Generalized multi-mode multi-angle static time sequence analysis method and system

Abstract

The invention discloses a generalizable multi-mode multi-angle static time sequence analysis method, which comprises the following steps: in multi-mode multi-angle analysis of electronic design, obtaining a basic model with generalization ability in multiple operation modes through pre-training; the basic model is expanded to any different process angles and parasitic angles by utilizing a lightweight fine tuning strategy, the angles comprise new angles which are not observed, and a final generalization model is obtained; and applying the generalization model to a plurality of to-be-evaluated process corners and parasitic corner combinations, and performing multi-mode multi-corner static time sequence prediction and analysis. According to the application, generalization pre-training under different process parasitic angles is realized, and after fine tuning, wider generalization can be realized under all operation modes, process angles and parasitic angles.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chip design, in particular, to a generalizable multi-mode multi-corner static timing analysis technical solution. BACKGROUND

[0002] With the continuous progress of process nodes, the Power, Voltage, and Temperature (PVT) fluctuations, parasitic effects, and the uncertainty of timing under multiple operating modes are significantly amplified. Therefore, the industry generally adopts a multi-mode multi-corner analysis process to strive for timing convergence under all expected combinations of power, voltage, temperature, and parasitic information. However, the existing process lacks cross-corner sharing and unified representation capabilities: the same design needs to be modeled and calculated separately in different corners, resulting in insufficient consistency and portability of local delay prediction results, further amplifying path-level cumulative errors and cross-scenario precision fluctuations, thereby increasing iteration overhead. In fact, EDA tools usually require independent analysis for each corner, which is inefficient and difficult to scale in chip design iterations.

[0003] Currently, the technical status of multi-mode multi-corner static timing analysis can be reviewed from two aspects: industry and academia.

[0004] The industry's contributions are mainly focused on mature electronic design automation tools, such as Synopsy's ICCompiler, etc. These tools form the cornerstone of the current chip design process. Their mainstream solution is based on deterministic algorithms, and in multi-mode multi-corner analysis, each process corner and parasitic corner is modeled, calculated, and evaluated as an independent scene. This "corner-by-corner analysis" approach, while able to cover all specified scenarios, has a fundamental flaw in lacking cross-corner knowledge sharing and unified representation. Each corner analysis is an independent, from-scratch calculation process, resulting in a large amount of redundant calculations. As chip design complexity increases and the number of corners that need to be verified increases dramatically, the computational cost and time overhead of this approach becomes extremely high, severely restricting the iteration efficiency of complex system-on-chip designs.

[0005] To cope with the challenges of traditional EDA tools in efficiency and accuracy, the academic community has started to actively explore machine learning based static timing analysis methods. Existing research has proposed a series of solutions from two complementary perspectives. The first class of methods aims to improve efficiency by simplifying and accelerating the multi-mode multi-corner analysis process. Specific techniques include pruning the search space, performing parallel analysis on multiple corners, and utilizing prior knowledge to approximate corners. These methods perform well when dealing with known, observable corners. The second class of methods utilizes machine learning algorithms to predict the timing results of target corners based on the information of observable corners. For example, Kahng et al. predict the timing under unobserved corners through machine learning to reduce the analysis workload; Bao et al. and Cao et al. construct a learning-based timing prediction framework to adapt to a wide range of supply voltage designs; Song et al. use a bidirectional long short-term memory network (BiLSTM) to predict the critical path under multiple corners. Previous work in this field has made some progress. For example, the APPSTA method significantly improves the accuracy of pre-layout STA by introducing pin-level wire length predictors and a dynamic incremental delay prediction scheme. On this basis, the GM 2 The GMSTA method introduces a multi-mode pre-training framework and achieves generalized prediction of multiple parasitic corners. These studies demonstrate the great potential of machine learning in improving the accuracy and efficiency of STA. However, existing academic research still has obvious limitations when dealing with complex multi-mode multi-corner scenarios. Some methods use transfer learning to achieve cross-corner generalization to some extent, but usually still require additional model transfer training for new, unseen corners, and their generalization ability is limited.

[0006] Although existing EDA tools and machine learning methods have made some progress in static timing analysis, they still have obvious shortcomings when faced with multi-mode multi-corner scenarios. First, traditional EDA tools need to perform analysis for each process and parasitic corner separately, and the repetitive calculation process leads to low efficiency, which cannot quickly cope with complex designs containing multiple corners. Second, previous machine learning solutions lack the ability to generalize to unseen process / parasitic corners: models often only work accurately under specific conditions trained, and the accuracy decreases significantly once the process conditions change. For example, the GM 2Although the STA model is more accurate than Synopsys IC Compiler under most process corners, the prediction accuracy is lacking under the worst process (slow process corner) condition, and the generalization ability is insufficient. To maintain accuracy under all process corners, additional fine-tuning of the model is still needed for extreme slow corners. This reflects that the existing model is not comprehensive enough in adapting to different process corners. APPSTA exhibits stable multi-mode generalization in circuit implementation of different operating modes, and significantly improves the local delay and slack accuracy compared to industrial tools (the average error reduction is 50% and 64% respectively). It should be noted that APPSTA focuses on multiple operating modes and does not solve the problem of integrated generalization of multiple process corners and multiple parasitic corners.

