A multi-model fusion memory macro cell design quality prediction system and method

Abstract

The application provides a multi-model fusion memory macro cell design quality prediction system and method. The system includes a feature extraction and coding module, a multi-model fusion prediction module, and a continuous learning module. By integrating multiple AI sub-models and using stacking or voting methods for result fusion, high-precision prediction of memory macro cell power consumption, performance, and area is achieved. The system has closed-loop learning capability and can adaptively update with feedback data, and can be integrated into a memory compiler or EDA flow to provide fast and accurate PPA prediction and optimization support in the early stages of chip design.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design and automation (EDA), and particularly relates to a multi-model fusion memory macrocell design quality prediction system and method. Background Technology

[0002] As semiconductor processes advance to the nanometer node, memory macrocells (such as SRAM and eDRAM) have become a critical component in terms of area and power consumption within a System-on-Chip (SoC). In the early design stages, especially during architecture exploration and physical planning, designers need to quickly and accurately assess the impact of different memory structures, configurations, and layout schemes on the overall power-to-appearance (PPA) of the chip. However, since detailed physical layout information is not yet generated at this stage, traditional EDA tools primarily rely on experience-based statistical models or time-consuming circuit-level simulations for estimation, resulting in limited prediction accuracy, low computational efficiency, and weak generalization capabilities.

[0003] In existing technologies, there are several methods that use machine learning for power consumption or area prediction. For example, some tools use linear regression or simple neural network models to model cell libraries for specific processes. However, these methods typically have the following limitations:

[0004] The model is simple and has limited expressive power: it is difficult to fully capture the complex nonlinear relationship between the complex circuit structure of memory (such as sensitive amplifiers, decoders, bit cell arrays) and multi-dimensional design parameters (such as bit width, depth, process deviation, voltage).

[0005] The feature engineering is crude: there is a lack of systematic and unified encoding methods for the structural features unique to memory macrocells, resulting in incomplete or redundant model input information.

[0006] Lack of adaptability and generalization ability: The model is usually trained on data from a specific process node or design team. When faced with new processes, new design specifications or different application scenarios (such as high-performance computing and low-power IoT), the prediction accuracy drops significantly, requiring the collection of data and retraining of the model, which is costly.

[0007] Disconnected from the design flow: Predictive models often exist as standalone tools and fail to be seamlessly integrated into MemoryCompiler or the front-end physical design flow, making it impossible to achieve real-time, interactive design space exploration.

[0008] Therefore, there is an urgent need for a high-precision, high-efficiency, highly generalizable and easily integrated memory macrocell PPA prediction solution to support rapid decision-making and optimization in advanced chip design. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-model fusion memory macrocell design quality prediction system and method. It can utilize the advantages of multiple AI models to achieve high-precision and rapid prediction of PPA indicators, and has continuous learning and cross-platform generalization capabilities, which can effectively assist in optimization decisions in the early stages of design.

[0010] The present invention is implemented as follows: In a first aspect, the present invention provides a multi-model fusion memory macrocell design quality prediction system, comprising:

[0011] The feature extraction and encoding module is used to extract the design parameters of memory macrocells and perform vectorized encoding.

[0012] The multi-model fusion prediction module is used to input vectorized and encoded design parameters into multiple different artificial intelligence sub-models in parallel, and output multiple design quality prediction results respectively; the design quality prediction results output by the multiple different artificial intelligence sub-models are fused by an ensemble learning strategy to obtain a comprehensive design quality prediction result.

[0013] The continuous learning and adaptive update module is used to dynamically update the parameters of multiple artificial intelligence sub-models based on the comprehensive design quality prediction results and the actual comprehensive design quality results obtained through subsequent simulation or tape-out.

[0014] Preferably, the design parameters include at least one of the following: process node, bit width, depth, memory cell structure, and auxiliary circuit topology.

[0015] Preferably, the ensemble learning strategy is stacking or voting.

[0016] Preferably, the multiple design quality prediction results include three indicators: power consumption, performance, and area.

[0017] Preferably, it also includes a prediction interface module that supports standard application programming interface calls and can be integrated into a memory compiler or EDA design flow.

[0018] Preferably, the multiple different artificial intelligence sub-models include, but are not limited to, Gradient Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), and Recurrent Neural Network (RNN).

