Buffer memory and computing device and system using the buffer memory

By designing a buffer memory and a computing unit in a computing device, dividing the input data into multiple feature groups and storing and processing separately, the problem of inefficiency of deep neural network circuits in the prior art in matrix operations and convolution operations is solved. Achieve more efficient computing and storage utilization.

CN111767238BActive Publication Date: 2024-11-26SK HYNIX INC +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201911017681.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-04-01
Filing Date
2019-10-24
Publication Date
2024-11-26
Estimated Expiration
2039-10-24

AI Technical Summary

Technical Problem

When performing matrix operations, especially convolution operations, existing deep neural network (DNN) circuits have low storage and processing efficiency of data feature groups, resulting in low computing efficiency and low memory utilization rate, which cannot Make full use of the data bandwidth of the system bus.

Method used

A computing device is designed, including a buffer memory and a computing unit. By dividing the input data into multiple feature groups and storing them separately in multiple memories, and performing batch processing operations using a buffer circuit and a computing unit, the multiple is realized. Simultaneous processing and calculation of feature groups.

Benefits of technology

The computing efficiency of the computing device and the utilization rate of the memory are improved, the data bandwidth of the system bus is fully utilized, and the processing speed of matrix operations and convolution operations is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111767238B_ABST
    Figure CN111767238B_ABST
Patent Text Reader

Abstract

The present application discloses a computing device, which includes a buffer memory that provides first to nth input feature groups to a computing unit. The buffer memory includes first to nth memories, and is configured to divide the first to nth input feature groups and store the first to nth input feature groups in the first to nth memories, respectively, wherein each input feature group includes multiple features. Multiple features of an input feature group are divided and stored in the first to nth memories. Features having the same order in the first to nth input feature groups are stored one by one in the first to nth memories.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of Korean Application No. 10-2019-0037832 filed on April 1, 2019, in the Korean Intellectual Property Office, which is incorporated herein by reference in its entirety. Technical Field

[0003] Various embodiments generally relate to an integrated circuit technology, and more particularly, to a buffer memory and a computing device and system using the buffer memory. Background Art

[0004] As the fourth industrial revolution progresses, technologies related to the Internet of Things (IoT), mobile, cloud storage, big data analysis, and artificial intelligence (AI) continue to develop. In order to implement artificial intelligence, technologies such as machine learning and deep learning have been developed. Because both machine learning and deep learning calculate or classify received inputs to generate outputs, the two technologies are similar to each other. At the same time, machine learning can analyze and infer data through algorithms input by humans, while deep learning can analyze and infer data through deep neural networks.

[0005] A deep neural network (DNN) circuit may include a matrix operation circuit that performs a MAC (multiply-accumulate) operation on received inputs to generate an output. The matrix operation circuit may perform a convolution operation and may typically include a systolic array. The systolic array may perform a MAC operation on a set of input features such as activation inputs and weight data. Summary of the invention

[0006] In an embodiment, the computing device may include a buffer memory and a computing unit. The buffer memory may include first to nth memories, and is configured to store first to nth input feature groups each including a plurality of features in the first to nth memories, and to store a plurality of features of one input feature group separately in the first to nth memories. Features having the same order in the first to nth input feature groups may be stored one by one in the first to nth memories, and "n" may be an integer equal to or greater than 2. The computing unit may be configured to receive the first to nth input feature groups stored in the buffer memory, and perform a computing operation on the received input feature groups.

[0007] In an embodiment, the computing device may include a buffer memory and a computing unit. The buffer memory may include first to fourth memories, and is configured to store first to fourth input feature groups each having at least first to fourth features in the first to fourth memories, and to divide and store the first to fourth features of one input feature group in the first to fourth memories. Features having the same order in the first to fourth input feature groups may be divided and stored in the first to fourth memories so as not to overlap with each other. The computing unit may be configured to receive the first to fourth input feature groups stored in the buffer memory, and perform a computing operation on the received input feature groups.

[0008] In an embodiment, a system may include a host device and a computing device. The host device may be configured to transmit first input data via a system bus. The computing device may be configured to receive the first input data and perform a computing operation on the first input data. The computing device may include a buffer, a buffer memory, and a computing unit. The buffer may be configured to classify the first input data into first to nth input feature groups, and store the first to nth input feature groups, each of which has at least first to nth features, where n is an integer equal to or greater than 2. The buffer memory may include first to nth memories, and is configured to store the first to nth features of the first to nth input feature groups in the first to nth memories. The first to nth features of an input feature group may be stored separately in the first to nth memories. The computing unit may be configured to receive the first to nth input feature groups stored in the buffer memory, and perform a computing operation on the first to nth input feature groups.

[0009] In an embodiment, an operating method of a system may include classifying first input data transmitted from a host into first to nth input feature groups each having first to nth features to temporarily store the first to nth input feature groups, where n is an integer equal to or greater than 2. The operating method may include storing the first to nth features of the first to nth input feature groups in first to nth memories, wherein the first to nth features of one input feature group are separately stored in the first to nth memories, and features having the same order in the first to nth input feature groups are uniformly stored in the first to nth memories. The operating method may include simultaneously outputting features having the same order in the first to nth input feature groups. And the operating method may include performing a calculation operation on the output features of the first to nth input feature groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a diagram showing a configuration of a system according to an embodiment.

[0011] Figure 2A and Figure 2B 2 is a diagram illustrating a configuration and a calculation operation of a calculation circuit according to an embodiment.

[0012] Figure 3 is a diagram showing the flow of data in the system according to the present embodiment.

[0013] Figure 4 is a diagram illustrating a configuration of a buffer memory and a method for storing a plurality of input feature groups in the buffer memory according to an embodiment. DETAILED DESCRIPTION

[0014] The various embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. However, the present invention can be implemented in different forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present invention to those skilled in the art. Throughout the disclosure, the same reference numerals represent the same components throughout the various drawings and embodiments of the present invention.

[0015] Note that references to "an embodiment," "another embodiment," etc. are not necessarily to only one embodiment, and different references to any such phrases are not necessarily to the same embodiment.

[0016] It will be understood that although the terms "first", "second", "third", etc. may be used herein to describe various elements, these elements are not limited by these terms. These terms are used to distinguish one element from another element. Therefore, without departing from the spirit and scope of the present invention, the first element described below may also be referred to as the second element or the third element.

[0017] It will be further understood that when used in this specification, the terms "comprises," "comprising," "including," and "comprising" indicate the presence of stated elements without excluding the presence or addition of one or more other elements. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0018] As used herein, unless the context clearly indicates otherwise, the singular form may also include the plural form, and vice versa. Unless otherwise specified or clearly directed to the singular form from the context, the articles "a" and "an" used in this specification and the appended claims should generally be interpreted as meaning "one or more".

[0019] Hereinafter, a buffer memory, a computing device, and a system using the same according to the present disclosure will be described below with reference to the accompanying drawings through exemplary embodiments.

[0020] Figure 1 FIG. 1 shows a configuration of a system 1 according to an embodiment. Figure 1In the system 1, the system 1 may include a host device 101 and a computing device 102. The host device 101 may provide data to the computing device 102 and receive data output from the computing device 102. The host device 101 may be connected to the computing device 102 via a system bus 103. The host device 101 may transmit input data IND1 and IND2 to the computing device 102 via the system bus 103. The host device 101 may receive output data OD from the computing device 102 via the system bus 103. The output data OD may be generated by a computing operation performed by the computing device 102 on the input data IND1 and IND2. The host device 101 may provide various control signals to the computing device 102 via the system bus 103 for the computing operation of the computing device 102. The input data may include first input data IND1 and second input data IND2. For example, the first input data IND1 may include information related to a set of input features. The second input data IND2 may include information related to weight data.