[0007] More importantly, there is a lack of effective means to quickly expand the model to new unobserved corners. The training of machine learning models relies on a large amount of labeled timing data, but for many unobserved process / parasitic corners, there is often a lack of labeled data available for training. Supervised learning requires a large amount of labeled data support, and the cost of obtaining such data may even exceed the entire chip design process itself. In the prior art, if the model is to adapt to new processes or environmental conditions, it is usually necessary to collect data and train the model again, which not only increases the development cost but also takes a lot of time. The most advanced existing machine learning timing analysis method needs to train the model for several days to adapt to a new unobserved process corner, which is unacceptable in a fast iteration design process. In addition, adjusting the model hyperparameters for each corner is also an inefficient approach: traditional methods may need to optimize and train hyperparameters for different process / parasitic corners one by one, resulting in a lot of redundant work and low overall efficiency. In summary, the existing technology cannot meet the needs of multi-mode and multi-corner static timing analysis in terms of generalization ability and analysis efficiency.

[0008] Therefore, it is urgent to study an analysis and prediction method that can share information between corners, uniformly represent multi-corner characteristics, and stably generalize at an early stage of design, to improve accuracy and computational efficiency at the same time. SUMMARY

[0009] To solve the problem in the prior art that static timing analysis needs to cover multiple operating modes and multiple process parasitic corners in digital circuit design. However, the existing method has poor generalization in different modes and corners, and usually needs to train the model and optimize the hyperparameters for each corner separately, which not only brings redundant overhead, but also leads to the need for re-tuning when facing unobserved corners, which takes a long time, often several days.

[0010] The traditional machine learning method needs to be trained respectively for each process angle and parasitic angle when processing multiple angles, which brings extremely high calculation and time cost. When encountering unobserved angles, the existing method often cannot directly predict, and needs to be fine-tuned or trained again, which affects the efficiency of the design process. The existing method often needs several days of training time to obtain relatively accurate timing results on new angles, which is difficult to meet the demand for rapid iteration in circuit design.

[0011] In a first aspect, the embodiments of the present application provide a generalizable multi-mode multi-angle static timing analysis method, which comprises:

[0012] Model pre-training step: in the multi-mode multi-angle analysis of electronic design, a basic model with generalization ability under multiple operation modes is obtained through pre-training;

[0013] Lightweight fine-tuning step: using a lightweight fine-tuning strategy, the basic model is extended to any different process angle and parasitic angle, including new angles that have never been observed, to obtain a final generalization model;

[0014] Generalization analysis step: the generalization model is used for multiple process angles and parasitic angle combinations to be evaluated for multi-mode multi-angle static timing prediction and analysis.

[0015] In the embodiments of the present application, the above-mentioned generalizable multi-mode multi-angle static timing analysis method further comprises:

[0016] Circuit decomposition step: the entire circuit of the electronic design is decomposed into multiple local circuits, the local circuit contains a ternary relationship of input pin to output pin to receiving pin, and the generalization model is used to predict the difference between the real local delay and the reference delay by learning the local delay at the ternary relationship granularity.

[0017] In the embodiments of the present application, the above-mentioned model pre-training step comprises:

[0018] A dataset containing multiple operation modes, process angles and parasitic angles is constructed;

[0019] Under a preset typical reference process angle, all preset operation modes are traversed, and each operation mode generates a local sample set containing circuit features and real delay;

[0020] The difference data between the predicted real delay and the reference delay under all operation modes is collected to form a total training set, and a regressor is trained from the total training set to generate a basic model with cross-mode generalization ability.

[0021] In the embodiments of the present application, the above-mentioned lightweight fine-tuning step comprises:

[0022] The lightweight fine-tuning strategy includes: fine-tuning the base model with a preset learning rate, and all angles used for fine-tuning share the same set of hyperparameters;

[0023] Once hyperparameter tuning is performed on the observed angles, the pre-trained model is fine-tuned as a whole with a preset learning rate to generate a final generalization model; for unobserved angles, the base model is used for inference.

[0024] In the embodiment of the application, the generalization analysis step of the multi-angle mode further includes:

[0025] The generalization model is used for a plurality of process angle and parasitic angle combinations that do not participate in training, the errors of local delay and endpoint relaxation are respectively evaluated, and the evaluation results are compared with the pre-routing analysis results under the same setting and the fitting level on the reference angle, to realize generalization prediction and analysis.

[0026] In a second aspect, the embodiments of the application provide a generalizable multi-mode multi-angle static timing analysis system, which adopts the generalizable multi-mode multi-angle static timing analysis method described above, and the system includes:

[0027] The model pre-training module: in multi-mode multi-angle analysis of electronic design, a base model with generalization capability under multiple operating modes is obtained through pre-training;

[0028] The lightweight fine-tuning module: the base model is expanded to any different process angle and parasitic angle by using the lightweight fine-tuning strategy, the angles include new angles that have never been observed, and a final generalization model is obtained;

[0029] The generalization analysis module: the generalization model is used for a plurality of process angle and parasitic angle combinations to be evaluated, and multi-mode multi-angle static timing prediction and analysis are performed;

[0030] The circuit decomposition module: the entire circuit of the electronic design is decomposed into a plurality of local circuits, the local circuits include a ternary relationship from the input pin to the output pin to the receiving pin, the generalization model is used to predict the difference between the real local delay and the reference delay by learning the local delay at the ternary relationship granularity.

