阅读说明：本技术 神经形态计算设备及其操作方法 (Neuromorphic computing device and method of operation thereof ) 是由 黄荣南 于 2021-04-22 设计创作，主要内容包括：一种神经形态计算设备,包括：第一存储单元阵列,包含多个电阻存储单元并且被配置为通过多条位线或源极线输出多个读取电流；第二存储单元阵列,包含多个参考电阻存储单元并且被配置为通过至少一条参考位线或至少一条参考源极线输出至少一个参考电流；电流电压转换电路,被配置为输出分别对应于所述多个读取电流的多个信号电压并且输出对应于所述至少一个参考电流的至少一个参考电压；以及模数转换电路,被配置为使用所述至少一个参考电压将所述多个信号电压转换为多个数字信号并且输出所述多个数字信号。(A neuromorphic computing device comprising: a first memory cell array including a plurality of resistive memory cells and configured to output a plurality of read currents through a plurality of bit lines or source lines; a second memory cell array including a plurality of reference resistive memory cells and configured to output at least one reference current through at least one reference bit line or at least one reference source line; a current-voltage conversion circuit configured to output a plurality of signal voltages respectively corresponding to the plurality of read currents and output at least one reference voltage corresponding to the at least one reference current; and an analog-to-digital conversion circuit configured to convert the plurality of signal voltages into a plurality of digital signals using the at least one reference voltage and output the plurality of digital signals.)

1. A neuromorphic computing device comprising:

a first memory cell array including a plurality of resistive memory cells and configured to output a plurality of read currents through a plurality of bit lines or source lines;

a second memory cell array including a plurality of reference resistive memory cells and configured to output at least one reference current through at least one reference bit line or at least one reference source line;

a current-voltage conversion circuit configured to output a plurality of signal voltages respectively corresponding to the plurality of read currents and output at least one reference voltage corresponding to the at least one reference current; and

an analog-to-digital conversion circuit configured to convert the plurality of signal voltages into a plurality of digital signals using the at least one reference voltage and output the plurality of digital signals.

2. The neuromorphic computing device of claim 1, wherein the plurality of reference resistive memory cells and the plurality of resistive memory cells comprise the same resistive material.

3. The neuromorphic computing device of claim 1, wherein the number of at least one reference bit line is the same as a number obtained by exponentially subtracting 1 from the base-2 number of bits of each of the plurality of digital signals.

4. The neuromorphic computing device of claim 1, wherein the at least one reference bit line includes a first reference bit line and a second reference bit line, an

Wherein a ratio of the number of memory cells in a high resistance state to the number of memory cells in a low resistance state of a first reference memory cell connected to the first reference bit line among the plurality of reference resistance memory cells is different from a ratio of the number of memory cells in a high resistance state to the number of memory cells in a low resistance state of a second reference memory cell connected to the second reference bit line among the plurality of reference resistance memory cells.

5. The neuromorphic computing device of claim 1, wherein the at least one reference voltage equally quantifies an operating voltage range of the analog-to-digital conversion circuitry, an

Wherein the resistive memory cells connected to each of the at least one reference bit line have a distribution of resistance states to form a corresponding one of the at least one reference voltage.

6. The neuromorphic computing device of claim 1, wherein the at least one reference voltage unequally quantifies an operating voltage range of the analog-to-digital conversion circuitry,

wherein the quantization interval adjacent to the center of the operating voltage range is narrower than the quantization interval adjacent to the sides of the operating voltage range, an

Wherein the resistive memory cells connected to each of the at least one reference bit line have a distribution of resistance states to form a corresponding one of the at least one reference voltage.

7. The neuromorphic computing device of claim 1, wherein the analog-to-digital conversion circuitry comprises a plurality of analog-to-digital converters corresponding to each of the plurality of signal voltages and configured to convert the corresponding signal voltage to a digital signal.

8. The neuromorphic computing device of claim 1, further comprising: a multiplexing circuit configured to select a signal voltage from a plurality of signal voltages and output the selected signal voltage,

wherein the analog-to-digital conversion circuit comprises a plurality of analog-to-digital converters configured to convert the selected signal voltages into digital signals, wherein the number of analog-to-digital converters is less than the number of signal voltages.

9. The neuromorphic computing device of claim 1, wherein the analog-to-digital conversion circuitry comprises a plurality of analog-to-digital converters configured to convert corresponding ones of a plurality of signal voltages to digital signals, an

Wherein each of the plurality of analog-to-digital converters is configured to output a digital signal corresponding to the corresponding signal voltage by comparing the corresponding signal voltage with the at least one reference voltage.

10. The neuromorphic computing device of claim 9, wherein each of the plurality of analog-to-digital converters comprises:

a comparison circuit configured to output at least one comparison signal by comparing the corresponding signal voltage with the at least one reference voltage; and

an encoding circuit configured to output a digital signal corresponding to the corresponding signal voltage based on the at least one comparison signal.

11. The neuromorphic computing device of claim 1, wherein between the at least one reference source line and a ground node, a load resistor having a resistance value between a low-resistance state and a high-resistance state of a reference resistive memory cell is connected.

12. The neuromorphic computing device of claim 1, further comprising: a word line driver configured to drive a plurality of word lines connected to the first memory cell array, to drive a plurality of reference word lines connected to the second memory cell array, and to drive the plurality of reference word lines such that all of the plurality of word lines are activated during a computing operation of the neuromorphic computing device.

13. The neuromorphic computing device of claim 12, wherein states corresponding to elements of a weight matrix used in computational operations of the neuromorphic computing device are written to a plurality of resistive memory cells, and

wherein the word line driver is configured to activate at least one of the plurality of word lines in accordance with the element values of the input feature vector in a computing operation of the neuromorphic computing device.

14. The neuromorphic computing device of claim 13, further comprising: an adder circuit configured to generate output data by performing an accumulation and/or summation operation using the plurality of digital signals.

15. A method of operating a neuromorphic computing device having a first array of memory cells with main resistive memory cells and a second array of memory cells with reference resistive memory cells, the method comprising:

activating all reference word lines connected to the second memory cell array;

obtaining at least one reference voltage value based on at least one reference current output from the second memory cell array; and

a plurality of digital signals are output by converting a signal voltage value corresponding to a read current output from the first memory cell array using at least one reference voltage value.

16. The method of claim 15, wherein each of the reference resistive memory cells comprises a resistive material that is the same as a material of the main resistive memory cell.

17. The method of claim 15, wherein said outputting a plurality of digital signals comprises:

generating at least one comparison signal by comparing a first one of the signal voltage values with the at least one reference voltage value; and

outputting a first digital signal corresponding to the first signal voltage value based on the at least one comparison signal.

18. The method of claim 15, further comprising: output data is generated by performing an accumulation and/or summation operation using the plurality of digital signals.

19. A neuromorphic computing device comprising:

a first resistive memory cell array including a plurality of resistive memory cells arranged in a region where a plurality of word lines intersect a plurality of bit lines;

a second resistive memory cell array including a plurality of reference resistive memory cells arranged in a region where the plurality of reference word lines intersect the plurality of reference bit lines;

a word line driver configured to drive a plurality of word lines and a plurality of reference word lines, activate at least one selection word line among the plurality of word lines according to an element value of an input feature vector, and activate all the plurality of reference word lines;

a current-voltage conversion circuit configured to convert a plurality of reference currents output from the second resistive memory cell array into a plurality of reference voltages through an electrical path including a plurality of reference bit lines, and convert a plurality of read currents output from the first resistive memory cell array into a plurality of signal voltages through an electrical path including a plurality of bit lines according to activation of at least one selected word line;

an analog-to-digital conversion circuit configured to convert the plurality of signal voltages into a plurality of digital signals and output the plurality of digital signals using the plurality of reference voltages as references for analog-to-digital conversion; and

an adder circuit configured to generate at least one piece of output data by performing an accumulation and/or summation operation using the plurality of digital signals.

