Method and apparatus for characterizing a standard integrated cell library using artificial intelligence
A machine learning-based approach efficiently generates and analyzes standard cell libraries by transforming and mixing PVT features and channels, addressing the complexity of PVT analysis and enhancing the accuracy and efficiency of cell library characterization.
Patent Information
- Application Number
- KR1020250108502
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2045-08-06
AI Technical Summary
Existing methods for characterizing standard integrated cell libraries face challenges in efficiently handling the complexity of Process-Voltage-Temperature (PVT) analysis, leading to increased computational costs and engineering effort, especially at advanced semiconductor nodes, which affects the accuracy and efficiency of cell library characterization.
A machine learning-based data generation model is employed to generate prediction data by mixing features and channels based on target PVT and a reference library, using multiple mixing layers and dimension transformation, followed by formatting the data into a library format, enabling efficient generation and analysis of cell libraries compatible with EDA tools.
The method achieves accurate and efficient characterization of standard cell libraries, reducing runtime and optimizing design space while ensuring compatibility with EDA tools, thereby improving the balance between accuracy and efficiency in PVT analysis.
Smart Images

Figure R1020250108502_ABST
Abstract
Description
Technology Field
[0001] Below, a description of the AI-based standard integrated cell library characterization methodology is provided. Background Technology
[0003] A standard cell can refer to a cell-unit block containing logic gates such as INVERTER, NAND, and NOR, as well as storage elements such as flip-flops or latches. A standard cell library refers to a collection of these standard cells and may include the cell's overall layout, circuit information, timing, and power information. Such standard cell libraries are used to efficiently implement complex digital circuits and, for example, can be used in the design of Application-Specific Integrated Circuits (ASICs).
[0004] The background technology described above is possessed or acquired by the inventor in the process of deriving the content of the disclosure of the present application, and cannot necessarily be considered as prior art disclosed to the general public prior to the filing of this application. Prior art literature
[0006] Korean Published Patent Application No. 10-2013-0035578 (published April 9, 2013) presents a method for generating a standard cell library for a DPL process and a method for generating a DPL mask using the same. Korean Registered Patent Application No. 10-2049103 (published November 20, 2019) presents a standard cell within a standard cell library and a variation thereof.
[0007] A method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment includes the steps of generating prediction data based on a target process-voltage-temperature (PVT) and a reference library using a machine learning-based data generation model, and generating a cell library by formatting the generated prediction data into a library format. The step of generating the prediction data may include the step of mixing features and channels based on the target PVT and the reference library using a plurality of mixing layers of the data generation model, and the step of generating the prediction data by adding the reference library to the mixed features and channels.
[0008] The step of generating the prediction data may further include, prior to mixing the features and channels, a step of converting the features and channels of the target PVT and the reference library into a predetermined dimension using a dimension projection block and a dimension transformation unit of the data generation model, and a step of aligning the dimensionally transformed features of the target PVT and the reference library for each using a first feature mixing layer and a second feature mixing layer of the data generation model.
[0009] The step of converting to the predetermined dimension above may include the step of generating transformed reference library features and channels in which the reference library is converted to the predetermined dimension using the dimension projection block, and generating transformed PVT features and channels in which the target PVT is converted to the predetermined dimension using the dimension transformation unit.
[0010] The above-described dimension projection block can generate the transformed reference library features and channels by inputting the matrix of the reference library, where the X-axis is a feature and the Y-axis is a channel, into the dimension projection layer, and then transposing the output matrix of the reference library.
[0011] The above dimension transformation unit can generate the transformed PVT features and channels by applying an iterative function that copies the strings of the target PVT, which is a matrix where the X-axis is a feature and the Y-axis is a channel.
[0012] The step of aligning the above features for each of them may include aligning the features of the reference library with respect to the features of the PVT using the first feature mixing layer, and aligning the features of the PVT with respect to the features of the reference library using the second feature mixing layer.
[0013] The step of mixing the above features and channels may include the step of generating a feature-channel combination matrix by combining the aligned target PVT and the features and channels of the reference library, the step of mixing the channels of the feature-channel combination matrix through the channel mixing area of the mixing layer, and the step of mixing the features of the feature-channel combination matrix through the feature mixing area of the mixing layer.