[0031] In the embodiment of the application, the model pre-training module includes:

[0032] A dataset containing multiple operating modes, process angles and parasitic angles is constructed;

[0033] Under a preset typical reference process angle, all preset operating modes are traversed, and a local sample set containing circuit features and real delay is generated for each operating mode;

[0034] Collect the difference data between the predicted true delay and the reference delay under all operating modes to form a total training set. Train a regressor on the total training set to generate a base model with cross-mode generalization ability.

[0035] In this embodiment of the invention, the lightweight fine-tuning module includes:

[0036] The lightweight fine-tuning strategy includes: fine-tuning the base model using a preset learning rate, and all angles used for fine-tuning sharing the same set of hyperparameters;

[0037] Hyperparameter tuning is performed on the observed angles, and the pre-trained model is fine-tuned as a whole with a preset learning rate to generate the final generalized model; for the unobserved angles, the base model is used for inference.

[0038] Thirdly, embodiments of this application provide a circuit whose design employs the steps of the aforementioned generalizable multi-mode multi-angle static timing analysis method.

[0039] Fourthly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the generalizable multi-mode multi-angle static timing analysis method described above.

[0040] Compared with existing technologies, it has the following outstanding advantages:

[0041] 1) The method and system of this invention propose GM 3 STA is a multimodal, multi-angle analysis method based on transfer learning. To achieve generalized pre-training under different parasitic angles of various processes, step GM... 2 STA achieves generalization across multiple operating modes, and with fine-tuning, it can achieve even broader generalization across all operating modes, process angles, and parasitic angles.

[0042] 2) The method and system of this invention summarize the types of typical process corners and analyze the pre-trained model GM. 2 The accuracy of STA under different process parasitic angles. Based on this, this chapter discusses GM. 2 The generalization ability of STA under different process corner parasites was classified, and it was found that except for slow process corners, GM 2 STA can achieve good generalization ability under various process parasitic angles.

[0043] 3) The method and system of this invention are designed to achieve generalization under different parasitic angles in the process, GM 3STA employs a lightweight fine-tuning strategy, performing hyperparameter tuning only once for all angles. This approach not only eliminates the need for repeated hyperparameter tuning for parasitic angles of different processes, but also avoids fine-tuning unobserved angles in the absence of labeled data.

[0044] 4) The construction of the dataset for the method and system of this invention provides a basis for evaluating the pre-trained method GM. 2 STA and supports GM 3 This invention utilizes STA transfer learning and fine-tuning to construct a circuit dataset in a pre-routing static timing analysis scenario. This dataset covers multiple operating modes and can be used for generalization studies across multiple processes and parasitic angles. The data generation process is as follows: Synopsys DesignCompiler is used for logic synthesis and clock cycle constraint configuration; Synopsys IC Compiler completes placement and routing. Based on this, features and connectivity relationships are extracted from the pre-routing circuit implementation and design report, while label values ​​are derived from the post-routing design and report. Attached Figure Description

[0045] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0046] Figure 1 This is a schematic diagram of the multi-mode, multi-angle static timing analysis method of the present invention;

[0047] Figure 2 Pre-trained GM for embodiments of the present invention 2 STA method diagram;

[0048] Figure 3 The multi-mode, multi-angle static timing analysis method GM is an embodiment of the present invention. 3 STA method diagram;

[0049] Figure 4 GM is an embodiment of the present invention. 2 A schematic diagram illustrating the size of the STA training set, validation set, and test set;

[0050] Figure 5 GM is an embodiment of the present invention. 2 A schematic diagram illustrating the training details of STA and baseline methods;

[0051] Figure 6 This is a schematic diagram of the multi-mode, multi-angle static timing analysis system of the present invention;

[0052] Figure 7 This is a schematic diagram of the computer hardware of the present invention. Detailed Implementation

[0053] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0054] It should also be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0055] It should also be understood that, in various embodiments of the present invention, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0056] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0057] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0058] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0059] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] To make the above-mentioned features and effects of the present invention clearer and easier to understand, specific embodiments are described below in conjunction with the accompanying drawings. This specification discloses one or more embodiments incorporating the features of the present invention. The disclosed embodiments are merely illustrative. The scope of protection of the present invention is not limited to the disclosed embodiments, but is defined by the appended claims.

[0061] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0062] In conducting multi-mode, multi-angle static time-series analysis, this invention first systematically evaluated the engineering applicability of existing industrial processes and academic methods. On the one hand, it is difficult to obtain sufficient labeled data for unobserved process / parasitic angles, making it difficult to directly transfer existing supervised machine learning models. On the other hand, repeated training or tuning of models angle by angle is costly and has poor scalability in engineering. Existing alternative paths (including angle-by-angle analysis relying on traditional EDA tools, or machine learning schemes that require additional fine-tuning of the target angle) cannot simultaneously ensure both the comprehensiveness and efficiency of the analysis.

