Neural network operational circuit, control circuit thereof and control method thereof
By introducing word line selection management circuit and timing generation circuit into neural network operation circuit, dynamically adjusting the delay time, the problem of word line driving time proportional to the number of word lines in the prior art is solved, and high-speed operation of neurons is realized and improve computing efficiency.
Patent Information
- Application Number
- CN202380023404.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-02-28
- Filing Date
- 2023-02-03
- Publication Date
- 2024-11-05
AI Technical Summary
The existing neural network computing circuits are difficult to realize high-speed computing of neurons. This is mainly because the setting time when word line is driven is proportional to the number of word lines selected, which makes it difficult to meet the high-speed computing requirements for fixed delay time settings.
A neural network computing circuit including word line selection management circuit and timing generation circuit is designed. By dynamically setting the delay time, the delay time from word line driving to operation circuit startup is adjusted according to the selected word line number, avoiding fixed delay time limitations.
High-speed computing of neurons is realized, which reduces the delay time before the operation circuit is started, and improves the efficiency of neural network computing.
Smart Images

Figure CN118901067A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to a neural network operation circuit using a semiconductor memory element that can achieve low power consumption and large-scale integration, a control circuit thereof, and a control method thereof. Background Art
[0002] With the development of information and communication technology, IoT (Internet of Things) technology, which connects all things to the Internet, has attracted attention. In IoT technology, various electronic devices are connected to the Internet, and the high performance of the devices is expected to be improved. As a technology to achieve further high performance, the research and development of artificial intelligence (AI) technology, which allows electronic devices to learn and judge by themselves, has been actively carried out in recent years.
[0003] In artificial intelligence technology, neural network technology is used, which is an engineering imitation of human brain-type information processing, and semiconductor integrated circuits that can perform neural network operations at high speed and low power consumption are being actively researched and developed.
[0004] Neural networks are composed of basic elements called neurons (sometimes also called perceptrons) that connect multiple inputs through couplings called synapses with different coupling weight coefficients. By connecting multiple neurons, advanced computing processes such as image recognition and voice recognition can be performed. In neurons, a multiplication and accumulation operation is performed to add up all the multiplication results of each input and each coupling weight coefficient.
[0005] Patent Document 1 discloses an example of a neural network operation circuit using a resistance variable nonvolatile memory (hereinafter, also simply referred to as a "resistance variable memory element"). In this technology, a resistance variable nonvolatile memory capable of setting an analog resistance value (conductance) is used to form a neural network operation circuit, an analog resistance value (conductance) equivalent to a coupling weight coefficient is stored in a nonvolatile memory element, and an analog current value flowing through the memory element is used according to the selection state of a word line equivalent to an input.
[0006] The multiplication and accumulation operation performed by the neuron is performed in the following manner: an analog current value is obtained as a multiplication and accumulation operation result according to the selection states of multiple word lines corresponding to the input, and the above analog current value is obtained by adding the current values flowing through multiple non-volatile memory elements storing analog resistance values (conductance) corresponding to the coupling weight coefficient. Such a neural network operation circuit using non-volatile memory elements can achieve low power consumption, and in recent years, resistance variable non-volatile memories that can set analog resistance values (conductance) are being actively developed.
[0007] Patent Document 2 discloses an example of a semiconductor integrated circuit capable of realizing a highly integrated memory cell for holding coupling weight coefficients. The semiconductor integrated circuit includes a network configuration information holding circuit for holding network configuration information, wherein the network configuration information includes address information of memory cells to which coupling weight coefficients of the neural network are allocated in a memory array, and controls a word line driving circuit and a column selecting circuit when the neural network is operated based on the network configuration information, thereby operating an operation circuit, thereby being able to perform operations of an arbitrary neural network configuration.
[0008] Prior art literature
[0009] Patent Literature
[0010] Patent Document 1: International Publication No. 2019 / 049741
[0011] Patent Document 2: International Publication No. 2019 / 049686 Summary of the invention
[0012] Problems to be solved by the invention
[0013] However, in the configurations disclosed in Patent Documents 1 and 2, it is difficult to increase the speed of the computational operation of neurons.
[0014] Therefore, the present disclosure is made in view of the above-mentioned problems, and an object of the present disclosure is to provide a neural network operation circuit, a control circuit and a control method thereof that can perform neuron operations at high speed.
[0015] Means used to solve problems
[0016] In order to solve the above-mentioned problems, a neural network operation circuit related to a technical solution of the present disclosure comprises: a plurality of word lines; a plurality of bit lines arranged in a form of crossing the word lines; a plurality of memory cells composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines and the plurality of bit lines, respectively holding coupling weight coefficients of the neural network; a word line driving circuit capable of driving any one or more of the plurality of word lines to a selected state; a column selection circuit capable of selecting any bit line of the plurality of bit lines; an operation circuit, by using a current flowing through the bit line selected by the column selection circuit, performs a multiplication and addition operation of a plurality of coupling weight coefficients held by the plurality of memory cells connected to the bit line selected by the column selection circuit and input data represented by the driving state of the plurality of word lines, and performs an activation function operation on the result of the multiplication and addition operation. A neuron operation; a word line selection state signal generating circuit, which generates a word line selection state signal, and the word line selection state signal indicates the word line set to the selected state by the above-mentioned word line driving circuit; a timing generating circuit, which outputs a word line start signal for starting the above-mentioned word line driving circuit based on the above-mentioned word line selection state signal, a column selection signal for driving the above-mentioned column selection circuit, and an operation circuit control signal for starting the above-mentioned operation circuit; an operation result processing circuit, which processes the output of the above-mentioned operation circuit, that is, the operation result; and a word line selection number management circuit, which manages the number of selected word lines and sends it to the above-mentioned timing generating circuit, the above-mentioned selected word line number is information related to the number of word lines that become the selected state when performing the above-mentioned multiplication and addition operation, and the above-mentioned timing generating circuit sets the delay time from the output of the above-mentioned word line start signal to the output of the above-mentioned operation circuit control signal according to the above-mentioned selected word line number.
[0017] In order to solve the above-mentioned problems, a control circuit of a neural network operation circuit related to a technical solution of the present disclosure is a control circuit for controlling the neural network operation circuit, and comprises the word line selection state signal generating circuit, the timing generating circuit, the operation result processing circuit, and the word line selection number management circuit constituting the above-mentioned neural network operation circuit; the above-mentioned timing generating circuit sets the delay time from the output of the above-mentioned word line start signal to the output of the above-mentioned operation circuit control signal according to the above-mentioned selected word line number.
