analysis of the complete high-speed serial channel, providing guidance for the design of high-speed serial channels. Compared with traditional simulation tools that are time-consuming, it can effectively improve the efficiency of simulation and design. Attached Figure Description
[0022] Figure 1 This is a flowchart of the serial channel transmitter modeling process of the present invention;
[0023] Figure 2 This is the equivalent circuit structure of the transmitting end proposed in this invention;
[0024] Figure 3 This is the neural network structure used in this invention;
[0025] Figure 4 This is a comparison chart of the prediction results of the model proposed in this invention and the simulation results of the simulation tool. Detailed Implementation
[0026] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0027] like Figure 1 As shown, the present invention provides a high-speed serial channel transmitter modeling method based on neural networks, and the specific implementation steps are as follows:
[0028] 1) Construct the equivalent circuit of the transmitting end
[0029] A high-speed serial channel consists of a transmitter, a propagation medium, and a receiver connected in series. The transmitter is a nonlinear device, and its output port is connected to a grounded load capacitor. The receiver is a linear circuit. Considering that the load, propagation medium, and receiver all affect the nonlinear behavior of the transmitter, modeling the transmitter alone is not suitable for signal integrity analysis of the high-speed serial channel. Therefore, the transmitter, its output port, the load, the propagation medium, and the receiver are equivalently modeled, and the entire equivalent circuit is modeled to reflect the behavior of the transmitter in the high-speed serial channel.
[0030] like Figure 2 As shown, the equivalent circuit consists of a transmitter, a load capacitor, a 0.1-inch lossless transmission line, and a pull-up matching resistor. One end of the load capacitor is connected to the transmitter output port, and the other end is grounded. One end of the lossless transmission line is connected to the transmitter output port, and the other end is connected in series with the matching resistor. The other end of the matching resistor is connected to a pull-up voltage source. The characteristic impedance of the lossless transmission line is equal to the characteristic impedance of the propagation medium in the high-speed serial channel. The magnitude of the matching resistor is also equal to the characteristic impedance of the propagation medium. The magnitude of the pull-up voltage source is the Thevenin equivalent open-circuit voltage of the receiver in the high-speed serial channel.
[0031] 2) Constructing the dataset
[0032] HSPICE transient simulation was performed on the equivalent circuit of the transmitting end. The input signal was a square wave with an initial value of 0, a length of at least one period, and a certain rise / fall time. The simulation time was the same as the length of the input signal. The simulated output signal of the transmitting end was used as a label. The input signal, along with the load capacitance value, the matching resistor value, and the pull-up voltage, constituted the features. By changing the amplitude and rise / fall time of the input signal, the load capacitance value, the characteristic impedance of the transmission line (i.e., the matching resistor value), and the pull-up voltage, multiple simulations were performed to obtain multiple sets of data, forming a dataset. In this example, the square wave period used for the input signal was 10ns, the duration of the input signal was 25ns, the simulation step size was 0.1ns, and when constructing the dataset, the amplitude range was 1.0-1.8V, the rise / fall time range was 0.1-0.5ns, the load capacitance value range was 10-60pF, the matching resistor value range was 50-70Ω, and the pull-up voltage range was 0.5-1.0V. The dataset was divided into training and test sets proportionally.
[0033] 3) Training the neural network model
[0034] The transmitting end is modeled using a neural network model. The network input consists of an input signal of length p {x(t-p+1),x(t-p+2),…,x(t-1),x(t)}, the load capacitance, the matching resistor, and the pull-up voltage. The length p of the input signal depends on the electrical characteristics of the transmitting device itself. The network output is the output signal y(t) at time t. Figure 3 As shown, the structure of the neural network includes an input layer, several hidden layers, and an output layer. The input layer has p+3 neurons, the hidden layers have 18 neurons, and the output layer has 1 neuron. The output of each hidden layer is activated by the tanh activation function.
[0035] In practical applications, to obtain accurate simulation results, the simulation step size is usually small, resulting in long input and output signal lengths. To reduce the computational load on the network without losing key information, uniform sampling of the input and output signals can be performed. In this example, the input and output signals are downsampled by one-fifth, and the input signal length p = 45. Leading the input signal with zeros allows the network to predict the output signal sequentially starting from t = 0. The mean square error between the network's predicted output signal and the simulated output signal is used as the loss function. The network is trained on the training set to obtain the transmitter model, and the model's performance is tested on the test set. The formula for calculating the loss function is:
[0036]
[0037] Where m represents the length of the input signal, and y(i) represents the value of the i-th output signal obtained from the simulation. This represents the value of the i-th output signal of the neural network.