[0063] Based on this, with generalization as the primary goal, the inventors proposed a multi-modal, multi-angle static time series analysis method and system (GM) based on transfer learning. 3 STA): Model training is completed on observed processes / parasitic angles, and the acquired knowledge is transferred to unobserved angles; the technical approach is to first complete pre-training of multiple operation modes according to the APPSTA approach (i.e., preliminary work GM). 2(STA), and then design one-time hyperparameter tuning and lightweight fine-tuning to achieve rapid adaptation across process corners and parasitic corners. Without sacrificing accuracy, robust feature representation and training strategies are used to improve robustness to process, voltage, and parasitic differences, thereby achieving true unified generalization for multi-mode and multi-corner scenarios. Applying the above-implemented static timing analysis method that generalizes under multiple operating modes to the generalization of multiple process corners and multiple parasitic corners still faces two major challenges from the multi-mode and multi-corner analysis problem:

[0064] First, the high time cost of a single hyperparameter tuning session makes it impractical to retrain machine learning models for all process parasitic corners. Even for nonparametric models like Random Forest and XGBoost, training a machine learning model from scratch typically requires a week or more of hyperparameter tuning. Second, the lack of labeled data for unobserved corners makes it impossible to train machine learning models for these corners, as supervised learning requires a large amount of labeled data for training, and the cost of obtaining these labels may even exceed the cost of the digital circuit design itself.

[0065] Therefore, this invention proposes GM 3 STA, a multimodal, multi-faceted analysis method based on transfer learning, boasts excellent speed and accuracy. GM 3 The core of STA lies in utilizing a transfer learning framework to train on observed angles in order to predict the temporal outcomes of unobserved angles. To address the two challenges mentioned above, GM... 3 STA employs a lightweight fine-tuning strategy, performing hyperparameter tuning only once for all angles. This eliminates the need for repeated hyperparameter tuning of parasitic angles from different processes and avoids fine-tuning of unobserved angles lacking labeled data. GM 3 Static time series analysis (STA) can be roughly divided into two steps: First, pre-training the model using a method similar to APPSTA to achieve static time series analysis under multiple operating modes; this step is called GM. 2 STA; Second, a customized transfer learning framework is used to quickly fine-tune the pre-trained model for all process parasitic angles.

[0066] The application scenario of this invention is set as static timing analysis before routing. In this scenario, a GM (Gateway Management System) is designed. 3 The STA method details were explained, and its performance was thoroughly evaluated. Experimental results show that GM... 3 STA outperforms Synopsys IC Compiler, one of the most advanced electronic design automation tools, in terms of accuracy across different operating modes, process angles, and parasitic angles. Compared to Synopsys IC Compiler, GM... 3STA achieved an average reduction of 70.5% in local delay error and an average reduction of 48.8% in relaxation error across different circuit implementations at 8 process corners and 8 parasitic corners. Meanwhile, GM... 3 STA eliminates the need for repeated model training and has a short runtime. In contrast, existing state-of-the-art machine learning-based pre-wiring static timing analysis methods require several days of training to generalize to unobserved angles, while GM... 3 STA requires only a few seconds of inference time.

[0067] The methods of this application will be described in detail below with reference to specific embodiments:

[0068] like Figure 1 As shown, this invention proposes a generalizable multi-mode multi-angle static time series analysis method GM 3 STA is a generalizable static time series analysis method designed to address the multiple challenges faced by existing technologies in handling static time series analysis for multi-mode multi-corner (MMMC) scenarios, including efficiency, accuracy, and generalization. The core of this invention is a two-stage analysis framework based on transfer learning. First, a base model with strong generalization capabilities across multiple operating modes is obtained through pre-training. Then, a lightweight fine-tuning strategy is used to quickly and efficiently extend this model to different process corners and parasitic corners, including unobserved new corners.

[0069] The generalizable multi-mode multi-angle static time series analysis method proposed in this invention includes:

[0070] Model pre-training step 101: In the multi-mode multi-angle analysis of electronic design, a basic model with generalization ability in multiple operating modes is obtained through pre-training;

[0071] Lightweight fine-tuning step 102: Using a lightweight fine-tuning strategy, the base model is extended to any different process angles and parasitic angles, including new angles that have not been observed, to obtain the final generalized model;

[0072] Generalization analysis step 103: Apply the generalization model to multiple combinations of process corners and parasitic corners to be evaluated, and perform multi-mode multi-corner static time series prediction and analysis.

[0073] In this embodiment of the invention, the above-described generalizable multi-mode multi-angle static time series analysis method further includes the following steps before the model pre-training step 101:

[0074] Circuit decomposition step 104: Decompose the entire circuit of the electronic design into multiple local circuits. Each local circuit contains a ternary relationship from input pin to output pin to receive pin. By learning the local delay at the ternary relationship granularity, the generalization model is used to predict the difference between the true local delay and the reference delay.