[0018] In order to solve the above-mentioned problems, a control method of a neural network operation circuit related to a technical solution of the present disclosure is provided, wherein the neural network operation circuit comprises: a plurality of word lines; a plurality of bit lines arranged in a form of crossing the word lines; a plurality of memory cells composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines and the plurality of bit lines, respectively maintaining coupling weight coefficients of the neural network; a word line driving circuit capable of driving any one or more of the plurality of word lines into a selected state; a column selection circuit capable of selecting any bit line of the plurality of bit lines; and an operation circuit that implements an input operation represented by the driving state of the plurality of word lines by using a current flowing through the bit line selected by the column selection circuit. The multiplication and addition operation of data is performed, and the operation of neurons is performed by performing the operation processing of the activation function on the result of the above-mentioned multiplication and addition operation; the above-mentioned control method includes: a step of setting the word line selection state when driving any one or more word lines among the above-mentioned multiple word lines to a selected state; a delay time setting step, setting the delay time according to the number of selected word lines, and the above-mentioned number of selected word lines is information associated with the number of word lines that become the selected state when performing the above-mentioned multiplication and addition operation; a word line driving step, driving the above-mentioned word line according to the above-mentioned word line selection state; and an operation circuit starting step, starting the above-mentioned operation circuit, and the above-mentioned delay time is the period from the above-mentioned word line driving step to the above-mentioned operation circuit starting step.
[0019] Effects of the Invention
[0020] By using the neural network operation circuit, its control circuit and its control method related to the present disclosure, it is possible to perform neuron operations at a high speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 This is a diagram showing an example of the configuration of a neural network.
[0022] Figure 2 It is a diagram showing neurons in a neural network operation.
[0023] Figure 3 Yes means Figure 2 Diagram of the computation performed by the neuron shown.
[0024] Figure 4 is a diagram showing a step function as an example of an activation function of a neuron in a neural network operation.
[0025] Figure 5 This is a block diagram showing the structure of a neural network operation circuit according to the first embodiment.
[0026] Figure 6 Yes means Figure 5 A diagram showing an example of the configuration of a memory array included in a neural network operation circuit.
[0027] Figure 7 Yes means Figure 6 A diagram showing an example of the configuration of a memory cell.
[0028] Figure 8 Yes means Figure 5 A block diagram showing an example of the configuration of a control circuit included in a neural network operation circuit shown in FIG.
[0029] Fig. 9 Yes means Figure 5 FIG. 2 is a diagram showing an example configuration of a word line driving circuit included in a neural network operation circuit and a plurality of word lines and a driver power supply.
[0030] Fig.10 It means by Fig. 9 The diagram shows the operation of the word line driving circuit when one word line is selected.
[0031] Fig.11 It means by Fig. 9 The diagram shown is a diagram of the operation of the word line driving circuit when k word lines are selected.
[0032] Fig.12 It means by Fig. 9 FIG. 1 is a diagram showing the relationship between the number of word lines selected by the word line driving circuit (number of selected WL lines) and the delay time Tset for word line setting.
[0033] Fig.13 It means in Figure 5 The neural network operation circuit shown is a diagram of a configuration example in which four operation circuits are installed.
[0034] Fig.14 This is a diagram showing a specific example of a neural network configuration using the neural network operation circuit according to the first embodiment.
[0035] Fig.15 It means from Fig.14 Figure 2 shows an example of output data of the input layer, hidden layer, and output layer of the neural network shown.
[0036] Fig.16 It is a diagram showing an operation waveform of a hidden layer using the neural network operation circuit according to the first embodiment.
[0037] Fig.17 It is a diagram showing an operation waveform of an output layer using the neural network operation circuit according to the first embodiment.
[0038] Fig.18 It is a diagram showing the operation flow of the neural network operation circuit according to the first embodiment.
[0039] Fig.19It is a diagram showing a configuration example of a control circuit included in the neural network operation circuit according to the second embodiment.
[0040] Fig. 20 Yes means Fig.19 FIG. 1 is a diagram showing an example of network configuration information stored in a network configuration information storage circuit included in the control circuit shown.
[0041] Fig.21 It is a diagram showing an operation waveform of a hidden layer using the neural network operation circuit according to the second embodiment.
[0042] Fig. 22 It is a diagram showing an operation waveform of an output layer using the neural network operation circuit according to the second embodiment.
[0043] Fig.23 It is a diagram showing the operation flow of the neural network operation circuit according to the second embodiment. DETAILED DESCRIPTION
[0044] (Knowledge obtained by the inventors of this application)
[0045] The configurations disclosed in Patent Documents 1 and 2 described above have the following problems.
[0046] That is, when operating a neural network, it is necessary to select a plurality of word lines corresponding to the input state of the neuron at a time, but due to the operation of a plurality of word line drivers when driving the plurality of word lines, a word line setup time is required until the power supply potential of the word line driver drops and the charge is supplied from the power supply circuit. The more word lines are selected at the same time, the longer the setup time is.
[0047] Therefore, in order to realize an action based on an arbitrary neural network structure and arbitrary input to a neuron, it is necessary to set a larger fixed delay time equivalent to the setup time of the word line corresponding to the total number of word lines of the mounted memory array, and then drive the operation circuit, so it is difficult to realize high-speed operation of the neuron.
[0048] Therefore, the inventors of the present application have set up a control circuit in a neural network operation circuit, and the control circuit comprises: a word line selection number management circuit, which manages the number of selected word line lines, and the above-mentioned number of selected word line lines is information associated with the number of word lines that become a selected state when performing multiplication and addition operations; and a timing generation circuit, which sets the delay time from the output of the word line start signal for starting the word line driving circuit to the output of the operation circuit control signal for starting the operation circuit according to the number of selected word line lines.
[0049] Thus, the delay time is dynamically determined according to the actual number of selected word lines, so it is no longer necessary to set a large fixed delay time as in the past, thereby realizing a neural network operation circuit, its control circuit and its control method that can perform neuron operations at high speed.
[0050] (Implementation Method)
[0051] Hereinafter, the embodiments of the present disclosure will be described with reference to the accompanying drawings. In addition, the embodiments described below all represent a specific example of the present disclosure. The numerical values, shapes, materials, constituent elements, configuration positions of constituent elements and connection forms, steps, order of steps, etc. shown in the following embodiments are examples and are not intended to limit the present disclosure. In addition, the figures are not necessarily strictly illustrated. In each figure, the same reference numerals are given to substantially the same configurations, and repeated descriptions are omitted or simplified. In addition, "connection" refers to electrical connection, which includes not only the case where two circuit elements are directly connected, but also the case where two circuit elements are indirectly connected in a state where other circuit elements are inserted between the two circuit elements.