[0014] The above channel mixing region can generate a feature-channel mixing matrix in which channels are mixed by transposing the feature-channel combination matrix, which is a matrix where the X-axis is a feature and the Y-axis is a channel, and inputting it into a channel mixing module, and then transposing the output feature-channel combination matrix again.
[0015] The above feature mixing area can generate the mixed features and channels by inputting the above feature-channel mixing matrix into the feature mixing module.
[0016] The above method may further include the step of analyzing the characteristics of the cell library generated using an EDA (Electronic Design Automation) tool and the step of inputting the characteristics of the cell library into a parameter generation algorithm to generate a parameter set applied to the data generation model.
[0017] An apparatus for characterizing a standard integrated cell library using artificial intelligence according to one embodiment includes a processor that generates prediction data based on a target process-voltage-temperature (PVT) and a stored reference library using a machine learning-based data generation model, and generates a cell library by formatting the generated prediction data into a library format. In generating the prediction data, the processor mixes features and channels based on the target PVT and the reference library using a plurality of mixing layers of the data generation model, and generates the prediction data by adding the reference library to the mixed features and channels.
[0018] Before combining the features and channels, the processor can convert the features and channels of the target PVT and the reference library into a predetermined dimension using the dimension projection block and dimension transformation unit of the data generation model, and can align the dimensionally transformed features of the target PVT and the reference library for each using the first feature mixing layer and the second feature mixing layer of the data generation model.
[0019] The processor can, in converting to the predetermined dimension, generate transformed reference library features and channels in which the reference library is converted to the predetermined dimension using the dimension projection block, and generate transformed PVT features and channels in which the target PVT is converted to the predetermined dimension using the dimension transformation unit.
[0020] The above-described dimension projection block can generate the transformed reference library features and channels by inputting the matrix of the reference library, where the X-axis is a feature and the Y-axis is a channel, into the dimension projection layer, and then transposing the output matrix of the reference library.
[0021] The above dimension transformation unit can generate the transformed PVT features and channels by applying an iterative function that copies the strings of the target PVT, which is a matrix where the X-axis is a feature and the Y-axis is a channel.
[0022] The processor can align the features of the reference library with respect to the features of the PVT using the first feature mixing layer and align the features of the PVT with respect to the features of the reference library using the second feature mixing layer when aligning the features for each of the above features.
[0023] The processor can, in mixing the features and channels, combine the aligned target PVT and the features and channels of the reference library to generate a feature-channel combination matrix, mix the channels of the feature-channel combination matrix through the channel mixing area of the mixing layer, and mix the features of the feature-channel combination matrix through the feature mixing area of the mixing layer.
[0024] The above channel mixing region can generate a feature-channel mixing matrix in which channels are mixed by transposing the feature-channel combination matrix, which is a matrix where the X-axis is a feature and the Y-axis is a channel, and inputting it into a channel mixing module, and then transposing the output feature-channel combination matrix again.
[0025] The above feature mixing area can generate the mixed features and channels by inputting the channel mixing matrix into the feature mixing module.
[0026] The above processor analyzes the characteristics of the cell library generated using an EDA (Electronic Design Automation) tool and inputs the characteristics of the cell library into a parameter generation algorithm to generate a set of parameters applied to the data generation model. Brief explanation of the drawing
[0028] FIG. 1 illustrates a block diagram of a device for characterizing a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 2 exemplarily illustrates the inference process of a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 3 exemplarily illustrates a dimensional projection block within a data generation model of a device characterizing a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 4 exemplarily illustrates a mixing layer within a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 5 exemplarily illustrates the training process of a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 6 illustrates a flowchart of a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment. FIG. 7 shows a scatter plot of the inference results of a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment and comparative embodiments. FIG. 8 shows a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment and Pareto points and Pareto fronts for the inference results of comparative embodiments. Specific details for implementing the invention
[0029] Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, actual implementations are not limited to the specific embodiments disclosed, and the scope of this specification includes modifications, equivalents, or substitutions included in the technical concept described by the embodiments.
[0030] Terms such as "first" or "second" may be used to describe various components, but these terms should be interpreted solely for the purpose of distinguishing one component from another. For example, the first component may be named the second component, and similarly, the second component may be named the first component.
