Chip performance modeling method and system for adaptive depth migration

Abstract

The invention provides a self-adaptive deep migration chip performance modeling method and system, and relates to the technical field of chips, and the method comprises the steps: obtaining a to-be-modeled chip performance parameter data set, and constructing a deep neural network model; carrying out migration weight adaptive adjustment based on the feature distribution difference of the chip samples in the source domain and the target domain, and constructing an adaptive weight matrix and migration layer weight dynamic fusion; model training is carried out by considering prediction accuracy and feature alignment degree; and predicting the performance of the to-be-tested chip by using the trained model. According to the method, the performance prediction precision of the cross-chip series is improved, and the modeling data demand is reduced.

Description

Technical Field

[0001] This invention relates to chip technology, and more particularly to a chip performance modeling method and system based on adaptive deep migration. Background Technology

[0002] With the rapid development of integrated circuit technology, performance evaluation and optimization in chip design and manufacturing processes have become increasingly important. Chip performance modeling refers to predicting chip performance parameters, such as power consumption, timing, and temperature, by establishing mathematical models. These models help designers evaluate design schemes before chip manufacturing, reducing design risks and trial-and-error costs. Traditional chip performance modeling methods mainly include physical modeling, statistical modeling, and machine learning methods. With the advancement of deep learning technology, deep neural network-based chip performance modeling methods have become a hot research topic due to their excellent nonlinear feature extraction and fitting capabilities.

[0003] Traditional modeling methods have limited generalization ability for complex chip architectures and process nodes, making it difficult to accurately capture high-dimensional nonlinear relationships, resulting in insufficient prediction accuracy for novel chips. Acquiring chip performance data is often costly and time-consuming, especially for novel chips, where it is difficult to obtain enough samples to build accurate prediction models. This data scarcity severely restricts the training effect of the models. Existing transfer learning methods, when applied to the chip field, fail to effectively handle the differences in feature distribution between the source and target domains, leading to negative transfer during knowledge transfer. This results in poor model adaptability in the target domain, affecting the accuracy of performance predictions. Summary of the Invention

[0004] The embodiments of the present invention provide a chip performance modeling method and system for adaptive depth migration, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a chip performance modeling method for adaptive depth migration, comprising: Obtain the performance parameter dataset of the chip to be modeled, and construct a deep neural network model based on the performance parameter dataset; Based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the transfer weights of the deep neural network model are adaptively adjusted, and the difference metric between the source domain chip samples and the target domain chip samples is calculated. An adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during the training process, to obtain a trained deep neural network model. The trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction result of the chip under test is obtained.

[0006] Based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the deep neural network model undergoes adaptive adjustment of transfer weights. The calculation of the difference metric between the source domain chip samples and the target domain chip samples includes: Calculate the Mahalanobis distance between the feature distribution of the source domain and the feature distribution of the target domain, generate a feature distribution difference matrix, and adaptively adjust the transfer weights of the deep neural network model based on the feature distribution difference matrix; Based on the deep neural network model with adaptively adjusted transfer weights, the Euclidean distance between the source domain chip samples and the target domain chip samples is calculated to obtain the difference metric.

[0007] An adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights, including: An adaptive weight matrix is ​​constructed based on the difference metric, the difference metric is mapped to each dimension of the adaptive weight matrix according to a preset weight allocation strategy, and the adaptive weight matrix is ​​standardized. The standardized adaptive weight matrix is ​​dynamically fused with the migration layer weights. The fusion ratio is determined by adjusting the influence factor of the difference metric to obtain the optimized migration layer weights.

[0008] The model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during training. The resulting trained deep neural network model includes: The deep neural network model is trained based on the optimized transfer layer weights, and the performance prediction results and transfer feature distribution during the training process are obtained. The performance prediction accuracy is calculated based on the performance prediction results, and the transfer feature alignment degree is calculated based on the transfer feature distribution. Using the performance prediction accuracy and the transfer feature alignment as training metrics, a trained deep neural network model is obtained when the performance prediction accuracy reaches a first preset threshold and the transfer feature alignment reaches a second preset threshold.

[0009] Obtaining performance prediction results and transfer feature distribution during the training process, calculating performance prediction accuracy based on the performance prediction results, and calculating transfer feature alignment based on the transfer feature distribution include: The intermediate output data during the training process of the deep neural network model is obtained, and the performance prediction results and transfer feature distribution are extracted from the intermediate output data. The performance prediction accuracy is calculated based on the matching relationship between the performance prediction results and the real labels, and the transfer feature alignment degree is calculated based on the similarity between the transfer feature distribution and the reference feature distribution.