[0052] <Neural network operation>
[0053] First, the basic theory of neural network operation is explained.
[0054] Figure 1 This is a diagram showing an example of the configuration of a neural network (here, a deep neural network). The neural network is composed of an input layer 101 for inputting input data, a hidden layer 102 (sometimes referred to as an intermediate layer) for receiving input data from the input layer 101 and performing calculations, and an output layer 103 for receiving output data from the hidden layer 102 and performing calculations.
[0055] In each of the input layer 101, the hidden layer 102, and the output layer 103, there are many basic elements of the neural network called neurons 100, and each neuron 100 is connected via coupling weights 104. The plurality of coupling weights 104 have different coupling weight coefficients to connect the neurons 100. A plurality of input data is input to the neuron 100, and the multiplication and accumulation operation of the plurality of input data and the corresponding coupling weight coefficient is performed in the neuron 100, and the output data is output as output data. Here, the hidden layer 102 is a layer (in the Figure 1 In terms of forming a deep neural network, it is composed of 4 layers of neurons. Figure 1 The neural network shown is called a deep neural network.
[0056] Figure 2 It is a diagram showing neurons in a neural network operation. Figure 3 Yes means Figure 2The diagram of the calculation formula executed by the neuron 100 shown in FIG. The k inputs x1 to xk of the neuron 100 are connected with coupling weights having coupling weight coefficients w1 to wk, respectively, and the multiplication and addition operation of the inputs x1 to xk and the coupling weight coefficients w1 to wk is performed. In addition, the neuron 100 has an activation function f, and the activation function is operated on the result of the multiplication and addition operation, and the output y is output.
[0057] Figure 4 is a graph showing a step function as an example of an activation function f of a neuron in a neural network operation. The horizontal axis is the input u of the activation function f, and the vertical axis is the output f(u) of the activation function f. Figure 4 As shown, the step function is a function that outputs f(u)=0 when the input u is a negative value (<0), and outputs f(u)=1 when the input u is a positive value (≥0). Figure 2 In the neuron 100, when the activation function f of the step function is used, when the result of the multiplication and addition operation of the input x0~xn and the coupling weight coefficients w0~wn is a negative value, the output y=0 is output, and when the result of the multiplication and addition operation is a positive value, the output y=1 is output.
[0058] <First Embodiment>
[0059] Figure 5 1 is a block diagram showing the structure of the neural network operation circuit according to the first embodiment. The neural network operation circuit is composed of the following parts: a memory array 500, in which memory cells connected to a plurality of word lines 501 and a plurality of bit lines 502 are arranged in a matrix; a word line driving circuit 503, which sets any word line among the plurality of word lines 501 to be selected or unselected; an integer number n of operation circuits 5051 to 505n greater than 1; a column selection circuit 504, which selects any one or more of the plurality of bit lines 502 and is connected to the n operation circuits 5051 to 505n respectively; and a control circuit 506.
[0060] The operation circuits 5051 to 505n use the current flowing through the bit line 502 selected by the column selection circuit 504 to perform a multiplication-addition operation on a plurality of coupling weight coefficients maintained by a plurality of memory cells connected to the bit line 502 selected by the column selection circuit 504 and input data represented by the driving states of a plurality of word lines 501, and perform neuron operations by performing activation function operations on the results of the multiplication-addition operations.
[0061] Figure 6 Yes means Figure 5FIG. 1 is a diagram showing an example of the configuration of a memory array 500 provided in a neural network operation circuit. A semiconductor storage element, namely a memory cell 600, is arranged in a matrix shape and is connected to a plurality of word lines 501 (WLk, ··, WLk-1, ··, WL2, WL1) and a plurality of bit lines 502 (BL11, BL12, ··, BL1u, BL21, ··BLnu).
[0062] The plurality of word lines 501 are composed of an integer number k greater than or equal to 1.
[0063] The plurality of bit lines 502 are logically divided into n units using an integer u greater than or equal to 1 as one unit, and one or more of the divided bit lines are connected to n operation circuits 5051 to 505 n respectively via a column selection circuit 504 .
[0064] For example, one or more bit lines out of u bit lines BL11 to BL1u are connected to the operation circuit 5051, one or more bit lines out of an integer number u bit lines BL21 to BL2u are connected to the operation circuit 5052, and similarly one or more bit lines out of u bit lines BLn1 to BLnu are connected to the operation circuit 505n.
[0065] Figure 7 Yes means Figure 6 The figure shows an example of the structure of a memory cell 600. It is a memory cell (ReRAM) formed by connecting a transistor T0 having a word line connected to the gate and a resistance variable memory element R in series, and connecting a bit line BL to one end of the resistance variable memory element R and connecting a source line SL to one end of the transistor T0.
[0066] In the resistance variable memory element R, the coupling weight coefficient of the neuron is stored as a resistance value, and the resistance value is determined by Figure 5 The writing circuit setting and saving are not shown in the figure. Figure 7 In the example, a resistance change memory element (ReRAM) is used as the memory unit 600, but other non-volatile memory elements, such as a magnetoresistive memory element (MRAM) and a phase change resistance memory element (PRAM), may also be used. In addition, a charge storage element such as a flash memory whose threshold changes according to the amount of charge stored may also be used.
[0067] Figure 8 Yes means Figure 5 The control circuit 506 is a block diagram of the structure of the neural network operation circuit shown in FIG. The control circuit 506 is composed of a word line selection state signal generation circuit 801, a timing generation circuit 802, an operation result processing circuit 803, and a word line selection number management circuit 804.
[0068] The word line selection state signal generation circuit 801 generates a word line selection state signal 507 (WL_IN[k:1]) selected by the word line driving circuit 503 .
[0069] The calculation result processing circuit 803 uses the calculation result 511 (Y[n:1]) of n "0" or "1" obtained by the calculation circuits 5051~505n as the calculation result of the hidden layer or output layer of the neural network. When it is the calculation result of the hidden layer, it is processed as the input data of the next layer connected to the hidden layer, and when it is the calculation result of the output layer, it is output as the action result of the neural network.
[0070] The word line selection number management circuit 804 receives the input data equivalent to the input layer of the neural network or the result processed by the operation result processing circuit 803 as input data, and sends it to the word line selection state signal generation circuit 801 as data indicating the selected word line during the multiplication and addition operation. In addition, the word line selection number management circuit 804 counts the number of bits indicating the selection state of the word line contained in the data (here, it is set to be selected in the case of "1") as the number of selected word lines, and sends it to the timing generation circuit 802.