[0031] When it is stated that a component is "connected" to another component, it should be understood that it may be directly connected to or coupled with that other component, or that there may be other components in between.
[0032] The singular expression includes the plural expression unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to specify the existence of the described features, numbers, steps, actions, components, parts, or combinations thereof, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0033] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification.
[0034] Hereinafter, embodiments will be described in detail with reference to the attached drawings. In the description with reference to the attached drawings, identical components are given the same reference numeral regardless of the drawing number, and redundant descriptions thereof will be omitted.
[0035] Cell libraries can contain important information about standard cells, including electrical characteristics such as rise and fall delays, power consumption, constraints, and other specifications. Characterizing such cell libraries requires extensive simulation of various behavioral patterns and design parameters, and this process becomes increasingly complex at advanced nodes.
[0036] In particular, since Process-Voltage-Temperature (PVT) characterization is necessary to accurately model behavior under various conditions, cell libraries require characterization of multiple PVT corners. PVT refers to essential data for generating timing and power models used for static timing analysis and power estimation. As semiconductor technology advances to smaller nodes, the importance of analyzing such PVT is growing as sensitivity to change increases and design margins shrink.
[0037] PVT analysis is a critical process for evaluating overall circuit performance across various manufacturing process variations, supply voltage fluctuations, and operating temperature ranges. In integrated circuit design, PVT analysis enables the assurance of robust chip performance and reliability under diverse conditions. From an optimization perspective, PVT analysis can be used to optimize yield by identifying performance boundaries, evaluating trade-offs between PPA metrics, and ensuring the design meets specifications across various scenarios. However, previous design space optimization (DSO) methods, which involved characterizing cell libraries, faced challenges in terms of increasing computational costs and engineering effort due to the growing complexity of PVT analysis and the complexity of various scenarios.
[0038] Accordingly, the apparatus and method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment efficiently generate various types of standard cell libraries including data on timing and power using a data generation model, and since the generated standard cell library can be applied to EDA tools, the results can be analyzed and verified without difficulty, and a balance between accuracy and efficiency can be achieved in a characterization process that is difficult to balance due to the complexity of PVT analysis.
[0039] FIG. 1 illustrates a block diagram of a device for characterizing a standard integrated cell library using artificial intelligence according to one embodiment.
[0040] FIG. 2 exemplarily illustrates the inference process of a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment.
[0041] The device (100) may include a processor (110) and a memory (120).
[0042] Referring to FIG. 2, in the Design Space Optimization stage, the processor (110) can generate predicted data (e.g., predicted values) based on the target Process-Voltage-Temperature (PVT) and a stored reference library (e.g., Ref lib) using a machine learning-based data generation model (210) (e.g., LIBMixer), and can generate a cell library (e.g., Inferred Library) by formatting the generated predicted data into a library file format (e.g., .lib) (e.g., using a formatter) (220). The training of the data generation model (210) is described later in FIG. 5. In generating prediction data, the processor (110) can generate prediction data (e.g., predicted values) by using a plurality of mixing layers (215) of the data generation model (210) to mix features and channels based on the target PVT and the reference library, and by adding the reference library (e.g., Ref lib) to the mixed features and channels. The processor (110) can generate a more accurate cell library with a short runtime using the machine learning-based data generation model (210) and the formatter (220), expand the scope of the cell library by including complex standard cells with all functions in the generated cell library, and build the generated cell library into a library file compatible with EDA tools.
[0043] The processor (110) can analyze the characteristics of a generated cell library (e.g., inferred library) using an electronic design automation (EDA) tool. The processor (110) can ensure design quality by modeling and simulating the operation of a circuit designed based on the cell library through the EDA tool. The processor (110) can generate a set of parameters applied to a data generation model (210) by inputting the characteristics of the cell library into a parameter generation model (240) (e.g., optimizer). The processor (110) can derive improved design results by applying the generated set of parameters to the data generation model (210).
[0044] In generating prediction data, the processor (110) converts the features and channels of the target PVT and reference library to a predetermined dimension using the dimension projection block (212) and dimension transformation unit (211) of the data generation model (210) before mixing the features and channels, and can align the dimensionally transformed features of the target PVT and reference library to each other using the first feature mixing layer (213) and the second feature mixing layer (214) of the data generation model (210).