20. The neuromorphic computing device of claim 19, wherein each of the plurality of reference resistive memory cells comprises a same resistive material as a material of the plurality of resistive memory cells, an

Wherein the number of reference bit lines corresponds to a number obtained by subtracting 1 from a base-2 number by an exponent of the number of bits of each of the plurality of digital signals.

Technical Field

The present inventive concept relates to a neuromorphic computing device and an operating method thereof, and more particularly, to a neuromorphic computing device and an operating method thereof for performing an operation using a resistive memory cell.

Background

Applications including deep learning Neural Networks (NN) or neuromorphic computing (neuromorphic computing), such as image recognition, natural language processing, and various pattern matching or classification tasks, may become as important as general-purpose computing. The core computational elements or neurons of the NN multiply a set of input signals with a set of weights and sum the products. Thus, the neurons perform vector matrix multiply or Multiply Accumulate (MAC) operations. The NN generally includes a number of interconnected neurons, and each neuron performs a MAC operation. Therefore, the operation of the NN is computationally intensive.

By manufacturing the artificial nervous system at a neuron level, the semiconductor circuit manufactured by simulating the information processing method processed by the brain is a neuromorphic computing device or a neuromorphic chip, and can be effectively used for realizing an intelligent system adaptive to unspecified environments.

Disclosure of Invention

The present disclosure provides a neuromorphic computing device and a method of operation thereof that minimizes inaccurate inferences generated by the neuromorphic computing device due to temperature and/or time dependencies of resistive memory cells.

According to an exemplary embodiment of the inventive concept, there is provided a neuromorphic computing device including: a first memory cell array including a plurality of resistive memory cells and configured to output a plurality of read currents through a plurality of bit lines or source lines; a second memory cell array including a plurality of reference resistive memory cells and configured to output at least one reference current through at least one reference bit line or at least one reference source line; a current-voltage conversion circuit configured to output a plurality of signal voltages respectively corresponding to the plurality of read currents and output at least one reference voltage corresponding to the at least one reference current; and an analog-to-digital conversion circuit configured to convert the plurality of signal voltages into a plurality of digital signals using the at least one reference voltage and output the plurality of digital signals.

According to an exemplary embodiment of the inventive concept, there is provided an operating method of a neuromorphic computing device including a first memory cell array including main resistive memory cells and a second memory cell array including reference memory cells, the operating method including: activating all reference word lines connected to the second memory cell array; obtaining at least one reference voltage value based on at least one reference current output from the second memory cell array; and outputting a plurality of digital signals by converting signal voltage values corresponding to the read current output from the first memory cell array using at least one reference voltage value.

According to an exemplary embodiment of the inventive concept, there is provided a neuromorphic computing device including: a first resistive memory cell array including a plurality of resistive memory cells arranged in a region where a plurality of word lines intersect a plurality of bit lines; a second resistive memory cell array including a plurality of reference resistive memory cells arranged in a region where the plurality of reference word lines intersect the plurality of reference bit lines; a word line driver configured to drive a plurality of word lines and a plurality of reference word lines, activate at least one selection word line among the plurality of word lines according to an element value of an input feature vector, and activate all the plurality of reference word lines; a current-voltage conversion circuit configured to convert a plurality of reference currents output from the second resistive memory cell array into a plurality of reference voltages through an electrical path including a plurality of reference bit lines, and convert a plurality of read currents output from the first resistive memory cell array into a plurality of signal voltages through an electrical path including a plurality of bit lines according to activation of at least one selection word line; an analog-to-digital conversion circuit configured to convert the plurality of signal voltages into a plurality of digital signals and output the plurality of digital signals using the plurality of reference voltages as references for analog-to-digital conversion; and an adder circuit configured to generate at least one piece of output data by performing an accumulation and/or summation operation using the plurality of digital signals.

According to an exemplary embodiment of the inventive concept, there is provided an operating method of a neuromorphic computing device including a first memory cell array having main resistive memory cells connected to a plurality of word lines and a second memory cell array having reference resistive memory cells connected to a plurality of reference word lines, the operating method including: activating a selected word line of the plurality of word lines according to an element value of an input feature vector used in calculation of the neuromorphic computing device; activating all the plurality of reference word lines; and outputting a plurality of digital signals by converting a signal voltage value corresponding to the read current output from the first memory cell array into a plurality of digital signals according to activation of the selected word line using at least one reference voltage value corresponding to at least one reference current output from the second memory cell array.

According to an exemplary embodiment of the inventive concept, there is provided a neuromorphic computing device that converts a current output from a crossbar array into a digital signal, including: a first memory cell array having a plurality of resistive memory cells respectively connected to a plurality of word lines; a second memory cell array having a plurality of reference resistive memory cells arranged in a region where the plurality of reference word lines and the plurality of reference bit lines intersect, and having a same resistance material as that of the plurality of resistive memory cells; and a word line driver configured to drive the plurality of word lines and the plurality of reference word lines and to activate all of the plurality of reference word lines during operation of the neuromorphic computing device, and the number of reference bit lines may be the same as a number obtained by subtracting 1 from a base-2 number raised to the power of the number of bits of the digital signal.

According to an exemplary embodiment of the inventive concept, a neuromorphic circuit comprises: a first memory cell array having a plurality of memory cells, each memory cell connected between a respective one of the first plurality of word lines and a respective one of the second plurality of bit lines; and a second memory cell array having a plurality of reference memory cells, each reference memory cell connected between a respective one of the first plurality of reference word lines and a respective one of the third plurality of reference bit lines; the number of the third plurality of reference bit lines is based on the number of bits of each memory cell and is decreased by one by the power of the number of states of each bit.

Drawings

Embodiments of the inventive concept will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings, in which:

fig. 1 is a block diagram illustrating a neuromorphic computing device according to an embodiment of the present inventive concept;

fig. 2A and 2B are block diagrams illustrating an example of a neural network system driven by a neuromorphic computing device, according to an embodiment of the present inventive concept;

fig. 3A, 3B, and 3C are circuit diagrams illustrating a first memory cell array according to an embodiment of the inventive concept;

fig. 4A and 4B are circuit diagrams illustrating a second memory cell array according to an embodiment of the inventive concept;

fig. 5 is a graph illustrating an operating voltage range and a plurality of reference voltages of an analog-to-digital conversion circuit according to an embodiment of the inventive concept;

fig. 6 is a circuit diagram illustrating states of a second memory cell array and a reference memory cell according to an embodiment of the inventive concept;

fig. 7A is a block diagram illustrating an analog-to-digital conversion circuit according to an embodiment of the inventive concept;

fig. 7B is a block diagram illustrating a neuromorphic computing device according to an embodiment of the present inventive concept;

fig. 8 is a block diagram illustrating an analog-to-digital converter according to an embodiment of the inventive concept;

fig. 9 is a block diagram illustrating an analog-to-digital converter according to an embodiment of the inventive concept.

Fig. 10 is a graph illustrating an operating voltage range and a plurality of reference voltages of an analog-to-digital conversion circuit according to an embodiment of the inventive concept;

fig. 11 is a circuit diagram illustrating states of a second memory cell array and a reference memory cell according to an embodiment of the inventive concept;

fig. 12 is a circuit diagram illustrating a second memory cell array according to an embodiment of the inventive concept;

fig. 13 is a block diagram illustrating an analog-to-digital converter according to an embodiment of the inventive concept;

fig. 14 is a flow chart of a method of operation of a neuromorphic computing device according to embodiments of the present inventive concept; and

fig. 15 is a block diagram illustrating an electronic system according to an embodiment of the inventive concept.

Detailed Description

Hereinafter, exemplary embodiments of the inventive concept will be described in detail with reference to the accompanying drawings.

Fig. 1 illustrates a neuromorphic computing device 10 according to an embodiment of the present inventive concept.