[0021] The computing device 102 may be connected to the host device 101 via the system bus 103. The computing device 102 may receive input data IND1 and IND2 transmitted from the host device 101 via the system bus 103. The computing device 102 may perform a computing operation on the input data IND1 and IND2, and may transmit output data OD generated by the computing operation to the host device 101 via the system bus 103. The computing device 102 may be a neural network computing device, such as a DNN accelerator. The computing device 102 may receive first and second input data IND1 and IND2 from the host device 101, and perform a computing operation on a set of input features and weight data. The computing operation may be a convolution operation or a MAC (multiply-accumulate) operation on the set of input features and weight data.

[0022] The computing device 102 may include a host interface 110, a buffer circuit 120, a memory 130, and a computing unit 140. The host interface 110 may be connected to the system bus 103 to receive input data IND1 and IND2 transmitted from the host device 101, and provide the received input data IND1 and IND2 to the internal circuit of the computing device 102. The host interface 110 may transmit output data OD generated from the computing device 102 to the host device 101 through the system bus 103. The host interface 110 may receive first and second input data IND1 and IND2 through the system bus 103. The host interface 110 may provide the received first input data IND1 to the buffer circuit 120, and provide the received second input data IND2 to the memory 130. The host interface 110 may receive calculation result data CRD output from the computing unit 140 via the buffer circuit 120. The host interface 110 may transmit the output data OD to the host device 101 by driving the system bus 103 based on the calculation result data CRD generated from the computing unit 140.

[0023] The buffer circuit 120 may receive the first input data IND1 from the host interface 110. The buffer circuit 120 may classify the first input data IND1 into a plurality of input feature groups FD, and store the plurality of input feature groups FD. Each of the input feature groups FD may include a plurality of (input) features. Each of the features may indicate unit data of the input feature group FD, and the unit data may include any number of bits. The system bus 103 may have a data bandwidth different from the size of one feature. That is, the data bandwidth of the system bus 103 may be greater than the size of one feature. The first input data IND1 may be transmitted synchronously with the clock signal through the system bus 103, the number of bits included in the first input data IND1 transmitted once through the system bus 103 may be greater than the number of bits included in one feature, and the first input data IND1 transmitted once through the system bus 103 may include information about a plurality of features. Therefore, the buffer circuit 120 may classify the first input data IND1 into a plurality of input feature groups FD, and store the classified input feature groups FD. The buffer circuit 120 may provide the plurality of input feature groups FD to the calculation unit 140. The buffer circuit 120 may perform a batch operation to simultaneously provide a plurality of input feature groups FD to the computing unit 140 so as to improve the computing efficiency of the computing unit 140. The buffer circuit 120 may receive the computing result data CRD generated by the computing unit 140 through the computing operation and provide the computing result data CRD to the host interface 110.

[0024] In particular, the buffer circuit 120 may include a buffer 121 and a buffer memory 122. The buffer 121 may receive the first input data IND1 and store the received first input data IND1. The buffer 121 may classify the first input data IND1 into a plurality of input feature groups FD and store the plurality of input feature groups FD separately. The buffer 121 may include a plurality of FIFO (First In First Out) circuits. At least the number of FIFO circuits may correspond to the number of input feature groups FD that can be included in the first input data IND1. The plurality of FIFO circuits may store a plurality of features of the respective input feature groups FD.

[0025] The buffer memory 122 may store the input feature group FD output from the buffer 121, and output the stored input feature group FD to the calculation unit 140. The buffer memory 122 may store a plurality of input feature groups FD, and perform a batch operation to simultaneously output the features of each input feature group FD to the calculation unit 140. The buffer memory 122 may include a plurality of memories. The minimum number of memories included in the buffer memory 122 may be the number of input feature groups FD. The buffer memory 122 may store the features of each input feature group FD separately in a plurality of memories. The plurality of memories may store one or more features of each input feature group FD. The features of each input feature group FD stored in each memory may be arranged in a different order. Features having the same order in a plurality of input feature groups FD may not be stored in the same memory. For example, a memory storing the lth feature of one input feature group FD may not store the lth feature of another input feature group FD. Here, "l" may be an integer between 1 and the number of features included in one input feature group. The buffer memory 122 may simultaneously output the features of each input feature group FD, and provide the features to the calculation unit 140. The buffer memory 122 can simultaneously output features having the same order in a plurality of input feature groups FD during a single read operation. For example, the buffer memory 122 can simultaneously output the first feature of a plurality of input feature groups FD during a single read operation. The buffer 121 and the buffer memory 122 will be described in detail below.

[0026] In an embodiment, each of the memories in the buffer memory 122 may include a dual-port SRAM (static random access memory). The plurality of memories may have a storage depth corresponding to the number of bits included in at least one feature. For example, the plurality of memories may include a plurality of storage areas, and one storage area may store a number of bits equal to or greater than the number of bits included in one feature. The plurality of storage areas may be accessed based on different addresses.

[0027] The memory 130 may receive the second input data IND2 from the host interface 110. The memory 130 may classify the second input data IND2 into a plurality of weight data WD and store the plurality of weight data WD. The memory 130 may provide the plurality of weight data WD to the calculation unit 140. Each of the weight data WD may include a plurality of weights. The weight may indicate unit data of the weight data WD, and the unit data may include any number of bits. In an embodiment, the memory 130 may include a DRAM (dynamic random access memory) or an SRAM (static random access memory).

[0028] The calculation unit 140 may receive a plurality of input feature groups FD output from the buffer circuit 120 and a plurality of weight data WD output from the memory 130. The calculation unit 140 may perform a calculation operation on the input feature group FD and the weight data WD. For example, the calculation unit 140 may perform a convolution operation on the input feature group FD and the weight data WD. The calculation unit 140 may include a plurality of calculation circuits 141, 142, ... 14n. The plurality of calculation circuits 141, 142, ... 14n may be MAC calculation circuits having the same structure. Each of the calculation circuits 141, 142, ... 14n may perform a calculation operation on the input feature group FD and the weight data WD to output a calculation operation result. Each of the calculation circuits 141, 142, ... 14n may include a plurality of processing elements, and the plurality of processing elements may receive the input feature group FD and the weight data WD in synchronization with the clock signal, and perform a MAC operation on the received input feature group FD and the received weight data WD. Each of the computation circuits 141 , 142 , . . . 14n may include a systolic array configured to perform a MAC operation on the input feature group FD and the weight data WD.

[0029] Figure 2A and Figure 2B 2 is a diagram showing the configuration and calculation operation of the calculation circuit 200 according to the embodiment. Figure 2A In the embodiment, the computing circuit 200 may be a systolic array including a plurality of processing elements PE. The computing circuit 200 may be used as Figure 1 Each of the calculation circuits 141 to 14n in. Figure 2A In the embodiment, the calculation circuit 200 may receive three input feature groups each having three features as input signals, and receive three weight data each having three weight signals. Figure 2AThe calculation circuit 200 is shown to include nine processing elements PE1 to PE9 to perform MAC operations on three input feature groups and three weight data, each input feature group having three features and each weight data having three weight signals. However, the number of input feature groups, the number of features in one input feature group, the number of weight data, and the number of weight signals in one weight data may be less than or greater than three. Therefore, the number of processing elements in the calculation circuit 200 may also be reduced or increased.