[0045] In converting to a predetermined dimension, the processor (110) can use a dimension projection block (212) to generate transformed reference library features and channels in which the reference library is converted to a predetermined dimension, and use a dimension conversion unit (211) to generate transformed PVT features and channels in which the target PVT is converted to a predetermined dimension. The features may represent, for example, behavioral scenarios that distinguish what kind of information (e.g., characteristics) the data contains, and the channels may represent, for example, functionality variations that distinguish what environment a component operates in.
[0046] The dimension transformation unit (211) can generate transformed PVT features and channels by applying a repeat function that copies strings of a target PVT matrix in which the X-axis is a feature and the Y-axis is a channel. The dimension transformation unit (211) can, for example, apply the repeat function to strings of a target PVT having one row and five columns. The transformed PVT features and channels to which the repeat function is applied may have three rows and five columns by copying the rows of the target PVT strings. This may mean that a (1X5) matrix composed of one channel and five features is copied by applying the repeat function to generate a (3X5) matrix composed of three channels and five features.
[0047] FIG. 3 exemplarily illustrates a dimensional projection block within a data generation model of a device characterizing a standard integrated cell library using artificial intelligence according to one embodiment.
[0048] The dimensional projection block (212) can generate transformed reference library features and channels (350) by transposing (320) the matrix of a reference library (310) in which the X-axis is a feature and the Y-axis is a channel, inputting it into a dimensional projection layer (330) which is a fully connected layer, and then transposing (340) the output matrix of the reference library again. The processor (110) can train cell library data patterns on the dimensional projection layer (330) during training. The processor (110) can map the input channel size to the hidden channel size for the channel domain using the trained dimensional projection layer (330).
[0049] In aligning features for each, the processor (110) may align features of a reference library with respect to features of a PVT using a first feature mixing layer (213) and align features of a PVT with respect to features of a reference library using a second feature mixing layer (214). The processor (110) may ensure dimensional consistency through the first feature mixing layer (213) and the second feature mixing layer (214).
[0050] FIG. 4 exemplarily illustrates a mixing layer within a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment.
[0051] The processor (110) can mix features and channels by combining the features and channels of the aligned target PVT and reference library to create a feature-channel combination matrix, mix the channels of the feature-channel combination matrix through the channel mixing area (410) of the mixing layer, and mix the features of the feature-channel combination matrix through the feature mixing area (420) of the mixing layer.
[0052] The channel mixing area (410) can generate a feature-channel mixing matrix in which channels are mixed by transposing (411) a feature-channel combination matrix, in which the X-axis is a feature and the Y-axis is a channel, and inputting it into the channel mixing module (412), and then transposing (413) the output output feature-channel combination matrix again. The feature mixing area (420) can generate mixed features and channels by inputting the feature-channel mixing matrix into the feature mixing module (421). The channel mixing area (410) can perform 2D Batch Normalization on the feature-channel combination matrix before transposing (411) the feature-channel combination matrix. Additionally, the feature mixing area (420) can perform 2D batch normalization on the feature-channel mixing matrix before inputting the feature-channel mixing matrix into the feature mixing module (421). The channel mixing area (410) and the feature mixing area (420) can preprocess the input data through normalization to make it easier to analyze by processing it into a form with a constant distribution according to predetermined rules.
[0053] The channel mixing module (412) and the feature mixing module (421) may be multi-layer perceptrons (MLPs). A multi-layer perceptron is a multi-layer structure classification module composed of multiple hidden layers, capable of solving complex non-linear problems. The channel mixing module (412) may include a fully connected (FC) layer, a GeLU layer, and a dropout layer. During training, the processor (110) may use the channel mixing module (412) to model dependencies between values in each library table and output values that reflect specification changes according to design conditions during inference. The feature mixing module (421) may include a fully connected layer, a GeLU layer, a dropout layer, a GeLU layer, and a dropout layer. The feature mixing module (421) may include dual GeLU layers and dropout layers to handle more complex feature relationships than the channel mixing module (412). The processor (110) can model the difference between feature dimensions using the feature mixing module (421) during training and output a value that reflects the internal characteristics of the cell during inference.