The neuromorphic computing device 10 may include a first memory cell array 100, a second memory cell array 200, a word line driver 150, a current-to-voltage conversion circuit 300, an analog-to-digital conversion circuit 400, and an adder circuit 500. Similarly, the neuromorphic circuit device may include: a current-voltage conversion circuit 300 coupled to the first memory cell array 100 and the second memory cell array 200; an analog-to-digital conversion circuit 400 coupled to the current-to-voltage conversion circuit; and adder circuit 500 coupled to the analog-to-digital conversion circuit, but is not so limited.

In one embodiment, as described below with reference to fig. 2A and 2B, the neuromorphic computing device 10 may be used to drive any neural network system, such as an Artificial Neural Network (ANN) system, a Convolutional Neural Network (CNN) system, a Deep Neural Network (DNN) system, a deep learning system, and/or a machine learning system. For example, various services and/or applications, such as image classification services, biometric-based user authentication services, Advanced Driver Assistance System (ADAS) services, voice assistant services, Automatic Speech Recognition (ASR) services, and so forth, may be executed and processed by the neuromorphic computing device 10. In this case, the data stored in the first memory cell array 100 may be weights included in a plurality of layers constituting the neural network system, and the plurality of read currents (Iread _1, Iread _ 2.., Iread _ M) and the plurality of signal voltages (Vsig _1, Vsig _ 2.., Vsig _ M) may indicate the results of multiply-accumulate operations performed by the neural network system. In other words, the first memory cell array 100 can perform data storage and calculation operations at once, as described later with reference to fig. 3B. The neuromorphic circuit of an example embodiment may include the memory cells RMC of the first array 100 and the reference memory cells RRMC of the second array 200 formed as resistive memory cells. The value stored by the memory unit RMC may correspond to a weight to be applied to the input signal in the neuromorphic circuit.

The first memory cell array 100 may include a plurality of resistive memory cells RMC arranged in a region where a plurality of word lines WL1, WL 2., WLN (N is a natural number of 2 or more) cross a plurality of bit lines BL1, BL 2., BLM (M is a natural number of 2 or more). Each of the plurality of resistive memory cells may include a resistive element RE. The detailed structure of the first memory cell array 100 will be described later with reference to fig. 3A and 3C.

The first memory cell array 100 may store a plurality of pieces of data. For example, a plurality of pieces of data may be stored in the plurality of resistance memory cells RMC by using the resistance change of the resistance element RE included in each of the plurality of resistance memory cells RMC. Likewise, a plurality of word lines WL1, WL2, which are connected to the first memory cell array 100, WLN may be driven by the word line driver 150. The first memory cell array 100 may output a plurality of read currents Iread _1, Iread _2,. and Iread _ M corresponding to a plurality of signal voltages Vsig _1, Vsig _2,. and Vsig _ M, respectively. For example, the first memory cell array 100 may output a plurality of read currents Iread _1, Iread _2, Iread _ M through an electrical path including bit lines BL1, BL 2. The first memory cell array 100 may provide the first to mth read currents Iread _1 to Iread _ M to the current-voltage conversion circuit 300.

The second memory cell array 200 may include a plurality of reference resistance memory cells RRMC arranged in regions where a plurality of reference word lines RWL1, RWL 2.. the RWLN intersects with a plurality of reference bit lines RBL1, RBL 2.. the RBLn (n is a natural number of 2 or more). Fig. 1 illustrates an embodiment in which the second memory cell array 200 is connected to a plurality of reference bit lines RBL1, RBL 2. The number of reference bit lines connected to the second memory cell array 200 will be described below. Each of the plurality of reference resistance memory cells RRMC may include a resistance element RE. The specific structure of the second memory cell array 200 will be described later with reference to fig. 4A and 4B.

In one embodiment, each of the plurality of reference resistive memory cells RRMC included in the second memory cell array 200 may include the same resistive material as the plurality of resistive memory cells RMC included in the first memory cell array 100.

Further, in one embodiment, the number of reference bit lines connected to the second memory cell array 200 may be determined based on the number of bits of the digital signals DS _1, DS _ 2. For example, when the number of bits of each of the digital signals DS _1, DS _2, …, DS _ M is k bits, the number of reference bit lines connected to the second memory cell array 200 may correspond to the number obtained by subtracting 1 from the k power of 2. For example, when the digital signals DS _1, DS _ 2., DS _ M are 1-bit digital signals, the second memory cell array 200 may be connected to one reference bit line, and when the digital signals DS _1, DS _ 2., DS _ M are 2-bit digital signals, the second memory cell array 200 may be connected to three reference bit lines.

The second memory cell array 200 can output a plurality of reference currents Iref _1, Iref _2, Iref _ n corresponding to a plurality of reference voltages Vref _1, Vref _ 2. For example, the second memory cell array 200 may output a plurality of reference currents Iref _1, Iref _2, and Iref _ n via an electrical path including reference bit lines RBL1, RBL 2. The second memory cell array 200 may provide the first to nth reference currents Iref _1 to Iref _ n to the current-voltage conversion circuit 300.

The word line driver 150 may drive a plurality of word lines WL1, WL 2.., WLN connected to the first memory cell array 100 and a plurality of reference word lines RWL1, RLW 2.., RWLN connected to the second memory cell array 200.

For example, depending on the element values of the input eigenvector used in the operation of the neuromorphic computing device 10, the word line driver 150 may drive the plurality of word lines WL1, WL 2.., WLN such that at least one word line selected from the plurality of word lines WL1, WL 2.., WLN is activated. For example, when each element value of the input feature vector of length N is "1" or "0", the word line driver 150 may drive a plurality of word lines WL1, WL 2.

Word line drivers 150 may drive the plurality of reference word lines RWL1, RWL 2.., RWLN such that the plurality of reference word lines RWL1, RWL 2.., RWLN are all activated during an operating period of the neuromorphic computing device 10.

The current-to-voltage conversion circuit 300 may convert an input current signal into a voltage signal. For example, the current-voltage conversion circuit 300 may convert the first read current Iread _1 to output the first signal voltage Vsig _1, convert the second read current Iread _2 to output the second signal voltage Vsig _2, and convert the mth read current Iread _ M to output the mth signal voltage Vsig _ M in the same manner. Also, for example, the current-voltage conversion circuit 300 may convert the first reference current Iref _1 to output the first reference voltage Vref _1, convert the second reference current Iref _2 to output the second reference voltage Vref _2, and convert the nth reference current Iref _ n to output the nth reference voltage Vref _ n in the same manner. Thus, the current-to-voltage conversion circuit 300 may include a plurality of current-to-voltage converters, wherein the number of current-to-voltage converters may be the same as the number of input current signals, and, according to one embodiment, may be less than the number of input current signals. In the latter case (when the number of current-to-voltage converters is smaller than the number of input current signals), the current-to-voltage converters may sequentially convert the input current signals into voltage signals. The current-to-voltage conversion circuit 300 may provide a plurality of signal voltages Vsig _1, Vsig _ 2.., Vsig _ M and a plurality of reference voltages Vref _1, Vref _ 2.., Vref _ n to the analog-to-digital conversion circuit 400.

The analog-to-digital conversion circuit 400 may convert the plurality of signal voltages Vsig _1, Vsig _2,.. the Vsig _ M into the plurality of digital signals DS _1, DS _2, …, DS _ M, respectively, using the plurality of reference voltages Vref _1, Vref _2,. the Vref _ n, and may output the plurality of digital signals DS _1, DS _2, …, DS _ M. For example, the analog-to-digital conversion circuit 400 may use the plurality of reference voltages Vref _1, Vref _ 2., Vref _ n as references for analog-to-digital conversion to convert the plurality of signal voltages Vsig _1, Vsig _2, …, Vsig _ M into the plurality of digital signals DS _1, DS _2, …, DS _ M. To this end, the analog-to-digital conversion circuit 400 may include a plurality of analog-to-digital converters, and the number of analog-to-digital converters may be the same as the number of input signal voltages, and, according to one embodiment, may be smaller than the number of input signal voltages. More detailed description will be made later with reference to fig. 7A and 7B. Analog-to-digital conversion circuit 400 may provide a plurality of digital signals DS _1, DS _ 2.