[0010] The trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction results of the chip under test are as follows: The trained deep neural network model and the chip under test are obtained. Feature parameters are extracted from the chip under test and input into the deep neural network model. The input feature parameters are processed based on the deep neural network model to generate a performance feature representation of the chip under test. The performance feature representation is analyzed using the deep neural network model to obtain the performance prediction results of the chip under test.

[0011] A second aspect of the present invention provides a chip performance modeling system for adaptive depth migration, comprising: The first module is used to obtain the performance parameter dataset of the chip to be modeled, and to construct a deep neural network model based on the performance parameter dataset. The second module is used to adaptively adjust the transfer weights of the deep neural network model based on the feature distribution differences between the source domain chip samples and the target domain chip samples, calculate the difference metric between the source domain chip samples and the target domain chip samples, construct an adaptive weight matrix based on the difference metric, and dynamically fuse the adaptive weight matrix with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The third module is used to train the model based on the optimized transfer layer weights. During the training process, the performance prediction accuracy and the degree of transfer feature alignment are considered simultaneously to obtain the trained deep neural network model. The fourth module is used to predict the performance of the chip under test using the trained deep neural network model, and obtain the performance prediction result of the chip under test.

[0012] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: By adaptively adjusting the transfer weights of the deep neural network model based on the feature distribution differences between chip samples in the source and target domains, the problem of poor transfer learning performance caused by domain differences in traditional chip performance modeling methods is effectively solved.

[0015] The innovative introduction of a difference metric and adaptive weight matrix construction mechanism enables dynamic fusion of transfer layer weights, allowing the model to automatically adjust its transfer strategy based on the feature distribution of different chip domains, significantly improving the model's generalization ability and adaptability.

[0016] By simultaneously considering performance prediction accuracy and transfer feature alignment during model training, a dual optimization objective is formed, which ensures both prediction accuracy and consistency of cross-domain feature representation, thus solving the transfer bias problem caused by a single optimization objective in traditional methods.

[0017] This method can effectively improve the accuracy and reliability of performance prediction through knowledge transfer when the number of chip samples is limited, providing a more efficient and accurate technical means for the performance evaluation of new chips. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the chip performance modeling method for adaptive depth migration according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the chip performance modeling method for adaptive depth migration according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Obtain the performance parameter dataset of the chip to be modeled, and construct a deep neural network model based on the performance parameter dataset; Based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the transfer weights of the deep neural network model are adaptively adjusted, and the difference metric between the source domain chip samples and the target domain chip samples is calculated. An adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during the training process, to obtain a trained deep neural network model. The trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction result of the chip under test is obtained.

[0022] In one optional implementation, based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the deep neural network model undergoes adaptive adjustment of transfer weights, and the calculation of the difference metric between the source domain chip samples and the target domain chip samples includes: Calculate the Mahalanobis distance between the feature distribution of the source domain and the feature distribution of the target domain, generate a feature distribution difference matrix, and adaptively adjust the transfer weights of the deep neural network model based on the feature distribution difference matrix; Based on the deep neural network model with adaptively adjusted transfer weights, the Euclidean distance between the source domain chip samples and the target domain chip samples is calculated to obtain the difference metric.

[0023] We acquire source and target domain chip sample datasets. The source domain dataset contains labeled chip defect samples, which are typically numerous and fully labeled. The target domain dataset consists of chip samples for defect detection, which may have insufficient labels or exhibit distributional differences from the source domain. These datasets are then input into a pre-trained deep neural network model for feature extraction.

[0024] Features of chip samples in the source and target domains are extracted using a deep neural network model. This model can be a pre-trained convolutional neural network in the source domain, such as ResNet or VGG. For each chip sample image, the intermediate layer feature representation is obtained through the model's forward propagation. Specifically, the output of a specific convolutional or fully connected layer in the network is selected as the feature representation; these features describe important information such as the texture and shape of the chip sample. After extraction, the feature distributions of the source and target domains are obtained.

[0025] To calculate the Mahalanobis distance between the feature distributions of the source and target domains, assume the mean vector of the source domain feature distribution is μs and its covariance matrix is ​​Σs, and the mean vector of the target domain feature distribution is μt and its covariance matrix is ​​Σt. The Mahalanobis distance is calculated as follows: First, calculate the difference vector between the means of the two distributions Δμ = μs - μt. Then, calculate the average of the covariance matrices of the two distributions Σ = (Σs + Σt) / 2. Finally, calculate the Mahalanobis distance d = sqrt((Δμ)T · Σ-1· (Δμ)). Here, Σ-1 represents the inverse of matrix Σ, and sqrt represents the square root operation.