[0030] The computing circuit 200 may include first to ninth processing elements PE1 to PE9. The processing elements PE1 to PE9 may sequentially receive features and weight signals in each input feature group and weight data in synchronization with a clock signal CLK defining a unit cycle, and perform a MAC operation on the received features and weights. Each of the first to ninth processing elements PE1 to PE9 may include multiplication logic, accumulation logic, and registers. The first to ninth processing elements PE1 to PE9 may perform multiplication through multiplication logic in parallel, and perform accumulation through accumulation logic. Figure 2B 1 shows an input feature group FD and weight data WD represented as a matrix, and a result FD×WD obtained by performing a convolution operation on the input feature group FD and weight data WD. Figure 2A and Figure 2B , the operation of the calculation circuit 200 will be described as follows.

[0031] When the first clock signal CLK1 switches during the first unit cycle, the first processing element PE1 may receive the first feature F11 of the first input feature group and the first weight signal W11 of the first weight data. The first processing element PE1 may generate F11W11 by multiplying the first feature F11 of the first input feature group by the first weight signal W11 of the first weight data. The first clock signal CLK1 may be a pulse signal that switches synchronously with the first clock cycle of the clock signal CLK. Similarly, the second to ninth clock signals CLK2 to CLK9 may be pulse signals that switch synchronously with the second to eighth clock cycles of the clock signal CLK, respectively.

[0032] When the second clock signal CLK2 switches during the second unit cycle, the first processing element PE1 may accumulate the previously calculated F11W11 and 0 and store the calculation result. The first processing element PE1 may receive the second feature F12 of the first input feature group and the first weight signal W21 of the second weight data. The first processing element PE1 may output the first feature F11 of the first input feature group to the second processing element PE2, and output the first weight signal W11 of the first weight data to the third processing element PE3. The first processing element PE1 may generate F12W21 by multiplying the second feature F12 of the first input feature group by the first weight signal W21 of the second weight data. The second processing element PE2 may receive the second weight signal W12 of the first weight data in synchronization with the second clock signal CLK2 and receive the first feature F11 of the first input feature group from the first processing element PE1. The second processing element PE2 may generate F11W12 by multiplying the first feature F11 of the first input feature group by the second weight signal W12 of the first weight data. The third processing element PE3 may receive the first feature F21 of the second input feature group in synchronization with the second clock signal CLK2 and receive the first weight signal W11 of the first weight data from the first processing element PE1. The third processing element PE3 may generate F21W11 by multiplying the first feature F21 of the second input feature group by the first weight signal W11 of the first weight data.

[0033] When the third clock signal CLK3 switches during the third unit cycle, the first processing element PE1 may accumulate the previously calculated F11W11 and F12W21, and store the calculation result F11W11+F12W21. The first processing element PE1 may receive the third feature F13 of the first input feature group and the first weight signal W31 of the third weight data. The first processing element PE1 may output the second feature F12 of the first input feature group to the second processing element PE2, and output the first weight signal W21 of the second weight data to the third processing element PE3. The first processing element PE1 may generate F13W31 by multiplying the third feature F13 of the first input feature group by the first weight signal W31 of the third weight data. The second processing element PE2 may accumulate the previously calculated F11W12 and 0, and store the calculation result. The second processing element PE2 may receive the second weight signal W22 of the second weight data, and receive the second feature F12 of the first input feature group from the first processing element PE1. The second processing element PE2 may output the first feature F11 of the first input feature group to the fourth processing element PE4, and output the second weight signal W12 of the first weight data to the fifth processing element PE5. The second processing element PE2 may generate F12W22 by multiplying the second feature F12 of the first input feature group by the second weight signal W22 of the second weight data. The third processing element PE3 may accumulate the previously calculated F21W11 and 0, and store the calculation result. The third processing element PE3 may receive the second feature F22 of the second input feature group, and receive the first weight signal W21 of the second weight data from the first processing element PE1. The third processing element PE3 may output the first feature F21 of the second input feature group to the fifth processing element PE5, and output the first weight signal W11 of the first weight data to the sixth processing element PE6. The third processing element PE3 may generate F22W21 by multiplying the second feature F22 of the second input feature group by the first weight signal W21 of the second weight data. The fourth processing element PE4 may receive the third weight signal W13 of the first weight data in synchronization with the third clock signal CLK3 and receive the first feature F11 of the first input feature group from the second processing element PE2. The fourth processing element PE4 may generate F11W13 by multiplying the first feature F11 of the first input feature group by the third weight signal W13 of the first weight data. The fifth processing element PE5 may receive the second weight signal W12 of the first weight data from the second processing element PE2 and receive the first feature F21 of the second input feature group from the third processing element PE3. The fifth processing element PE5 may generate F21W12 by multiplying the first feature F21 of the second input feature group by the second weight signal W12 of the first weight data.The sixth processing element PE6 may receive the first feature F31 of the third input feature group in synchronization with the third clock signal CLK3 and receive the first weight signal W11 of the first weight data from the third processing element PE3. The sixth processing element PE6 may generate F31W11 by multiplying the first feature F31 of the third input feature group by the first weight signal W11 of the first weight data.

[0034] When the fourth clock signal CLK4 switches during the fourth unit cycle, the first processing element PE1 may accumulate the previously calculated F11W11+F12W21 and F13W31, and store the calculation result F11W11+F12W21+F13W31. The first processing element PE1 may output the third feature F13 of the first input feature group to the second processing element PE2, and output the first weight signal W31 of the third weight data to the third processing element PE3. The second processing element PE2 may accumulate the previously calculated F11W12 and F12W22, and store the calculation result F11W12+F12W22. The second processing element PE2 may receive the second weight signal W32 of the third weight data, and receive the third feature F13 of the first input feature group from the first processing element PE1. The second processing element PE2 may output the second feature F12 of the first input feature group to the fourth processing element PE4, and output the second weight signal W22 of the second weight data to the fifth processing element PE5. The second processing element PE2 may generate F13W32 by multiplying the third feature F13 of the first input feature group by the second weight signal W32 of the third weight data. The third processing element PE3 may accumulate the previously calculated F21W11 and F22W21, and store the calculation result F21W11+F22W21. The third processing element PE3 may receive the third feature F23 of the second input feature group, and receive the first weight signal W31 of the third weight data from the first processing element PE1. The third processing element PE3 may output the second feature F22 of the second input feature group to the fifth processing element PE5, and output the first weight signal W21 of the second weight data to the sixth processing element PE6. The third processing element PE3 may generate F23W31 by multiplying the third feature F23 of the second input feature group by the first weight signal W31 of the third weight data. The fourth processing element PE4 may receive the third weight signal W23 of the second weight data in synchronization with the fourth clock signal CLK4 and receive the second feature F12 of the first input feature group from the second processing element PE2. The fourth processing element PE4 may accumulate the previously calculated F11W13 and 0, and store the calculation result. The fourth processing element PE4 may output the third weight signal W13 of the first weight data to the seventh processing element PE7. The fourth processing element PE4 may generate F12W23 by multiplying the second feature F12 of the first input feature group by the third weight signal W23 of the second weight data. The fifth processing element PE5 may accumulate the previously calculated F21W12 and 0, and store the calculation result. The fifth processing element PE5 may receive the second weight signal W22 of the second weight data from the second processing element PE2, and receive the second feature F22 of the second input feature group from the third processing element PE3.The fifth processing element PE5 may output the first feature F21 of the second input feature group to the seventh processing element PE7, and output the second weight signal W12 of the first weight data to the eighth processing element PE8. The fifth processing element PE5 may generate F22W22 by multiplying the second feature F22 of the second input feature group by the second weight signal W22 of the second weight data. The sixth processing element PE6 may accumulate the previously calculated F31W11 and 0, and store the calculation result. The sixth processing element PE6 may receive the second feature F32 of the third input feature group, and receive the first weight signal W21 of the second weight data from the third processing element PE3. The sixth processing element PE6 may output the first feature F31 of the third input feature group to the eighth processing element PE8. The sixth processing element PE6 may generate F32W21 by multiplying the second feature F32 of the third input feature group by the first weight signal W21 of the second weight data. The seventh processing element PE7 may receive the third weight signal W13 of the first weight data from the fourth processing element PE4, and receive the first feature F21 of the second input feature group from the fifth processing element PE5. The seventh processing element PE7 may generate F21W13 by multiplying the first feature F21 of the second input feature group by the third weight signal W13 of the first weight data. The eighth processing element PE8 may receive the second weight signal W12 of the first weight data from the fifth processing element PE5 and the first feature F31 of the third input feature group from the sixth processing element PE6. The eighth processing element PE8 may generate F31W12 by multiplying the first feature F31 of the third input feature group by the second weight signal W12 of the first weight data.