The adder circuit 500 may generate the output data ODAT by performing an accumulation and/or summation operation using a plurality of digital signals DS _1, DS _ 2. For example, the adder circuit 500 may output the first output data ODAT _1, the second output data ODAT _2 to the mth output data ODAT _ M using a plurality of digital signals DS _1, DS _ 2. In one embodiment, adder circuit 500 may include at least one adder and at least one shift register.

In general, the plurality of resistive memory cells RMC included in the first memory cell array 100 may have temperature dependency and time dependency. For example, the resistance element RE included in each of the plurality of resistance memory cells RMC may have temperature dependency in which the resistance decreases when the temperature increases and the resistance increases when the temperature decreases. Further, the resistance element RE may have time dependency such as a retention characteristic in which the resistance decreases with time or, in some cases, a drift characteristic in which the resistance increases after a certain time has elapsed after data writing. Thus, the plurality of read currents Iread _1, Iread _2,. output from the first memory cell array 100, Iread _ M may depend on temperature and time, and it may be desirable to reduce or eliminate the effects of temperature dependence and time dependence on accurate data storage and operation.

The neuromorphic circuit embodiment may include a word line driver circuit 150 configured to simultaneously or substantially simultaneously drive a respective one of the first plurality of word lines WLi and a respective one of the first plurality of reference word lines RWLi. The plurality of memory cells may include substantially the same cell material as the plurality of reference memory cells, and the plurality of memory cells may have at least one of substantially the same temperature characteristics or substantially the same time decay characteristics as the plurality of reference memory cells.

The neuromorphic computing device 10 according to an embodiment of the inventive concept may convert the plurality of signal voltages Vsig _1, Vsig _2,.. Vref _ M into the plurality of digital signals DS _1, DS _2,.. DS _ M, which are obtained from the plurality of reference currents Iref _1, Iref _2,. ere.. Iref _ n output from the second memory cell array 200 by using the plurality of reference voltages Vref _1, Vref _2,. ere.. Vref _ n, the second memory cell array 200 including the same resistance material as that of the resistance element RE included in the first memory cell array 100.

The second memory cell array 200 has the same temperature dependence and/or time dependence as the first memory cell array 100, thereby preventing a decrease in the inference accuracy of the neuromorphic computing device 10 from the temperature dependence and/or time dependence of the plurality of resistive memory cells RMC. That is, according to embodiments of the present inventive concept, the reliability and accuracy of the operation or inference of the neuromorphic computing device 10 may be improved.

Fig. 2A and 2B are diagrams illustrating an example of a neural network system driven by a neuromorphic computing device according to an embodiment of the inventive concept.

Referring to fig. 2A, a network structure of a general neural network may include an input layer IL, a plurality of hidden layers HL1, HL 2.

The input layer IL may include i (where i is a natural number) input nodes x1, x 2.., xi, and vector input data IDAT of length i may be input to each input node.

Multiple hidden layersHL1, HL 2., HLn includes n (where n is a natural number) hidden layers, and hidden nodes h¹ ₁，h¹ ₂，h¹ ₃，...，h¹ _m，h² ₁，h² ₂，h² ₃，...，h² _m，hⁿ ₁，hⁿ ₂，hⁿ ₃，...， hⁿ _m. For example, the hidden layer HL1 may include m (where m is a natural number) hidden nodes h¹ ₁，h¹ ₂， h¹ ₃，...，h¹ _mThe hidden layer HL2 may comprise m hidden nodes h² ₁，h² ₂，h² ₃，...，h² _mAnd the hidden layer HLn may comprise m hidden nodes hⁿ ₁，hⁿ ₂，hⁿ ₃，...，hⁿ _m。

The output layer OL may include j (where j is a natural number) output nodes y1, y 2.., yj corresponding to categories to be classified, and may output the result (e.g., a score or a category score) of each category as output data ODAT for the input data IDAT. The output layer OL may be referred to as a full connection layer, and for example, the probability that the input data IDAT corresponds to a vehicle may be represented by a number.

The network structure shown in fig. 2A may include branches between nodes shown as straight lines between two nodes, and, although not shown, weights used in each of the respective branches. At this time, nodes in the same layer do not need to be branched, and nodes included in different layers may be fully or partially branched.

Each node in FIG. 2A (e.g., h)¹ ₁) The output of a previous node (e.g., x1) may be received and manipulated, and may be directed to a subsequent node (e.g., h)² ₁) And outputting the result of the operation. At this time, each node may operate a value to be output, for example, a nonlinear function, by applying an input value to a specific function.

In general, the network structure of the neural network is previously determined, and an appropriate value is calculated using data with known correct answers according to the weights of branches between nodes to determine to which category the data belongs. Data with known correct answers is referred to as "learning data" and the process of determining weights is referred to as "learning". In addition, it is assumed that a batch of structures and weights that can be independently learned are "models", and a process in which a model with a certain weight predicts which category input data belongs to and outputs a predicted value is referred to as a "test" process.

Referring to fig. 2B, an example of an operation performed in one node ND included in the network configuration of fig. 2A is specifically shown.

When inputting N number of inputs a₁，a₂，a₃，...，a_NWhen provided to a node ND, the node ND may provide N inputs a₁，a₂，a₃，...，a_NAnd respectively correspond to the input a₁，a₂，a₃，...，a_NN weights w₁，w₂，w₃，…，w_NMultiply and sum, add the offset b to the summed input value, and apply the sum as an input to a particular function σ to generate an output value (e.g., z). It should be understood that the output z may be branched to be output to one or more other nodes.

When one layer included in the network structure shown in fig. 2A includes M nodes ND shown in fig. 2B, an output value of one layer may be obtained by equation 1 shown below.

[ equation 1]

W*A＝Z

In equation 1 above, W represents a weight with respect to branches included in one layer, and may be implemented in the form of an M × N matrix. A denotes N inputs a received from one layer₁，a₂，a₃，...， a_NAnd may be implemented in the form of an N x1 matrix. Z represents M outputs Z output from one layer₁，z₂， z₃，...，z_MAnd may be implemented in the form of an M x1 matrix.

Fig. 3A, 3B, and 3C illustrate first memory cell arrays 100a and 100B according to an embodiment of the inventive concept. The first memory cell arrays 100a and 100b shown in fig. 3A to 3C may correspond to the first memory cell array 100 of fig. 1.

Referring to fig. 3A, the first memory cell array 100a may include a plurality of resistive memory cells RMC arranged in a region where a plurality of word lines WL1, WL 2. Each of the plurality of resistive memory cells RMC may include a resistive element RE and may be connected to one of a plurality of word lines WL1, WL2, WLN and one of a plurality of bit lines BL1, BL2, BLM.

The resistance value of the resistive element RE may be changed by a write voltage applied by the plurality of word lines WL1, WL 2.., WLN and/or the plurality of bit lines BL1, BL 2.., BLM. The plurality of resistive memory cells RMC may store data due to a resistance change. For example, when a write voltage is applied to a selected word line and a ground voltage (e.g., about 0V) is applied to a selected bit line, data "1" may be written to the selected resistive memory cell, and when a ground voltage is applied to a selected word line and a write voltage is applied to a selected bit line, data "0" may be written to the selected resistive memory cell. Further, when a read voltage is applied to the selected word line and a ground voltage is applied to the selected bit line, data written to the selected resistive memory cell may be read.

