Multi-chip model temperature distribution prediction method based on meta learning
Through a meta-learning method, using symbol distance function and convolutional neural network, a multi-chip model temperature distribution prediction model is constructed, which solves the problems of low accuracy and low efficiency of multi-chip model temperature distribution prediction in the existing technology, and realizes the Efficient and accurate prediction under small sample conditions.
Patent Information
- Application Number
- CN202411163843.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-12-13
AI Technical Summary
The existing multi-chip model temperature distribution prediction method has low accuracy and low efficiency. It requires multiple iterations to solve and the calculation is complex, which affects the accuracy and efficiency of the method.
Using a meta-learning-based method, a task space data set is constructed through numerical simulation, and a symbol distance function and a convolutional neural network are used to establish a temperature distribution prediction model under the encoding-decoding structure, and train it on the model-independent meta-learning framework to achieve Accurate and efficient prediction of temperature distribution of multi-chip models.
The accurate temperature distribution prediction of the new task is achieved based on a small number of samples, which reduces the calculation cost and improves the training efficiency, and improves the accuracy and efficiency of the temperature distribution prediction of multi-chip models.
Smart Images
Figure CN119129393A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of large-scale integrated circuits, and in particular to a method for predicting temperature distribution of a multi-chip model based on meta-learning. Background Art
[0002] Multi-Chip Module (MCM) refers to a chip packaging model that packages multiple integrated circuit chips horizontally or vertically in a common module;
[0003] Multi-chip model temperature distribution prediction refers to the process of using a large amount of data related to the multi-chip model (such as power consumption data, ambient temperature data, heat dissipation condition data, etc.) through data analysis, machine learning or deep learning and other technical means to predict the temperature distribution of the multi-chip model under different working conditions.
[0004] Temperature distribution prediction of multi-chip models is the basis for system stability and reliability assessment, and is also a prerequisite for optimizing design and improving performance.
[0005] At present, for the prediction of multi-chip temperature distribution, as described in the paper “Thermal placement optimization of multichip modules using a sequential metamodeling-based optimization approach” (HCCheng, YHTsai, KNChen, and J.Fang, “Thermal placement optimization of multichip modules using a sequential metamodeling-based optimization approach,” Appl. Therm. Eng., vol. 30, no. 17–18, pp. 2632–2642, 2010, DOI: 10.1016 / j.applthermaleng.2010.07.004.), a multi-chip component thermal layout optimization method based on sequential metamodeling is proposed, including: constructing a chip junction temperature performance evaluation system for thermal optimization design of multi-chip model package; performing concurrent subspace optimization by using movement restrictions and response surface method (RSM) to achieve weighted transformation of multi-objective problems into single-objective optimization analysis; and using a dynamic experimental design strategy based on movement restrictions and different sampling techniques to improve the speed of experimental design (DOE) and optimization solution, and obtain the optimal chip layout design.
[0006] However, the above method requires multiple iterative solutions for the thermal layout optimization design of the multi-chip model, which increases the computational complexity and has a significant impact on the final solution accuracy, thereby affecting the accuracy of the method.
[0007] For example, regarding the prediction of multi-chip temperature distribution, as described in the paper "Data-Driven Modeling of Geometry-Adaptive Steady Heat Transfer based on Convolutional Neural Networks: Heat Conduction" (JZPeng, X.Liu, ZDXia, N.Aubry, Z.Chen, and WTWu, "Data-drivenmodeling of geometry-adaptive steady heat convection based on convolutionalneural networks," Fluids, vol.6, no.12, 2021, doi:10.3390 / fluids6120436.), it discloses a data-driven model for predicting the steady-state heat conduction of thermal objects of arbitrary geometric shapes in two-dimensional space, including: constructing a heat conduction convolutional neural network; by introducing the signed distance function (SDF), the physical and mathematical information characteristics of the geometric object are output; combined with the results of numerical simulation of the temperature field, the neural network is trained and the network structure is optimized; the conditional function is used to optimize the model performance to achieve efficient prediction of the two-dimensional steady-state temperature field.