[0035] When the fifth clock signal CLK5 switches during the fifth unit cycle, the second processing element PE2 may accumulate the previously calculated F11W12+F12W22 and F13W32, and store the calculation result F11W12+F12W22+F13W32. The second processing element PE2 may output the third feature F13 of the first input feature group to the fourth processing element PE4, and output the second weight signal W32 of the third weight data to the fifth processing element PE5. The third processing element PE3 may accumulate the previously calculated F21W11+F22W21 and F23W31, and store the calculation result F21W11+F22W21+F23W31. The third processing element PE3 may output the third feature F23 of the second input feature group to the fifth processing element PE5, and output the first weight signal W31 of the third weight data to the sixth processing element PE6. The fourth processing element PE4 may accumulate previously calculated F11W13 and F12W23, and store the calculation result F11W13+F12W23. The fourth processing element PE4 may receive the third weight signal W33 of the third weight data, and receive the third feature F13 of the first input feature group from the second processing element PE2. The fourth processing element PE4 may output the third weight signal W23 of the second weight data to the seventh processing element PE7. The fourth processing element PE4 may generate F13W33 by multiplying the third feature F13 of the first input feature group by the third weight signal W33 of the third weight data. The fifth processing element PE5 may accumulate previously calculated F21W12 and F22W22, and store the calculation result F21W12+F22W22. The fifth processing element PE5 may receive the second weight signal W32 of the third weight data from the second processing element PE2, and receive the third feature F23 of the second input feature group from the third processing element PE3. The fifth processing element PE5 may output the second feature F22 of the second input feature group to the seventh processing element PE7, and output the second weight signal W22 of the second weight data to the eighth processing element PE8. The fifth processing element PE5 may generate F23W32 by multiplying the third feature F23 of the second input feature group by the second weight signal W32 of the third weight data. The sixth processing element PE6 may accumulate the previously calculated F31W11 and F32W21, and store the calculation result F31W11+F32W21. The sixth processing element PE6 may receive the third feature F33 of the third input feature group, and receive the first weight signal W31 of the third weight data from the third processing element PE3. The sixth processing element PE6 may output the second feature F32 of the third input feature group to the eighth processing element PE8. The sixth processing element PE6 may generate F33W31 by multiplying the third feature F33 of the third input feature group by the first weight signal W31 of the third weight data. The seventh processing element PE7 may accumulate previously calculated F21W13 and 0, and store the calculation result.The seventh processing element PE7 may receive the third weight signal W23 of the second weight data from the fourth processing element PE4, and receive the second feature F22 of the second input feature group from the fifth processing element PE5. The seventh processing element PE7 may output the third weight signal W13 of the first weight data to the ninth processing element PE9. The seventh processing element PE7 may generate F22W23 by multiplying the second feature F22 of the second input feature group by the third weight signal W23 of the second weight data. The eighth processing element PE8 may accumulate the previously calculated F31W12 and 0, and store the calculation result. The eighth processing element PE8 may receive the second weight signal W22 of the second weight data from the fifth processing element PE5, and receive the second feature F32 of the third input feature group from the sixth processing element PE6. The eighth processing element PE8 may output the first feature F31 of the third input feature group to the ninth processing element PE9. The eighth processing element PE8 may generate F32W22 by multiplying the second feature F32 of the third input feature group by the second weight signal W22 of the second weight data. The ninth processing element PE9 may receive the third weight signal W13 of the first weight data from the seventh processing element PE7 and the first feature F31 of the third input feature group from the eighth processing element PE8. The ninth processing element PE9 may generate F31W13 by multiplying the first feature F31 of the third input feature group by the third weight signal W13 of the first weight data.

[0036] When the sixth clock signal CLK6 switches during the sixth unit cycle, the fourth processing element PE4 may accumulate the previously calculated F11W13+F12W23 and F13W33, and store the calculation result F11W13+F12W23+F13W33. The fourth processing element PE4 may output the third weight signal W33 of the third weight data to the seventh processing element PE7. The fifth processing element PE5 may accumulate the previously calculated F21W12+F22W22 and F23W32, and store the calculation result F21W12+F22W22+F23W32. The fifth processing element PE5 may output the third feature F23 of the second input feature group to the seventh processing element PE7, and output the second weight signal W32 of the third weight data to the eighth processing element PE8. The sixth processing element PE6 may accumulate the previously calculated F31W11+F32W21 and F33W31, and store the calculation result F31W11+F32W21+F33W31. The sixth processing element PE6 may output the third feature F33 of the third input feature group to the eighth processing element PE8. The seventh processing element PE7 may accumulate the previously calculated F21W13 and F22W23, and store the calculation result F21W13+F22W23. The seventh processing element PE7 may receive the third weight signal W33 of the third weight data from the fourth processing element PE4, and receive the third feature F23 of the second input feature group from the fifth processing element PE5. The seventh processing element PE7 may output the third weight signal W23 of the second weight data to the ninth processing element PE9. The seventh processing element PE7 may generate F23W33 by multiplying the third feature F23 of the second input feature group by the third weight signal W33 of the third weight data. The eighth processing element PE8 may accumulate the previously calculated F31W12 and F32W22, and store the calculation result F31W12+F32W22. The eighth processing element PE8 may receive the second weight signal W32 of the third weight data from the fifth processing element PE5, and receive the third feature F33 of the third input feature group from the sixth processing element PE6. The eighth processing element PE8 may output the second feature F32 of the third input feature group to the ninth processing element PE9. The eighth processing element PE8 may generate F33W32 by multiplying the third feature F33 of the third input feature group by the second weight signal W32 of the third weight data. The ninth processing element PE9 may accumulate the previously calculated F31W13 and 0, and store the calculation result. The ninth processing element PE9 may receive the third weight signal W23 of the second weight data from the seventh processing element PE7, and receive the second feature F32 of the third input feature group from the eighth processing element PE8. The ninth processing element PE9 may generate F32W23 by multiplying the second feature F32 of the third input feature group by the third weight signal W23 of the second weight data.

