Generator allocation method for semiconductor layout, allocation training method, and operation device
The method automates generator allocation in semiconductor layout design by using a convolutional neural network to infer and assign generators for sub-cells within the semiconductor layout, addressing the productivity issues caused by complex designs and reducing the time and manpower required for allocation.
Patent Information
- Application Number
- PCT/KR2023/014782
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-04
- Filing Date
- 2023-09-26
- Publication Date
- 2024-11-07
AI Technical Summary
The increasing complexity of semiconductor circuit designs with multi-layer hierarchical structures and numerous sub-cells leads to a significant increase in time and manpower required for generator allocation, reducing productivity in semiconductor layout design.
A method involving an arithmetic unit that receives a semiconductor layout, divides it into channels, converts each channel into a matrix, and uses a trained convolutional neural network to infer and assign appropriate generators, displaying candidate generators with corresponding probabilities for each sub-cell, thereby automating the generator allocation process.
This approach significantly improves semiconductor layout design productivity by reducing the time and manpower needed for generator allocation, enabling efficient handling of complex layouts with numerous sub-cells and varying designs.
Smart Images

Figure KR2023014782_07112024_PF_FP_ABST
Abstract
Description
Generator allocation method, allocation learning method, and computational unit in semiconductor layout
[0001] The present disclosure generally relates to a generator allocation method, an allocation learning method, and a computational device in a semiconductor layout.
[0002] Analog circuits, digital circuits, or hybrid signal processing circuits that process both analog and digital signals are ultimately implemented using semiconductor devices, which are then laid out by designers. Traditionally, designers would draw the elements of a circuit, then modify the layout according to design rules to complete the circuit.
[0003] The process of designing semiconductor circuits, once performed manually by designers, is gradually shifting toward generator-based semiconductor circuit layout design methods. Generator-based layout design technology receives parameters from designers and automatically generates a layout that corresponds to the input parameters and adheres to design rules.
[0004] While hand-designed layouts by designers can only produce layouts that correspond to a single fixed parameter, generator technology can generate multiple variable sizes, features, and designs in a single operation through coding.
[0005] This generator-based layout design is effectively used in circuits that require high-speed operation, circuits that require low power consumption, and semiconductor layout designs that require high performance for high-speed processing of large amounts of data.
[0006] When a designer converts a semiconductor layout into a generator, the designer needs to assign a generator to the layout or its subcells that can be generated by the generator.
[0007] However, as semiconductor circuit technology has developed rapidly, semiconductors have become multi-layered hierarchical structures, and a single top-level cell can contain countless sub-cells of lower layers. Therefore, in the process of assigning a corresponding generator to each of these sub-cells, the complex names of the sub-cells must be remembered, and at the same time, the names of the generator libraries, class names, function argument names, etc. must be specified. As the types of sub-cells increase and the types of previously designed generator libraries increase, the time required to assign generators increases exponentially, resulting in a lot of time and manpower being consumed, which in turn lowers productivity.
[0008] One of the challenges this embodiment seeks to address is to overcome these difficulties in prior art. One of the challenges this embodiment seeks to address is to improve productivity by reducing the time and manpower required for semiconductor layout design.
[0009] The present embodiment is a generator allocation method of a semiconductor layout performed by an operation device, the allocation method including: an input step in which the operation device receives the layout, a division step in which the layout is divided into a plurality of channels, a conversion step in which each of the divided channels is converted into a matrix, and an inference step in which a generator capable of generating the layout is inferred from the matrix, wherein the inference step is performed by displaying one or more generator candidates corresponding to the input layout.
[0010] In one aspect of the present embodiment, the method further comprises an assignment step for assigning a generator to the layout after the inference step.
[0011] In one aspect of the present embodiment, the input layout is a layout including a plurality of sub-cells included in a semiconductor circuit, and the inference step is performed by displaying generator candidates for each layout of the plurality of input sub-cells.
[0012] In one aspect of the present embodiment, the layout received by the computing device is a layout of a plurality of user-specified sub-cells included in a semiconductor circuit.