In one embodiment, each of the plurality of resistive memory cells RMC may be implemented by including any resistive memory cell, such as a phase change random access memory (PRAM) cell, a Resistive Random Access Memory (RRAM) cell, a Magnetic Random Access Memory (MRAM) cell, a Ferroelectric Random Access Memory (FRAM) cell, or the like.

In one embodiment, the resistive element RE may include a phase change material whose crystalline state changes according to the amount of current. Various materials may be used for the phase change material, such as 2-element compounds GaSb, InSb, InSe, Sb2Te3 and/or GeTe, 3-element compounds GeSbTe, GaSeTe, InSbTe, SnSb2Te4 and/or InSbGe, and 4-element compounds AgInSbTe, (GeSn) SbTe, GeSb (SeTe) and/or Te81Ge15Sb2S 2. In another embodiment, the resistive element RE may comprise a perovskite compound, a transition metal oxide, a magnetic material, a ferromagnetic material or an antiferromagnetic material. However, the resistance material included in the resistance element RE is not limited to the above-described material.

Referring to fig. 3B, an example in which the first memory cell array 100a of fig. 3A performs the above-described operation with reference to fig. 2A and 2B is shown.

Each resistive memory cell RMC may correspond to a synapse or branch of a neural network system and may store a weight. Accordingly, the M × N data stored in the first memory cell array 100a may correspond to the weight matrix implemented in the form of an M × N matrix, i.e., W of [ equation 1] above, included in one layer described above with reference to fig. 2A and 2B.

N input voltages V1, V2, VN applied by a plurality of word lines WL1, WL2₁，a₂，a₃，...，a_NAnd may correspond to an input matrix implemented in the form of an N x1 matrix, i.e. the above-mentioned [ equation 1]]A of (1).

Through a plurality of bit lines BL1, BL 2.., M read currents Iread _1, Iread _ 2.. output by the BLM, Iread _ M may correspond to M outputs z output from one layer₁，z₂，z₃，...，z_MAnd may correspond to an output matrix implemented in the form of an M x1 matrix, i.e. the above-mentioned [ equation 1]]Z of (1).

In other words, in a state where the first memory cell array 100a is implemented by storing a plurality of weights in a matrix form in the plurality of resistive memory cells RMC, when the input voltages V1, V2, are inputted through the plurality of word lines WL1, WL 2.. WLN corresponding to a plurality of input values VN, the read currents Iread _1, Iread _2,. that are outputted through the plurality of bit lines BL1, BL 2.. BLM may be the result of a multiply-accumulate operation performed by the neural network system. When the multiple layers of the neural network system are all implemented as described above, a neuromorphic computing device that performs both data storage and computational operations may be implemented.

Referring to fig. 3C, the first memory cell array 100b may include a plurality of resistive memory cells RMC' arranged in a region where a plurality of word lines WL1, WL 2.., WLN, a plurality of bit lines BL1, BL 2.., BLM, and a plurality of source lines SL1, SL2, …, SLM cross each other.

Each of the plurality of resistive memory cells RMC may include a cell transistor CT and a resistive element RE, and may be connected to one of a plurality of word lines WL1, WL 2.., WLN, one of a plurality of bit lines BL1, BL 2.., BLM, and one of a plurality of source lines SL1, SL 2.., one of a plurality of source lines SLM. For example, the cell transistor CT may include a first electrode connected to one of a plurality of source lines SL1, SL2, …, the SLM, a second electrode, and a gate electrode connected to one of a plurality of word lines WL1, WL 2. The resistance element RE may be connected between the second electrode of the cell transistor CT and one of a plurality of bit lines BL1, BL 2.

For example, data "1" may be written to the selected resistive memory cell when a power supply voltage (e.g., VCC) is applied to the selected word line, a write voltage is applied to the selected bit line, and a ground voltage is applied to the selected source line, and data "0" may be written to the selected resistive memory cell when a power supply voltage is applied to the selected word line, a ground voltage is applied to the selected bit line, and a write voltage is applied to the selected source line. In addition, when a power supply voltage is applied to a selected word line, a read voltage is applied to a selected bit line, and a ground voltage is applied to a selected source line, data written to a selected resistive memory cell may be read.

The first memory cell array 100b may output a plurality of read currents Iread _1, Iread _2, Iread _ M through a circuit including a plurality of bit lines BL1, BL 2. In one embodiment, the first memory cell array 100b may output a first read current Iread _1 flowing through the first bit line BL1 and flowing through the first source line SL1, output a second read current Iread _2 flowing through the second bit line BL2 and flowing through the second source line SL2, and output an M read current Iread _ M flowing through the M bit line BLM and flowing through the M source line SLM.

Further, the case where the first memory cell arrays 100a and 100B are formed in a two-dimensional array structure is described with reference to fig. 3A, 3B, and 3C, but the inventive concept is not limited thereto, and according to one embodiment, the first memory cell array may be formed in a three-dimensional vertical array structure. The structures of the resistive memory cells RMC and RMC' may also vary according to embodiments.

Fig. 4A and 4B illustrate second memory cell arrays 200a and 200B according to an embodiment of the inventive concept. The second memory cell arrays 200a and 200B described with respect to fig. 4A and 4B may correspond to the second memory cell array 200 of fig. 1.

Referring to fig. 4A, the second memory cell array 200a may include a plurality of reference resistive memory cells RRMC arranged in a region where a plurality of reference word lines RWL1, RWL 2.., RWLN intersect a plurality of reference bit lines RBL1, RBL 2.., RBLn. Each of the plurality of reference resistive memory cells RRMC may include a resistive element RE and may be connected to one of a plurality of reference word lines RWL1, RWL 2.., one of RWLN and one of a plurality of reference bit lines RBL1, RBL 2.., RBLn. In one embodiment, the resistive element RE may include the same resistive material as the resistive element RE of the first memory cell array (100 of fig. 1), and a detailed description thereof may be understood from the description of fig. 1 and 3A.

The second memory cell array 200a may output a plurality of reference currents Iref _1, Iref _2, Iref _ n via an electrical path including a plurality of reference bit lines RBL1, RBL 2. For example, the second memory cell array 200a may output the first reference current Iref _1 through the first reference bit line RBL1, the second reference current Iref _2 through the second reference bit line RBL2, and the nth reference current Iref _ n through the nth reference bit line RBLn.

Referring to fig. 4B, the second memory cell array 200B may include a plurality of reference resistance memory cells RRMC' arranged in a region where a plurality of reference word lines RWL1, RWL 2., RWLN intersect with a plurality of reference bit lines RBL1, RBL2, …, RBLn and a plurality of reference source lines RSL1, RSL 2., RSLn.

Each of the plurality of reference resistive memory cells RRMC' may include a cell transistor CT and a resistance element RE, and may be connected to one of a plurality of reference word lines RWL1, RWL 2., RWLN, one of the reference word lines RBL1, RBL 2., one of the reference bit lines RBLM, and one of a plurality of reference source lines RSL1, RSL 2., RSLM. For example, the cell transistor CT may include a first electrode connected to one of the plurality of reference source lines RSL1, RSL2, …, RSLM, a second electrode, and a gate electrode connected to one of the plurality of reference word lines RWL1, RWL 2. The resistance element RE may be connected between the second electrode of the cell transistor CT and one of the plurality of reference bit lines RBL1, RBL 2. The detailed description thereof can be understood from the description of fig. 1 and 3C.

The second memory cell array 200b can output a plurality of reference currents Iref _1, Iref _2, Iref _ n through an electrical path including a plurality of reference bit lines RBL1, RBL 2. In one embodiment, the second memory cell array 200b may output a first reference current Iref _1 flowing through the first reference bit line RBL1 and flowing out through the first reference source line RSL1, output a second reference current Iref _2 flowing through the second reference bit line RBL2 and flowing out through the second reference source line RSL2, and output an nth reference current Iref _ n flowing through the nth reference bit line RBLn and flowing out through the nth reference source line RSLn.