[0075] Specifically, in a specific embodiment of the present invention, the problem decomposition includes:

[0076] The goal of this invention is to achieve accurate timing prediction for the entire circuit under any combination of "operating mode × process angle × parasitic angle". Directly modeling the end-to-end timing margin of the entire circuit is both complex and impractical within the vast multi-mode, multi-angle space. Therefore, this invention decomposes the complex relaxation estimation problem into predicting local delays of a local circuit composed of "a single standard cell and its driven interconnect branches", and then accumulating the arrival time and relaxation at the path level using a graph traversal approach.

[0077] To further reduce errors and improve portability, pin-level differentiation is introduced at the local circuit level:

[0078] A local circuit is defined as a ternary relationship (input pin → output pin → receive pin). The local delay on each pin pair is defined as the time taken for a signal to travel from the input pin through the timing arc within the cell to the output pin, and then along a specific interconnect branch from the output pin to the specific receive pin, under different operating modes and corner-domain processes / parasitic angles. This quantity includes "gate delay + the netting and load effects of that fan-out branch," and has a consistent physical meaning across modes and corners. Feature extraction and annotation are both tied to specific pin pairs, without merging or averaging different fanouts of the same driving net, to avoid systematic bias caused by "smoothing out" differences in geometric length, coupling capacitance, and endpoint load.

[0079] Fan-out branches within the same drive network often differ significantly in geometric paths, parasitic parameters, and endpoint loads. Modeling these branches using "network-level averaging" or "cell-level summarization" methods can easily lead to mean bias and accumulate into larger relaxation errors at the path level. By learning local delays at the granularity of (input pin, output pin, specific receiver pin), the model can focus on the inherent timing patterns of propagation within standard cells and specific fan-out branches, thereby significantly reducing local prediction errors. Furthermore, at the graph-level accumulation stage, the error is not amplified by fan-out confusion. This fine-grained prediction objective with consistent physical meaning lays the foundation for learning cross-mode general knowledge during the pre-training phase and robustly generalizing this knowledge to different processes / parasitic angles during the transfer learning phase.

[0080] In this embodiment of the invention, the above-mentioned model pre-training step 101 includes:

[0081] A dataset containing multiple operating modes, process angles, and parasitic angles was constructed;

[0082] Under a preset typical reference process corner, all preset operating modes are traversed, and each operating mode generates a local sample set containing circuit features and actual delays;

[0083] Collect the difference data between the predicted true delay and the reference delay under all operating modes to form a total training set. Train a regressor on the total training set to generate a base model with cross-mode generalization ability.

[0084] Specifically, such as Figure 2 As shown, the specific process for this stage is as follows: Figure 2 As shown in Algorithm 1. In a specific embodiment of the present invention, multi-mode pre-training (GM) 2 STA includes:

[0085] The first stage of this invention is multi-mode pre-training, the goal of which is to build a time series prediction model with strong basic generalization ability under multiple operating modes.

[0086] The system first iterates through all given operating modes under a selected typical reference process corner, generating a local sample set containing circuit features and actual delays for each mode. To enable the model to learn more generalizable timing variation patterns, this invention does not directly predict the absolute delay value, but rather predicts the difference between the actual delay and a reference delay. The difference data from all modes are collected into a total training set and used to train an XGBoost regressor, ultimately generating a pre-trained base model with cross-mode generalization capabilities.

[0087] In this embodiment of the invention, the above-mentioned lightweight fine-tuning step 102 includes:

[0088] The lightweight fine-tuning strategy includes: fine-tuning the base model using a preset learning rate, and all angles used for fine-tuning sharing the same set of hyperparameters;

[0089] Hyperparameter tuning is performed on the observed angles, and the pre-trained model is fine-tuned as a whole with a preset learning rate to generate the final generalized model; for the unobserved angles, the base model is used for inference.

[0090] Specifically, such as Figure 3 As shown, the process for this stage is as follows: Figure 3 As shown in Algorithm 2. In a specific embodiment of the present invention, Multi-angle migration fine-tuning (GM) 3 STA specifically includes:

[0091] After obtaining the pre-trained model, the second stage is multi-angle transfer fine-tuning, which aims to efficiently extend the capabilities of the base model to arbitrary process angles and parasitic angles, including unobserved angles.

[0092] This invention utilizes a small amount of observed angle data to fine-tune a pre-trained model. Its core lies in its unique lightweight fine-tuning strategy: the fine-tuning process uses a small learning rate, which is a proportional coefficient used to control the magnitude of model parameter updates during fine-tuning. This learning rate adjusts the step size of parameter updates along the optimization direction in each iteration, thereby reducing disturbance to pre-trained knowledge while ensuring convergence stability. The learning rate can be between 0.001 and 0.1. Furthermore, all angles used for fine-tuning share the same set of hyperparameters. Hyperparameters are control parameters pre-set before model training / fine-tuning and not directly learned from training data. They constrain model structural complexity, sampling strategies, and the training iteration process. In this embodiment, all angles used for fine-tuning share the same set of hyperparameters; these hyperparameters include at least: number of iterations, number of training iterations, learning rate, model complexity control parameters, sample sampling ratio, leaf node constraint parameters, and column sampling rate parameters. This strategy avoids independent and time-consuming model training or hyperparameter search for each angle, thereby reducing the time cost of adapting to new angles from several days to seconds. After fine-tuning, the system obtains a final model capable of accurately predicting the timing of any target angle.