[0013] In one aspect of the present embodiment, the dividing step is performed by dividing each of the physical layers included in the semiconductor circuit into corresponding channels, and the same physical layer included in the semiconductor circuit is performed by dividing it into the same channel.
[0014] In one aspect of the present embodiment, the division step is performed by dividing physical layers performing the same function in the semiconductor circuit into the same channel. In one aspect of the present embodiment, the inference step is performed using a learned convolutional neural network including a feature extraction layer and a fully connected layer.
[0015] In one aspect of the present embodiment, the inference step is performed using a learned convolutional neural network that further includes a channel reduction layer.
[0016] In one aspect of the present embodiment, the assignment step is performed by further indicating the input layout and the probability that the one or more generator candidates correspond.
[0017] Another embodiment is a method for learning a computational unit for assigning a generator to a semiconductor layout, the method for assigning learning includes: a preprocessing step of receiving the layout as input and dividing the layout into a plurality of channels, a conversion step of converting each of the divided channels into a matrix, and a learning step of training a neural network with the converted channels, wherein the preprocessing step is performed by inputting at least one layout from among at least one layout manually designed by a designer and at least one layout formed by inputting parameters to the generator into the neural network.
[0018] In one aspect, the preprocessing step is performed by dividing each of the physical layers included in the semiconductor circuit into corresponding channels, and the same physical layers included in the semiconductor circuit are divided into the same channels.
[0019] In one aspect, in the above preprocessing step, the physical layers performing the same function in the semiconductor circuit are divided into the same channel and performed.
[0020] In one aspect, the neural network is a convolutional neural network, wherein the convolutional neural network includes one or more convolutional layers, one or more pooling layers, and one or more fully connected layers.
[0021] Another embodiment is a computing device for allocating a generator to a semiconductor layout, the computing device including: at least one processor; and a memory storing one or more programs executed by the processor, wherein when the programs are executed by the one or more processors, a method for allocating a generator to a semiconductor layout is performed on the one or more processors, the method for allocating a generator to the semiconductor layout including: an input step for receiving the semiconductor layout, a division step for dividing the semiconductor layout into a plurality of channels, a conversion step for converting each of the divided channels into a matrix, and an inference step for inferring a generator to be allocated to the layout from the matrix, wherein the inference step is performed by displaying one or more generator candidates corresponding to the input layout.
[0022] In one aspect, the inference step is performed by further indicating the input layout and the corresponding probabilities for the one or more generator candidates.
[0023] In one aspect, the layout input to the terminal is a layout of a plurality of sub-cells included in a semiconductor circuit, and the inference step is performed by displaying generator candidates for each of the layouts of the plurality of input sub-cells.
[0024] In one aspect, the layout received by the terminal is a layout of a plurality of user-specified sub-cells included in a semiconductor circuit.
[0025] In one aspect, the division step is performed by dividing each of the physical layers included in the semiconductor circuit into corresponding channels, and the same physical layers included in the semiconductor circuit are divided into the same channels.
[0026] In one aspect, in the above division step, the physical layers performing the same function in the semiconductor circuit are divided into the same channel and performed.
[0027] In one aspect, the inference step is performed using a trained convolutional neural network including one or more convolutional operation layers, one or more pooling layers, and one or more fully connected layers.
[0028] In one aspect, the computing device further performs a generator allocation learning method of the semiconductor layout before performing the generator allocation method of the semiconductor layout, wherein the allocation learning method includes: a preprocessing step of receiving the layout as input and dividing the layout into a plurality of channels, and a conversion step of converting each of the divided channels into a matrix, wherein the input step is performed by inputting at least one layout from among at least one layout manually designed by a designer and at least one layout formed by inputting parameters to the generator into the neural network.
[0029] According to this embodiment, the advantage of improving the productivity of layout design of semiconductor circuits is provided.
[0030] FIG. 1 is a flowchart illustrating an outline of a generator allocation method of a semiconductor layout according to the present embodiment.
[0031] Fig. 2 is a flowchart illustrating an outline of a generator allocation learning method of a semiconductor layout according to the present embodiment.
