Method and system for using machine learning to predict cell characteristics in circuit design procedure

Abstract

A method for a circuit design process includes: obtaining information about a specific circuit element from a circuit element database; generating an initial dataset based on the information about the specific circuit element; establishing or optimizing a machine learning model based on the initial dataset; and generating one or more target characteristic data about the specific circuit element based on the established or optimized machine learning model.

Description

Technical Field

[0001] This invention relates to the field of circuit design, and more particularly to a method and system for predicting the characteristics of circuit components using machine learning methods in the circuit design process. Prior Technology

[0002] In the field of integrated circuit design, accurately predicting and controlling the leakage current of circuit cells is a key challenge. Leakage current not only directly affects the power consumption of integrated circuits but is also closely related to heat generation, thus impacting the performance, reliability, and lifespan of integrated circuits. However, there is a complex nonlinear relationship between leakage current and various operating conditions and environmental factors (such as temperature), making accurate prediction of leakage current under different operating conditions and environments particularly difficult. Current leakage current estimation methods mainly fall into two categories: circuit simulation software, such as the Simulation Program with Integrated Circuit Emphasis (SPICE) simulation method, and linear interpolation. Software simulation is relatively accurate, performing detailed circuit simulations by adjusting different setpoints. However, this method has significant drawbacks, such as high time cost: each different setpoint requires restarting the entire simulation program, which can significantly extend the design cycle in large-scale integrated circuit designs. On the other hand, linear interpolation attempts to quickly estimate leakage current by interpolating a finite number of sample points. Although this method is fast, its accuracy is low for the following reasons: 1) Ignoring nonlinear relationships: The relationship between leakage current and temperature is usually nonlinear, and linear interpolation cannot accurately capture this complex relationship; 2) Sample limitations: The number of available sample points is usually limited, which limits the accuracy of the interpolation method, especially in areas outside the sample points.

[0003] On the other hand, the limitations of cell libraries further exacerbate this problem. Supplier-provided cell libraries typically contain only a limited number of leakage current data points, which are often insufficient to cover all the operating conditions that integrated circuits might encounter in practical applications. This is especially true when performing thermal analysis, which requires accurate leakage current data at multiple temperature points, exceeding the capabilities of cell libraries.

[0004] This situation forces designers to make trade-offs between accuracy and efficiency. Comprehensive software simulations can yield highly accurate results, but at the cost of lengthy simulation times. Conversely, fast but less accurate linear interpolation methods can lead to significant errors in thermal analysis and power consumption estimation, thus impacting the overall quality and reliability of the circuit design. Therefore, there is an urgent need in the field for a new method that can accurately predict the leakage current of circuit components under different operating conditions without unduly increasing computational complexity. Summary of the Invention

[0005] In view of the challenges of predicting leakage current of circuit elements in integrated circuit design, this invention proposes an innovative method and system. The core of this invention is to establish a data analysis model (e.g., a machine learning model) that can accurately simulate the complex nonlinear relationship between leakage power consumption and a multidimensional parameter space (including operating condition parameters and state parameters). Furthermore, this invention not only considers operating factors such as the temperature of circuit elements but also incorporates the potential levels of all input pins of the circuit elements as state features, thus providing a more comprehensive prediction of leakage behavior. Moreover, this invention generates an initial dataset based on limited information provided by a limited database of circuit elements and utilizes model training and optimization to generate leakage power consumption predictions covering a wide range of operating conditions and state parameters based on relatively little known data. Through these features, this invention enables integrated circuit designers to perform more accurate power consumption assessment and optimization during the design phase.

[0006] An embodiment of the present invention provides a method for a circuit design process, the method comprising: obtaining information about a specific circuit element from a circuit element database; generating an initial dataset based on the information about the specific circuit element; establishing or optimizing a machine learning model based on the initial dataset; and generating one or more target characteristic data about the specific circuit element based on the established or optimized machine learning model.