[0026] A feature distribution difference matrix is ​​generated by organizing the calculated Mahalanobis distances into a matrix form. Each element of this matrix represents the degree of difference between the i-th dimension in the source domain feature space and the j-th dimension in the target domain feature space. Specifically, for each dimension in the feature space, the Mahalanobis distance of that dimension between the source and target domains is calculated, and these distance values ​​are filled into the corresponding positions in the difference matrix.

[0027] The transfer weights of the deep neural network model are adaptively adjusted based on the feature distribution difference matrix. Dimensions with large differences in the feature distribution difference matrix are identified; these dimensions represent features with significant distributional differences between the source and target domains. Then, the weights of relevant layers in the deep neural network are adjusted accordingly to reduce the impact of these differences on the prediction results. Specifically, for feature dimensions with large differences, the weight contribution of the corresponding neurons is reduced; for feature dimensions with small differences, their weight contribution is maintained or appropriately increased.

[0028] The formula for calculating the adjusted weights uses an inverse relationship with the degree of difference, i.e., the weight adjustment coefficient α = 1 / (1 + λ·d), where d is the Mahalanobis distance of the corresponding dimension, and λ is an adjustable hyperparameter used to control the intensity of the adjustment. Multiplying the original weight W by the adjustment coefficient α yields the adjusted weight W' = W·α. For feature dimensions with particularly significant differences, a threshold can be set for truncation to prevent information loss due to excessively small weights.

[0029] Based on the deep neural network model with adaptively adjusted transfer weights, features are extracted again from the source and target domain chip samples to obtain adjusted feature representations. These feature representations reflect the new understanding of samples from different domains after the model weight adjustment.

[0030] The Euclidean distance between the adjusted source domain chip sample features and the target domain chip sample features is calculated to obtain the difference metric. Specifically, for each sample s in the source domain sample set S and each sample t in the target domain sample set T, the Euclidean distance between their feature vectors is calculated, i.e., d(s,t) = ||f(s) - f(t)||2, where f(·) represents the feature vector extracted by the adjusted deep neural network model, and ||·||2 represents the L2 norm.

[0031] The final difference metric, D = Σ w(s,t)·d(s,t) / N, is obtained by calculating the mean or weighted sum of the Euclidean distances between all sample pairs. Here, w(s,t) is the weighting coefficient, which can be set according to the importance of the samples, and N is the total number of sample pairs. This difference metric reflects the overall degree of difference between the source and target domains after adjustment of the transfer weights; a smaller value indicates a better model transfer effect.

[0032] The difference metric can be used to further evaluate the effectiveness of model transfer and determine whether additional model fine-tuning or domain adaptation is needed. If the difference metric is below a preset threshold, the model is considered to have adapted well to the target domain; otherwise, more samples from the target domain need to be collected or other domain adaptation techniques need to be used to further improve the model.

[0033] In practical applications, this method can be used for defect detection in the semiconductor chip manufacturing process. For example, a detection model trained on one type of chip can be transferred to a new type of chip. Through the aforementioned feature distribution difference analysis and adaptive weight adjustment, the model can accurately identify various defects on the new type of chip, thereby improving detection efficiency and accuracy.

[0034] In one optional implementation, an adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain optimized transfer layer weights, including: An adaptive weight matrix is ​​constructed based on the difference metric, the difference metric is mapped to each dimension of the adaptive weight matrix according to a preset weight allocation strategy, and the adaptive weight matrix is ​​standardized. The standardized adaptive weight matrix is ​​dynamically fused with the migration layer weights. The fusion ratio is determined by adjusting the influence factor of the difference metric to obtain the optimized migration layer weights.

[0035] Obtain a difference metric, which reflects the distributional difference between the source and target domain data. This metric can be calculated using methods such as Maximum Mean Difference (MMD), A-distance, or Wasserstein distance. Taking MMD as an example, it calculates the statistical distance between the source and target domain feature representations in a high-dimensional feature space, yielding a scalar value that measures the degree of difference between the two domains.