[0054] The channel mixing area (410) can perform a residual connection by connecting the feature-channel combination matrix input to the channel mixing area (410) to the feature-channel mixing matrix. The feature mixing area (420) can perform a residual connection by connecting the feature-channel mixing matrix input to the output of the feature mixing module (421). The processor (110) can increase the depth of the network through the residual connection while ignoring unnecessary mixing operations.
[0055] Since cell library data varies depending on the operation scenario (e.g., features) and function (e.g., channels), the processor (110) can process each piece of information individually through the channel mixing area (410) and the feature mixing area (420). Subsequently, the processor (110) can generate a composite matrix by iteratively aggregating the information obtained from each of the multiple mixing layers (215). The processor (110) can input the composite matrix into a mismatch layer (216), which is a fully connected layer, to generate a mismatch matrix (e.g., mixed features and channels) that reflects the mismatch between the reference values of the reference library and the composite matrix. The processor (110) can reflect the mismatch results while simultaneously matching the dimensions with the library matrix to be finally predicted through the mismatch layer (216). This means that the mismatch layer (216) can increase the size of the composite matrix to match the dimensions with the library matrix. Additionally, the mismatch layer (216) may be an MLP layer that can be trained to reflect the mismatch results. For example, if the size of the composite matrix is N by H and the size of the final prediction library is A by B, the processor (110) can transform the size of the composite matrix to A by B. If the values of N and A are the same, the processor can simply train the mismatch layer (216) based on a weight matrix of size H by B. If the values of N and A are different, a flatten layer can be added before the mismatch layer (216), and a dimension transformation layer can be added after the mismatch layer (216). Accordingly, the processor can flatten the N by H size of the composite matrix to a 1 by N*H size through the flatten layer. The processor can train the mismatch layer (216) for N*H by A*B mismatch based on 1 by N*H.The processor can convert the result of the mismatch layer (216), which is 1 by N*H in size, output through the dimension transformation layer, into an A by B size. The processor (110) can increase the dimension of the composite matrix while reducing the error through the mismatch layer (216).
[0056] The processor (110) can generate prediction data (e.g., predicted values) which is the final prediction by adding the reference libraries stored in the mismatch matrix (e.g., by resident linking). Through this, the processor (110) ensures that the output through the multiple mixing layers (215) matches the size of the input reference library table, while the precision of the data distribution can be improved through the final resident linking.
[0057] FIG. 5 exemplarily illustrates the training process of a data generation model of a device that characterizes a standard integrated cell library using artificial intelligence according to one embodiment.
[0058] The processor (110) can divide the foundry libraries into PVTs, target values, and reference libraries based on one corner selection. Additionally, the processor (110) can divide the reference libraries into reference values and a .lib format. The processor (110) can train an untrained data generation model (510) (e.g., LIBMixer) using the PVTs, target values, and reference values. The processor (110) can train by updating the weights of the data generation model (510) in a direction that minimizes the loss of Equation 1 based on the predicted values output by the data generation model (510). The processor (110) can terminate training after training for a predetermined number of epochs, or terminate training when the loss value does not change. Using the data generation model (510) trained in this way, the processor (110) can predict accurate library data.
[0059]
[0060] In mathematical formula 1, N represents the number of training samples in the batch, and x i wa y i can represent the i-th input pattern and the corresponding target library value, respectively. θ can represent the model parameter.
[0061] Additionally, the formatter (220) may be an algorithm that converts the input into a .lib format. The processor (110) may not only convert the predicted values of the previously trained data generation model (210) into a library file format through the formatter (220) configured to convert into a .lib format format, but may also concatenate the library files in a predetermined order.
[0062] According to various embodiments, the processor (110) may execute computer-readable code (e.g., software) stored in memory (120) and instructions triggered by the processor (110). The processor (110) may be a data processing device implemented in hardware having a circuit having a physical structure for executing desired operations. Desired operations may include, for example, code or instructions included in a program. The data processing device implemented in hardware may include, for example, a microprocessor, a central processing unit, a processor core, a multi-core processor, a multiprocessor, an Application-Specific Integrated Circuit (ASIC), or a Field Programmable Gate Array (FPGA).
[0063] The memory (120) can store at least one instruction (e.g., a program) that can be executed by the processor (110).
[0064] FIG. 6 illustrates a flowchart of a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment.