[0019] Secondly, the present invention provides a method for predicting the design quality of memory macrocells, comprising the following steps:

[0020] Used to extract design parameters of memory macrocells and perform vectorized encoding;

[0021] The vectorized design parameters are input in parallel to multiple different artificial intelligence sub-models, which output multiple design quality prediction results respectively. An ensemble learning strategy is used to fuse the design quality prediction results output by the multiple different artificial intelligence sub-models to obtain a comprehensive design quality prediction result.

[0022] The parameters of the artificial intelligence sub-model are dynamically updated based on the comprehensive design quality prediction results and the actual comprehensive design quality results obtained through subsequent simulation or tape-out.

[0023] Preferably, the multiple different artificial intelligence sub-models include, but are not limited to, Gradient Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), and Recurrent Neural Network (RNN).

[0024] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed, implements the method described thereon.

[0025] Fourthly, the present invention provides an electronic device including a processor and a memory, the memory storing a computer program, wherein the processor executes the program to implement the method described thereon.

[0026] The beneficial effects of this invention are at least as follows:

[0027] This invention employs a multi-model fusion prediction module, integrating various heterogeneous artificial intelligence sub-models such as GBDT, MLP, and RNN, and using ensemble learning strategies such as stacking or voting to fuse the results. It captures the complex nonlinear relationships and interaction effects between design parameters and PPA indicators from multiple dimensions, thereby overcoming the limitation of the expressive power of a single model and obtaining more accurate and stable comprehensive prediction results than any single model.

[0028] This invention utilizes a feature extraction and encoding module to perform structured and vectorized unified encoding of heterogeneous design parameters from multiple sources, such as process nodes and circuit topologies, providing the model with a high-quality, transferable input representation. A continuous learning and adaptive update module constructs a closed-loop feedback mechanism, enabling dynamic updates of model parameters using real data from subsequent simulations or tape-outs. This combination allows the system to effectively adapt to varying data distributions across different process nodes, design specifications, and application scenarios, reducing the model's over-reliance on specific training data and enhancing its cross-platform and cross-task generalization capabilities. Attached Figure Description

[0029] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the memory macrocell design quality prediction system provided in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the principle of the multi-model fusion prediction module provided in this embodiment of the invention.

[0032] Figure 3 This is a flowchart of the memory macrocell design quality prediction method provided in an embodiment of the present invention.

[0033] Figure 4 This is a comparison of prediction error curves for different models.

[0034] Definitions of abbreviations and key terms:

[0035] Memory: storage device;

[0036] PPA: Power, Performance, Area;

[0037] Macro: Macro unit, usually referring to memory modules such as SRAM and ROM;

[0038] GBDT: Gradient Boosting Decision Tree;

[0039] Ensemble Learning;

[0040] Netlist / Layout: Netlist / physical layout data. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0042] See appendix Figure 1 This embodiment provides a multi-model fusion memory macrocell design quality prediction system, deployed on a cloud computing platform or local server cluster, including:

[0043] The feature extraction and encoding module receives multi-dimensional raw design parameters of memory macrocells in JSON or specific script formats from users or design tools, and transforms them into standardized, machine-readable feature vectors. These raw design parameters include, but are not limited to: process technology node, memory width, depth, bitcell type (e.g., 6T, 8T), peripheral auxiliary circuit topology (e.g., decoder architecture, sensitive amplifier type, pre-charge scheme), operating voltage, temperature conditions, and target frequency. This module employs structured encoding, performing one-hot encoding on categorical variables (e.g., process node, cell type), standardizing and scaling continuous variables (e.g., bit width, voltage), and constructing combined features based on domain knowledge to form a unified high-dimensional feature vector, providing high-quality input for subsequent AI models. The encoded feature vector is sent to the multi-model fusion prediction module via an internal high-speed bus. All data is obtained through legal and compliant channels.

[0044] For the multi-model fusion prediction module, see [link / reference] Figure 2 include:

[0045] Heterogeneous Sub-model Pool: This pool integrates multiple AI prediction sub-models with diverse structures. Leveraging their different inductive biases, each sub-model learns different patterns from the feature data output by the feature extraction and encoding modules. Each sub-model outputs a corresponding preliminary PPA prediction result. These AI sub-models have been pre-trained using historical project data, simulation data, or publicly available benchmark datasets. Specifically, typical AI sub-models include:

[0046] Gradient Boosting Decision Tree (GBDT) excels at handling structured features and capturing complex interactions and decision boundaries between features.