[0036] An adaptive weight matrix is ​​constructed based on the acquired difference metric values. Assume the shape of the transfer layer's weight tensor is (c_out, c_in, k, k), where c_out represents the number of output channels, c_in represents the number of input channels, and k represents the kernel size. First, an adaptive weight matrix with the same shape as the transfer layer's weights is initialized. Then, according to a preset weight allocation strategy, the difference metric values ​​are mapped to each dimension of this matrix.

[0037] Weighting strategies can be determined based on the degree of influence of different dimensions on transfer learning. For example, for convolutional layers, the following strategy can be adopted: assign higher weights to the output channel dimension, as they directly affect feature extraction capabilities; assign medium weights to the input channel dimension; and assign lower weights to the spatial dimension of the convolutional kernel. In practice, the difference metric can be multiplied by different coefficients and assigned to different dimensions respectively.

[0038] After assigning weights to the matrix, the adaptive weight matrix needs to be standardized to ensure that the weight values ​​are within a reasonable range. Standardization can be performed using max-min normalization or Z-score normalization. Taking max-min normalization as an example, the element values ​​in the adaptive weight matrix are mapped to the interval [0, 1]: For each element in the adaptive weight matrix, calculate its standardized value: (original value - minimum value) / (maximum value - minimum value). After this processing, the element values ​​of the adaptive weight matrix are all restricted to between 0 and 1, which facilitates subsequent fusion with the transfer layer weights.

[0039] After standardization, the standardized adaptive weight matrix is ​​dynamically fused with the transfer layer weights. The fusion process involves adjusting the influence factor of the difference metric to determine the fusion ratio. This influence factor can be preset based on the characteristics of the transfer learning task or optimized using a validation set.

[0040] The dynamic fusion formula can be expressed as: Optimized weight = Original weight × (1 - Fusion coefficient × Adaptive weight), where the fusion coefficient is a scalar value calculated based on the difference metric and the influence factor. When the difference between the source and target domains is large, the fusion coefficient increases, and the influence of the adaptive weights is enhanced; when the difference is small, the fusion coefficient decreases, retaining more information from the original weights.

[0041] In practical applications, an adaptive mechanism can be designed to dynamically adjust the fusion coefficient. For example, the fusion coefficient can be automatically adjusted based on the trend of loss changes during training. If the training loss of the target task decreases slowly, the fusion coefficient can be increased to strengthen the influence of the adaptive weights; conversely, the fusion coefficient can be decreased.

[0042] After dynamic fusion, optimized transfer layer weights are obtained. These optimized weights are better adapted to the data distribution characteristics of the target domain, thus improving the model's performance in the target domain.

[0043] In practical applications, such as transferring computer vision models pre-trained on large-scale natural images to medical image analysis tasks, the data distribution differences between the source domain (natural images) and the target domain (medical images) are significant. By constructing an adaptive weight matrix and dynamically fusing the data, the impact of inter-domain differences on model performance can be effectively reduced.

[0044] Another application scenario is the transfer of speech recognition models from one language to another. By calculating the differences in acoustic features between the two languages, an adaptive weight matrix is ​​constructed and dynamically fused with the transfer layer weights of the pre-trained model, which can adapt to the characteristics of the target language more quickly.

[0045] It is important to note that the construction and dynamic fusion of the adaptive weight matrix should be performed during the model fine-tuning phase, not the inference phase. This ensures that the model has stable weights after training, facilitating deployment and use.

[0046] Through the above steps, an adaptive weight matrix is ​​constructed based on the difference metric, and then dynamically fused with the transfer layer weights of the deep neural network model to obtain optimized transfer layer weights. This method can effectively improve the efficiency of transfer learning and enable the model to better adapt to the data distribution characteristics of the target domain.

[0047] In one optional implementation, the model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during the training process, resulting in a trained deep neural network model including: The deep neural network model is trained based on the optimized transfer layer weights, and the performance prediction results and transfer feature distribution during the training process are obtained. The performance prediction accuracy is calculated based on the performance prediction results, and the transfer feature alignment degree is calculated based on the transfer feature distribution. Using the performance prediction accuracy and the transfer feature alignment as training metrics, a trained deep neural network model is obtained when the performance prediction accuracy reaches a first preset threshold and the transfer feature alignment reaches a second preset threshold.

[0048] At the start of model training, optimized transfer layer weights are applied to the corresponding layers of the deep neural network model. These transfer layer weights, obtained through the aforementioned weight optimization process, have already been preliminarily adapted to the data characteristics of the target task. Training is performed using batch processing, with each batch containing sample data from both the source and target domains.