[0065] In step (610), the processor can generate prediction data based on the target process-voltage-temperature (PVT) and reference library using a machine learning-based data generation model.
[0066] In step (620), the processor can create a cell library by formatting the generated prediction data into a library format.
[0067] After step (620), the processor can analyze the characteristics of the generated cell library using an EDA (Electronic Design Automation) tool and input the characteristics of the cell library into a parameter generation algorithm to generate a set of parameters applied to a data generation model.
[0068] In step (620), to generate prediction data, the processor may generate prediction data by using a plurality of mixing layers of a data generation model to mix features and channels based on a target PVT and a reference library, and by adding the reference library to the mixed features and channels. In mixing the features and channels, the processor may generate a feature-channel combination matrix by combining the aligned features and channels of the target PVT and the reference library, mix the channels of the feature-channel combination matrix through the channel mixing area of the mixing layer, and mix the features of the feature-channel combination matrix through the feature mixing area of the mixing layer. The channel mixing area may generate a feature-channel mixing matrix in which the channel is mixed by transposing a feature-channel combination matrix, in which the X-axis is a feature and the Y-axis is a channel, inputting it into a channel mixing module, and then transposing the output feature-channel combination matrix again. The feature mixing area can generate mixed features and channels by inputting a feature-channel mixing matrix into the feature mixing module.
[0069] Before mixing features and channels, the processor can convert the features and channels of the target PVT and reference library into predetermined dimensions using the dimension projection block and dimension transformation unit of the data generation model, and can align the dimensionally transformed features of the target PVT and reference library, respectively, using the first feature mixing layer and the second feature mixing layer of the data generation model. In aligning the features, the processor can align the features of the reference library with respect to the features of the PVT using the first feature mixing layer, and align the features of the PVT with respect to the features of the reference library using the second feature mixing layer.
[0070] In transforming to a predetermined dimension, the processor can generate transformed reference library features and channels by using a dimension projection block to transform the reference library into the predetermined dimension, and generate transformed PVT features and channels by using a dimension transformation unit to transform the target PVT into the predetermined dimension. The dimension projection block can generate transformed reference library features and channels by inputting a matrix of the reference library, where the X-axis is a feature and the Y-axis is a channel, into a dimension projection layer, and then transposing the output matrix of the reference library. The dimension transformation unit can generate transformed PVT features and channels by applying an iterative function that copies the strings of the target PVT, where the X-axis is a feature and the Y-axis is a channel matrix.
[0071] Additionally, the processor can generate tables in the cell library by dividing them by dimension according to characteristics. Tables in the cell library may include zero-dimensional tables, one-dimensional tables, and two-dimensional tables. Zero-dimensional tables represent data expressed as a single value and may, for example, include data on leakage power determined by transistor characteristics. One-dimensional tables represent data whose value changes according to a single variable (e.g., input voltage) and may include data on passive power that varies according to pin transitions. Two-dimensional tables represent complex data whose value changes according to two variables (e.g., input signal change rate, load capacity, etc.) and may include data on delay, transition time, power, and timing constraints.
[0072] FIG. 7 shows a scatter plot of the inference results of a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment and comparative embodiments.
[0073] In FIG. 7, the x-axis represents the reference value, the foundry library (e.g., Golden value), and the y-axis represents the predicted value of each model. FIG. 7 shows the results for passive power, leakage power, timing, and power portions, which are data of various dimensions. Referring to FIG. 7, it can be seen that the first embodiment shows results close to the diagonal, indicating superior accuracy compared to Comparative Example 1 and Comparative Example 2 in the overall portion. In contrast, Comparative Example 1 shows values relatively close to the diagonal for low-dimensional leakage power, but shows results that deviate significantly from the diagonal for passive power, which has slightly more complex dimensions. In addition, Comparative Example 1 shows scattered results, indicating almost inaccurate results for high-dimensional timing and power. Comparative Example 2, a model slightly improved over Comparative Example 1, shows better results than Comparative Example 1 overall, but results that fall short of the first embodiment.