[0032] FIG. 3 is a block diagram illustrating an outline of a semiconductor layout generator allocation learning method and a computational device that performs the semiconductor layout generator allocation method according to the present embodiment.
[0033] Figure 4 is a schematic plan view illustrating the layout of a planar MOS transistor.
[0034] FIG. 5 is a drawing illustrating that the MOS transistor layout illustrated in FIG. 4 is divided into each active channel (AL), poly channel (PL), metal channel (ML), and contact channel (CL), and is displayed in a matrix form with multiple rows and columns regularly arranged.
[0035] Figure 6 is a diagram illustrating the structure of a neural network used in the step of training or inferring a neural network with multiple divided layers.
[0036] Figure 7 is an exemplary layout of an inverter.
[0037] Figure 8 is a diagram illustrating a case where an inverter layout is input and an appropriate generator is inferred and assigned to each subcell of the inverter layout.
[0038] FIG. 9 is a diagram illustrating the layout of an exemplary RX0 circuit designed with the CMOS1 process used in the layout-generator experiment.
[0039] Figure 10 is a layout of a high-speed I / O transceiver circuit designed directly by the designer using a CMOS2 process different from the CMOS1 process.
[0040] Figures 11(a) and 11(b) are examples of resistor units designed by inputting different design parameters into the same generator to have different aspect ratios.
[0041] Hereinafter, the present embodiment will be described with reference to the attached drawings. Fig. 1 is a flowchart illustrating an outline of a method for allocating generators of a semiconductor layout according to the present embodiment. Referring to Fig. 1, the method for allocating generators of a semiconductor layout according to the present embodiment includes: an input step (S100) in which a calculation device (1, see Fig. 3) receives a layout; a division step (S200) in which the layout is divided into a plurality of channels; a conversion step (S300) in which each of the divided channels is converted into a matrix; and an inference step (S400) in which a generator to be allocated to the layout is inferred from the matrix, wherein the inference step is performed by displaying one or more generator candidates corresponding to the input layout.
[0042] FIG. 2 is a flowchart illustrating an outline of a generator allocation learning method for a semiconductor layout according to the present embodiment. Referring to FIG. 2, the generator allocation learning method for a semiconductor layout includes: a preprocessing step (S10) of receiving a layout as input and dividing the layout into a plurality of channels, a conversion step (S20) of converting each of the divided channels into a matrix, and a learning step (S30) of training a neural network with the converted channels, wherein the preprocessing step is performed by inputting at least one layout among at least one layout manually designed by a designer and at least one layout formed by inputting parameters to a generator into the neural network.
[0043] FIG. 3 is a block diagram illustrating an overview of a semiconductor layout generator allocation learning method and a computing device (1) that performs the semiconductor layout generator allocation method according to the present embodiment. Referring to FIG. 3, the computing device (1) according to the present embodiment may include an input unit (21), an output unit (22), a processor (25), a memory (24), and a data storage (23). The computing device (1) of FIG. 3 is according to one embodiment, and not all blocks illustrated in FIG. 3 are essential components, and some blocks included in the computing device (1) may be added, changed, or deleted in other embodiments. Meanwhile, the computing device (1) may be implemented as a computing device that performs the semiconductor layout generator allocation learning method and the semiconductor layout generator allocation method, and each component included in the computing device (1) may be implemented as a separate software device, or may be implemented as a separate hardware device combined with software.
[0044] The computing device (1) performs an input step (S100) of receiving a layout, a division step (S200) of dividing the layout into a plurality of channels, a conversion step (S300) of converting each of the divided channels into a matrix, and an inference step (S400) of inferring a generator to be allocated to the layout from the matrix, and the inference step (S400) performs a generator allocation method of a semiconductor layout, which is performed to display one or more generator candidates corresponding to the input layout.