[0007] Embodiments of the present invention provide a system for a circuit design process, the system comprising: a processor and a memory. The memory is used to store a plurality of instructions, which, when executed by the processor, cause the system to perform the following operations: retrieve information about a specific circuit element from a circuit element database; generate an initial dataset based on the information about the specific circuit element; establish or optimize a machine learning model based on the initial dataset; and generate one or more target characteristic data about the specific circuit element based on the established or optimized machine learning model. Simple Explanation of the Diagram

[0008] Figure 1 illustrates a flowchart of a method for a circuit design process according to an embodiment of the present invention. Figure 2 illustrates an architecture diagram of a system for a circuit design process according to an embodiment of the invention. Figure 3 illustrates the initial data set and extended data set of an embodiment of the present invention. Figure 4 illustrates a flowchart of a method for a circuit design process according to another embodiment of the present invention. Implementation

[0009] Numerous specific details are described below to provide the reader with a thorough understanding of embodiments of the invention. However, those skilled in the art will appreciate how the invention can be implemented in the absence of one or more specific details, or by utilizing other methods, elements, or materials. In other instances, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring the core concepts of the invention.

[0010] The phrase "an embodiment" in this specification means that a particular feature, state, or characteristic described in that embodiment may be included in at least one embodiment of the invention. Therefore, the phrase "in an embodiment" appearing throughout this specification does not necessarily mean the same embodiment. Furthermore, the aforementioned particular features, states, or characteristics may be combined in any suitable form in one or more embodiments.

[0011] Please refer to Figure 1, which illustrates a flowchart of a method for a circuit design process according to an embodiment of the present invention. In step S101, multiple characteristic data corresponding to multiple operating condition parameter values ​​and / or multiple state parameter values ​​of at least one state parameter of a target circuit element (cell) under at least one operating condition are obtained as an initial dataset. Specifically, step S101 obtains the characteristic data corresponding to the target circuit element under the multiple operating condition parameter values ​​of the at least one operating condition and / or the characteristic data corresponding to the multiple state parameter values ​​of the at least one state parameter of the target circuit element from a cell library to generate the initial dataset. More specifically, the at least one operating condition can be the operating temperature of the target circuit element, the at least one state parameter can be the potential of the input terminal (e.g., signal pins, leads, pads, etc.) of the target circuit element, the multiple operating condition parameter values ​​can be multiple specific operating temperature values, the multiple state parameter values ​​can be the potential magnitude of specific input terminals, and the characteristic data can be the leakage power corresponding to the multiple operating temperature values ​​and / or the potential values ​​corresponding to the multiple input terminal values. Furthermore, the circuit element database may contain pre-determined data and characteristics of various different circuit elements.

[0012] In step S102, one or more data processing operations are performed on the data pairs in the initial dataset. In some embodiments, these data processing operations include logarithmic and normalization processes. In some embodiments, these data processing operations may also include various types of nonlinear transformations. Note that step S102 is optional in this invention; in some embodiments, if the numerical distribution range of the characteristic data does not require processing or adjustment, then step S102 can be omitted.

[0013] In step S103, a machine learning model is established or optimized. More specifically, step S103 can establish the machine learning model based on the initial dataset processed by step S102, or based on the initial dataset not processed by step S102. On the other hand, if the machine learning model performs worse than expected in subsequent validation processes, the process can also return to step S103 to optimize the machine learning model. In some embodiments, the machine learning model can be a sequential model with a multi-layer structure. The multi-layer structure of the sequential model can include multiple dense layers, i.e., fully connected layers. Furthermore, the machine learning model is expected to capture the relationship between specific operating conditions and / or state parameters and characteristic data.

[0014] In step S104, within a predetermined computational resource limit, the machine learning model is trained or optimized to improve its fit to the processed initial dataset (or the unprocessed initial dataset). In some embodiments, the predetermined computational resource limit may correspond to a step limit or an iteration limit (e.g., 500 steps or 500 iterations). In some embodiments, the predetermined computational resource limit may correspond to a training time limit or an optimization time limit. Furthermore, the training or optimization of the machine learning model includes adjusting model parameters. For example, adjustments to parameters such as the number of units, kernel initializer, bias initializer, and activation function settings. Moreover, if the machine learning model uses the aforementioned sequential model structure, the initialization of the weight matrix of the first dense layer can use a Gaussian distribution to randomly initialize the weights (i.e., random_normal).