[0008] However, the above method performs thermal prediction by constructing a network data-driven model, which not only takes a lot of time to prepare labeled training data sets, but also takes up a lot of computing resources to train the model, which affects the efficiency and economy of the method to a certain extent.
[0009] In summary, the problems existing in the prior art are that the multi-chip model temperature distribution prediction method is not accurate enough and has a relatively low efficiency. Summary of the invention
[0010] The purpose of the present invention is to provide a multi-chip model temperature distribution prediction method based on meta-learning with high accuracy and efficiency.
[0011] The technical solution to achieve the purpose of the present invention is:
[0012] A method for predicting temperature distribution of a multi-chip model based on meta-learning comprises the following steps:
[0013] (10) Construction of task space data set: The temperature distribution training data set of the multi-chip model is obtained through numerical simulation, and the input matrix is constructed using the signed distance function. After standardization, the chip power data set and chip quantity data set, which are composed of the unitless numerical input matrix and the unitless numerical output matrix and classified, are obtained to form the task space data set;
[0014] (20) Acquiring a temperature distribution prediction model: adopting a model-independent meta-learning method, using the task space dataset, training a convolutional neural network, and obtaining a temperature distribution prediction model based on meta-learning;
[0015] (30) Prediction of temperature distribution of multi-chip model: Based on the chip power and number of chips required to be predicted for the multi-chip model, a prediction input data set is constructed, and the temperature distribution prediction model based on meta-learning is combined to predict the temperature distribution of the multi-chip model.
[0016] Furthermore, the (10) task space data set construction step includes:
[0017] (11) Construction of multi-chip model temperature distribution data set: numerical simulation is used to solve the temperature distribution of the multi-chip model and generate a multi-chip model temperature distribution training data set;
[0018] (12) Input matrix construction: The center coordinates (x c ,y c ) is converted into a signed distance function matrix, and the chip area value is set to power P k The opposite of -P k , and obtain the unitless numerical input matrix t(x c ,y c );
[0019] (13) Output matrix construction: construct a temperature distribution matrix T(x, y) based on the multi-chip model temperature distribution training data set, and obtain a unitless numerical output matrix t(x, y) through standardization processing;
[0020] (14) Task space data set construction: The unitless numerical input matrix is combined with the unitless numerical output matrix, and the chips are reclassified according to chip power and chip quantity to obtain a chip power data set and a chip quantity data set, thereby obtaining a task space data set.
[0021] Furthermore, the (30) multi-chip model temperature distribution prediction step includes:
[0022] (31) Prediction input data set construction: obtaining the chip power and number of the required prediction multi-chip model, and constructing a prediction input data set consisting of a variable chip power input data set and a variable chip number input data set;
[0023] (32) Prediction of temperature distribution of multi-chip model: The predicted input data set is used as the input of the temperature distribution prediction model based on meta-learning, and its output is the required temperature distribution of the multi-chip model.
[0024] Compared with the prior art, the present invention has the following significant advantages:
[0025] 1. More accurate prediction: The present invention introduces the signed distance function (SDF) to express the geometric features of the multi-chip model, constructs a task space data set, establishes a convolutional neural network (CNN) model under the encoding-decoding structure, and trains it on the model-independent meta-learning (MAML) framework to achieve accurate temperature distribution prediction of the multi-chip model under the new task based on a small number of samples;
[0026] 2. More efficient prediction: The present invention introduces a model-independent meta-learning model, establishes a convolutional neural network model under a coding-decoding structure, and performs training on a model-independent meta-learning framework to obtain an initialization meta-parameter θ0 with good learning ability, and provides a small number of samples for updating based on the meta-parameter θ0. On the basis of ensuring the calculation accuracy, the computing cost is reduced and the training efficiency is improved, thereby realizing the effective temperature distribution prediction of the multi-chip model. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is the main flow chart of the multi-chip model temperature distribution prediction method based on meta-learning of the present invention.
[0028] Figure 2 yes Figure 1 Flowchart of the steps for constructing the task space dataset.
[0029] Figure 3 yes Figure 1 Flowchart of the steps for obtaining the temperature distribution prediction model based on meta-learning.