[0045] In addition, the computing device (1) includes a preprocessing step (S10) of receiving the layout and dividing the layout into a plurality of channels, a conversion step (S20) of converting each of the divided channels into a matrix, and a learning step (S30) of training a neural network with the converted matrices, wherein the preprocessing step performs a generator allocation learning method of a semiconductor layout by inputting at least one layout among at least one layout manually designed by a designer and at least one layout formed by inputting parameters to the generator into the neural network.
[0046] The input unit (21) refers to a means for receiving a layout of a semiconductor in order to train a neural network (200) or to assign generators to subcells. In addition, the input unit (21) may be linked with a processor (25) to receive a layout designed by a user or a layout generated by a generator, or may be linked with an external device to directly acquire data and transmit it to the processor (25). The input unit (21) may be a device or server for inputting or receiving log information (log), various condition information, or control signals, but is not necessarily limited thereto.
[0047] The output unit (22) can display the input layout, generated generator, and information thereof in conjunction with the processor (25). It is preferable that the output unit (22) display various information through a display (not shown), speaker, etc. provided in the computing device (1) to output predetermined information, but is not necessarily limited thereto.
[0048] The processor (25) executes at least one command or program stored in the memory (24). The processor (25) according to the present embodiment calculates data for performing each step based on data obtained from the input unit (21) or data storage (23).
[0049] The memory (24) includes at least one instruction or program executable by the processor (25). The memory (24) can fetch and store an instruction or program for performing processing stored in the data storage (23). The memory (24) can store associated values such as results performed at each stage and intermediate values.
[0050] Data storage (23) refers to a general data structure implemented in the storage space (hard disk or memory) of a computer system using a data storage management program (DBMS). Data storage (23) can freely perform data search (extraction), deletion, editing, addition, etc. In one embodiment, data storage (23) can store compiled instructions so that the generator allocation method according to the present embodiment is executed and driven by the processor (25). Memory (24) can fetch the corresponding instructions at the request of the processor (25).
[0051] The data storage (23) according to the present embodiment can receive and store the generator allocation results and user constraints of sub-cells in the layout provided through the input unit (21), and can provide the stored data as needed. Meanwhile, the data storage (23) is described as being implemented within the computing device (1), but is not necessarily limited thereto, and may be implemented as a separate data storage device.
[0052] Hereinafter, a generator allocation learning method of a semiconductor layout according to the present embodiment will be described with reference to FIG. 4. FIG. 4 is a plan view schematically illustrating a layout (100) of a planar MOS transistor. Referring to FIGS. 2 to 4, the MOS transistor (100) illustrated in FIG. 4 includes elements such as a channel, a gate, a drain, and a source. The layout (100) of the MOS transistor can be expressed as an overlapping of physical layers corresponding to each element, namely an active physical layer (AL), a poly physical layer (PL), a metal physical layer (ML), and a contact physical layer (CL).
[0053] In the preprocessing step (S10), the input layout is divided into multiple channels (S10). In addition to MOS transistors, various unexplained components and circuits may include numerous unexplained elements, and these may be implemented in different unexplained physical layers, in addition to the described active physical layer (AL) and poly physical layer (PL).
[0054] Fig. 5 is a diagram illustrating a state in which the layout of the MOS transistor illustrated in Fig. 4 is divided into multiple channels. Referring to Fig. 5, multiple physical layers included in the layout and displayed overlappingly are each divided into different channels. The divided channels are converted into a matrix form as illustrated (S20).
[0055] For example, recently, a technology has emerged that forms the gate of a transistor with metal. If the polysilicon physical layer and the metal gate physical layer are classified as different channels, as in this embodiment, the two transistor layouts that can be generated with the same generator can be trained to assign different generators to each other. However, since both the polysilicon gate and the metal gate perform the same function as the gate in a transistor, the physical layers that perform the same function are represented as the same channel, and neural network training is performed. Therefore, it is possible to train to determine whether layouts that include different physical layers that perform the same function can be generated with the same generator.