[0015] In step S105, the training or optimization process of the machine learning model is evaluated to determine whether a predetermined termination condition has been met. More specifically, step S105 checks whether a predetermined computational resource limit has been reached. In some embodiments, step S105 checks whether the training or optimization process has reached the step limit or the iteration limit. In some embodiments, step S105 checks whether the training or optimization process has reached the training time limit or the optimization time limit. If the evaluation result in step S105 is negative (i.e., the predetermined termination condition has not been met), the process proceeds to step S106. In step S106, the learning rate of the machine learning model is adjusted. In some embodiments, step S106 may adjust the learning rate in each step or each iteration according to an exponential decay method or other possible learning rate adjustment methods. Alternatively, the learning rate may be adjusted every few steps or every few iterations. By adjusting the learning rate, the training or optimization efficiency of the machine learning model can be improved.

[0016] If the evaluation result in step S105 is yes (i.e., the predetermined termination condition has been met), the process proceeds to step S107. In step S107, one or more target characteristic data are generated based on the established or optimized machine learning model. Further, if the characteristic data corresponds to the leakage power consumption of the target circuit element, step S107 outputs one or more leakage power consumption values ​​corresponding to multiple operating condition parameter values ​​and / or multiple state parameter values. In step S108, the consistency between the one or more target characteristic data generated by the machine learning model and the processed initial dataset (or the unprocessed initial dataset) is verified. In some embodiments, step S108 evaluates a loss (e.g., a loss function) between the one or more target characteristic data and the processed initial dataset, for example, calculating the loss based on mean squared error, root mean square error, or mean absolute error, thereby quantifying the goodness of fit between the one or more target characteristic data and the processed initial dataset to determine the reliability and accuracy of the machine learning model.

[0017] In step S109, it is determined whether the loss is less than a predetermined value. More specifically, the loss can be the percentage error between the generated target characteristic data and the initial dataset, or other error quantification metrics. Step S109 can determine whether the machine learning model meets the expected accuracy requirements. If the determination result of step S109 is no, the process returns to step S103 to optimize the machine learning model. If the determination result of step S109 is yes, the process proceeds to step S110. In step S110, the generated target characteristic data is output, further generating an extended characteristic dataset. More specifically, the extended characteristic dataset covers a wider range of operating condition parameter values ​​and / or state parameter values, and their relationship with the characteristic data, or a more granular range of operating condition parameter values ​​and / or state parameter intervals, and their relationship with the characteristic data. This extended characteristic dataset can be generated by integrating or merging the target characteristic data into the initial dataset, and is used to further update the circuit element database, thereby enhancing the comprehensiveness and accuracy of the existing data in the circuit element database. Furthermore, this extended feature set can also be applied in a circuit design process to accurately simulate the overall circuit characteristics.

[0018] Figure 2 illustrates an architecture diagram of a system for a circuit design flow according to an embodiment of the present invention. As shown, system 100 can be used to execute the method shown in Figure 1, thereby accurately predicting and generating characteristic data of individual circuit elements and simulating overall circuit characteristics in the circuit design flow. System 100 can be designed to focus on various hardware and software configuration combinations to meet the needs of training and optimizing machine learning models, as shown in Figure 2. System 100 typically includes at least one processor 110 and one or more memory devices 120. Alternatively, for specific machine learning tasks, system 100 can employ an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). In some embodiments, system 100 can utilize a graphics processing unit (GPU) 130 to handle parallel processing tasks common in machine learning workloads. The memory subsystem of system 100 may include various hierarchical structures, such as: fast static random access memory (SRAM) for immediate data access; high bandwidth memory (HBM) for large-scale high-speed data transfer; and large-capacity dynamic random access memory (DRAM) for storing large amounts of model parameters and intermediate results. In advanced configurations, system 100 may also incorporate storage device 140, such as a non-volatile memory high-speed (NVMe) solid-state drive for fast storage and retrieval of large model files and datasets. Storage device 140 may be used to store circuit element libraries, initial datasets, and extended datasets. During model building / training or optimization, the circuit element libraries, initial datasets, and extended datasets may be loaded into one or more memory devices 120 to accelerate the processing and use of this data and information by processor 110 or GPU 130.