[0030] Figure 4 yes Figure 1 Flowchart of the temperature distribution prediction steps for the multi-chip model. DETAILED DESCRIPTION
[0031] like Figure 1 As shown, the multi-chip model temperature distribution prediction method based on meta-learning of the present invention includes the following steps:
[0032] (10) Construction of task space data set: The temperature distribution training data set of the multi-chip model is obtained through numerical simulation, and the input matrix is constructed using the signed distance function. After standardization, the chip power data set and chip quantity data set, which are composed of the unitless numerical input matrix and the unitless numerical output matrix and classified, are obtained to form the task space data set;
[0033] like Figure 2 As shown, the (10) task space data set construction acquisition includes:
[0034] (11) Construction of multi-chip model temperature distribution data set: numerical simulation is used to solve the temperature distribution of the multi-chip model and generate a multi-chip model temperature distribution training data set;
[0035] The temperature distribution data set of the multi-chip model (MCM) includes the center coordinates of the chip (x c ,y c ) and temperature distribution. Among them, the center coordinate of the chip (x c ,y c ) is used to construct the input matrix, and the temperature distribution of the chip is used to construct the output matrix;
[0036] (12) Input matrix construction: The center coordinates (x c ,y c ) is converted into a signed distance function matrix, and the chip area value is set to power P k The opposite of -P k , and obtain the unitless numerical input matrix t(x c ,y c );
[0037] (13) Output matrix construction: construct a temperature distribution matrix T(x, y) based on the multi-chip model temperature distribution training data set, and obtain a unitless numerical output matrix t(x, y) through standardization processing;
[0038] The (13) output matrix construction step includes:
[0039] (131) Temperature distribution matrix construction: constructing a temperature distribution matrix T(x, y) according to the input matrix for the temperature distribution of the multi-chip model;
[0040] (132) Output matrix acquisition: The temperature distribution matrix T(x, y) is subjected to normalization processing to obtain an output matrix t(x, y);
[0041] 1. Standardization processing:
[0042]
[0043] in represents the average temperature, and σ represents the standard deviation of each sample:
[0044]
[0045]
[0046] Where N X ,N Y Represent the resolution of the matrix in the x and y directions respectively. After standardization, the mean of all numerical matrices is 0 and the variance is 1;
[0047] (14) Task space data set construction: The unitless numerical input matrix is combined with the unitless numerical output matrix, and the chips are reclassified according to chip power and chip quantity to obtain a chip power data set and a chip quantity data set, thereby obtaining a task space data set.
[0048] The (14) task space data set construction step includes:
[0049] (141) Matrix combination: combining the unitless numerical input matrix (x, y) and the unitless numerical output matrix t(x, y) according to the center of the chip model;
[0050] (142) Data classification: classify the combined matrix according to chip power and chip quantity;
[0051] (143) Acquisition of task space data set: The classified data is subjected to a small sample classification experiment method to obtain a chip power data set. By limiting the variation range of the number of chips, a chip quantity data set is obtained, thereby obtaining a task space data set with chip power and chip quantity as prediction tasks.
[0052] The (143) task space data set construction step includes:
[0053] (1431) Chip power data set acquisition: A small sample classification experimental method is used to obtain the chip power configuration of the classified data, and data enhancement is used to expand the task set to obtain the chip power data set.
[0054] 1. Power configuration acquisition: A small sample classification experiment method is used to specify the number of chips and power distribution range in the chip power data set, and the power configuration is obtained by random selection;
[0055] Chip quantity regulation: The chip quantity under the specified chip power space is 2;
[0056] Chip power range regulations: power distribution range is {0.1, 0.2, 0.3, 0.4, 0.5}, unit is W;
[0057] Power configuration acquisition: Randomly select two power configurations from the power range (allowing two chips to have the same power) as the prediction task in the power space to obtain the power configuration;
[0058] 2. Construction of chip power data set: With the symmetry of the multi-chip model, data enhancement is used to expand the data set under each task pool and obtain the chip power data set;
[0059] Task sample acquisition: After the power configuration is acquired, each task sample is 400;
[0060] Dataset expansion: The data was expanded by using symmetry and rotation methods to obtain 2,800 task samples;
[0061] (1432) Chip quantity data set acquisition: The chip quantity data set is obtained by randomly setting the chip power, limiting the chip quantity variation range, and specifying the chip distribution law;
[0062] 1. Chip power is randomly set: the power is randomly set to 0.3W or 0.4W;
[0063] 2. Limitation of chip quantity variation range: According to the size restrictions of the substrate and the chip, the chip quantity variation range is limited to {1,2,3,4,5};
[0064] 3. Chip distribution rules: In each task, the chips follow the same distribution rule. In the chip power space, the tasks in the meta-training and meta-testing phases satisfy the same distribution.