[0037] When the seventh clock signal switches during the seventh unit cycle, the seventh processing element PE7 may accumulate the previously calculated F21W13+F22W23 and F23W33, and store the calculation result F21W13+F22W23+F23W33. The seventh processing element PE7 may output the third weight signal W33 of the third weight data to the ninth processing element PE9. The eighth processing element PE8 may accumulate the previously calculated F31W12+F32W22 and F33W32, and store the calculation result F31W12+F32W22+F33W32. The eighth processing element PE8 may output the third feature F33 of the third input feature group to the ninth processing element PE9. The ninth processing element PE9 may accumulate the previously calculated F31W13 and F32W23, and store the calculation result F31W13+F32W23. The ninth processing element PE9 may generate F33W33 by multiplying the third feature F33 of the third input feature group by the third weight signal W33 of the third weight data.

[0038] When the eighth clock signal switches during the eighth unit cycle, the ninth processing element PE9 may accumulate the previously calculated F31W13+F32W23 and F33W33 and store the calculation result F31W13+F32W23+F33W33. When the ninth clock signal switches during the ninth unit cycle, the calculation results respectively stored in the first to ninth processing elements PE1 to PE9 may be output. Figure 2B As shown, convolution operation results (FD×WD) of the first to third input feature groups FD and the first to third weight data WD may be output from the first to ninth processing elements PE1 to PE9.

[0039] Figure 3 is a diagram showing the flow of data in the system 1 according to the present embodiment. Figures 1 to 3 , the host device 101 may transmit the first input data IND1 through the system bus 103. The first input data IND1 may include a plurality of input feature groups. The data bandwidth of the system bus 103 may be greater than the size of one feature, and the data bits transmitted at one time through the system bus 103 may include a plurality of features. For example, when the system bus 103 has a data bandwidth of 32 bits and one feature has a size of 8 bits, four features may be transmitted at one time through the system bus 103. When the first input data IND1 has 256 bits, the first input data IND1 may include four input feature groups, each of which includes eight features. Figure 3 The system bus 103 is shown to have a data bandwidth of 32 bits, but the system bus 103 may have a data bandwidth of 16 bits or less or 64 bits or more. Figure 3 A feature is shown to have a size of eight bits, but a feature may have a size of four bits or less or 16 bits or more.

[0040] The system bus 103 may include a plurality of buses. Hereinafter, for clarity of description, it is assumed that four features can be transmitted at a time through the system bus 103. The system bus 103 may include first to fourth buses. Each of the first to fourth buses may include a plurality of signal transmission lines. One signal transmission line may be used to transmit one bit of the first input data IND1. One bus may indicate a signal transmission line capable of transmitting a plurality of bits corresponding to one feature. For example, one bus may include eight signal transmission lines, and one feature may include eight bits.

[0041] The first to fourth buses may serially transmit bits of the first input data IND1. The first input data IND1 may be transmitted more than eight times in synchronization with the clock signal through the first to fourth buses and include a total of 256 bits. The first bus may sequentially transmit the 1st to 8th bits (I11), the 33rd to 40th bits (I15), the 65th to 72nd bits (I21), the 97th to 104th bits (I25), the 129th to 136th bits (I31), the 161st to 168th bits (I35), the 193rd to 200th bits (I41), and the 225th to 232nd bits (I45) of the first input data in synchronization with the clock signal. The 1st to 8th bits (I11) of the first input data IND1 may correspond to the first feature F11 of the first input feature group FD1. The 33rd to 40th bits (I15) of the first input data IND1 may correspond to the fifth feature F15 of the first input feature group FD1. The 65th to 72nd bits (I21) of the first input data IND1 may correspond to the first feature F21 of the second input feature group FD2. The 97th to 104th bits (I25) of the first input data IND1 may correspond to the fifth feature F25 of the second input feature group FD2. The 129th to 136th bits (I31) of the first input data IND1 may correspond to the first feature F31 of the third input feature group FD3. The 161st to 168th bits (I35) of the first input data IND1 may correspond to the fifth feature F35 of the third input feature group FD3. The 193rd to 200th bits (I41) of the first input data IND1 may correspond to the first feature F41 of the fourth input feature group FD4. The 225th to 232nd bits (I45) of the first input data IND1 may correspond to the fifth feature F45 of the fourth input feature group FD4.

[0042] The second bus may sequentially transmit the 9th to 16th bits (I12), the 41st to 48th bits (I16), the 73rd to 80th bits (I22), the 105th to 112th bits (I26), the 137th to 144th bits (I32), the 169th to 176th bits (I36), the 201st to 208th bits (I42), and the 233rd to 240th bits (I46) of the first input data IND1 in synchronization with the clock signal. The 9th to 16th bits (I12) of the first input data IND1 may correspond to the second feature F12 of the first input feature group FD1. The 41st to 48th bits (I16) of the first input data IND1 may correspond to the sixth feature F16 of the first input feature group FD1. The 73rd to 80th bits (I22) of the first input data IND1 may correspond to the second feature F22 of the second input feature group FD2. The 105th to 112th bits (I26) of the first input data IND1 may correspond to the sixth feature F26 of the second input feature group FD2. The 137th to 144th bits (I32) of the first input data IND1 may correspond to the second feature F32 of the third input feature group FD3. The 169th to 176th bits (I36) of the first input data IND1 may correspond to the sixth feature F36 of the third input feature group FD3. The 201st to 208th bits (I42) of the first input data IND1 may correspond to the second feature F42 of the fourth input feature group FD4. The 233rd to 240th bits (I46) of the first input data IND1 may correspond to the sixth feature F46 of the fourth input feature group FD4.

[0043] The third bus may sequentially transmit the 17th to 24th bits (I13), the 49th to 56th bits (I17), the 81st to 88th bits (I23), the 113th to 120th bits (I27), the 145th to 152nd bits (I33), the 177th to 184th bits (I37), the 209th to 216th bits (I43), and the 241st to 248th bits (I47) of the first input data IND1 in synchronization with the clock signal. The 17th to 24th bits (I13) of the first input data IND1 may correspond to the third feature F13 of the first input feature group FD1. The 49th to 56th bits (I17) of the first input data IND1 may correspond to the seventh feature F17 of the first input feature group FD1. The 81st to 88th bits (I23) of the first input data IND1 may correspond to the third feature F23 of the second input feature group FD2. The 113th to 120th bits (I27) of the first input data IND1 may correspond to the seventh feature F27 of the second input feature group FD2. The 145th to 152nd bits (I33) of the first input data IND1 may correspond to the third feature F33 of the third input feature group FD3. The 177th to 184th bits (I37) of the first input data IND1 may correspond to the seventh feature F37 of the third input feature group FD3. The 209th to 216th bits (I43) of the first input data IND1 may correspond to the third feature F43 of the fourth input feature group FD4. The 241st to 248th bits (I47) of the first input data IND1 may correspond to the seventh feature F47 of the fourth input feature group FD4.