[0074] In addition, Tables 1 and 2 allow for a comparison of the EDA tool compatibility and performance between the exemplary embodiment and the comparative embodiments. Table 1 shows performance results for power (e.g., Dynamic Power and Leakage Power) and area, while Table 2 shows performance results for timing with worst-case slack (WS). In Tables 1 and 2, the robustness and reliability of library characterization can be confirmed as the exemplary embodiment (e.g., Ours) shows results nearly identical to the reference foundry library (e.g., Golden) across various benchmarks compared to the comparative embodiments (e.g., Ridge and Ridge+).
[0075]
[0076]
[0077] FIG. 8 shows a method for characterizing a standard integrated cell library using artificial intelligence according to one embodiment and Pareto points and Pareto fronts for the inference results of comparative embodiments.
[0078] In FIG. 8, a Pareto point represents the optimal solution of the cell library, and a Pareto front is a set of Pareto points, which is represented as a line. FIG. 8(a) shows a Pareto front analysis showing the trade-off between power and performance (pt) of an ibex core, and FIG. 8(b) shows a Pareto front analysis showing the trade-off between power and area (pa) of an ibex core. FIG. 8(c) shows a Pareto front analysis showing the trade-off between power and performance (pt) of a rocket core, and FIG. 8(d) shows a Pareto front analysis showing the trade-off between power and area (pa) of a rocket core. In FIG. 8, it can be seen that the reference value line (810) is nearly identical to the line (830) of one embodiment. Additionally, in FIG. 8(a) and FIG. 8(c), the x-axis indicates that the lower the value, the higher the performance level, and the y-axis indicates that the lower the value, the lower the power consumption. In FIG. 8(b) and FIG. 8(d), the x-axis indicates that the smaller the value, the more compact the design. Accordingly, it can be seen that the reference value line (810) and the line (830) of one embodiment are located at a lower value overall than the line (820) of the comparative embodiment. This may mean that the design of the cell library of one embodiment has been optimized and the runtime has been significantly reduced.
[0079] The embodiments described above may be implemented as hardware components, software components, and / or combinations of hardware and software components. For example, the devices, methods, and components described in the embodiments may be implemented using a general-purpose computer or a special-purpose computer, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.
[0080] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be stored on any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on computer-readable recording media.
[0081] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination, and the program instructions recorded on the medium may be those specifically designed and configured for the embodiment or those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.
[0082] The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.
[0083] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based thereon. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.
[0084] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.
Claims
Claim 1 A method for characterizing a standard integrated cell library using artificial intelligence comprises: a step of generating prediction data based on a target process-voltage-temperature (PVT) and a reference library using a machine learning-based data generation model; and a step of generating a cell library by formatting the generated prediction data into a library format, wherein the step of generating the prediction data comprises: a step of mixing features and channels based on the target PVT and the reference library using a plurality of mixing layers of the data generation model; and a step of generating the prediction data by adding the reference library to the mixed features and channels. Claim 2 A method according to claim 1, wherein the step of generating the prediction data further comprises: a step of converting the features and channels of the target PVT and the reference library to a predetermined dimension using a dimension projection block and a dimension transformation unit of the data generation model before mixing the features and channels; and a step of aligning the dimensionally transformed features of the target PVT and the reference library, respectively, using a first feature mixing layer and a second feature mixing layer of the data generation model. Claim 3 In paragraph 2, the step of converting to the previously determined dimension comprises the step of generating transformed reference library features and channels in which the reference library is converted to the previously determined dimension using the dimension projection block, and generating transformed PVT features and channels in which the target PVT is converted to the previously determined dimension using the dimension transformation unit. Claim 4 In paragraph 3, the method comprises: the dimension projection block transposing a matrix of the reference library, where the X-axis is a feature and the Y-axis is a channel, inputting it into a dimension projection layer, and then transposing the output matrix of the reference library again to generate the transformed reference library features and channels. Claim 5 In paragraph 3, the dimension transformation unit generates the transformed PVT features and channels by applying an iterative function that copies the strings of the target PVT, which is a matrix where the X-axis is a feature and the Y-axis is a channel. Claim 6 In claim 2, the step of aligning the features for each of the above features comprises the step of aligning the features of the reference library with respect to the features of the PVT using the first feature mixing