[0071] here, Fig. 9 It indicated Figure 5 The neural network operation circuit shown includes a word line driving circuit 503, a plurality of word lines 501, and a memory array 900 in which a word line load driven by the word line driving circuit 503 in the memory array 500 is simplified as a capacitive load.
[0072] The word line driving circuit 503 includes a NAND gate 903 and a word line driver 902 corresponding to each of the plurality of word lines 501. Based on a word line selection state signal 507 (WL_IN[k:1]) and a word line enable signal 508 (WL_EN), the word line driver 902 selects any word line of the word lines WL1 to WLk. The power supply (WL_P) of the word line driver 902 during word line selection is supplied by a power supply circuit 901 inside or outside the chip.
[0073] exist Fig.10 Indicated in Fig. 9 The operation of the word line driving circuit 503 when one word line WL1 is selected in the configuration of Fig.11 2 shows the operation of the word line driving circuit 503 when a maximum of k word lines WL1 to WLk of WL1 to WLk are selected.
[0074] exist Fig.10In the figure, the word line enable signal 508 (WL_EN) is transferred to High, the word line driver 902 charges the word line WL1, and consumes the charge of the driver power supply (WL_P), so that the power supply potential of the power supply WL_P decreases, and then recovers through the charge supply of the power supply circuit 901. It can be seen that the delay time t1 is required as the setting time for selecting the word line WL1.
[0075] The amount of power supply potential drop and the recovery time based on charge supply vary depending on the number of word lines selected. Fig.11 As shown, when k word lines, which is the maximum number, are selected, the amount of drop in the power supply potential increases, and a delay time tk longer than t1 is required for the setup time to select the word lines WL1 to WLk.
[0076] Fig.12 Indicated by Fig. 9 The relationship between the number of selected word lines ("number of selected word lines") selected by the word line driving circuit 503 and the delay time ("Tset") corresponding to the required word line setup is shown. The delay time Tset increases as the number of selected word lines increases.
[0077] The timing generation circuit 802 generates a word line start signal 508 (WL_EN), and a column selection signal 509 (BL_IN[u:1]) that connects a plurality of bit lines 502 with operation circuits 5051 to 505n via a column selection circuit 504, and generates an operation circuit control signal 510 (OP_EN) after a delay time Tset corresponding to the word line setup time.
[0078] Here, the timing generation circuit 802 generates a timing signal based on the relationship between the number of selected word lines counted by the word line selection number management circuit 804 and the required rise time ( Fig.12 ) to set the delay time Tset.
[0079] Next, use Figure 8 and Figure 13 to Figure 17 A specific operation example using the neural network operation circuit according to the first embodiment will be described.
[0080] Fig.13 Indicated in Figure 5 The structure of the neural network operation circuit shown in FIG. 1 is a case where four operation circuits (5051 to 5054) are installed. Figure 5 It's the same.
[0081] Fig.14This is a diagram showing a specific example of a neural network configuration using a neural network operation circuit according to the first embodiment. Here, a specific neural network configuration example in the operation example described later is shown, where the input layer 1401 has four nodes a1 to a4, the hidden layer 1402 has two nodes b1 and b2, and the output layer 1403 has two nodes c1 and c2. In each of the input layer 1401, the hidden layer 1402, and the output layer 1403, there are many basic elements of the neural network called neurons 1400, and each neuron 1400 is connected via a coupling weight 1404. The plurality of coupling weights 1404 have different coupling weight coefficients and connect the neurons 1400.
[0082] Fig.15 It means from Fig.14 The diagram shows an example of output data of the input layer 1401, hidden layer 1402, and output layer 1403 of the neural network shown. Here, the calculation results of the nodes a1 to a4 of the input layer 1401, the nodes b1 and b2 of the hidden layer 1402, and the nodes c1 and c2 of the output layer 1403 are shown.
[0083] Fig.16 14 is a diagram showing an operation waveform (i.e., a signal waveform) of the hidden layer 1402 using the neural network operation circuit according to the first embodiment. Fig.14 The operation of the hidden layer 1402 of the neural network is used Fig.13 The operation waveform of the neural network operation circuit shows that the input data of the hidden layer 1402 during the operation is Fig.15 The case shown is (a4, a3, a2, a1) = (1, 1, 1, 1).
[0084] The signal BL_sel1 indicates a bit line selected during the calculation operation of the hidden layer 1402 among the plurality of bit lines 502 .
[0085] Signal Y[4:1] represents the operation result 511 of "0" or "1" of the four operation circuits 5051 to 5054. Among the four signals of Y[4:1], Y[4] is equivalent to the operation result 511 of the operation circuit 5054, Y[3] is equivalent to the operation result 511 of the operation circuit 5053, Y[2] is equivalent to the operation result 511 of the operation circuit 5052, and Y[1] is equivalent to the operation result 511 of the operation circuit 5051.
[0086] The word line selection number management circuit 804 receives data corresponding to the input data a1 to a4 from the external input, and outputs the word line selection state signal WL_IN[k:1]=0xf to the word line driving circuit 503 via the word line selection state signal generation circuit 801. In addition, the word line selection number management circuit 804 counts the number of data ("1") corresponding to the WL selection included in the input data a1 to a4, and sends it to the timing generation circuit 802 as the number of selected word line lines (4).
[0087] The timing generation circuit 802 sets a delay time t4 corresponding to the setup time of the number of selected word lines (4), and causes the word line enable signal WL_EN to transition to High, thereby causing the word line driver circuit 503 to start selecting the four word lines WL1 to WL4, and causes the column select circuit 504 to start selecting the bit line BL_sel1 by transitioning a signal corresponding to one or more bit lines BL_sel1 to be selected in the column select signal BL_IN[u:1] to transition to High. Then, the timing generation circuit 802 causes the operation circuit control signal OP_EN to transition to High after the set delay time t4, thereby enabling the operation circuits 5051 to 5054 to start.
[0088] The operation circuits 5051 to 5054 perform multiplication and addition operations using the flowing current according to the resistance value stored in the memory cell 600 as a coupling weight coefficient, and then output the operation result Y[4:1]=0x1 by performing operation processing of the activation function. The above-mentioned memory cell 600 is configured at the intersection of the selected word lines WL1~WL4 and the selected bit line BL_sel1.
[0089] Fig.17 is a diagram showing an operation waveform (i.e., a signal waveform) of the output layer 1403 using the neural network operation circuit according to the first embodiment. Fig.14 The operation of the output layer 1403 composed of the neural network of Fig.13 The operation waveform of the output layer 1403 in the case of the neural network operation circuit shown in FIG. 1 shows that the input data during the operation of the output layer 1403 is the operation result of the hidden layer 1402. Fig.15 The case where (b2, b1) = (0, 1) is shown.