[0038] The present invention provides a method for predicting the output signal of a high-speed serial channel, which includes: for a high-speed serial channel, obtaining the load capacitance value, matching resistor value, and pull-up voltage by performing equivalent circuit modeling on the transmitting end, and predicting the corresponding transmitting end output signal under a given input signal using the trained transmitting end model.
[0039] In this example, the model's predictions on the test set showed an average relative error of 0.67% and a maximum relative error of 7.04% compared to HSPICE simulation results, representing a 19.9-fold improvement in simulation speed. Figure 4 The comparison between the model's prediction results on a single test sample and the HSPICE simulation results shows that the proposed neural network-based serial channel transmitter model has high accuracy and efficiency, and can provide effective assistance for signal integrity analysis of serial channels.
[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for modeling a high-speed serial channel transmitter based on a neural network, characterized in that, The method includes the following steps: 1) Construct the equivalent circuit of the transmitter: The transmitter is equivalent to the load, propagation medium and receiver connected to its output port. The equivalent circuit includes the transmitter, load capacitor, lossless transmission line and matching resistor connected in sequence. 2) Constructing a dataset: Perform transient simulation on the equivalent circuit of the transmitter. The input signal is a square wave with an initial value of 0, a length of at least one cycle, and a certain rise / fall time. The output signal of the transmitter obtained from the simulation is used as a label. The input signal, load capacitance value, matching resistor value, and pull-up voltage together constitute the features. By changing the amplitude of the input signal, the rise / fall time, the load capacitance value, the matching resistor value, and the pull-up voltage, multiple simulations are performed to obtain multiple sets of data, which constitute the dataset. The dataset is divided into training set and test set according to the proportion. 3) Training the neural network model: The transmitter is modeled using a neural network model. The network input consists of an input signal of length p {x(t-p+1),x(t-p+2),…,x(t-1),x(t)}, the load capacitance value, the matching resistor value, and the pull-up voltage. The network output is the output signal y(t) at time t. Zeros are padded in front of the input signal so that the network can predict the output signal sequentially starting from t=0. The error between the network's predicted output signal and the simulated output signal is used as the loss function. The network is then trained on the training set to obtain the transmitter model.
2. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, In step 1), one end of the load capacitor is connected to the output port of the transmitting end, and the other end is grounded. One end of the lossless transmission line is connected to the output port of the transmitting end, and the other end is connected in series with the matching resistor. The other end of the matching resistor is connected to the pull-up voltage source. The characteristic impedance of the lossless transmission line is equal to the characteristic impedance of the propagation medium in the high-speed serial channel. The magnitude of the matching resistor is equal to the characteristic impedance of the propagation medium. The magnitude of the pull-up voltage source is the equivalent open-circuit voltage of the receiving end of the high-speed serial channel.
3. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, The transmitting end is a nonlinear device, and its output port is also connected to a load capacitor, which is grounded. The receiving end is a linear circuit.
4. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, In step 2), the transient simulation is implemented using the simulation tool HSPICE, and the simulation time is the same as the length of the input signal.
5. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, In step 2), the test set is used to test the model after modeling.
6. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, The neural network structure in step 3) includes an input layer, several hidden layers, and an output layer. The input layer has p+3 neurons, the hidden layer has 18 neurons, and the output layer has 1 neuron. The output of each hidden layer is activated by the tanh activation function.
7. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, In step 3), the length p of the input signal depends on the electrical characteristics of the transmitting device itself.
8. The high-speed serial channel transmitter modeling method based on neural networks according to claim 1, characterized in that, In step 3), the mean square error between the output signal predicted by the network and the output signal obtained from the simulation is used as the loss function. The formula for calculating the loss function is as follows: Where m represents the length of the input signal, and y(i) represents the value of the i-th output signal obtained from the simulation. This represents the value of the i-th output signal of the neural network.
9. A method for predicting output signals of a high-speed serial channel, characterized in that, include: For a high-speed serial channel, by performing equivalent circuit modeling on its transmitting end, the load capacitance value, matching resistor value, and pull-up voltage are obtained. The trained transmitting end model as described in any of claims 1-8 is used to predict the corresponding transmitting end output signal of the high-speed serial channel under a given input signal.