[0065] 4. Construction of chip quantity data set: The chip quantity data set is obtained by designing chip power, chip quantity variation range, and chip distribution law;
[0066] (1433) Task space data set acquisition: The task space data set consists of a chip power data set and a chip quantity data set;
[0067] The prediction task of the present invention is divided into a chip power prediction task and a chip quantity prediction task according to chip power and chip quantity;
[0068] (20) Acquiring a temperature distribution prediction model: adopting a model-independent meta-learning method, using the task space dataset, training a convolutional neural network, and obtaining a temperature distribution prediction model based on meta-learning;
[0069] like Figure 3 As shown, the (20) temperature distribution prediction model acquisition includes:
[0070] (21) Convolutional neural network construction: Use the convolution layer to construct the encoding layer, and use transposed convolution to construct the decoding layer to obtain the convolutional neural network;
[0071] (211) Coding layer construction: Using the convolutional layer, the dimension of the input data is reduced, the signed distance function (SDF) matrix and the latent space features between adjacent points of the convolutional layer are extracted, and the coding layer is constructed;
[0072] (212) Decoding layer construction: Use transposed convolution to construct the decoding layer;
[0073] (213) Convolutional neural network construction: A convolutional neural network is obtained by constructing encoding layers and decoding layers.
[0074] (22) Acquisition of a temperature distribution model based on meta-learning: The chip power task space dataset and the chip quantity task space dataset obtained by dividing the task space dataset are used, and model-independent meta-learning is adopted to train the convolutional neural network to obtain a temperature distribution prediction model based on meta-learning.
[0075] (221) Data set division: dividing the task space data set according to the chip power task space data set and the chip quantity task space data set to achieve task space data set division;
[0076] (2211) Chip power task space data set partitioning: randomly select 12 chip power task space data sets, divide 9 of them into meta-training training sets and 3 into meta-training validation sets, and realize chip power task space data set partitioning;
[0077] (2212) Chip number task space data set partitioning: The three tasks with chip numbers of 2, 3, and 4 are used as convolutional neural network training data sets to achieve chip number task space data set partitioning;
[0078] Each task contains 10,000 sample instances.
[0079] (2213) Task space data set partitioning: partitioning the task space data set by chip power task space data set and chip quantity task space data set;
[0080] (222) Neural network training: using the divided data set to perform meta-training to implement neural network training.
[0081] (2221) Meta-training: Based on model-agnostic meta-learning (MAML), the meta-training process is realized through task-level training and meta-training;
[0082] (22211) Task-level training: When extracting tasks of size B from the task space, gradient update operations are performed on all tasks simultaneously;
[0083] 1. Gradient update method:
[0084]
[0085] where α is the learning rate used for internal task updates, For the task τ bThe weight of the network model after the last i-time gradient update, b is the index of the task in the task batch, and B represents the size of the task batch extracted from the task pool.
[0086] 2. Loss Function
[0087]
[0088] Where K represents the task τ b The number of samples used for internal update, t represents the standardized temperature field obtained by numerical simulation, represents the temperature field predicted by the convolution model, W represents the weight of the convolutional neural network, λ is the regularization coefficient, and λ||W||2 is the L2 regularization term.