[0090] BL_sel2 indicates a bit line selected in the operation of the output layer 1403 among the plurality of bit lines 502 .
[0091] The word line selection number management circuit 804 receives data corresponding to the input data b1 and b2 from the calculation result processing circuit 803, and outputs the word line selection state signal WL_IN[k:1]=0x1 to the word line driving circuit 503 via the word line selection state signal generation circuit 801. In addition, the word line selection number management circuit 804 counts the number of data corresponding to WL selection ("1") included in the input data b1 and b2, and sends it to the timing generation circuit 802 as the number of selected word lines (1).
[0092] The timing generation circuit 802 sets a delay time t1 corresponding to the setup time of the number of selected word lines (1), and starts the word line selection of WL2 by turning the word line enable signal WL_EN High, and starts the selection of BL_sel2 by turning the signal corresponding to one or more bit lines (BL_sel2) to be selected in the column select signal BL_IN[u:1] High. Then, the timing generation circuit 802 starts the operation circuits 5051 to 5054 by turning the operation circuit control signal OP_EN High after the set delay time t1.
[0093] The operation circuits 5051 to 5054 perform multiplication and addition operations using the flowing current according to the resistance value stored in the memory cell 600 as a coupling weight coefficient, and then perform operation processing of the activation function, thereby outputting the operation result Y[4:1]=0x2. The above-mentioned memory cell 600 is configured at the intersection of the selected word line WL1 and the selected bit line BL_sel2.
[0094] The operation result processing circuit 803 receives the operation result 511 as Fig.14 The result of the neural network action is output to the outside.
[0095] Fig.18 This is a diagram showing an operation flow of a neural network operation using the neural network operation circuit according to the first embodiment (that is, a control method of the neural network operation circuit).
[0096] After the operation starts (1800), the word line selection number management circuit 804 checks the word line selection state information determined based on the external input or the calculation result of the previous layer as the input of the calculation operation target layer (1801).
[0097] Based on the result, the word line selection state signal generation circuit 801 generates a word line selection state signal 507 and outputs it to the word line driving circuit 503. The word line selection state signal 507 determines which word line among the plurality of word lines of the memory array is selected (1802).
[0098] Next, the word line selection number management circuit 804 counts the number of word lines to be selected according to the word line selection state information, and outputs it as the number of selected word line lines to the timing generation circuit 802. Therefore, the timing generation circuit 802 sets the delay time Tset as the word line setting time according to the received number of selected word line lines (1803). In addition, the timing generation circuit 802 outputs a word line start signal 508 to the word line driving circuit 503, and outputs a column selection signal 509 to the column selection circuit 504, and outputs an operation circuit control signal 510 to the operation circuits 5051 to 5054 after the set delay time Tset.
[0099] As a result, the word line driving circuit 503 receiving the word line selection state signal 507 and the word line activation signal 508 and the column selection circuit 504 receiving the bit line selection information and the column selection signal 509 (not shown) select the word line and the bit line, respectively (1804).
[0100] Next, after the standby delay time Tset (1805), the operation circuits 5051 to 5054 operate (1806), output the operation result 511, and end (1807).
[0101] Through the above actions, select Fig.14 After any number of word lines required for the operation of the neural network are connected, the delay time Tset until the operation circuits 5051 to 5054 are activated is changed according to the number of word lines selected during the actual operation. Therefore, it is not necessary to set a larger fixed setup time taking into account the total number of word lines of the memory array 900 as the delay time as in the past, and the operation of the neuron 1400 can be performed at a high speed.
[0102] As described above, the neural network operation circuit related to the present embodiment comprises: a plurality of word lines 501; a plurality of bit lines 502 arranged in a form of crossing the word lines 501; a plurality of memory cells 600 composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines 501 and the plurality of bit lines 502, respectively maintaining coupling weight coefficients of the neural network; a word line driving circuit 503 capable of driving any one or more of the plurality of word lines 501 to a selected state; a column selection circuit 504 capable of selecting any bit line 502 of the plurality of bit lines 502; operation circuits 5051 to 505n, by using the current flowing through the bit line 502 selected by the column selection circuit 504, perform multiplication and accumulation operations of the plurality of coupling weight coefficients maintained by the plurality of memory cells 600 connected to the bit line 502 selected by the column selection circuit 504 and the input data represented by the driving state of the plurality of word lines 501, and perform operation processing of the activation function on the result of the multiplication and accumulation operations, thereby performing neuron operation. operation; a word line selection state signal generating circuit 801, generating a word line selection state signal 507 indicating that the word line 501 is set to a selected state by the word line driving circuit 503; a timing generating circuit 802, outputting a word line starting signal 508 for starting the word line driving circuit 503 based on the word line selection state signal 507, a column selection signal 509 for driving the column selection circuit 504, and an operation circuit control signal 510 for starting the operation circuits 5051~505n; an operation result processing circuit 803, processing the operation result 511 as the output of the operation circuits 5051~505n; and a word line selection number management circuit 804, managing the number of selected word line bars and sending it to the timing generating circuit 802, the above-mentioned number of selected word line bars is information associated with the number of word lines 501 that become the selected state when performing the multiplication and addition operation; the timing generating circuit 802 sets the delay time from the output of the word line starting signal 508 to the output of the operation circuit control signal 510 according to the number of selected word line bars.
[0103] Therefore, after the word line is selected, the delay time until the operation circuit is activated is changed according to the information associated with the number of word lines that actually become selected, that is, the number of selected word lines. Therefore, it is no longer necessary to set a larger fixed setup time corresponding to the total number of word lines in the memory array as the delay time as in the past, and the neuron operation can be performed at a high speed.
[0104] Here, the number of selected word lines is specifically a value obtained by counting the number of word lines 501 that are in a selected state, which is included in the data input from the outside of the neural network operation circuit or the output data from the operation result processing circuit 803, by the word line selection number management circuit 804. Thus, a delay time corresponding to the number of word lines that are actually in a selected state is set.
[0105] In addition, the memory cell 600 is composed of a resistance change memory element, a magnetoresistive memory element, a phase change resistance memory element, or a charge memory element whose threshold value changes according to the amount of stored charge. Thus, the memory cell is implemented by a non-volatile memory element, and the stored information is continuously retained even if the power supply is disconnected.
[0106] In addition, the control circuit 506 for controlling the neural network operation circuit according to the present embodiment comprises a word line selection state signal generating circuit 801, a timing generating circuit 802, an operation result processing circuit 803 and a word line selection number management circuit 804 which constitute the above-mentioned neural network operation circuit. The timing generating circuit 802 sets the delay time from the output of the word line start signal 508 to the output of the operation circuit control signal 510 according to the number of selected word line lines.