[0093] In this embodiment of the invention, the generalization analysis step 103 of the multi-angle pattern further includes:

[0094] The generalization model is applied to multiple combinations of process corners and parasitic corners that were not trained, and the errors of local delay and endpoint relaxation are evaluated respectively. The results are compared with the pre-routing analysis results and the fitting level on the reference corner under the same settings to achieve generalization prediction and analysis.

[0095] Specifically, in a specific embodiment of the present invention, the above-mentioned pre-training generalization analysis in multi-angle mode specifically includes:

[0096] To test the multi-mode pre-trained model (GM) 2 To assess the adaptability of the model under different process and parasitic angles, this invention trains the model on a selected typical reference process and reference parasitic angle. The training objective is to measure the difference between the true local delay and the reference delay, thereby reducing systematic offsets between different modes / angles. After training, without any additional fine-tuning, the model is directly applied to multiple untrained process / parasitic angle combinations to evaluate the errors in local delay and endpoint relaxation. The results are then compared with the pre-wiring analysis results and the fitting level on the reference angle using industrial tools under the same settings to examine its cross-angle generalization ability and stability.

[0097] Evaluation results show that, thanks to the "multi-operation mode training + difference learning" setup, GM 2 STA maintains an error level close to the reference angle across most process and parasitic angles, demonstrating a good angle-domain migration trend. However, in scenarios with extremely slow process angles and those superimposed with unfavorable parasitic angles, the error of cross-angle extrapolation increases more significantly. This deviation accumulates at the path level and is further amplified until the endpoint relaxation, affecting the judgment of critical paths and timing margins. Mechanistic analysis shows that these angles are often accompanied by significant shifts in characteristic distributions such as increased input slew rate, expanded load capacitance, and extended equivalent parasitic angle range. Simultaneously, the nonlinear sensitivity of local delay to these characteristics increases, and relying solely on reference angle experience for extrapolation easily leads to systematic biases. Therefore, relying solely on multi-mode pre-training at the reference angle is insufficient to maintain the same accuracy as the signing-out process across the entire angle domain. An angle-adaptive mechanism is needed to compensate for the errors caused by the aforementioned distribution shifts and enhanced nonlinearity without increasing the cost of angle-by-angle retraining and parameter tuning.

[0098] Based on the above analysis, this invention proposes GM in the second stage. 3 STA's lightweight transfer learning involves uniformly fine-tuning hyperparameters on a small number of observed angles and then fine-tuning the pre-trained model with a small learning rate. This allows the model to achieve stable adaptation to the slow-process angle family without changing its structure and interface. For unobserved angles, the model is used directly for inference without additional fine-tuning. This conclusion, together with experimental observations, constitutes the methodological basis for a natural transition from "multi-operation mode generalization" to "multi-process / multi-parasitic angle generalization," laying the foundation for achieving consistent and efficient time series prediction across the entire angle domain.

[0099] Specifically, one specific embodiment of the present invention includes:

[0100] To verify the GM invention 3 To assess the effectiveness of STA, the inventors constructed a dedicated circuit dataset and conducted comprehensive training and evaluation in specific hardware and software environments.

[0101] 1. Dataset Construction

[0102] Because there is a lack of benchmark datasets in the public domain suitable for static timing analysis of multiple modes and angles, this invention constructs a dataset covering multiple operating modes, process angles, and parasitic angles. This dataset is based on 15 circuit designs from the open-source ITC'99 benchmark and the OpenROAD project.

[0103] For each circuit design, 40 corresponding circuit implementations were generated by setting 40 different clock cycle constraints (representing different operating modes). The logic synthesis and physical design processes of these circuits were completed using industry-standard EDA tools and TSMC's N45GS standard cell library.

[0104] For each generated circuit implementation, the present invention further analyzes it under 8 typical process corners and 8 parasitic corners, thereby constructing a large multi-mode, multi-corner static timing analysis dataset.

[0105] like Figure 4 As shown, in terms of data acquisition, features used for the machine learning model (such as reference latency, input slew rate, physical coordinates, etc.) were extracted from the pre-wiring design report; while the label values ​​used for model training (i.e., the true local latency) were precisely obtained from the post-wiring design report. The entire dataset was divided into independent training, validation, and test sets to ensure an objective and fair evaluation of model performance. Figure 4 For GM 2 The training size of the STA training set, validation set, and test set.

[0106] 2. Implementation Environment and Model Training

[0107] like Figure 5 As shown, Figure 5 For GM 2 Training details of the STA and baseline methods. The algorithm and model of this invention were implemented and validated on a server equipped with a 16-core Intel(R) Xeon(R) Gold 6254 CPU (operating frequency 3.10GHz). The machine learning model was mainly implemented using Python and the XGBoost library. In the multi-mode pre-training stage, the system used the training set to train the XGBoost model under typical reference angles (TC process angle and 9m_cbest parasitic angle) to generate the base model.