[0089] (22212) Meta-level training: For B task batches, set a meta-objective and minimize the meta-objective to obtain an excellent initialization parameter θ0;
[0090] 1. Meta-goal representation:
[0091]
[0092] 2. The final meta-parameter θ0 is updated as follows:
[0093]
[0094] Where β is the meta-learning rate, Representative task τ b losses;
[0095] (22213) Meta-training: Meta-training is achieved through task-level training and meta-training;
[0096] (2222) Meta-verification: By calculating the relative error and chip junction temperature t j , implement meta-verification;
[0097] (22221) Relative error acquisition: For task τ b In test case i, the relative error of a point (x, y) in space is calculated;
[0098] 1. Relative error:
[0099]
[0100] (22222) Chip junction temperature t j Get: Using R 2 Measure the chip junction temperature prediction and obtain the chip junction temperature index;
[0101] 1. R 2 Calculation method:
[0102]
[0103] Where N t represents the number of chips in the test batch, k represents the chip index in the batch, Represents the average junction temperature of the chips in this batch, t j represents the maximum chip junction temperature obtained by numerical simulation, The representative model predicts the maximum junction temperature of the chip;
[0104] (22223) Meta-verification: Meta-verification is achieved by calculating the relative error and chip junction temperature indicators;
[0105] (223) Acquiring a temperature distribution prediction model based on meta-learning: acquiring a temperature distribution prediction model based on meta-learning according to the neural network training method;
[0106] (30) Prediction of temperature distribution of multi-chip model: Based on the chip power and number of chips required to be predicted for the multi-chip model, a prediction input data set is constructed, and the temperature distribution prediction model based on meta-learning is combined to predict the temperature distribution of the multi-chip model.
[0107] like Figure 4 As shown, the (30) multi-chip model temperature distribution prediction step includes:
[0108] (31) Prediction input data set construction: obtaining the chip power and number of the required prediction multi-chip model, and constructing a prediction input data set consisting of a variable chip power input data set and a variable chip number input data set;
[0109] The (31) prediction input data set construction steps are specifically as follows:
[0110] According to the chip power, number and chip center coordinates (x c ,y c ), the input matrix is constructed using the signed distance function, and the variable chip power input data set and the variable chip quantity input data set classified by the unitless numerical input matrix are obtained through standardization to form the prediction input data set.
[0111] (32) Prediction of temperature distribution of multi-chip model: The predicted input data set is used as the input of the temperature distribution prediction model based on meta-learning, and its output is the required temperature distribution of the multi-chip model.
[0112] The temperature distribution prediction of multi-chip model includes the temperature distribution prediction of variable chip power task space and variable chip number task space.
[0113] The (32) multi-chip model temperature distribution prediction step comprises:
[0114] (321) Prediction of spatial temperature distribution of variable chip power task: using the variable chip power input data set as the input of the meta-learning-based temperature distribution prediction model, and its output data set is the temperature distribution of each node when the multi-chip model changes chip power;
[0115] (322) Prediction of temperature distribution in task space with variable chip quantity: The variable chip quantity input data set is used as the input of the temperature distribution prediction model based on meta-learning, and its output data set is the temperature distribution of each node when the chip power of the multi-chip model is changed.
[0116] The present invention combines convolutional neural networks with model-independent meta-learning (MAML) to construct a multi-chip temperature distribution prediction model based on meta-learning, introduces a signed distance function (SDF) to express the geometric features of the multi-chip model, constructs a task space data set, establishes a convolutional neural network (CNN) model under a coding-decoding structure, and trains it in model-independent meta-learning (MAML), thereby obtaining an initialization meta-parameter θ0 with good learning ability, and provides a small number of samples for updating based on the meta-parameter θ0, so as to solve the temperature field prediction problem of the multi-chip model under cross-task conditions under the condition of few samples, including a variable chip power space and a variable chip quantity space.