[0047] Multilayer perceptrons (MLPs) have powerful general function approximation capabilities and are suitable for modeling high-dimensional nonlinear mappings.

[0048] Recurrent Neural Networks (RNNs) are suitable for processing features with temporal or topological dependencies (such as partial representations of layout connectivity features).

[0049] Model selection and routing unit: Based on user-specified design optimization objectives (e.g., "high performance priority," "ultra-low power priority," or "high density priority"), this unit automatically selects or adjusts the combination of sub-models and their weights involved in prediction. Different optimization objectives may correspond to different training data distributions and feature importance; this unit achieves adaptive prediction strategies.

[0050] Fusion Decision Unit: Employing advanced ensemble learning strategies, such as stacking or weighted voting, the initial PPA predictions from the selected sub-models are fused. Stacking typically uses a meta-learner (such as linear regression or a shallow neural network) to learn how to optimally combine the outputs of each sub-model to obtain a final composite prediction, which is usually more accurate and stable (more robust) than any single sub-model.

[0051] Continuous Learning and Adaptive Update Module: This system constructs a closed-loop learning mechanism. After the design is completed based on the prediction results and real PPA data is obtained through subsequent simulation or tape-out, this data can be fed back to this system. This module is responsible for: managing the feedback data pool and performing data cleaning and verification; periodically or trigger-based incremental training or fine-tuning of the sub-model using new data to adapt the model to process drift, design style changes, etc.; monitoring model performance, and if the prediction error continues to increase (possibly due to concept drift caused by changes in feature distribution), triggering an alarm or automatically initiating the model retraining process to ensure the long-term prediction accuracy of the system. The closed-loop feedback and dynamic model update mechanism implemented in this module ensures that the system can continuously absorb new real data as the project iterates, automatically adjusting the model to track process drift and design trends, thereby maintaining and optimizing long-term prediction performance and reducing long-term maintenance costs.

[0052] Lightweight Prediction Service and Integration Interface Module: This system is encapsulated as a deployable service, providing a standard application programming interface (API). This interface is designed to be lightweight and responsive, allowing: the Memory Compiler to call this service in real-time to obtain PPA estimates when generating macrocells with different configurations, assisting the compiler's internal optimization algorithms. During the floorplanning phase, EDA physical design tools can quickly assess the impact of different macrocell types, sizes, and placements on the overall chip PPA, enabling data-driven placement exploration. Designers can perform "What-if" analysis through a graphical user interface (GUI) or scripts, i.e., scanning one or more design parameters (such as changing bit width or voltage) to instantly obtain PPA change curves, thereby determining the optimal design point. This module embeds the system into the memory compiler or EDA design flow, enabling designers to obtain millisecond-level PPA predictions even when physical information is incomplete in the early stages of design. This replaces traditional time-consuming circuit-level simulation, realizing data-driven real-time design space exploration, thereby accelerating the chip architecture optimization and physical planning process. Meanwhile, the modular system architecture and standard application programming interface (API) design constitute the technical features that facilitate integration, enabling the system to be flexibly deployed in a service-oriented manner and seamlessly integrated into various EDA toolchains or enterprise internal design platforms without disrupting existing workflows, thereby improving the level of design automation and lowering the barrier to entry.

[0053] This embodiment also provides a method for predicting the design quality of memory macrocells, including the following steps:

[0054] Used to extract design parameters of memory macrocells and perform vectorized encoding;

[0055] The vectorized design parameters are input in parallel to multiple different artificial intelligence sub-models, which output multiple design quality prediction results respectively. An ensemble learning strategy is used to fuse the design quality prediction results output by the multiple different artificial intelligence sub-models to obtain a comprehensive design quality prediction result.

[0056] The parameters of the artificial intelligence sub-model are dynamically updated based on the comprehensive design quality prediction results and the actual comprehensive design quality results obtained through subsequent simulation or tape-out.

[0057] Specifically, refer to Figure 3 When the system receives a prediction request, the process is as follows:

[0058] Step S201: Parse the input and obtain the Memory macrocell parameters.

[0059] Step S202: The feature extraction and encoding module performs the encoding operation.

[0060] Step S203: Based on the optimization target marker attached to the request, the model selection unit selects the combination of sub-models that performs best on low-power data from the model pool (e.g., it may assign higher weights to RNN models).