[0107] In addition, the control method of the neural network operation circuit of the present embodiment is a control method of the neural network operation circuit, and the neural network operation circuit comprises: a plurality of word lines 501; a plurality of bit lines 502 arranged in a form of crossing the word lines 501; a plurality of memory cells 600 composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines 501 and the plurality of bit lines 502, respectively maintaining the coupling weight coefficients of the neural network; a word line driving circuit 503 capable of driving any one or more word lines 501 among the plurality of word lines 501 to a selected state; a column selection circuit 504 capable of selecting any bit line 502 among the plurality of bit lines 502; and operation circuits 5051 to 505n, which implement the driving of the plurality of word lines 501 by using the current flowing through the bit line 502 selected by the column selection circuit 504. The control method comprises: a step of setting a word line 501 selection state when any one or more word lines 501 among a plurality of word lines 501 are driven to a selection state; a delay time setting step of setting a delay time according to the number of selected word lines, wherein the number of selected word lines is information associated with the number of word lines 501 that become a selection state when performing a multiplication and accumulation operation; a word line 501 driving step of driving the word line 501 according to the word line 501 selection state; and a step of starting the operation circuits 5051~505n, starting the operation circuits 5051~505n; the delay time is the period from the word line 501 driving step to the operation circuit 5051~505n starting step.
[0108] Therefore, after the word line is selected, the delay time until the operation circuit is activated is changed according to the information associated with the number of word lines that actually become selected, that is, the number of selected word lines. Therefore, it is no longer necessary to set a larger fixed setup time corresponding to the total number of word lines in the memory array as the delay time as in the past, and the neuron operation can be performed at a high speed.
[0109] Here, the number of selected word lines is the number of word lines 501 in the selected state included in the selection state of the word lines 501. Thus, a delay time is set according to the number of word lines actually in the selected state.
[0110] <Second Embodiment>
[0111] Next, a neural network operation circuit according to the second embodiment will be described. The neural network operation circuit according to this embodiment basically has the following features: Figure 5 The structure of the neural network operation circuit of the first embodiment is the same as that of the neural network operation circuit of the first embodiment shown in FIG. However, the structure of the control circuit is different from that of the first embodiment. Hereinafter, the control circuit of this embodiment is referred to as control circuit 506a, and the description will be centered on the differences from the first embodiment.
[0112] Fig.19 FIG. 5 is a diagram showing the structure of a control circuit 506a included in the neural network operation circuit according to the second embodiment. Figure 8 The control circuit 506 is newly equipped with a network configuration information holding circuit 1906.
[0113] In the network configuration information holding circuit 1906, the number of layers of the neural network and the number of nodes in each layer are stored as network configuration information.
[0114] Fig. 20 It means in Fig.19 FIG. 19 is a diagram showing an example of network configuration information stored in the network configuration information storage circuit 1906 of the control circuit shown in FIG. Fig.14 An example of information stored in the network configuration information holding circuit 1906 in the case of a neural network is that 1, 2, and 3 are assigned as layer IDs for the input layer 1401, the hidden layer 1402, and the output layer 1403, respectively, and the number of nodes in each layer is stored (4, 2, 2).
[0115] The word line selection number management circuit 1904 receives input data equivalent to the input layer of the neural network or the result of the previous layer processed by the operation result processing circuit 803 as input data, and sends it to the word line selection state signal generation circuit 801 as data representing the selected word line during the multiplication and addition operation.
[0116] In addition, the word line selection number management circuit 1904 receives information corresponding to the number of nodes of the neural network layer of the action target in the network configuration information from the network configuration information holding circuit 1906, and sends it to the timing generation circuit 802 as the number of selected word line lines.
[0117] Other configurations are similar to those of the first embodiment. Figure 8The control circuit 506 is the same.
[0118] Next, use Figures 13-15 and Figure 19 to Figure 22 A specific operation example using the neural network operation circuit according to the second embodiment will be described.
[0119] Fig.21 14 is a diagram showing an operation waveform (i.e., a signal waveform) of the hidden layer 1402 using the neural network operation circuit according to the second embodiment. Fig.14 The operation of the hidden layer 1402 composed of the neural network of Fig.13 The operation waveform of the neural network operation circuit is shown, and the input data when the operation of the hidden layer 1402 is Fig.15 The case shown is (a4, a3, a2, a1) = (1, 1, 1, 1).
[0120] The word line selection number management circuit 1904 receives data corresponding to the input data a1 to a4 from the external input, and outputs the word line selection state signal WL_IN[k:1]=0xf to the word line driving circuit 503 via the word line selection state signal generation circuit 801. In addition, the word line selection number management circuit 1904 receives 4 as the number of nodes of the layer ID1 corresponding to the input layer 1401 stored in the network configuration information holding circuit 1906, and sends it as the number of selected word line bars (4) to the timing generation circuit 802. In addition, the timing generation circuit 802 sets a delay time t4 corresponding to the setting time of the number of selected word line bars (4).
[0121] Other circuit actions and waveforms are Figure 8 , Fig.16 It's the same.
[0122] Fig. 22 is a diagram showing an operation waveform (i.e., a signal waveform) of the output layer 1403 using the neural network operation circuit according to the second embodiment. Fig.14 The operation of the output layer 1403 composed of the neural network of Figure 5 The operation waveform of the neural network operation circuit is shown in FIG. 1403, and the input data of the operation operation of the output layer 1403 is the operation result of the hidden layer 1402. Fig.15 The case where (b2, b1) = (0, 1) is shown.
[0123] The word line selection number management circuit 1904 receives data corresponding to the input data b1 and b2 from the calculation result processing circuit 803, and outputs the word line selection state signal WL_IN[k:1]=0x1 to the word line driving circuit 503 via the word line selection state signal generation circuit 801. In addition, the word line selection number management circuit 1904 receives 2, which is the number of nodes corresponding to the layer ID2 of the hidden layer 1402, stored in the network configuration information holding circuit 1906, and sends it to the timing generation circuit 802 as the number of selected word line lines (2).
[0124] The timing generation circuit 802 sets the delay time t2 corresponding to the number of selected word line bars (two).
[0125] Other circuit actions and waveforms are Figure 8 , Fig.17 It's the same.
[0126] Fig.23 This is a diagram showing an operation flow of a neural network operation using the neural network operation circuit according to the second embodiment (a control method of the neural network operation circuit).
[0127] After the operation starts (2300), the word line selection number management circuit 1904 checks the word line selection state information (2301) determined based on the external input or the calculation result of the previous layer as the input of the calculation operation target layer.