[0044] The fourth bus may sequentially transmit the 25th to 32nd bits (I14), the 57th to 64th bits (I18), the 89th to 96th bits (I24), the 121st to 128th bits (I28), the 153rd to 160th bits (I34), the 185th to 192nd bits (I38), the 217th to 224th bits (I44), and the 249th to 256th bits (I48) of the first input data IND1 in synchronization with the clock signal. The 25th to 32nd bits (I14) of the first input data IND1 may correspond to the fourth feature F14 of the first input feature group FD1. The 57th to 64th bits (I18) of the first input data IND1 may correspond to the eighth feature F18 of the first input feature group FD1. The 89th to 96th bits (I24) of the first input data IND1 may correspond to the fourth feature F24 of the second input feature group FD2. The 121st to 128th bits (I28) of the first input data IND1 may correspond to the eighth feature F28 of the second input feature group FD2. The 153rd to 160th bits (I34) of the first input data IND1 may correspond to the fourth feature F34 of the third input feature group FD3. The 185th to 192nd bits (I38) of the first input data IND1 may correspond to the eighth feature F38 of the third input feature group FD3. The 217th to 224th bits (I44) of the first input data IND1 may correspond to the fourth feature F44 of the fourth input feature group FD4. The 249th to 256th bits (I48) of the first input data IND1 may correspond to the eighth feature F48 of the fourth input feature group FD4.

[0045] The computing device 102 may include a buffer circuit 120 to classify the first input data IND1 into a plurality of input feature groups and provide the plurality of input feature groups FD to the computing unit 140. The buffer 121 of the buffer circuit 120 may include a plurality of FIFO circuits. The first FIFO circuit FIFO1 may store the 1st to 64th bits (I11 to I18) of the first input data IND1, and the information stored in the first FIFO circuit FIFO1 may correspond to the first input feature group FD1. The second FIFO circuit FIFO2 may store the 65th to 128th bits (I21 to I28) of the first input data IND1, and the information stored in the second FIFO circuit FIFO2 may correspond to the second input feature group FD2. The third FIFO circuit FIFO3 may store the 129th to 192nd bits (I31 to I38) of the first input data IND1, and the information stored in the third FIFO circuit FIFO3 may correspond to the third input feature group FD3. The fourth FIFO circuit FIFO4 may store 193rd to 256th bits ( I41 to 148 ) of the first input data IND1 , and information stored in the fourth FIFO circuit FIFO4 may correspond to a fourth input feature group FD4 .

[0046] The buffer memory 122 of the buffer circuit 120 may perform a batch operation to simultaneously provide the first to fourth input feature groups FD1 to FD4 to the calculation unit 140. Since the number of input feature groups is four, the buffer memory 122 may include at least four memories. Figure 3 As shown, the buffer memory 122 needs to store the features F11 to F18, F21 to F28, F31 to F38 and F41 to F48 of the first to fourth input feature groups FD1 to FD4 separately in four memories, so that the features F11 to F18, F21 to F28, F31 to F38 and F41 to F48 of the first to fourth input feature groups FD1 to FD4 can be output simultaneously. At this time, when one input feature group is stored in one memory, the batch operation can be fully performed. However, according to the type and characteristics of the data or application, the size of the input feature group can be changed differently. However, the memory may have a fixed size and / or depth. Therefore, when a memory with a very small size is used, the memory may not store input feature groups with various sizes, and when a memory with a very large size is used, most of the storage space of the memory may be wasted. Therefore, an effective method that can store input feature groups by using a memory with a minimum size may be required.

[0047] Figure 4 4 is a diagram showing a configuration of a buffer memory 400 and a method of storing a plurality of input feature groups in the buffer memory 400 according to an embodiment. The buffer memory 400 may be used as Figure 1 The buffer memory 122 is shown. Figure 4 In the embodiment, the buffer memory 400 may include first to fourth memories 410 to 440. The features F11 to F18, F21 to F28, F31 to F38, and F41 to F48 of the first to fourth input feature groups FD1 to FD4 may each be separately stored in the first to fourth memories 410 to 440. The first to fourth memories 410 to 440 may not store features having the same order in the first to fourth input feature groups FD1 to FD4. Features of one input feature group may be separately stored in the first to fourth memories 410 to 440. Features having the same order in the first to fourth input feature groups FD1 to FD4 may be separately stored in the first to fourth memories 410 to 440 one by one (i.e., uniformly). Features having the same order in the first to fourth input feature groups FD1 to FD4 may be separately stored in the first to fourth memories 410 to 440 so as not to overlap with each other.

[0048] The order in which the features F11 to F18, F21 to F28, F31 to F38, and F41 to F48 of the first to fourth input feature groups FD1 to FD4 are respectively stored in the first to fourth memories 410 to 440 may be different from each other. The features F11 to F18, F21 to F28, F31 to F38, and F41 to F48 of the first to fourth input feature groups FD1 to FD4 may be separately stored in the first to fourth memories 410 to 440 in a changed order. For example, the first feature of the k-th input feature group may be stored in the k-th memory. Here, "k" may be an integer between 1 and 4. The first to fourth features F11 to F14 of the first input feature group FD1 may be respectively stored in the first to fourth memories 410 to 440, and the fifth to eighth features F15 to F18 of the first input feature group FD1 may be respectively stored in the first to fourth memories 410 to 440. The order of the memories storing the first to eighth features F21 to F28 of the second input feature group FD2 may be changed. The first to third features F21 to F23 of the second input feature group FD2 may be stored in the second to fourth memories 420 to 440, respectively, and the fourth feature F24 of the second input feature group FD2 may be stored in the first memory 410. The fifth to seventh features F25 to F27 of the second input feature group FD2 may be stored in the second to fourth memories 420 to 440, respectively, and the eighth feature F28 of the second input feature group FD2 may be stored in the first memory 410. The order of the memories storing the first to eighth features F31 to F38 of the third input feature group FD3 may be changed. The first and second features F31 and F32 of the third input feature group FD3 may be stored in the third and fourth memories 430 and 440, respectively, and the third and fourth features F33 and F34 of the third input feature group FD3 may be stored in the first and second memories 410 and 420, respectively. The fifth and sixth features F35 and F36 of the third input feature group FD3 may be stored in the third and fourth memories 430 and 440, respectively, and the seventh and eighth features F37 and F38 of the third input feature group FD3 may be stored in the first and second memories 410 and 420, respectively. The order of the memories storing the first to eighth features F41 to F48 of the fourth input feature group FD4 may be changed. The first feature F41 of the fourth input feature group FD4 may be stored in the fourth memory 440, and the second to fourth features F42 to F44 of the fourth input feature group FD4 may be stored in the first to third memories 410 to 430, respectively. The fifth feature F45 of the fourth input feature group FD4 may be stored in the fourth memory 440, and the sixth to eighth features F46 to F48 of the fourth input feature group FD4 may be stored in the first to third memories 410 to 430, respectively.

[0049] Each of the first to fourth memories 410 to 440 may include a plurality of storage areas. Each of the storage areas may store one feature therein. The number of the plurality of areas may be equal to or greater than the number of features included in one input feature group. Figure 4 It is shown that each of the first to fourth memories 410 to 440 includes eight storage areas. However, each of the first to fourth memories 410 to 440 may include a larger number of storage areas to store an input feature group including more features. Different addresses may be respectively allocated and / or mapped to the plurality of storage areas, and the plurality of storage areas may be accessed based on the allocated and / or mapped addresses.