[0019] Figure 3 illustrates the initial and extended datasets of an embodiment of the present invention. As shown, the initial dataset may include leakage current LP1, LP2, and LP3 of a specific circuit element at specific operating temperature values ​​K1, K2, and K3 (e.g., -40°C, 0°C, and 125°C), obtained from a circuit element database. The machine learning model established or optimized using the method of the present invention can further predict the leakage current LP4 to LPn of the specific circuit element at specific operating temperature values ​​K4 to Kn. The specific operating temperature values ​​K4 to Kn can be any temperature value within the temperature range defined by operating temperature values ​​K1, K2, and K3 (e.g., any temperature value between -40°C and 125°C), any temperature value outside the temperature range defined by operating temperature values ​​K1, K2, and K3 (e.g., any temperature value less than -40°C or greater than 125°C), or both. The leakage current LP4~LPn at specific operating temperatures K4~Kn can be merged or integrated with the initial dataset to generate an extended dataset. This extended dataset can then be used to update information and data in the circuit component database, and subsequently applied in the circuit design process.

[0020] Figure 4 illustrates a flowchart of a method for a circuit design process according to another embodiment of the present invention. The flowchart shown in Figure 4 can be considered a simplified version of the flowchart shown in Figure 1, illustrating the key steps for implementing the concept of the present invention. As shown, this simplified version includes the following flow:

[0021] Step S201: Obtain information about a specific circuit element from a circuit element database;

[0022] Step S202: Generate an initial data set based on the information of the specific circuit element;

[0023] Step S203: Establish or optimize a machine learning model based on the initial dataset; and

[0024] Step S204: Based on the established or optimized machine learning model, generate one or more target characteristic data about the specific circuit element.

[0025] Since the principles and details of the above steps have been explained in detail through the above embodiments, they will not be repeated here. It is worth noting that the above process can be improved by adding other additional steps or making appropriate modifications and adjustments to better enhance the reliability, accuracy, and efficiency of the simulation and analysis stages in the circuit design process.

[0026] In summary, the method and system provided by this invention can be applied to accurately predict and optimize the leakage power consumption of circuit elements during the integrated circuit design process. Specifically, the method of this invention cleverly combines high-precision nonlinear modeling techniques, enabling designers to quickly generate detailed leakage power consumption estimates covering a wide range of variables based on limited known data. This invention not only considers operating conditions such as the temperature of circuit elements but also incorporates the state parameters of the circuit elements, thus providing a more comprehensive and accurate description of leakage behavior. By applying this invention, integrated circuit designers can perform more accurate power consumption assessments and hotspot analyses in the early stages of design, significantly reducing the number of design iterations and significantly improving design efficiency. This not only accelerates the product development cycle but also helps improve the performance, reliability, and energy efficiency of the final product.

[0027] Embodiments of the present invention can be embodied as devices, methods, or computer program products. Accordingly, embodiments of the present invention can take the form of entities implemented entirely in hardware, entities implemented entirely in software (including firmware, resident software, microcode, etc.), or entities combining software and hardware aspects, which can be collectively referred to as "modules" or "systems." Furthermore, embodiments of the present invention can take the form of computer program products embodied in any tangible medium, and such medium has computer-usable code embodied therein. In terms of hardware, the present invention can be implemented by applying any of the following techniques or related combinations: individual operating logic of logic gates capable of executing logic functions according to data signals, and application-specific integrated circuits (ASICs), programmable gate arrays (PGAs), or field-programmable gate arrays (FPGAs) with appropriate combinational logic.