[0056] The active physical layer (AL), poly physical layer (PL), metal physical layer (ML) and contact physical layer (CL) illustrated in FIG. 4 are divided and displayed as an active channel (AC), a poly channel (PC), a metal channel (MC) and a contact channel (CC), as illustrated in FIG. 5, and each of these channels is displayed in the form of a matrix in which a plurality of rows and columns are regularly arranged. In addition, in the channels corresponding to the physical layers included in each element, an appropriate number may be assigned to the matrix element corresponding to the position of the pattern drawn on the physical layer.
[0057] In one embodiment, the provided layout relates to a single transistor, and may be designed by the designer who designed the transistor taking into account the maximum current that the transistor is intended to conduct and the maximum operating voltage (enduring voltage).
[0058] In another embodiment, the layout may be a layout of a device formed by inputting device parameters into a generator, for example, the provided layout may be a layout of a transistor formed by inputting design parameters of the transistor, such as a channel width (W), a channel length (L), and the number of contacts, into the generator. In another example, the provided layout may be a layout formed by generating a random combination of values of the types of parameters required for each generator using a random parameter generator.
[0059] In this way, a neural network can be trained using layouts formed by inputting parameters to a generator and layouts designed by a designer. In one embodiment, the formed layout is preprocessed and converted to train a neural network (S30).
[0060] Figure 6 is a diagram illustrating the structure of a neural network used in the step of training (S30) or inferring (S400) a neural network with multiple divided layers. Referring to Figures 2 to 6, the neural network may include a channel reduction layer (210), a feature extraction layer (220), and a fully connected layer (230).
[0061] An embodiment of the illustrated neural network (200) exemplifies one channel reduction layer (210) that reduces the number of channels by performing convolution operations with one or more 1x1 kernels, a feature extraction layer (220) including four convolution operation layers (CO1, CO2, CO3, CO4) and three pooling layers (PO1, PO2, PO3) respectively positioned between the three convolution operation layers, and two fully connected layers (230).
[0062] The first convolutional layer uses 1x1 kernels to reduce the number of channels and model size. The four convolutional layers (CO1, CO2, CO3, and CO4) all perform convolution operations with 3x3 kernels, and the three pooling layers (PO1, PO2, and PO3) perform operations with 2x2 kernels to extract input features. As the learning process progresses, the weight values of each convolutional kernel are continuously updated.
[0063] Hereinafter, a method of assigning a generator to a layout using a neural network (200) trained through the above process is described. Fig. 7 is an exemplary layout of an inverter (inverter, 300), and Fig. 8 is a diagram illustrating a case in which an inverter layout is input and an appropriate generator is inferred and assigned to each subcell of the inverter layout. Referring to Figs. 1 to 8, an arithmetic unit (1) receives an inverter layout through an input unit (21). The input layout may include a plurality of subcells. In the illustrated embodiment, the inverter (300) includes a plurality of MOS transistors, which are subcells, and a plurality of vias.
[0064] The illustrated embodiment illustrates providing the layout of a single inverter (300) to the input (21) of the computing device (1), but in another embodiment, the user may select a portion of the inverter (300) and provide only the selected portion to the input (21) to convert the sub-cells included in the selected portion into generators.
[0065] The computational device (1) divides the input layout into multiple channels (S200). As described in the learning process, the multiple elements included in the layout can each be divided into multiple physical layers, and the divided physical layers can be divided into multiple channels.
[0066] The segmented channels are converted into a matrix format (S300) and provided to a neural network (200). In one embodiment, the trained neural network (200) reduces the number of channels from the input and extracts features. Generator candidates that can be assigned to the input layout from the extracted features can be displayed as output units (22).
[0067] Fig. 8 is a diagram illustrating a state in which generators are allocated to subcells of an input inverter layout. In the embodiment illustrated in Fig. 8, generators are allocated and output to PMOS transistors (PMOS), NMOS transistors (NMOS), vias between metal 1 and metal 2 (Met12_via), vias between poly and metal 1 (PolyMet1_via), first supply rail (Nbody_Supply), and second supply rail (Pbody_Supply), which are subcells of the layout of an input inverter (300).