[0108] In the multi-angle transfer fine-tuning stage, using a subset of observed angle data from the validation set, the lightweight fine-tuning strategy of this invention is employed to fine-tune the base model, resulting in the final generalized model. Through the above implementation process, this invention verifies the proposed GM... 3 The STA method and system can run effectively in standard EDA processes and hardware environments, and its technical effectiveness in terms of high accuracy, high efficiency and strong generalization ability is demonstrated through the construction of a dedicated dataset.

[0109] As described above, the system and method of the present invention can be implemented well.

[0110] Compared with the prior art, the present invention has the following outstanding advantages and beneficial effects:

[0111] The GM of this invention 3The STA method outperforms existing technologies in both accuracy and efficiency, providing a breakthrough improvement for timing analysis in chip design processes. Specifically, under different operating modes and different process / parasitic angle combinations in various practical circuit designs, compared to Synopsys IC Compiler, GM... 3 The mean square error of local delay in STA was reduced by an average of approximately 70.5%, and the mean absolute error of endpoint relaxation was reduced by an average of approximately 48.8%, significantly improving the accuracy of timing analysis in the pre-placement / routing stage. Methodologically, GM... 3 STA achieves a unified approach from "multi-operation mode generalization" to "multi-process / parasitic angle generalization" using a two-step framework of "pre-training (GM2STA) + transfer learning": training is completed first on observed angles, and then transferred to unobserved angles. The core only requires one hyperparameter tuning and lightweight fine-tuning for "all angles." The efficiency advantage of this design lies in avoiding the days-long time overhead of repeated angle-by-angle training / parameter tuning, enabling unobserved angles to obtain timing results in seconds. The source of its generalization ability is clearly explained in the research summary: first, problem decomposition—transforming the overall circuit relaxation estimation into local delay prediction at the local circuit level; second, introducing a reference delay as an input feature, transforming delay value prediction into "incremental" prediction, reducing learning difficulty, and enhancing cross-angle stability. In terms of contributions, GM... 2 STA has achieved generalization across multiple operating modes; combined with system analysis and classification of typical power voltage-temperature angles, it was found that, except for slow process-related angles, the pre-trained model exhibits good generalization ability under various process / parasitic angles, which also confirms GM's performance in engineering. 3 The feasibility and necessity of STA requiring only "one-time parameter tuning + lightweight fine-tuning" are highlighted. In summary, as an integrated machine learning solution for multiple modes and perspectives, GM... 3 STA completed a system evaluation covering 8 process corners and 8 parasitic corners in the pre-routing scenario, and significantly outperformed industrial tools in terms of accuracy and timeliness, providing a generalized path that can be directly incorporated into the design process.

[0112] Secondly, such as Figure 6 As shown, this application provides a generalizable multi-mode multi-angle static timing analysis system, employing the generalizable multi-mode multi-angle static timing analysis method described above. The system includes:

[0113] Model pre-training module 201: In multi-mode and multi-angle analysis of electronic design, a basic model with generalization ability in multiple operating modes is obtained through pre-training;

[0114] Lightweight fine-tuning module 202: Using a lightweight fine-tuning strategy, the base model is extended to any different process angles and parasitic angles, including new angles that have not been observed, to obtain the final generalized model;

[0115] Generalization Analysis Module 203: The generalization model is applied to multiple combinations of process corners and parasitic corners to be evaluated to perform multi-mode multi-corner static time series prediction and analysis;

[0116] Circuit decomposition module 204: Decomposes the entire circuit of the electronic design into multiple local circuits. The local circuits contain ternary relationships from input pins to output pins to receive pins. By learning the local delay at the granularity of the ternary relationships, the generalization model is used to predict the difference between the true local delay and the reference delay.

[0117] In this embodiment of the invention, the model pre-training module 201 includes:

[0118] A dataset containing multiple operating modes, process angles, and parasitic angles was constructed;

[0119] Under a preset typical reference process corner, all preset operating modes are traversed, and each operating mode generates a local sample set containing circuit features and actual delays;

[0120] Collect the difference data between the predicted true delay and the reference delay under all operating modes to form a total training set. Train a regressor on the total training set to generate a base model with cross-mode generalization ability.

[0121] In this embodiment of the invention, the lightweight fine-tuning module 202 includes:

[0122] The lightweight fine-tuning strategy includes: fine-tuning the base model using a preset learning rate, and all angles used for fine-tuning sharing the same set of hyperparameters;

[0123] Hyperparameter tuning is performed on the observed angles, and the pre-trained model is fine-tuned as a whole with a preset learning rate to generate the final generalized model; for the unobserved angles, the base model is used for inference.

[0124] Thirdly, embodiments of this application provide a circuit whose design employs the steps of the aforementioned generalizable multi-mode multi-angle static timing analysis method.

[0125] Fourthly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the generalizable multi-mode multi-angle static timing analysis method described above.