[0049] In each training iteration, the input data is fed into the deep neural network model, and the model's predicted output is obtained through forward propagation. For performance prediction tasks, the difference between the model output and the true label is used to calculate the loss function value. Simultaneously, the output of the transfer feature layer is extracted; these features represent the distribution of source and target domain data in the feature space.

[0050] Performance prediction accuracy is calculated using evaluation metrics appropriate to the specific task type. For classification tasks, accuracy, F1 score, or AUC can be used; for regression tasks, mean squared error (MSE) or mean absolute error (MAE) can be used. For example, in chip performance prediction tasks, the relative error between predicted and actual power consumption can be used as an evaluation metric. Specifically, a prediction is considered accurate when the relative error between the predicted and actual values ​​is less than 10%.

[0051] The degree of feature alignment during transfer is measured based on the similarity of feature distributions between the source and target domains. One approach is to quantify the difference in feature distributions between the two domains by calculating the maximum mean difference (MMD). The smaller the MMD value, the closer the feature distributions of the two domains are, and the better the transfer effect. Another approach is to use a domain discriminator to evaluate the domain independence of features; a domain discrimination accuracy close to 50% indicates that the features have achieved good domain alignment.

[0052] During training, the model's overall loss function consists of the prediction task loss and the feature alignment loss: Total loss = Prediction task loss + λ × Feature alignment loss; Here, λ is the weighting coefficient that balances the two losses. It can be adjusted according to the specific task requirements. A larger λ value will make the model pay more attention to domain alignment, while a smaller λ value will pay more attention to prediction accuracy.

[0053] After a certain number of training iterations, the validation set is evaluated to calculate the current model's performance, prediction accuracy, and feature alignment. Prediction accuracy is obtained by comparing the predictions on the validation set with the true labels. Feature alignment is evaluated by extracting feature representations from the source and target domain data in the validation set and then calculating their distribution differences.

[0054] To determine whether a model has reached the completion of training, two preset thresholds are set: the first preset threshold measures the performance prediction accuracy, and the second preset threshold measures the degree of transfer feature alignment. For example, the first preset threshold can be set to a prediction accuracy of 90%, and the second preset threshold can be set to an MMD value of 0.05. When the model's performance prediction accuracy on the validation set reaches or exceeds the first preset threshold, and the degree of transfer feature alignment reaches or exceeds the second preset threshold, the model is considered to have completed training.

[0055] It is worth noting that the setting of the two thresholds needs to be determined based on the difficulty and requirements of the specific task. For simpler tasks, a higher accuracy threshold can be set; while for transfer learning scenarios with large differences between domains, the feature alignment requirements need to be appropriately reduced.

[0056] In practical applications, to avoid overfitting, an early stopping strategy can be adopted. When the model's performance no longer improves within several consecutive evaluation periods, or when the performance on the validation set declines while the performance on the training set continues to improve, the training process can be terminated early, and the model with the best validation performance can be selected as the final model.

[0057] After training is completed, a final evaluation is performed on the test set to verify the model's generalization ability. The test set usually contains data that was not used during training and validation, which can more objectively reflect the model's performance in real-world applications.

[0058] In the example of chip performance prediction, through the training method described above, the final model can effectively transfer chip data knowledge from different process nodes in the source domain to the new process node in the target domain, achieving accurate prediction of the performance of the new process chip with a prediction error controlled within 7%. At the same time, the MMD values ​​of the feature distributions in the source and target domains are less than 0.03, indicating that the model has successfully achieved knowledge transfer and feature alignment.

[0059] By employing this training method that comprehensively considers both performance prediction accuracy and transfer feature alignment, the resulting deep neural network model can not only accurately predict the performance metrics of the target task, but also effectively utilize source domain knowledge to improve the model's generalization ability in situations where data is scarce.

[0060] In one optional implementation, obtaining the performance prediction results and transfer feature distribution during the training process, calculating the performance prediction accuracy based on the performance prediction results, and calculating the transfer feature alignment degree based on the transfer feature distribution include: The intermediate output data during the training process of the deep neural network model is obtained, and the performance prediction results and transfer feature distribution are extracted from the intermediate output data. The performance prediction accuracy is calculated based on the matching relationship between the performance prediction results and the real labels, and the transfer feature alignment degree is calculated based on the similarity between the transfer feature distribution and the reference feature distribution.