Fig. 5 illustrates an operating voltage range of an analog-to-digital conversion circuit and a plurality of reference voltages Vref _1, Vref _2, and Vref _3 according to an embodiment of the inventive concept. Fig. 5 illustrates an embodiment in which the plurality of digital signals DS _1, DS _2,. and DS _ M output by the analog-to-digital conversion circuit 400 of fig. 1 are 2-bit digital signals. The number of bits of the digital signal and the number of reference voltages as described above are merely examples for convenience of description, and are not limited to those shown in fig. 5. Fig. 5 is described together with fig. 1.

The analog-to-digital conversion circuit 400 may convert a plurality of signal voltages into a plurality of digital signals using a plurality of reference voltages provided by the current-to-voltage conversion circuit 300. The plurality of reference voltages may be used as reference values for converting the signal voltage in the form of an analog signal into a digital signal.

For example, when the first signal voltage Vsig _1 is less than the first reference voltage Vref _1, the analog-to-digital conversion circuit 400 may output "00" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. In addition, for example, when the first signal voltage Vsig _1 is greater than or equal to the first reference voltage Vref _1 and less than the second reference voltage Vref _2, the analog-to-digital conversion circuit 400 may output "01" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. Also, for example, when the first signal voltage Vsig _1 is greater than or equal to the second reference voltage Vref _2 and less than the third reference voltage Vref _3, the analog-to-digital conversion circuit 400 may output "10" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. In addition, for example, when the first signal voltage Vsig _1 is greater than or equal to the third reference voltage Vref _3, the analog-to-digital conversion circuit 400 may output "11" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1.

Referring to fig. 5, a plurality of reference voltages including the first to third reference voltages Vref _1 to Vref _3 may substantially equally divide an operating voltage range of the analog-to-digital conversion circuit 400. However, the inventive concept is not limited thereto, and the plurality of reference voltages may divide the operating voltage range of the analog-to-digital conversion circuit 400 unequally. This embodiment will be described in detail with reference to fig. 10.

Fig. 6 illustrates states of a second memory cell array and a reference memory cell according to an embodiment of the inventive concept. Fig. 6 may illustrate an embodiment in which the number of reference bit lines in the second memory cell array 200B according to the embodiment of fig. 4B is 3. In particular, fig. 6 may illustrate an embodiment in which the operating voltage range of the analog-to-digital conversion circuit 400 is divided substantially equally as shown in fig. 5. Fig. 6 is described together with reference to fig. 1 and 5.

The second memory cell array 200 may output the first reference current Iref _1 corresponding to the first reference voltage Vref _1 through the first reference source line RSL 1. Referring to fig. 5, corresponding to the first reference source line RSL1, the first reference voltage Vref _1 may be generated by writing data such that more memory cells in the high resistance state HRS than those in the low resistance state LRS among the reference resistance memory cells connected to the first reference bit line RBL 1. As a non-limiting example, with respect to a reference resistive memory cell connected to first reference bit line RBL1, the ratio of the number of memory cells of the high resistance state HRS to the number of memory cells of the low resistance state LRS may be 8: 0.

The second memory cell array 200 may output the second reference current Iref _2 corresponding to the second reference voltage Vref _2 through the second reference source line RSL 2. Corresponding to the second reference source line RSL2, the second reference voltage Vref _2 may be generated by writing data such that the number of memory cells in the high resistance state HRS is the same as the number of memory cells in the low resistance state LRS among the reference resistive memory cells connected to the second reference bit line RBL 2. As a non-limiting example, with respect to a reference resistive memory cell connected to second reference bit line RBL2, the ratio of the number of memory cells in the high resistance state HRS to the number of memory cells in the low resistance state LRS may be 4: 4.

The second memory cell array 200 may output the third reference current Iref _3 corresponding to the third reference voltage Vref _3 through the third reference source line RSL 3. Corresponding to the third reference source line RSL3, the third reference voltage Vref _3 may be generated by writing data such that, among the reference resistive memory cells connected to the third reference bit line RBL3, there are fewer memory cells in the high resistance state HRS than those in the low resistance state LRS. As a non-limiting example, with respect to the reference resistive memory cells connected to third reference bit line RBL3, the ratio of the number of memory cells in the high resistance state HRS to the number of memory cells in the low resistance state LRS may be 0: 8.

As described above, in order for the second memory cell array 200 to output the reference currents corresponding to the reference voltages of different levels, respectively, the ratio of the memory cells in the high resistance state HRS to the memory cells in the low resistance state LRS may be different.

Fig. 7A illustrates an analog-to-digital conversion circuit 400a according to an embodiment of the inventive concept. The analog-to-digital conversion circuit 400a of fig. 7A corresponds to the analog-to-digital conversion circuit 400 of fig. 1.

The analog-to-digital conversion circuit 400a may include a plurality of analog-to-digital converters, and in one embodiment, the number of analog-to-digital converters may be the same as the number of signal voltages Vsig _1, Vsig _ 2.

For example, the analog-to-digital conversion circuit 400a may include a first analog-to-digital converter 401_1a and second to mth analog-to-digital converters 401_2a to 401_ Ma.

The first analog-to-digital converter 401_1a may convert the first signal voltage Vsig _1 by using the first to third reference voltages Vref _1 to Vref _3 to output the first digital signal DS _ 1.

The second analog-to-digital converter 401_2a may convert the second signal voltage Vsig _2 to output the second digital signal DS _2 by using the first to third reference voltages Vref _1 to Vref _ 3.

Similarly, the mth analog-to-digital converter 401_ Ma may convert the mth signal voltage Vsig _ M by using the first to third reference voltages Vref _1 to Vref _3 to output the mth digital signal DS _ M.

Each of the first to mth analog-to-digital converters 401_1a to 401_ Ma will be described in more detail with reference to fig. 8 and 9.

Fig. 7B illustrates a neuromorphic computing device 10 according to an embodiment of the present inventive concept. Fig. 7B illustrates an embodiment in which a multiplexing circuit 350 is added between the current-to-voltage conversion circuit 300 and the analog-to-digital conversion circuit 400 of the neuromorphic computing device 10 of fig. 1.

In contrast to fig. 1, the neuromorphic computing device 10 may further include a multiplexing circuit 350 and a MUX decoder 360.

The analog-to-digital conversion circuit 400b may include a plurality of analog-to-digital converters, and in one embodiment, the number of analog-to-digital converters may be less than the number of signal voltages Vsig _1, Vsig _ 2.

For example, the analog-to-digital conversion circuit 400b may include a first analog-to-digital converter 401_1b and second to kth analog-to-digital converters 401_2b to 401_ kb. Here, k is a natural number of 2 or more, and may be less than M.

The multiplexing circuit 350 may select k signal voltages from the M signal voltages based on a selection signal SEL and supply the selected k signal voltages to the analog-to-digital converting circuit 400 b. For example, the multiplexing circuit 350 may supply the first to k-th signal voltages Vsig _1 to Vsig _ k to the analog-to-digital conversion circuit 400b by selecting the first to k-th signal voltages Vsig _1 to Vsig _ k from the first to M-th signal voltages Vsig _1 to Vsig _ M based on the selection signal SEL.

The MUX decoder 360 may output a selection signal SEL to the multiplexing circuit 350 to select which signal voltage is output to the analog-to-digital converting circuit 400 b.

The first analog-to-digital converter 401_1b may convert the first signal voltage Vsig _1 by using the first to third reference voltages Vref _1 to Vref _3 to output the first digital signal DS _ 1.

The second analog-to-digital converter 401_2b may convert the second signal voltage Vsig _2 to output the second digital signal DS _2 by using the first to third reference voltages Vref _1 to Vref _ 3.

Similarly, the kth analog-to-digital converter 401_ kb may convert the kth signal voltage Vsig _ k by using the first to third reference voltages Vref _1 to Vref _3 to output the kth digital signal DS _ k.