[0126] In addition, combined Figure 1 The generalizable multi-mode multi-angle static timing analysis method described in this application embodiment can be implemented by electronic devices, such as computer devices. Figure 7This is a schematic diagram of the hardware structure of a computer device according to an embodiment of this application.

[0127] In some embodiments, the computer device may further include a communication interface 83 and a bus 80. For example, Figure 7 As shown, the processor 81, memory 82, and communication interface 83 are connected through bus 80 and complete communication with each other.

[0128] Specifically, the processor 81 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0129] The memory 82 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 81.

[0130] The processor 81 reads and executes computer program instructions stored in the memory 82 to implement any of the generalizable multi-mode multi-angle static timing analysis methods in the above embodiments.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A generalizable multi-mode, multi-angle static time series analysis method, characterized in that, The method includes: Model pre-training steps: In multi-mode and multi-angle analysis of electronic design, a basic model with generalization ability in multiple operating modes is obtained through pre-training; Lightweight fine-tuning steps: Using a lightweight fine-tuning strategy, the base model is extended to any different process angles and parasitic angles, including new angles that have not been observed, to obtain the final generalized model; Generalization analysis steps: The generalization model is applied to multiple combinations of process corners and parasitic corners to be evaluated to perform multi-mode, multi-corner static time series prediction and analysis.

2. The generalizable multi-mode multi-angle static time series analysis method according to claim 1, characterized in that, The method further includes: Circuit decomposition steps: The entire circuit of the electronic design is decomposed into multiple local circuits, each containing a ternary relationship from input pin to output pin to receive pin. Local delays are learned at the granularity of the ternary relationships, and the generalization model is used to predict the difference between the true local delay and the reference delay.

3. The generalizable multi-mode multi-angle static time series analysis method according to claim 1 or 2, characterized in that, The model pre-training step includes: A dataset containing multiple operating modes, process angles, and parasitic angles was constructed; Under a preset typical reference process angle, all preset operating modes are traversed, and each operating mode generates a local sample set containing circuit features and actual delays; Collect the difference data between the predicted true delay and the reference delay under all operating modes to form a total training set. Train a regressor on the total training set to generate the base model with cross-mode generalization ability.

4. The generalizable multi-mode multi-angle static time series analysis method according to claim 1 or 2, characterized in that, The lightweight fine-tuning steps include: The lightweight fine-tuning strategy includes: fine-tuning the base model using a preset learning rate, and all angles used for fine-tuning sharing the same set of hyperparameters; Hyperparameters are tuned once on the observed angles, and the pre-trained model is fine-tuned as a whole with a preset learning rate to generate the final generalized model; for the unobserved angles, the base model is used for inference.

5. The generalizable multi-mode multi-angle static time series analysis method according to claim 1 or 2, characterized in that, The generalization analysis step of the multi-angle pattern also includes: The generalization model is applied to multiple combinations of process corners and parasitic corners that were not trained, and the errors of local delay and endpoint relaxation are evaluated respectively. The results are compared with the pre-routing analysis results and the fitting level on the reference corner under the same settings to achieve generalization prediction and analysis.

6. A generalizable multi-mode multi-angle static time series analysis system, employing the generalizable multi-mode multi-angle static time series analysis method as described in any one of claims 1-5, characterized in that, The system includes: Model pre-training module: In multi-mode and multi-angle analysis of electronic design, a basic model with generalization ability in multiple operating modes is obtained through pre-training; Lightweight fine-tuning module: Using a lightweight fine-tuning strategy, the basic model is extended to any different process angles and parasitic angles, including new angles that have not been observed, to obtain the final generalized model; Generalization analysis module: The generalization model is applied to multiple combinations of process corners and parasitic corners to be evaluated to perform multi-mode multi-corner static time series prediction and analysis; Circuit decomposition module: Decomposes the entire circuit of the electronic design into multiple local circuits, each containing a ternary relationship from input pin to output pin to receive pin. By learning local delays at the granularity of these ternary relationships, the generalization model is used to predict the difference between the true local delay and a reference delay.

7. The generalizable multi-mode multi-angle static time series analysis system according to claim 6, characterized in that, The model pre-training module includes: A dataset containing multiple operating modes, process angles, and parasitic angles was constructed; Under a preset typical reference process angle, all preset operating modes are traversed, and each operating mode generates a local sample set containing circuit features and actual delays; Collect the difference data between the predicted true delay and the reference delay under all operating modes to form a total training set. Train a regressor on the total training set to generate the base model with cross-mode generalization ability.

8. The generalizable multi-mode multi-angle static time series analysis system according to claim 6, characterized in that, The lightweight fine-tuning module includes: The lightweight fine-tuning strategy includes: fine-tuning the base model using a preset learning rate, and all angles used for fine-tuning sharing the same set of hyperparameters; Hyperparameters are tuned once on the observed angles, and the pre-trained model is fine-tuned as a whole with a preset learning rate to generate the final generalized model; for the unobserved angles, the base model is used for inference.

9. A circuit, characterized in that, The circuit design employs the steps of the generalizable multi-mode multi-angle static timing analysis method as described in any one of claims 1-5.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the generalizable multi-mode multi-angle static timing analysis method as described in any one of claims 1-5.

Images