[0061] In each iteration of model training, intermediate output data is extracted from the deep neural network. This intermediate output data typically includes the activation values, gradient information, and prediction results of each layer of the neural network. Specifically, for a neural network with multiple hidden layers, the output of each layer is recorded during the forward propagation process, and these outputs constitute part of the intermediate output data.

[0062] From the acquired intermediate output data, two key types of information are extracted: performance prediction results and transfer feature distribution. Performance prediction results refer to the model's predicted output on the input data, such as the predicted probability of a class in a classification task or the predicted value in a regression task. Transfer feature distribution refers to the distribution of activation values ​​in a specific layer of the model (usually a hidden layer close to the output layer), which reflects the model's high-level representation ability of the input data.

[0063] The performance prediction results are extracted by obtaining the activation values ​​of the model's output layer. In classification tasks, this is typically the probability distribution processed by the softmax function; in regression tasks, it is simply the output layer values. These prediction results are then compared with the true labels in the dataset to calculate the performance prediction accuracy.

[0064] The method for calculating performance prediction accuracy depends on the task type. For classification tasks, metrics such as accuracy, precision, recall, or F1 score can be used; for regression tasks, metrics such as mean absolute error, mean squared error, or coefficient of determination can be used. During the calculation, the model's prediction results are compared with the true labels to obtain the matching relationship, and then the accuracy value is calculated according to the selected metric formula.

[0065] For example, in an image classification task, if a model correctly classifies 85 images out of a batch of 100 images, the accuracy is 85%. This accuracy value, as a quantitative representation of performance prediction accuracy, reflects the model's predictive ability at the current training stage.

[0066] The extraction of transfer feature distributions involves selecting the output of a specific layer (such as the penultimate fully connected layer) from the model as feature representations. These features are typically high-dimensional vectors that reflect the model's internal representation of the input data. To evaluate the quality of these features, they need to be compared with a reference feature distribution.

[0067] The reference feature distribution can be the feature distribution of the pre-trained model on the source domain data, or the ideal distribution that features should have. The degree of transfer feature alignment is obtained by calculating the similarity between the transfer feature distribution and the reference feature distribution.

[0068] Various distance metrics can be used to calculate the alignment of transfer features, such as KL divergence, JS divergence, maximum mean difference (MMD), or Wasserstein distance. Taking MMD as an example, it measures the distance between two distributions in the reproducing kernel Hilbert space. The smaller the MMD value, the more similar the two distributions are, and the higher the alignment of transfer features.

[0069] In the specific calculation, the feature distribution extracted by the current model from the target domain data and the reference feature distribution are first mapped to the kernel space, and then their distance in the kernel space is calculated. After normalization, the obtained distance value can be represented as an alignment degree between 0 and 1, where 1 represents perfect alignment and 0 represents complete misalignment.

[0070] In practical applications, a threshold can be set to determine whether the alignment of the transferred features meets the requirements. For example, if the alignment is greater than 0.8, the feature transfer is considered to be relatively successful; otherwise, the model structure or training strategy needs to be adjusted.

[0071] By continuously monitoring the performance prediction accuracy and transfer feature alignment during training, problems in model training can be identified in a timely manner. For example, if the performance prediction accuracy increases slowly or stagnates, it indicates that the learning rate setting is inappropriate or the model capacity is insufficient; if the transfer feature alignment is low, it indicates that there is a large difference between the source domain and the target domain, requiring the adoption of stronger domain adaptation techniques.

[0072] This monitoring mechanism can be integrated into a deep learning framework and implemented as a callback function in the training process. Every fixed number of iterations, it calculates the performance prediction accuracy and the degree of transfer feature alignment, records the results, and forms a quality curve of the training process.

[0073] Based on these quality curves, training parameters, such as learning rate and regularization strength, can be automatically adjusted to achieve adaptive optimization of the training process. At the same time, this data can also provide valuable feedback to model designers, helping them understand model behavior and improve model structure.

[0074] In summary, by obtaining the performance prediction results and transfer feature distribution during the training process, and calculating the performance prediction accuracy and transfer feature alignment, we can comprehensively evaluate the model training quality, provide data support for model optimization, and ultimately improve the performance and generalization ability of deep neural network models.

[0075] In one optional implementation, the trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction result of the chip under test is obtained by: The trained deep neural network model and the chip under test are obtained. Feature parameters are extracted from the chip under test and input into the deep neural network model. The input feature parameters are processed based on the deep neural network model to generate a performance feature representation of the chip under test. The performance feature representation is analyzed using the deep neural network model to obtain the performance prediction results of the chip under test.