[0068] In one embodiment, the computational device (1) can display the probability of corresponding to the generator candidates that are allocated among the subcells included in the input layout. For example, for any one subcell included in the input layout, the probability of corresponding to the generator candidates that can be allocated, such as (NMOS transistor generator, 89%), (PMOS transistor generator, 7%), and (via generator, 4%), can be displayed together. The designer can review the probability of corresponding to the presented generator candidates and assign the layout of the subcells included in the layout to the generator.
[0069] In another embodiment, the computational device (1) can assign a subcell included in the input layout to a generator having the highest matching probability among the generator candidates.
[0070]
[0071] Experimental example
[0072] Hereinafter, the experimental results of the present embodiment will be examined with reference to FIGS. 9, 10, and 11. FIG. 9 illustrates the layout of an exemplary RX0 circuit used in the layout-generator experiment. The layout of the RX0 circuit illustrated in FIG. 9 was designed directly by the designer using CMOS process 1. The generators to which multiple sub-cells used in the RX0 circuit will be allocated were stored in the generator library.
[0073] Table 1 below is an example of the accuracy of allocation.
[0074]
[0075] The allocation accuracy of this example was evaluated using a total of 800 subcells, including 799 subcells and one subcell to which a generator could not be assigned. An error occurred in assigning a generator to only one subcell, and generators were correctly assigned to the remaining 799 subcells, including one unassignable subcell.
[0076] Figure 10 illustrates the layout of a high-speed I / O transceiver circuit designed directly by the designer using a CMOS2 process, which is different from the CMOS1 process. The neural network used in this embodiment was trained using the layout designed using the CMOS1 process. The circuit illustrated in Figure 10 comprises a total of 5,214 subcells, including 284 unassignable subcells. Table 2 below illustrates the accuracy of the allocation of the circuit illustrated in Figure 10.
[0077]
[0078] This embodiment proposed accurate generator assignments to 4,947 sub-cells (positive true). Furthermore, this embodiment correctly identified 261 sub-cells for which generator assignment was not possible due to components such as duty cycle correction (DCC) circuits, pre-amplifiers, switches, and delay units.
[0079] Fig. 11(a) and Fig. 11(b) are examples of resistor units designed by inputting different design parameters to the same generator so that they have different aspect ratios. As exemplified in Fig. 11(a) and Fig. 11(b), resistor units with extreme aspect ratios were not included in the training data set, and therefore, the present embodiment was not trained on layouts with such extreme aspect ratios. However, the present embodiment recognized the resistor unit layouts exemplified in Fig. 11(a) and Fig. 11(b) as the same type and accurately assigned a generator capable of generating them. In other words, the present embodiment was able to correctly recognize layouts that were not included in the training data set on which learning was performed and assign an appropriate generator to them.
[0080]
[0081] While the present invention has been described with reference to the embodiments illustrated in the drawings to aid understanding, these are merely exemplary embodiments for practical purposes. Those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible. Therefore, the true technical protection scope of the present invention should be defined by the appended claims.
Claims
1. A generator allocation method of a semiconductor layout performed by an operation unit, wherein the allocation method is: An input stage in which the calculation device receives the above layout, A splitting step for splitting the above layout into multiple channels, A conversion step for converting each of the above divided channels into a matrix, A deduction step for inferring a generator capable of generating the above layout from the above matrix, An allocation method in which the above inference step is performed by displaying one or more generator candidates corresponding to the input layout.
2. In paragraph 1, After the above inference step An allocation method further comprising an allocation step of allocating a generator to the above layout.
3. In paragraph 1, The above input layout is, A layout including multiple sub-cells included in a semiconductor circuit, The above inference step is, An allocation method performed by displaying generator candidates for each layout of the input plurality of sub-cells.
4. In paragraph 1, The layout that the above-mentioned computing device receives as input is, An allocation method that is a layout of multiple user-specified sub-cells included in a semiconductor circuit.
5. In paragraph 1, The above division step is, It is performed by dividing each physical layer included in the semiconductor circuit into corresponding channels, An allocation method in which identical physical layers included in a semiconductor circuit are divided into identical channels.
6. In paragraph 1, The above division step is, An allocation method in which physical layers performing the same function in the above semiconductor circuit are divided into the same channel.