[0050] In particular, a method of storing the first to fourth input feature groups FD1 to FD4 in the first to fourth memories 410 to 440 will be described as follows. The first to fourth memories 410 to 440 may perform a write operation to store the respective features F11 to F18, F21 to F28, F31 to F38, and F41 to F48 of the first to fourth input feature groups FD1 to FD4. The first feature F11 of the first input feature group FD1 may be stored in the first storage area 411 of the first memory 410. The second to fourth features F12 to F14 of the first input feature group FD1 may be stored in the first storage areas 421 to 441 of the second to fourth memories 420 to 440, respectively. The fifth to eighth features F15 to F18 of the first input feature group FD1 may be stored in the second storage areas 412, 422, 432, and 442 of the first to fourth memories 410 to 440, respectively. The first feature F21 of the second input feature group FD2 may be stored in the third storage area 423 of the second memory 420. The second and third features F22 and F23 of the second input feature group FD2 may be stored in the third storage areas 433 and 443 of the third and fourth memories 430 and 440, respectively. The fourth feature F24 of the second input feature group FD2 may be stored in the third storage area 413 of the first memory 410. The fifth to seventh features F25 to F27 of the second input feature group FD2 may be stored in the fourth storage areas 424, 434, and 444 of the second to fourth memories 420 to 440, respectively, and the eighth feature F28 of the second input feature group FD2 may be stored in the fourth storage area 414 of the first memory 410. The first feature F31 of the third input feature group FD3 may be stored in the fifth storage area 435 of the third memory 430. The second feature F32 of the third input feature group FD3 may be stored in the fifth storage area 445 of the fourth memory 440, and the third and fourth features F33 and F34 of the third input feature group FD3 may be stored in the fifth storage areas 415 and 425 of the first and second memories 410 and 420, respectively. The fifth and sixth features F35 and F36 of the third input feature group FD3 may be stored in the sixth storage areas 436 and 446 of the third and fourth memories 430 and 440, respectively, and the seventh and eighth features F37 and F38 of the third input feature group FD3 may be stored in the sixth storage areas 416 and 426 of the first and second memories 410 and 420, respectively. The first feature F41 of the fourth input feature group FD4 may be stored in the seventh storage area 447 of the fourth memory 440. The second to fourth features F42 to F44 of the fourth input feature group FD4 may be stored in the seventh storage areas 417, 427, and 437 of the first to third memories 410 to 430, respectively.The fifth feature F45 of the fourth input feature group FD4 may be stored in the eighth storage area 448 of the fourth memory 440, and the sixth to eighth features F46 to F48 of the fourth input feature group FD4 may be stored in the eighth storage areas 418, 428 and 438 of the first to third memories 410 to 430, respectively.

[0051] The first to fourth memories 410 to 440 may simultaneously store the first to fourth features F11 to F14 of the first input feature group FD1 in the first storage areas 411, 421, 431, and 441 of the first to fourth memories 410 to 440, respectively, based on the first address ADD1. The first to fourth memories 410 to 440 may simultaneously store the fifth to eighth features F15 to F18 of the first input feature group FD1 in the second storage areas 412, 422, 432, and 442 of the first to fourth memories 410 to 440, respectively, based on the second address ADD2. The first to fourth memories 410 to 440 may simultaneously store the first to fourth features F21 to F24 of the second input feature group FD2 in the third storage areas 413, 423, 433, and 443 of the first to fourth memories 410 to 440, respectively, based on the third address ADD3. The first to fourth memories 410 to 440 may simultaneously store the fifth to eighth features F25 to F28 of the second input feature group FD2 in the fourth storage areas 414, 424, 434, and 444 of the first to fourth memories 410 to 440, respectively, based on the fourth address ADD4. The first to fourth memories 410 to 440 may simultaneously store the first to fourth features F31 to F34 of the third input feature group FD3 in the fifth storage areas 415, 425, 435, and 445 of the first to fourth memories 410 to 440, respectively, based on the fifth address ADD5. The first to fourth memories 410 to 440 may simultaneously store the fifth to eighth features F35 to F38 of the third input feature group FD3 in the sixth storage areas 416, 426, 436, and 446 of the first to fourth memories 410 to 440, respectively, based on the sixth address ADD6. The first to fourth memories 410 to 440 may simultaneously store the first to fourth features F41 to F44 of the fourth input feature group FD4 in the seventh storage areas 417, 427, 437, and 447 of the first to fourth memories 410 to 440, respectively, based on the seventh address ADD7. The first to fourth memories 410 to 440 may simultaneously store the fifth to eighth features F45 to F48 of the fourth input feature group FD4 in the eighth storage areas 418, 428, 438, and 448 of the first to fourth memories 410 to 440, respectively, based on the eighth address ADD8.

[0052] In an embodiment, when the number of features included in each of the first to fourth input feature groups FD1 to FD4 is less than eight, some storage areas of the first to fourth memories 410 to 440 may not be used. For example, when each of the first to fourth input feature groups FD1 to FD4 may include three features, the first to third features F11 to F13 of the first input feature group FD1 may be stored in the first storage areas 411, 421, and 431 of the first to third memories 410 to 430, respectively. The first to third features F21 to F23 of the second input feature group FD2 may be stored in the second storage areas 422 to 442 of the second to fourth memories 420 to 440, respectively. The first and second features F31 and F32 of the third input feature group FD3 may be stored in the third storage areas 433 and 443 of the third and fourth memories 430 and 440, respectively, and the third feature F33 of the third input feature group FD3 may be stored in the third storage area 413 of the first memory 410. The first feature F41 of the fourth input feature group FD4 may be stored in the fourth storage area 444 of the fourth memory 440 , and the second and third features F42 and F43 of the fourth input feature group FD4 may be stored in the fourth storage areas 414 and 424 of the first and second memories 410 and 420 , respectively.

[0053] During a single read operation, the first to fourth memories 410 to 440 may simultaneously output features having the same order in the first to fourth input feature groups FD1 to FD4. During the first read operation, the first to fourth memories 410 to 440 may simultaneously output first features F11, F21, F31, and F41 of the first to fourth input feature groups FD1 to FD4 based on the first address ADD1, the third address ADD3, the fifth address ADD5, and the seventh address ADD7, respectively. The first features F11, F21, F31, and F41 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits of the calculation unit 140, respectively. During the second read operation, the first to fourth memories 410 to 440 may simultaneously output second features F12 to F42 of the first to fourth input feature groups FD1 to FD4 based on the seventh address ADD7, the first address ADD1, the third address ADD3, and the fifth address ADD5, respectively. The second features F12 to F42 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits, respectively. During the third read operation, the first to fourth memories 410 to 440 may simultaneously output the third features F13 to F43 of the first to fourth input feature groups FD1 to FD4, respectively, based on the fifth address ADD5, the seventh address ADD7, the first address ADD1, and the third address ADD3. The third features F13 to F43 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits of the calculation unit 140, respectively. During the fourth read operation, the first to fourth memories 410 to 440 may simultaneously output the fourth features F14 to F44 of the first to fourth input feature groups FD1 to FD4, respectively, based on the third address ADD3, the fifth address ADD5, the seventh address ADD7, and the first address ADD1. The fourth features F14, F24, F34, and F44 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits, respectively. During the fifth read operation, the first to fourth memories 410 to 440 may simultaneously output the fifth features F15, F25, F35, and F45 of the first to fourth input feature groups FD1 to FD4 based on the second address ADD2, the fourth address ADD4, the sixth address ADD6, and the eighth address ADD8, respectively. The fifth features F15, F25, F35, and F45 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits, respectively. During the sixth read operation, the first to fourth memories 410 to 440 may simultaneously output the sixth features F16 to F46 of the first to fourth input feature groups FD1 to FD4 based on the eighth address ADD8, the second address ADD2, the fourth address ADD4, and the sixth address ADD6, respectively. The sixth features F16 to F46 of the first to fourth input feature groups FD1 to FD4 may be provided to four calculation circuits, respectively.During the seventh read operation, the first to fourth memories 410 to 440 may simultaneously output the seventh features F17 to F47 of the first to fourth input feature groups FD1 to FD4 based on the sixth address ADD6, the eighth address ADD8, the second address ADD2, and the fourth address ADD4, respectively. The seventh features F17 to F47 of the first to fourth input feature groups FD1 to FD4 may be provided to the four calculation circuits, respectively. During the eighth read operation, the first to fourth memories 410 to 440 may simultaneously output the eighth features F18 to F48 of the first to fourth input feature groups FD1 to FD4 based on the fourth address ADD4, the sixth address ADD6, the eighth address ADD8, and the second address ADD2, respectively. The eighth features F18 to F48 of the first to fourth input feature groups FD1 to FD4 may be provided to the four calculation circuits, respectively. The calculation unit 140 may sequentially receive the first to eighth features F11 to F18, F21 to F28, F31 to F38, and F41 to F48 of the first to fourth input feature groups FD1 to FD4 from the buffer memory 122, and simultaneously perform calculation operations on the first to fourth input feature groups FD1 to FD4. In this embodiment, the computing device and system can perform batch operations on multiple input feature groups by using a memory with a smaller size, thereby improving the operating efficiency of the system and reducing the manufacturing cost of the computing device.