[0076] When extracting characteristic parameters of the chip under test, a dedicated chip parameter acquisition tool is used to obtain basic chip information, such as architecture type, manufacturing process, and core configuration. For more refined physical characteristics, internal structural parameters of the chip, such as critical path delay and capacitance distribution, can be obtained through professional measurement equipment. Simultaneously, by applying standard test loads, dynamic performance data of the chip under different operating states is recorded, such as temperature changes and power consumption fluctuations under different loads.

[0077] After feature parameter extraction, data preprocessing is required, including data standardization, outlier handling, and feature selection. Data standardization uses a min-max standardization method to map each parameter value to the [0, 1] interval, eliminating dimensional differences. Outlier handling employs a standard deviation-based detection method to identify and correct abnormal data points. Feature selection uses correlation analysis and principal component analysis to select the most effective subset of parameters for predicting chip performance.

[0078] The preprocessed feature parameters are input into a deep neural network model, which consists of an input layer, multiple hidden layers, and an output layer. The hidden layers employ a fully connected layer structure, with each layer equipped with batch normalization and dropout mechanisms to improve the model's generalization ability. The ReLU activation function is used to effectively avoid the vanishing gradient problem. The number of nodes in the input layer matches the dimension of the feature parameters, while the output layer design is determined based on the target dimension of the performance prediction.

[0079] After receiving feature parameters, the deep neural network processes the input data using the forward propagation algorithm. The input layer receives the standardized feature parameter vector and passes it to the first hidden layer. In each hidden layer, the input data is multiplied by the layer's weight matrix and a bias vector is added, followed by a nonlinear transformation using the ReLU activation function. After multiple transformations, the neural network gradually extracts high-level representation features of the chip's performance, forming a performance feature representation of the chip under test. This representation contains the latent features and pattern information of the chip's performance, providing a foundation for subsequent performance prediction.

[0080] After obtaining the chip performance characteristics, the output layer of the deep neural network model analyzes and interprets these characteristics. The output layer employs different structures and activation functions depending on the prediction target. For regression problems (such as predicting specific performance values), the output layer typically does not use an activation function and directly outputs the predicted value; for classification problems (such as performance level classification), the softmax activation function is used for multi-class prediction.

[0081] During the performance prediction process, the model makes predictions for different performance metrics. For processing power prediction, the model analyzes the chip's core configuration, cache structure, and other characteristics to generate quantitative metrics such as MIPS (millions of instructions per second) or FLOPS (floating-point operations per second). For power consumption prediction, the model estimates power consumption under different load conditions based on parameters such as chip process technology, architecture, and operating frequency. For reliability prediction, factors such as thermal characteristics and voltage characteristics are comprehensively considered to evaluate the chip's stability and lifespan under long-term operating conditions.

[0082] To improve prediction accuracy, a confidence assessment mechanism can be introduced. By analyzing the probability distribution or variance of the model's output, a confidence score is assigned to each prediction, helping users determine the reliability of the prediction. For predictions with low confidence, the system will prompt that more test data or manual evaluation is needed.

[0083] Once the prediction results are generated, they will be output in a structured format, including core performance metrics, power consumption metrics, stability scores, and other dimensions. Simultaneously, to facilitate analysis, a visualization report will be generated, visually displaying the chip's performance across various dimensions using radar charts, bar charts, and other methods. This report will also provide a comparative analysis with a benchmark chip, highlighting the strengths and weaknesses of the chip under test.

[0084] In practical applications, model optimization can be performed on chips specific to particular domains. For example, for graphics processing chips, special attention is paid to their parallel computing capabilities and graphics rendering efficiency; for embedded system chips, the focus is on their power consumption performance and real-time response capabilities. Through domain-adaptive adjustments, the predictive accuracy of the model in specific scenarios can be improved.

[0085] A second aspect of the present invention provides a chip performance modeling system for adaptive depth migration, comprising: The first module is used to obtain the performance parameter dataset of the chip to be modeled, and to construct a deep neural network model based on the performance parameter dataset. The second module is used to adaptively adjust the transfer weights of the deep neural network model based on the feature distribution differences between the source domain chip samples and the target domain chip samples, calculate the difference metric between the source domain chip samples and the target domain chip samples, construct an adaptive weight matrix based on the difference metric, and dynamically fuse the adaptive weight matrix with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The third module is used to train the model based on the optimized transfer layer weights. During the training process, the performance prediction accuracy and the degree of transfer feature alignment are considered simultaneously to obtain the trained deep neural network model. The fourth module is used to predict the performance of the chip under test using the trained deep neural network model, and obtain the performance prediction result of the chip under test.