The first to kth analog-to-digital converters 401_1b to 401_ kb will be described in more detail with reference to fig. 8 and 9.

Fig. 8 illustrates an analog-to-digital converter 401 according to an embodiment of the inventive concept. The analog-to-digital converter 401 of fig. 8 may correspond to any one of the first to mth analog-to-digital converters 401_1a to 401_ Ma of fig. 7A, or any one of the first to kth analog-to-digital converters 401_1B to 401_ kb of fig. 7B.

The analog-to-digital converter 401 may convert the corresponding signal voltage Vsig into a digital signal DS using the first to third reference voltages Vref _1 to Vref _3 to output the digital signal DS. For example, the analog-to-digital converter 401 may compare the corresponding signal voltage Vsig with the first to third reference voltages Vref _1 to Vref _3 to output the digital signal DS.

To this end, the analog-to-digital converter 401 may include a comparison circuit 420 and an encoding circuit 440.

The comparison circuit 420 may compare the signal voltage Vsig with the first to third reference voltages Vref _1 to Vref _3 to output a plurality of comparison signals CS. To this end, as shown in fig. 9, the comparison circuit 420 may include a plurality of comparators. The comparison circuit 420 may provide a plurality of comparison signals CS to the encoding circuit 440.

The encode circuit 440 may generate a digital signal DS corresponding to the signal voltage Vsig based on the plurality of comparison signals CS and output the digital signal DS.

Fig. 9 illustrates an analog-to-digital converter 401 according to an embodiment of the inventive concept. Fig. 9 may illustrate a detailed block diagram of the analog-to-digital converter 401 of fig. 8.

The comparison circuit 420 may include a first comparator 421, a second comparator 422, and a third comparator 423.

The first comparator 421 may compare the signal voltage Vsig with the first reference voltage Vref _1 to output a first comparison signal CS 1. For example, when the signal voltage Vsig is greater than the first reference voltage Vref _1, the first comparator 421 may output the first comparison signal CS1 of a first logic level "1", and when the signal voltage Vsig is not greater than the first reference voltage Vref _1, the first comparator 421 may output the first comparison signal CS1 of a second logic level "0".

The second comparator 422 may compare the signal voltage Vsig with the second reference voltage Vref _2 to output a second comparison signal CS 2. For example, when the signal voltage Vsig is greater than the second reference voltage Vref _2, the second comparator 422 may output the second comparison signal CS2 of the first logic level "1", and when the signal voltage Vsig is not greater than the second reference voltage Vref _2, the second comparator 422 may output the second comparison signal CS2 of the second logic level "0".

The third comparator 423 may compare the signal voltage Vsig with the third reference voltage Vref _3 to output a third comparison signal CS 3. For example, when the signal voltage Vsig is greater than the third reference voltage Vref _3, the third comparator 423 may output the third comparison signal CS3 of the first logic level "1", and when the signal voltage Vsig is not greater than the third reference voltage Vref _3, the third comparator 423 may output the third comparison signal CS3 of the second logic level "0".

The encoding circuit 440 may include an encoder 442. The encoder 442 may receive the supply voltage VCC as an input and may receive the first comparison signal CS1 from the first comparator 421, the second comparison signal CS2 from the second comparator 422, and the third comparison signal CS3 from the third comparator 423. The encoder 442 may output a 2-bit digital signal DS based on the first comparison signal CS1, the second comparison signal CS2, and the third comparison signal CS 3.

Fig. 10 illustrates an operating voltage range of the analog-to-digital conversion circuit 400 and a plurality of reference voltages Vref _1, Vref _2, and Vref _3 according to an embodiment of the inventive concept. Fig. 10 illustrates an embodiment in which the plurality of digital signals DS _1, DS _2,. and DS _ M output by the analog-to-digital conversion circuit 400 of fig. 1 are 2-bit digital signals. The number of bits of the digital signal and the number of reference voltages as described above are merely examples for convenience of description, and are not limited to those shown in fig. 10. The number of reference voltages may be one less than the number of bits squared for the binary state; or more generally one less than the power of the number of bits of the multi-state, such as a tri-state, corresponding to the number of value states. Fig. 10 is described together with fig. 1.

The analog-to-digital conversion circuit 400 may convert a plurality of signal voltages into a plurality of digital signals using a plurality of reference voltages provided from the current-voltage conversion circuit 300. The plurality of reference voltages may be used as reference values for converting the signal voltage in the form of an analog signal into a digital signal.

For example, when the first signal voltage Vsig _1 is less than the first reference voltage Vref _1, the analog-to-digital conversion circuit 400 may output "00" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. In addition, for example, when the first signal voltage Vsig _1 is greater than or equal to the first reference voltage Vref _1 and less than the second reference voltage Vref _2, the analog-to-digital conversion circuit 400 may output "01" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. Also, for example, when the first signal voltage Vsig _1 is greater than or equal to the second reference voltage Vref _2 and less than the third reference voltage Vref _3, the analog-to-digital conversion circuit 400 may output "10" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1. In addition, for example, when the first signal voltage Vsig _1 is greater than or equal to the third reference voltage Vref _3, the analog-to-digital conversion circuit 400 may output "11" as the first digital signal DS _1 corresponding to the first signal voltage Vsig _ 1.

Referring to fig. 10, a plurality of reference voltages including the first to third reference voltages Vref _1 to Vref _3 may unequally divide an operating voltage range of the analog-to-digital conversion circuit 400.

In one embodiment, the plurality of reference voltages may divide the operating voltage range of the analog-to-digital conversion circuit 400 such that a voltage interval between the reference voltages near the center of the operating voltage range of the analog-to-digital conversion circuit 400 is formed to be narrower than a voltage interval near sides of the operating voltage range.

In performing analog-to-digital conversion in a general neuromorphic computing device, it may be desirable to perform relatively precise operations near the center of the operating voltage range. The neuromorphic computing device 10 according to an embodiment of the inventive concept may unequally quantify the operating voltage range according to the precision of the operation required within a particular voltage range, thereby improving the accuracy of the operation required according to the neuromorphic computing device 10 according to circumstances.

Fig. 11 illustrates states of the second memory cell array 200 and the reference memory cell according to an embodiment of the inventive concept. Fig. 11 may be an example in which the number of reference bit lines in the second memory cell array 200B according to the embodiment of fig. 4B is 3. In particular, fig. 11 may illustrate an embodiment in which the operating voltage range of the analog-to-digital conversion circuit is divided unequally as shown in fig. 10. Fig. 11 is described with reference to fig. 1, and mainly differences with respect to fig. 6 are described.

In contrast to the case shown in fig. 5, in the embodiment shown in fig. 10, since the reference voltage must be further aggregated near the center of the operating voltage range, the reference resistance memory cell can be written differently from the written state of the reference resistance memory cell in fig. 6.

For example, for a reference resistive memory cell connected to first reference bit line RBL1, the ratio of the number of memory cells in the high resistance state HRS to the number of memory cells in the low resistance state LRS may be 6: 2. However, the inventive concept is not limited thereto, and for example, a ratio of the number of memory cells of the high resistance state HRS to the number of memory cells of the low resistance state LRS may be 7:3 according to a design or requirement specification.

In addition, for example, for a reference resistive memory cell connected to third reference bit line RBL3, the ratio of the number of memory cells in the high resistance state HRS to the number of memory cells in the low resistance state LRS may be 2: 6. However, the inventive concept is not limited thereto, and for example, the ratio of the number of memory cells of the high resistance state HRS to the number of memory cells of the low resistance state LRS may be 3:7 according to a design or requirement specification.

In performing analog-to-digital conversion in a general neuromorphic computing device 10, relatively precise operation near the center of the operating voltage range may be required. The neuromorphic computing device 10 according to embodiments of the inventive concept may unequally quantify the operating voltage range according to the accuracy of the operation required within a particular voltage range, thereby improving the accuracy of the operation required according to the neuromorphic computing device 10 according to circumstances.