[0128] Based on the result, the word line selection state signal generation circuit 801 generates a word line selection state signal, and outputs it to the word line driving circuit 503. The word line selection state signal determines which word line among the plurality of word lines of the memory array is selected (2302).
[0129] Next, the word line selection number management circuit 1904 receives the node number of the layer ID1 corresponding to the input layer stored in the network configuration information holding circuit 1906, and sends it to the timing generation circuit 802 as the number of selected word line bars, so the timing generation circuit 802 sets the delay time Tset (2303) corresponding to the word line setting time based on the node number of the received network configuration information. Next, the timing generation circuit 802 outputs the word line start signal 508 to the word line driving circuit 503, and outputs the column selection signal 509 to the column selection circuit 504, and outputs the operation circuit control signal 510 to the operation circuits 5051 to 5054 after the set delay time Tset.
[0130] As a result, the word line driving circuit 503 receiving the word line selection state signal 507 and the word line activation signal 508 and the column selection circuit 504 receiving the bit line selection information and the column selection signal 509 (not shown) select the word line and the bit line respectively (2304).
[0131] Next, after the standby delay time Tset (2305), the operation circuits 5051 to 5054 operate (2306), output the operation result 511, and end (2307).
[0132] Through the above actions, select Fig.14 After any number of word lines required for the operation of the neural network are input, the delay time Tset until the operation circuits 5051 to 5054 are activated is changed according to the number of nodes input during the actual operation. Therefore, it is not necessary to set a larger fixed setup time considering the total number of word lines of the memory array 900 as the delay time as in the past, which can speed up the operation of the neuron 100.
[0133] In the first embodiment, if Figure 8 The control circuit and Fig.18 As shown in the action flow, the number of word lines selected in the actual action is counted, and the switching delay time Tset is switched. In contrast, in this embodiment, as shown in FIG. Fig.19 The control circuit and Fig.23 As shown in the action flow, the number of nodes using the network configuration information may increase the delay time Tset, but there is no need to count the number of "1" data contained in the input data. The control circuit 506a can be implemented with a simple circuit compared to the first embodiment.
[0134] As described above, the neural network operation circuit related to the present embodiment comprises: a plurality of word lines 501; a plurality of bit lines 502 arranged in a form of crossing the word lines 501; a plurality of memory cells 600 composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines 501 and the plurality of bit lines 502, respectively maintaining coupling weight coefficients of the neural network; a word line driving circuit 503 capable of driving any one or more of the plurality of word lines 501 to a selected state; a column selection circuit 504 capable of selecting any bit line 502 of the plurality of bit lines 502; operation circuits 5051 to 505n, by using the current flowing through the bit line 502 selected by the column selection circuit 504, perform multiplication and accumulation operations of the plurality of coupling weight coefficients maintained by the plurality of memory cells 600 connected to the bit line 502 selected by the column selection circuit 504 and the input data represented by the driving state of the plurality of word lines 501, and perform neural network operation by performing operation processing of the activation function on the result of the multiplication and accumulation operations. operation; a word line selection state signal generating circuit 801, generating a word line selection state signal 507 indicating that the word line 501 is set to the selected state by the word line driving circuit 503; a timing generating circuit 802, outputting a word line starting signal 508 for starting the word line driving circuit 503 based on the word line selection state signal 507, a column selection signal 509 for driving the column selection circuit 504, and an operation circuit control signal 510 for starting the operation circuits 5051~505n; an operation result processing circuit 803, processing the operation result 511 as the output of the operation circuits 5051~505n; and a word line selection number management circuit 1904, managing the number of selected word line bars and sending it to the timing generating circuit 802, the above-mentioned number of selected word line bars is information associated with the number of word lines 501 that become the selected state when performing the multiplication and addition operation; the timing generating circuit 802 sets the delay time from the output of the word line starting signal 508 to the output of the operation circuit control signal 510 according to the number of selected word line bars.
[0135] Here, the neural network operation circuit further includes a network configuration information holding circuit 1906 which stores the number of nodes which is the number of neurons in each layer of the neural network, and the number of selected word lines is the number of nodes stored in the network configuration information holding circuit 1906.
[0136] Thus, after a word line is selected, the delay time until the operation circuit is activated is changed according to the maximum number of word lines that can actually be selected, so it is not necessary to set a relatively large fixed setup time corresponding to the total number of word lines of the memory array as the delay time as in the past, and the neuron operation can be performed at a high speed. Furthermore, compared with the first embodiment, it is not necessary to count the number of "1" data included in the input data, so that the circuit is simplified.
[0137] In addition, the control method of the neural network operation circuit related to the present embodiment is a control method of the neural network operation circuit, the neural network operation circuit comprising: a plurality of word lines 501; a plurality of bit lines 502 arranged in a form of crossing the word lines 501; a plurality of memory cells 600 composed of semiconductor memory elements, arranged at the intersections of the plurality of word lines 501 and the plurality of bit lines 502, respectively maintaining the coupling weight coefficients of the neural network; a word line driving circuit 503 capable of driving any one or more word lines 501 among the plurality of word lines 501 to a selected state; a column selection circuit 504 capable of selecting any bit line 502 among the plurality of bit lines 502; operation circuits 5051 to 505n, by using the current flowing through the bit line 502 selected by the column selection circuit 504, implement the control of the plurality of word lines 501 with respect to the bit lines 502 selected by the column selection circuit 504. The control method comprises: a step of setting a word line 501 selection state when any one or more word lines 501 among a plurality of word lines 501 are driven to a selection state; a delay time setting step of setting a delay time according to the number of selected word lines which is information associated with the number of word lines 501 that become a selection state when performing the multiplication and accumulation operation; a word line 501 driving step of driving the word line 501 according to the word line 501 selection state; and a step of starting the operation circuits 5051~505n of starting the operation circuits 5051~505n; the delay time is a period from the word line 501 driving step to the operation circuit 5051~505n starting step.
[0138] Here, the neural network operation circuit stores the number of nodes, which is the number of neurons in each layer of the neural network, as network configuration information in the network configuration information storage circuit 1906, and the number of selected word lines is the number of nodes. Thus, a delay time corresponding to the maximum number of word lines that can actually be selected is set.
[0139] Thus, after a word line is selected, the delay time until the operation circuit is activated is changed according to the maximum number of word lines that can actually be selected, so it is not necessary to set a relatively large fixed setup time corresponding to the total number of word lines of the memory array as the delay time as in the past, and the neuron operation can be performed at a high speed. Furthermore, compared with the first embodiment, it is not necessary to count the number of "1" data included in the input data, so that the circuit is simplified.