Claims
1. A multi-chip model temperature distribution prediction method based on meta-learning, characterized in that: The steps include: (10) Construction of task space data set: The temperature distribution training data set of the multi-chip model is obtained through numerical simulation, and the input matrix is constructed using the signed distance function. After standardization, the chip power data set and chip quantity data set, which are composed of the unitless numerical input matrix and the unitless numerical output matrix and classified, are obtained to form the task space data set; (20) Acquiring a temperature distribution prediction model: adopting a model-independent meta-learning method, using the task space dataset, training a convolutional neural network, and obtaining a temperature distribution prediction model based on meta-learning; (30) Prediction of temperature distribution of multi-chip model: Based on the chip power and number of chips required to be predicted for the multi-chip model, a prediction input data set is constructed, and the temperature distribution prediction model based on meta-learning is combined to predict the temperature distribution of the multi-chip model.
2. The multi-chip model temperature distribution prediction method according to claim 1, characterized in that: The (10) task space data set construction step includes: (11) Construction of multi-chip model temperature distribution data set: numerical simulation is used to solve the temperature distribution of the multi-chip model and generate a multi-chip model temperature distribution training data set; (12) Input matrix construction: The center coordinates (x c ,y c ) is converted into a signed distance function matrix, and the chip area value is set to power P k The opposite of -P k , and obtain the unitless numerical input matrix t(x c ,y c ); (13) Output matrix construction: construct a temperature distribution matrix T(x, y) based on the multi-chip model temperature distribution training data set, and obtain a unitless numerical output matrix t(x, y) through standardization processing; (14) Task space data set construction: The unitless numerical input matrix is combined with the unitless numerical output matrix, and the chips are reclassified according to chip power and chip quantity to obtain a chip power data set and a chip quantity data set, thereby obtaining a task space data set.
3. The multi-chip model temperature distribution prediction method according to claim 2, characterized in that: The (14) task space data set construction step includes: (141) Matrix combination: Input the unitless value into the matrix t(x c ,y c ) and the unitless numerical output matrix t(x,y) are combined according to the center of the chip model; (142) Data classification: classify the combined matrix according to chip power and chip quantity; (143) Acquisition of task space data set: The classified data is subjected to a small sample classification experiment method to obtain a chip power data set. By limiting the variation range of the number of chips, a chip quantity data set is obtained, thereby obtaining a task space data set with chip power and chip quantity as prediction tasks.
4. The multi-chip model temperature distribution prediction method according to claim 3, characterized in that: The (20) temperature distribution prediction model acquisition includes: (21) Convolutional neural network construction: Use the convolution layer to construct the encoding layer, and use transposed convolution to construct the decoding layer to obtain the convolutional neural network; (22) Acquisition of a temperature distribution model based on meta-learning: The chip power task space dataset and the chip quantity task space dataset obtained by dividing the task space dataset are used, and model-independent meta-learning is adopted to train the convolutional neural network to obtain a temperature distribution prediction model based on meta-learning.
5. The multi-chip model temperature distribution prediction method according to claim 4, characterized in that: The (30) multi-chip model temperature distribution prediction step comprises: (31) Prediction input data set construction: obtaining the chip power and number of the required prediction multi-chip model, and constructing a prediction input data set consisting of a variable chip power input data set and a variable chip number input data set; (32) Prediction of temperature distribution of multi-chip model: The predicted input data set is used as the input of the temperature distribution prediction model based on meta-learning, and its output is the required temperature distribution of the multi-chip model.
6. The multi-chip model temperature distribution prediction method according to claim 5, characterized in that: The (31) prediction input data set construction steps are specifically as follows: According to the chip power, number and chip center coordinates (x c ,y c ), the input matrix is constructed using the signed distance function, and the variable chip power input data set and the variable chip quantity input data set classified by the unitless numerical input matrix are obtained through standardization to form the prediction input data set.
7. The multi-chip model temperature distribution prediction method according to claim 6, characterized in that: The (32) multi-chip model temperature distribution prediction step comprises: (321) Prediction of spatial temperature distribution of variable chip power task: using the variable chip power input data set as the input of the meta-learning-based temperature distribution prediction model, and its output data set is the temperature distribution of each node when the multi-chip model changes chip power; (322) Prediction of temperature distribution in task space with variable chip quantity: The variable chip quantity input data set is used as the input of the temperature distribution prediction model based on meta-learning, and its output data set is the temperature distribution of each node when the chip power of the multi-chip model is changed.