Fig. 12 illustrates a second memory cell array 200 according to an embodiment of the inventive concept. Fig. 12 may be an example in which a load resistor is added to the second memory cell array 200B according to the embodiment of fig. 4B. Fig. 12 will be described mainly with respect to differences from the embodiment of fig. 4B.

The load resistor may be connected to a reference bit line or a reference source line of the second memory cell array 200.

For example, the first reference source line RSL1 may be connected to a first terminal of the first source transistor ST1 gated by the READ signal READ, and the first load resistor Rload _1 may be connected between a second terminal of the first source transistor ST1 and a ground node. Similarly, the second reference source line RSL2 may be connected to a first terminal of the second source transistor ST2 gated by the READ signal READ, and the second load resistor Rload _2 may be connected between a second terminal of the second source transistor ST2 and a ground node. Similarly, the nth reference source line RSLn may be connected to a first terminal of the nth source transistor STn gated by the READ signal READ, and the nth load resistor Rload _ n may be connected between a second terminal of the nth source transistor STn and a ground node.

The first to nth load resistors Rload _1 to Rload _ n may each have the same resistance value, but is not limited thereto, and may have different resistance values from each other. In one embodiment, the resistance values of the first to nth load resistors Rload _1 to Rload _ n are values between the resistance value when the reference resistance storage unit is in the low resistance state LRS and the resistance value when in the high resistance state HRS.

As described above, in the resistive memory device, the resistance characteristics thereof vary with time or temperature change, and thus, the distributions of the reference voltages Vref _1, Vref _2, and Vref _3 with reference to fig. 5 or fig. 10 may vary. Therefore, some of the reference voltages Vref _1, Vref _2, and Vref _3 may coincide with boundary values between voltages to be actually classified. According to the second memory cell array 200 shown in fig. 12, it is possible to solve the problem that the boundaries of the reference voltage and the voltage are overlapped or shifted due to the drift of the reference voltage caused by the presence of the load resistance.

Fig. 13 illustrates an analog-to-digital converter 401 according to an embodiment of the inventive concept. The analog-to-digital converter 401 of fig. 13 may correspond to any one of the first to mth analog-to-digital converters 401_1a to 401_ Ma of fig. 7A, or any one of the first to kth analog-to-digital converters 401_1B to 401_ kb of fig. 7B. Fig. 13 shows an embodiment in which the digital signal DS is a 3-bit digital signal when compared to fig. 9.

The comparison circuit 420 may include first to seventh comparators 421 to 427, and the encoder 442 may output a 3-bit digital signal DS.

The operations of the first through seventh comparators 421 through 427 and the encoder 442 may be performed similarly as described with respect to fig. 9.

For example, the first comparator 421 may compare the signal voltage Vsig with the first reference voltage Vref _1 to output the first comparison signal CS 1. For example, when the signal voltage Vsig is greater than the first reference voltage Vref _1, the first comparator 421 may output the first comparison signal CS1 of a first logic level "1", and when the signal voltage Vsig is not greater than the first reference voltage Vref _1, the first comparator 421 may output the first comparison signal CS1 of a second logic level "0".

Similarly, the second comparator 422 may compare the signal voltage Vsig with the second reference voltage Vref _2 to output a second comparison signal CS2, the third comparator 423 may compare the signal voltage Vsig with the third reference voltage Vref _3 to output a third comparison signal CS3, the fourth comparator 424 may compare the signal voltage Vsig with the fourth reference voltage Vref _4 to output a fourth signal CS4, the fifth comparator 425 may compare the signal voltage Vsig with the fifth reference voltage Vref _5 to output a fifth comparison signal CS5, the sixth comparator 426 may compare the signal voltage Vsig with the sixth reference voltage Vref _6 to output a sixth comparison signal CS6, and the seventh comparator 427 may compare the signal voltage Vsig with the seventh reference voltage Vref _7 to output a seventh comparison signal CS 7.

The encoder 442 may receive the supply voltage VCC as an input, the first comparison signal CS1 from the first comparator 421, the second comparison signal CS2 from the second comparator 422, the third comparison signal CS3 from the third comparator 423, the fourth comparison signal CS4 from the fourth comparator 424, the fifth comparison signal CS5 from the fifth comparator 425, the sixth comparison signal CS6 from the sixth comparator 426, and the seventh comparison signal CS7 from the seventh comparator 427. The encoder 442 may output a 3-bit digital signal DS based on the first comparison signal CS1, the second comparison signal CS2, the third comparison signal CS3, the fourth comparison signal CS4, the fifth comparison signal CS5, the sixth comparison signal CS6, and the seventh comparison signal CS 7.

Fig. 14 is a flowchart of an operation method of the neuromorphic computing device 10 according to an embodiment of the present inventive concept. Fig. 14 is described together with fig. 1.

The neuromorphic computing device 10 may activate all reference word lines connected to the second memory cell array 200 (S120). For example, word line drivers 150 may drive a plurality of reference word lines RWL1, RWL 2.., RWLN such that the plurality of reference word lines RWL1, RWL 2.., RWLN are all activated.

The neuromorphic computing device 10 may obtain a reference voltage value based on the reference currents Iref _1, Iref _ 2., Iref _ n output from the second memory cell array 200 (S140).

The neuromorphic computing device 10 may convert the plurality of signal voltages Vsig _1, Vsig _2, · Vsig _ M, which correspond to the plurality of read currents Iread _1, Iread _2, · Iread _ M, respectively, into the plurality of digital signals DS _1, DS _2, ·. DS _ M using the obtained reference voltage values (S160).

Fig. 15 illustrates an electronic system 1000 according to an embodiment of the inventive concept.

The electronic system 1000 may include a processor 1010, a storage device 1020, a coupling 1030, an input/output (I/O) device 1040, a power supply 1050, and a neuromorphic computing device 1060. Electronic system 1000 may further include various ports that may communicate with video cards, sound cards, memory cards, USB devices, or other systems.

The processor 1010 may control all operations of the electronic system 1000, and may execute an operating system, applications, and the like. Memory device 1020 may store data necessary for the operation of electronic system 1000. Coupling 1030 may communicate with an external device. I/O devices 1040 may include input tools such as keyboards, keypads, touch pads, touch screens, mice, remote controls, etc., and output tools such as displays, speakers, printers, etc. Power supply 1050 may provide the power required for operation of electronic system 1000.

The neuromorphic computing device 1060 may drive and/or execute a neural network system and may be a neuromorphic computing device according to embodiments of the inventive concepts described with reference to the preceding figures. The neuromorphic computing device 1060 may include a main memory cell array 1100 corresponding to the first memory cell array 100 in the previous drawings, and a reference memory cell array 1200, a current-voltage conversion circuit 1300, and an analog-to-digital conversion circuit 1400 corresponding to the second memory cell array 200 in the previous drawings.

For example, the main memory cell array 1100 may provide a plurality of read currents Iread to the current-voltage conversion circuit 1300, and the reference memory cell array 1200 may provide a plurality of reference currents Iref to the current-voltage conversion circuit 1300.

The current-voltage conversion circuit 1300 may output a plurality of signal voltages Vsig by converting the plurality of read currents Iread and supply the plurality of signal voltages Vsig to the analog-to-digital conversion circuit 1400. The current-voltage conversion circuit 1300 may output a plurality of reference voltages Vref by converting the plurality of reference currents Iref and provide the plurality of reference voltages Vref to the analog-to-digital conversion circuit 1400.

The analog-to-digital conversion circuit 1400 can convert a plurality of signal voltages Vsig into a plurality of digital signals DS by using a plurality of reference voltages Vref as references for analog-to-digital conversion.

While the inventive concept has been particularly shown and described with reference to embodiments thereof, it will be understood that various changes in form and details may be made therein without departing from the spirit and scope of the appended claims.