[0086] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0087] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0088] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive depth-transfer chip performance modeling method, characterized in that, include: Obtain the performance parameter dataset of the chip to be modeled, and construct a deep neural network model based on the performance parameter dataset; Based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the transfer weights of the deep neural network model are adaptively adjusted, and the difference metric between the source domain chip samples and the target domain chip samples is calculated. An adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during the training process, to obtain a trained deep neural network model. The trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction result of the chip under test is obtained.

2. The method according to claim 1, characterized in that, Based on the feature distribution differences between the source domain chip samples and the target domain chip samples, the deep neural network model undergoes adaptive adjustment of transfer weights. The calculation of the difference metric between the source domain chip samples and the target domain chip samples includes: Calculate the Mahalanobis distance between the feature distribution of the source domain and the feature distribution of the target domain, generate a feature distribution difference matrix, and adaptively adjust the transfer weights of the deep neural network model based on the feature distribution difference matrix; Based on the deep neural network model with adaptively adjusted transfer weights, the Euclidean distance between the source domain chip samples and the target domain chip samples is calculated to obtain the difference metric.

3. The method according to claim 1, characterized in that, An adaptive weight matrix is ​​constructed based on the difference metric, and the adaptive weight matrix is ​​dynamically fused with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights, including: An adaptive weight matrix is ​​constructed based on the difference metric, the difference metric is mapped to each dimension of the adaptive weight matrix according to a preset weight allocation strategy, and the adaptive weight matrix is ​​standardized. The standardized adaptive weight matrix is ​​dynamically fused with the migration layer weights. The fusion ratio is determined by adjusting the influence factor of the difference metric to obtain the optimized migration layer weights.

4. The method according to claim 1, characterized in that, The model is trained based on the optimized transfer layer weights, taking into account both performance prediction accuracy and transfer feature alignment during training. The resulting trained deep neural network model includes: The deep neural network model is trained based on the optimized transfer layer weights, and the performance prediction results and transfer feature distribution during the training process are obtained. The performance prediction accuracy is calculated based on the performance prediction results, and the transfer feature alignment degree is calculated based on the transfer feature distribution. Using the performance prediction accuracy and the transfer feature alignment as training metrics, a trained deep neural network model is obtained when the performance prediction accuracy reaches a first preset threshold and the transfer feature alignment reaches a second preset threshold.

5. The method according to claim 4, characterized in that, Obtaining performance prediction results and transfer feature distribution during the training process, calculating performance prediction accuracy based on the performance prediction results, and calculating transfer feature alignment based on the transfer feature distribution include: The intermediate output data during the training process of the deep neural network model is obtained, and the performance prediction results and transfer feature distribution are extracted from the intermediate output data. The performance prediction accuracy is calculated based on the matching relationship between the performance prediction results and the real labels, and the transfer feature alignment degree is calculated based on the similarity between the transfer feature distribution and the reference feature distribution.

6. The method according to claim 1, characterized in that, The trained deep neural network model is used to predict the performance of the chip under test, and the performance prediction results of the chip under test are as follows: The trained deep neural network model and the chip under test are obtained. Feature parameters are extracted from the chip under test and input into the deep neural network model. The input feature parameters are processed based on the deep neural network model to generate a performance feature representation of the chip under test. The performance feature representation is analyzed using the deep neural network model to obtain the performance prediction results of the chip under test.

7. An adaptive depth-transfer chip performance modeling system, used to implement the method of any one of claims 1-6, characterized in that, include: The first module is used to obtain the performance parameter dataset of the chip to be modeled, and to construct a deep neural network model based on the performance parameter dataset. The second module is used to adaptively adjust the transfer weights of the deep neural network model based on the feature distribution differences between the source domain chip samples and the target domain chip samples, calculate the difference metric between the source domain chip samples and the target domain chip samples, construct an adaptive weight matrix based on the difference metric, and dynamically fuse the adaptive weight matrix with the transfer layer weights of the deep neural network model to obtain the optimized transfer layer weights. The third module is used to train the model based on the optimized transfer layer weights. During the training process, the performance prediction accuracy and the degree of transfer feature alignment are considered simultaneously to obtain the trained deep neural network model. The fourth module is used to predict the performance of the chip under test using the trained deep neural network model, and obtain the performance prediction result of the chip under test.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images