[0028] Flowcharts and block diagrams illustrate the architecture, functionality, and operation of systems, methods, and computer program products that represent different possible implementations of embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of program code, including one or more executable instructions to implement a specific logical function. It should also be noted that each block in a block diagram and / or flowchart, as well as combinations of blocks in block diagrams and / or flowcharts, can be implemented by a special-purpose hardware infrastructure system, or a combination of special hardware and computer program instructions. These computer program instructions may be stored in readable computer media to command a computer or other programmable data processing device to operate in a specific manner so that the instructions stored in the readable computer media produce the functions / operations specified in the blocks or combinations of blocks in the flowcharts and / or block diagrams. The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made in accordance with the claims of the present invention shall be covered by the present invention.

[0029] 100: System 110: Processor 120: Memory device 130: Graphics Processing Unit 140: Storage device S101~S110, S201~S204: Steps

Claims

1. A method for a circuit design flow, comprising: obtaining information about a specific circuit element (cell) from a cell library; generating an initial dataset based on the information about the specific circuit element; establishing or optimizing a machine learning model based on the initial dataset; generating one or more target characteristic data about the specific circuit element based on the established or optimized machine learning model; generating an extended characteristic dataset based on the one or more target characteristic data; updating the information about the specific circuit element in the cell library based on the extended characteristic dataset; and performing the circuit design flow using the updated cell library.

2. The method of claim 1, wherein the information of the particular circuit element includes a plurality of operating condition parameter values ​​of at least one operating condition of the particular circuit element and / or a plurality of state parameter values ​​of at least one state parameter of the particular circuit element, and characteristic data corresponding to the plurality of operating condition parameter values ​​and / or the plurality of state parameter values ​​respectively.

3. The method of claim 2, wherein the at least one operating condition corresponds to an operating temperature of the particular circuit element, the at least one state parameter corresponds to the potential of an input terminal of the particular circuit element, the plurality of operating condition parameter values ​​correspond to a plurality of operating temperature values ​​of the particular circuit element, the plurality of state parameter values ​​correspond to the potential values ​​of a plurality of input terminals of the particular circuit element, and the characteristic data corresponds to a plurality of leakage power.

4. The method as described in claim 1, wherein the step of establishing or optimizing the machine learning model based on the initial dataset comprises: performing one or more data processing operations on the initial dataset to obtain the processed initial dataset; and establishing or optimizing the machine learning model based on the processed initial dataset.

5. The method of claim 4, wherein the one or more data processing operations include at least one of logarithmic processing or a normalization process.

6. The method of claim 1, wherein the step of building or optimizing the machine learning model based on the initial dataset comprises: training or optimizing the machine learning model within a predetermined computing resource limit; evaluating whether a training process or an optimization process of the machine learning model has reached a predetermined termination condition related to the predetermined computing resource limit; and adjusting a learning rate of the machine learning model when the predetermined termination condition has not been reached.

7. The method as described in claim 6, wherein the predetermined computational resource limit corresponds to the number of steps, iterations, or time of the training or optimization process.

8. The method of claim 1, wherein the step of building or optimizing the machine learning model based on the initial dataset comprises: verifying the consistency between the one or more target feature data generated by the machine learning model and the initial dataset; determining whether a loss between the one or more target feature data and the initial dataset is less than a predetermined value; and optimizing the machine learning model when the loss is greater than the predetermined value.

9. A system for a circuit design flow, comprising: a processor; The system also includes a memory for storing a plurality of instructions that, when executed by the processor, cause the system to perform the following operations: retrieve information about a specific circuit element from a circuit element database; generate an initial dataset based on the information about the specific circuit element; build or optimize a machine learning model based on the initial dataset; and generate one or more target characteristic data about the specific circuit element based on the built or optimized machine learning model; generate an extended characteristic dataset based on the one or more target characteristic data; and update the information about the specific circuit element in the circuit element database based on the extended characteristic dataset. And the circuit design process is carried out using the updated circuit component database.

Images