7. In paragraph 1, The above inference step is, An allocation method performed using a learned convolutional neural network including a feature extraction layer and a fully connected layer.
8. In paragraph 7, The above inference step is, An allocation method performed using a learned convolutional neural network that further includes channel reduction layers.
9. In paragraph 2, The above allocation step is, An allocation method performed by further indicating the probability that the input layout and one or more generator candidates correspond.
10. A method for learning an operation unit that allocates a generator to a semiconductor layout, wherein the allocation learning method is: A preprocessing step of receiving the above layout and dividing the layout into multiple channels, A conversion step for converting each of the above divided channels into a matrix, and It includes a learning step that trains the neural network with the transformed channels, The above preprocessing step is, An assignment learning method performed by inputting one or more layouts manually designed by a designer and one or more layouts formed by inputting parameters to the generator into the neural network.
11. In paragraph 10, The above preprocessing step is, Each of the physical layers included in the above semiconductor circuit is divided into corresponding channels and performed, An allocation learning method in which the same physical layer included in the above semiconductor circuit is divided into the same channel.
12. In paragraph 10, In the above preprocessing step, An allocation learning method in which physical layers performing the same function in the above semiconductor circuit are divided into the same channel.
13. In paragraph 10, The above neural network, As a convolutional neural network, the convolutional neural network is One or more convolutional layers, One or more pooling layers and An assignment learning method that includes one or more fully connected layers.
14. An arithmetic device for allocating a generator to a semiconductor layout, said arithmetic device comprising: at least one processor; and A memory storing one or more programs executed by the processor, wherein when the programs are executed by the one or more processors, a generator allocation method of a semiconductor layout is performed in the one or more processors, wherein the generator allocation method is: An input step for receiving the above semiconductor layout, A division step for dividing the above semiconductor layout into multiple channels, A conversion step for converting each of the above divided channels into a matrix, A reasoning step for inferring a generator to be assigned to the layout from the above matrix, A computational device in which the above inference step is performed by displaying one or more generator candidates corresponding to the input layout.
15. In paragraph 14, The above inference step is, A computational device that further displays the input layout and the corresponding probabilities of the one or more generator candidates.
16. In paragraph 14, The above layout that the above terminal receives is, It is a layout of multiple sub-cells included in a semiconductor circuit, The above inference step is, A computational device that performs a calculation by displaying generator candidates for each layout of the input plurality of sub-cells.
17. In paragraph 14, The above layout that the above terminal receives is, A computational device that is a layout of multiple user-specified sub-cells included in a semiconductor circuit.
18. In paragraph 14, The above division step is, It is performed by dividing each physical layer included in the semiconductor circuit into corresponding channels, A computational device that performs operations by dividing the same physical layer included in a semiconductor circuit into identical channels.
19. In paragraph 14, In the above division step, An arithmetic device that performs the same function in the above semiconductor circuit by dividing it into identical channels.
20. In paragraph 14, The above inference step is, One or more convolutional operation layers, One or more pooling layers and A computational unit that performs computation using a learned convolutional neural network that includes fully connected layers.
21. In paragraph 14, The above computing device, Before performing the generator allocation method of the semiconductor layout, the generator allocation learning method of the semiconductor layout is further performed, wherein the allocation learning method is: A preprocessing step of receiving the above layout and dividing the layout into multiple channels, A conversion step for converting each of the above divided channels into a matrix, The above input steps are: A computational device that inputs one or more layouts manually designed by a designer and one or more layouts formed by inputting parameters to the generator into the neural network and performs the computation.
Citation Information
Patent Citations
Integrated circuit layout design method and device and readable storage medium
CN114662443A
Circuit simulation method, circuit simulation device, and program
JP2009087169A
Apparatus For Classifying Substance Using Photon Counting CT Based on Nearest Neighbor Algorithm And Method Thereof
KR102731710B1
Layout design system and layout design method
US20200184137A1
Physical verification workflow for semiconductor circuit designs
WO2022066952A1