[0140] Although the first and second embodiments of the present disclosure have been described above, the neural network operation circuit of the present disclosure is not limited to the above-mentioned examples, and various modifications are also effective without departing from the gist of the present disclosure.
[0141] For example, in the above embodiment, as a specific action example, the case where the number of computing circuits installed is 4, the maximum number of nodes of the neural network is 4, and the number of computing circuits is greater than the number of nodes, that is, the action of each layer of the input layer and the hidden layer is completed once, but the number of nodes of the neural network that can act is not limited by the number of computing circuits installed.
[0142] In addition, when the maximum number of nodes in the neural network is greater than the number of computing circuits mounted thereon, the computing operations of the neural network at each layer can be performed by performing multiple operations through the neural network computing circuit.
[0143] Industrial Applicability
[0144] The neural network operation circuit disclosed in the present invention can change the word line setting time according to the number of word line selections in the actual operation when multiple word lines are selected in the operation of any neural network structure, so that the operation of the neural network can be accelerated. The application of these technologies is practical for semiconductor integrated circuits equipped with artificial intelligence (AI) technology that can learn and judge by itself, and electronic devices equipped with them.
[0145] Description of symbols
[0146] 100, 1400 neurons
[0147] 101, 1401 Input layer
[0148] 102, 1402 hidden layers
[0149] 103, 1403 Output layer
[0150] 104, 1404 coupling weight
[0151] 500 Memory Array
[0152] 501 Word Line
[0153] 502 bit lines
[0154] 503 word line drive circuit
[0155] 504 Column Select Circuit
[0156] 5051~505n operation circuit
[0157] 506, 506a control circuit
[0158] 507 Word line selection status signal
[0159] 508 word line start signal
[0160] 509 Column selection signal
[0161] 510 Operation circuit control signal
[0162] 511 Operation result
[0163] 600 memory cells
[0164] 801 word line selection state signal generation circuit
[0165] 802 Timing Generation Circuit
[0166] 803 Operation result processing circuit
[0167] 804, 1904 Word line selection number management circuit
[0168] 900 Memory Array
[0169] 901 Power Circuit
[0170] 902 Word Line Driver
[0171] 903NAND Gate
[0172] 1906 Network configuration information holding circuit
[0173] T0 transistor
[0174] R resistance change memory element
Claims
1. A neural network operation circuit, wherein: have: multiple word lines; A plurality of bit lines are arranged in a manner crossing the word lines; A plurality of memory cells composed of semiconductor memory elements are arranged at intersections of the plurality of word lines and the plurality of bit lines, respectively maintaining coupling weight coefficients of the neural network; a word line driving circuit capable of driving any one or more word lines among the plurality of word lines to a selected state; A column selection circuit capable of selecting any bit line among the plurality of bit lines; The operation circuit performs a multiplication-accumulation operation of a plurality of coupling weight coefficients held by a plurality of the memory cells connected to the bit line selected by the column selection circuit and input data represented by the driving states of the plurality of word lines by using a current flowing through the bit line selected by the column selection circuit, and performs a neuron operation by performing an activation function operation on a result of the multiplication-accumulation operation; A word line selection state signal generating circuit that generates a word line selection state signal indicating the word line set to a selected state by the word line driving circuit; a timing generation circuit that outputs a word line activation signal for activating the word line driving circuit based on the word line selection state signal, a column selection signal for driving the column selection circuit, and an operation circuit control signal for activating the operation circuit; An operation result processing circuit processes the output of the operation circuit, that is, the operation result; and A word line selection number management circuit manages the number of selected word lines and sends it to the timing generation circuit. The number of selected word lines is information associated with the number of word lines that are selected when the multiplication and accumulation operation is performed. The timing generation circuit sets a delay time from output of the word line enable signal to output of the arithmetic circuit control signal according to the number of the selected word line lines.
2. The neural network operation circuit according to claim 1, wherein: The selected word line number is a value obtained by the word line selection number management circuit counting the number of word lines in a selected state included in data input from outside the neural network operation circuit or output data from the operation result processing circuit.
3. The neural network operation circuit according to claim 1, wherein: It also includes a network configuration information holding circuit, the network configuration information holding circuit storing the number of nodes as the number of neurons in each layer of the neural network, The number of selected word line bars is the number of nodes stored in the network configuration information holding circuit.
4. The neural network operation circuit according to any one of claims 1 to 3, wherein: The memory cell is composed of a resistance change memory element, a magnetoresistive memory element, a phase change resistance memory element, or a charge memory element whose threshold value changes according to the amount of stored charge.
5. A control circuit for controlling a neural network operation circuit, wherein: The neural network operation circuit according to any one of claims 1 to 4 comprises the word line selection state signal generation circuit, the timing generation circuit, the operation result processing circuit, and the word line selection number management circuit, The timing generation circuit sets a delay time from output of the word line enable signal to output of the arithmetic circuit control signal according to the number of the selected word line lines.
6. A control method for a neural network operation circuit, the neural network operation circuit comprising: multiple word lines; A plurality of bit lines are arranged in a manner of crossing the word lines; A plurality of memory cells composed of semiconductor memory elements are arranged at intersections of the plurality of word lines and the plurality of bit lines, respectively maintaining coupling weight coefficients of the neural network; a word line driving circuit capable of driving any one or more word lines among the plurality of word lines to a selected state; A column selection circuit capable of selecting any bit line among the plurality of bit lines; as well as The operation circuit performs a multiplication-addition operation with the input data represented by the driving states of the plurality of word lines by using the current flowing through the bit line selected by the column selection circuit, and performs a neuron operation by performing an activation function operation on the result of the multiplication-addition operation. The control method includes: A step of setting a word line selection state when driving any one or more word lines among the plurality of word lines to a selection state; a delay time setting step of setting the delay time according to the number of selected word lines, wherein the number of selected word lines is information associated with the number of word lines that are selected when performing the multiplication and accumulation operation; a word line driving step, driving the word line according to the word line selection state; and The operation circuit starting step starts the operation circuit. The delay time is a period from the word line driving step to the operation circuit activation step.
7. The control method of the neural network operation circuit according to claim 6, wherein: The number of selected word line lines is the number of word lines in the selected state included in the word line selection state.
8. The control method of the neural network operation circuit according to claim 6, wherein: The neural network operation circuit stores the number of neurons, i.e., the number of nodes, of each layer of the neural network as network configuration information in a network configuration information storage circuit. The number of selected word lines is the number of nodes.
Citation Information
Patent Citations
Neural network computation circuit using semiconductor storage element, and operation method
WO2019049686A1
Neural network arithmetic circuit using non-volatile semiconductor memory element
WO2019049741A1