[0054] Although various embodiments have been described above, those skilled in the art will appreciate that the described embodiments are only examples. Therefore, the computing device and system described herein should not be limited based on the described embodiments.

Claims

1. A computing device, comprising: A buffer memory, comprising first to nth memories, and storing first to nth input feature groups in the first to nth memories, each input feature group comprising a plurality of features, and the plurality of features of one input feature group are stored separately in the first to nth memories, wherein features having the same order in the first to nth input feature groups are stored one by one in the first to nth memories, and features having the same order in the first to nth input feature groups are respectively stored in different storage areas of the first to nth memories accessed based on different addresses from each other, wherein n is an integer equal to or greater than 2; as well as A computing unit receives the first to n-th input feature groups stored in the buffer memory, performs computing operations on the received input feature groups, and receives features having the same order in the first to n-th input feature groups from the first to n-th memories based on addresses during a single read operation. 2 . The computing device according to claim 1 , wherein the buffer memory stores a first feature among the plurality of features of the kth input feature group in a kth memory, wherein k is an integer between 1 and n. 3 . The computing device according to claim 2 , wherein the buffer memory stores the (n+1)th feature among the plurality of features of the kth input feature group in the kth memory.

4. The computing device according to claim 1, wherein the buffer memory simultaneously stores the first feature to the lth feature of the first input feature group in the first memory to the lth memory respectively, where l is an integer between 2 and n. 5 . The computing device of claim 1 , wherein during a single read operation, the buffer memory simultaneously outputs features having the same order in the first to nth input feature groups.

6. A computing device comprising: A buffer memory, comprising first to fourth memories, and storing first to fourth input feature groups in the first to fourth memories, each input feature group having at least first to fourth features, and dividing the first to fourth features of one input feature group and storing them in the first to fourth memories, wherein features having the same order in the first to fourth input feature groups are divided and stored in the first to fourth memories so as not to overlap with each other, and features having the same order in the first to fourth input feature groups are respectively stored in different storage areas of the first to fourth memories accessed based on different addresses from each other; as well as A computing unit receives the first to fourth input feature groups stored in the buffer memory, performs computing operations on the received input feature groups, and receives features having the same order in the first to fourth input feature groups from the first memory to the fourth memory based on the address during a single read operation.

7. The computing device of claim 6, wherein a first feature of the first input feature group is stored in a first storage area of ​​the first memory, a second feature of the first input feature group is stored in a first storage area of ​​a second memory, a third feature of the first input feature group is stored in a first storage area of ​​a third memory, and a fourth feature of the first input feature group is stored in a first storage area of ​​the fourth memory.

8. The computing device of claim 7, wherein the first feature of the second input feature group is stored in the second storage area of ​​the second memory, the second feature of the second input feature group is stored in the second storage area of ​​the third memory, the third feature of the second input feature group is stored in the second storage area of ​​the fourth memory, and the fourth feature of the second input feature group is stored in the second storage area of ​​the first memory.

9. The computing device of claim 8, wherein a first feature of a third input feature group is stored in a third storage area of ​​the third memory, a second feature of the third input feature group is stored in a third storage area of ​​the fourth memory, a third feature of the third input feature group is stored in a third storage area of ​​the first memory, and a fourth feature of the third input feature group is stored in a third storage area of ​​the second memory.

10. The computing device of claim 9, wherein a first feature of the fourth input feature group is stored in a fourth storage area of ​​the fourth memory, a second feature of the fourth input feature group is stored in a fourth storage area of ​​the first memory, a third feature of the fourth input feature group is stored in a fourth storage area of ​​the second memory, and a fourth feature of the fourth input feature group is stored in a fourth storage area of ​​the third memory. 11 . The computing device of claim 6 , wherein during a first read operation, the buffer memory simultaneously outputs first features of the first to fourth input feature groups. 12 . The computing device of claim 11 , wherein during a second read operation, the buffer memory simultaneously outputs second features of the first to fourth input feature groups. 13 . The computing device of claim 12 , wherein during a third read operation, the buffer memory simultaneously outputs third features of the first to fourth input feature groups. 14 . The computing device of claim 13 , wherein during a fourth read operation, the buffer memory simultaneously outputs fourth features of the first to fourth input feature groups.

15. A system comprising: The host device transmits first input data via the system bus; as well as a computing device, receiving the first input data and performing a computing operation on the first input data, The computing device comprises: a buffer, classifying the first input data into a first input feature group to an nth input feature group, and storing the first input feature group to the nth input feature group, each input feature group having at least first to nth features, wherein n is an integer equal to or greater than 2; a buffer memory comprising first to nth memories, and storing the first to nth features of the first to nth input feature groups in the first to nth memories, wherein the first to nth features of one input feature group are separately stored in the first to nth memories, wherein features having the same order in the first to nth input feature groups are respectively stored in different storage areas of the first to nth memories accessed based on different addresses from each other; and A computing unit receives the first to n-th input feature groups stored in the buffer memory, performs computing operations on the received input feature groups, and receives features having the same order in the first to n-th input feature groups from the first to n-th memories based on addresses during a single read operation. 16 . The system according to claim 15 , wherein the buffer comprises n FIFO circuits, ie, first-in-first-out circuits, and the FIFO circuits respectively store the first input feature group to the nth input feature group.

17. The system of claim 15, wherein each of the first to nth input feature groups further comprises an (n+1)th feature, The buffer memory stores each (n+1)th feature of the first to nth input feature groups in the memory storing the first features of the first to nth input feature groups.

18. The system of claim 15, wherein during a single read operation, the buffer memory simultaneously outputs features having the same order in the first to nth input feature groups.

19. The system of claim 15, wherein the host device further transmits second input data via the system bus, and The computing device further includes a memory configured to store the second input data and output the second input data as weight data to the computing unit.