[0061] Step S204: Input the feature vectors in parallel into the selected GBDT, MLP and RNN sub-models.

[0062] Step S205: Each sub-model performs inference independently and outputs preliminary predictions for the three indicators of power consumption, performance (such as access time), and area.

[0063] Step S206: The fusion decision unit (in this example, a pre-trained Stacking meta-learner) receives the three preliminary prediction vectors and outputs the final, more accurate integrated PPA prediction vector.

[0064] Step S207: Format the final predicted value and return it via API.

[0065] Step S208 (Asynchronous): After the design generates real PPA data and provides feedback, the continuous learning module adds the data pair to the training set and schedules the model to update and train when the system is under low load.

[0066] Reference Figure 4 By utilizing a dataset of thousands of SRAM macrocells with different configurations across multiple process nodes obtained through legal and compliant means, tests show that the prediction error (MSE of 0.062) of the ensemble model used in this invention is significantly lower than that of the best-performing single sub-models RNN (error 0.075), GBDT (error 0.081), and MLP (error 0.093). This verifies the effectiveness of the multi-model fusion strategy in improving prediction accuracy.

[0067] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-model fusion memory macrocell design quality prediction system, characterized in that, include: The feature extraction and encoding module is used to extract the design parameters of memory macrocells and perform vectorized encoding. The design parameters include categorical variables and continuous variables, specifically including at least one of the following: process node, bit width, depth, memory cell structure, and auxiliary circuit topology; The encoding adopts a structured encoding method, which encodes categorical variables one-hot, standardizes and scales continuous variables, and constructs combined features based on domain knowledge to form a unified high-dimensional feature vector. The multi-model fusion prediction module includes a heterogeneous sub-model pool, a model selection and routing unit, and a fusion decision unit. The heterogeneous sub-model pool includes gradient boosting decision trees, multilayer perceptrons, and recurrent neural networks. The model selection and routing unit selects combinations of sub-models and their weights to participate in the prediction based on the user-specified design optimization objective. The fusion decision unit uses an ensemble learning strategy to fuse the design quality prediction results output by the selected sub-models to obtain a comprehensive design quality prediction result. The continuous learning and adaptive update module is used to complete the design based on the comprehensive design quality prediction results and obtain the real comprehensive design quality results through subsequent simulation or tape-out. The results are then added to the feedback data pool as feedback data for data cleaning and verification. New data is used periodically or triggered to incrementally train or fine-tune multiple artificial intelligence sub-models and dynamically update the sub-model parameters. At the same time, the module monitors the model performance and triggers alarms or automatically starts the model retraining process when the prediction error continues to increase.

2. The system according to claim 1, characterized in that, The ensemble learning strategy is either a stacking method or a voting method.

3. The system according to claim 1, characterized in that, The multiple design quality prediction results include three indicators: power consumption, performance, and area.

4. The system according to claim 1, characterized in that, It also includes a prediction interface module that supports standard application programming interface calls and can be integrated into memory compilers or EDA design flows.

5. A method for predicting the design quality of memory macrocells, characterized in that, Includes the following steps: Used to extract design parameters of memory macrocells and perform vectorized encoding; the design parameters include categorical variables and continuous variables, specifically including at least one of the following: process node, bit width, depth, memory cell structure, and auxiliary circuit topology; The encoding adopts a structured encoding method, which encodes categorical variables one-hot, standardizes and scales continuous variables, and constructs combined features based on domain knowledge to form a unified high-dimensional feature vector. The vectorized design parameters are input in parallel to multiple different artificial intelligence sub-models in the heterogeneous sub-model pool. The model selection and routing unit selects the combination of sub-models and their weights to participate in the prediction according to the design optimization goal specified by the user. The fusion decision unit uses an ensemble learning strategy to fuse the design quality prediction results output by the selected sub-models to obtain the comprehensive design quality prediction result. The design is completed based on the comprehensive design quality prediction results, and the actual comprehensive design quality results are obtained through subsequent simulation or tape-out. These results are then added to the feedback data pool as feedback data for data cleaning and verification. New data is used periodically or triggered to incrementally train or fine-tune multiple artificial intelligence sub-models, dynamically updating the sub-model parameters. Simultaneously, model performance is monitored, and alarms are triggered or the model retraining process is automatically started when the prediction error continues to increase.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the method as described in claim 5.

7. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the program, it implements the method as described in claim 5.

Images