layer, and aligning the features of the PVT with respect to the features of the reference library using the second feature mixing layer. Claim 7 In claim 2, the step of mixing the features and channels comprises: a step of generating a feature-channel combination matrix by combining the aligned target PVT and the features and channels of the reference library; a step of mixing the channels of the feature-channel combination matrix through a channel mixing region of the mixing layer; and a step of mixing the features of the feature-channel combination matrix through a feature mixing region of the mixing layer. Claim 8 In claim 7, the channel mixing region is a method of inputting a feature-channel combination matrix, in which the X-axis is a feature and the Y-axis is a channel, into a channel mixing module by transposing the feature-channel combination matrix, and then transposing the output combination feature-channel output matrix again to generate a feature-channel mixing matrix in which the channels are mixed. Claim 9 In claim 8, the feature mixing region is a method of generating the mixed features and channels by inputting the feature-channel mixing matrix into a feature mixing module. Claim 10 A method according to claim 1, further comprising: a step of analyzing the characteristics of the cell library generated using an EDA (Electronic Design Automation) tool; and a step of inputting the characteristics of the cell library into a parameter generation algorithm to generate a set of parameters applied to the data generation model. Claim 11 An apparatus for characterizing a standard integrated cell library using artificial intelligence comprises a processor that generates prediction data based on a target process-voltage-temperature (PVT) and a stored reference library using a machine learning-based data generation model, and generates a cell library by formatting the generated prediction data into a library format. In generating the prediction data, the processor mixes features and channels based on the target PVT and the reference library using a plurality of mixing layers of the data generation model, and generates the prediction data by adding the reference library to the mixed features and channels. Claim 12 In claim 11, the processor converts the features and channels of the target PVT and the reference library to a predetermined dimension using a dimension projection block and a dimension transformation unit of the data generation model before combining the features and the channels, and aligns the dimensionally transformed features of the target PVT and the reference library to each other using a first feature mixing layer and a second feature mixing layer of the data generation model. Claim 13 In claim 12, the processor, in converting to the predetermined dimension, uses the dimension projection block to generate converted reference library features and channels in which the reference library is converted to the predetermined dimension, and uses the dimension conversion unit to generate converted PVT features and channels in which the target PVT is converted to the predetermined dimension. Claim 14 In claim 13, the device comprises a dimension projection block that transposes a matrix of the reference library, where the X-axis is a feature and the Y-axis is a channel, inputs it into a dimension projection layer, and then transposes the output matrix of the reference library again to generate the transformed reference library features and channels. Claim 15 In claim 13, the dimension transformation unit is a device that generates the transformed PVT features and channels by applying an iterative function that copies the strings of the target PVT, which is a matrix where the X-axis is a feature and the Y-axis is a channel. Claim 16 In claim 12, the processor, in aligning the features for each of the above features, aligns the features of the reference library with respect to the features of the PVT using the first feature mixing layer, and aligns the features of the PVT with respect to the features of the reference library using the second feature mixing layer. Claim 17 In claim 12, the processor, in mixing the features and channels, combines the aligned target PVT and the features and channels of the reference library to generate a feature-channel combination matrix, mixes the channels of the feature-channel combination matrix through the channel mixing region of the mixing layer, and mixes the features of the feature-channel combination matrix through the feature mixing region of the mixing layer. Claim 18 In claim 17, the channel mixing region is a device that inputs the feature-channel combination matrix, which is a matrix where the X-axis is a feature and the Y-axis is a channel, into a channel mixing module by transposing it, and then generates a feature-channel mixing matrix in which the channels are mixed by transposing the output feature-channel combination matrix again. Claim 19 In claim 18, the feature mixing region is a device that inputs the feature-channel mixing matrix into a feature mixing module to generate the mixed features and channels. Claim 20 In claim 11, the processor analyzes the characteristics of the cell library generated using an EDA (Electronic Design Automation) tool and inputs the characteristics of the cell library into a parameter generation algorithm to generate a set of parameters applied to the data generation model.
Citation Information
Patent Citations
Integrated circuit design for optimized manufacturing performance
JP2007529100A
Method for generating a standard cell library for DPL process and a method for generating a DPL mask using the same
KR1020130035578A
Methods for automated standard cell design
KR1020220161478A
Circuit Timing Analysis Incorporating the Effects of Temperature Inversion
US20120210287A1
Physical layout synthesis for standard cells using slice layouts
US20240281584A1
