阅读说明：本技术 半导体电路及其操作方法 (Semiconductor circuit and method of operating the same ) 是由 林榆瑄 王超鸿 于 2018-10-17 设计创作，主要内容包括：本发明公开了一种半导体电路及其操作方法。半导体电路的操作方法包括以下步骤：在第一时间操作存储器电路以获得第一存储器状态信号S1；第一时间之后的第二时间操作存储器电路以获得第二存储器状态信号S2；计算第一存储器状态信号S1和第二存储器状态信号S2的差异以获得状态差异信号SD；计算以得到未补偿输出数据信号OD,未补偿输出数据信号OD相关于输入数据信号ID和第二存储器状态信号S2；以及对状态差异信号SD与未补偿输出数据信号OD进行计算以获得补偿输出数据信号OD’。(The invention discloses a semiconductor circuit and an operating method thereof. The method of operating a semiconductor circuit comprises the steps of: operating the memory circuit at a first time to obtain a first memory state signal S1; operating the memory circuit at a second time after the first time to obtain a second memory state signal S2; calculating a difference of the first memory state signal S1 and the second memory state signal S2 to obtain a state difference signal SD; calculating to obtain an uncompensated output data signal OD, the uncompensated output data signal OD being related to the input data signal ID and the second memory state signal S2; and calculating the state difference signal SD and the uncompensated output data signal OD to obtain a compensated output data signal OD'.)

1. A method of operating a semiconductor circuit, comprising:

operating a memory circuit at a first time to obtain a first memory state signal S1;

operating the memory circuit at a second time after the first time to obtain a second memory state signal S2;

calculating a difference between the first memory state signal S1 and the second memory state signal S2 to obtain a state difference signal SD;

calculating to obtain an uncompensated output data signal OD related to an input data signal ID and the second memory status signal S2; and

the state difference signal SD and the uncompensated output data signal OD are calculated to obtain a compensated output data signal OD'.

2. The method of operating the semiconductor circuit of claim 1, comprising:

programming a reference memory array at the first time to obtain the first memory state signal; and

the first memory state signal is stored in a memory device.

3. The method of claim 2, wherein the reference Memory array comprises a resistive random access Memory (ReRAM), a phase change Memory (phase change Memory), or a bridge-type Memory (bridge-access Memory), and the Memory device comprises a flash Memory (flash Memory), a Read-Only Memory (ROM), or a one-time-programmable (OTP) Memory, wherein the reference Memory array and the main Memory array have the same Memory structure.

4. The method of operating the semiconductor circuit of claim 1, comprising:

reading cell state signals of a plurality of memory cells of a reference memory array at the second time; and

the memory cell state signals are calculated to obtain the second memory state signal S2.

5. The method of claim 4, wherein the second memory state signal S2 is less than the largest of the memory cell state signals and greater than the smallest of the memory cell state signals.

6. The method of claim 4, wherein the second memory state signal S2 is an arithmetic mean (mean), median (mean), or mode of the memory cell state signals.

7. The method of operating the semiconductor circuit of claim 1, wherein:

the first and second memory state signals S1 and S2 have a first electrical measurement unit,

the input data signal ID has a second electrical property measurement unit,

the uncompensated output data signal OD and the compensated output data signal OD' have a third electrical measurement unit,

the first electrical property measurement unit, the second electrical property measurement unit and the third electrical property measurement unit are different electrical property measurement units and accord with ohm law with each other.

8. A semiconductor circuit, comprising:

a main memory array;

a reference memory array;

a memory device for storing a first memory state signal S1 obtained by operating the reference memory array at a first time; and

a processing circuit for reading a second memory state signal S2 of the reference memory array at a second time after the first time and for calculating an uncompensated output data signal OD associated with an input data signal ID and another second memory state signal S2 of the primary memory array at the second time, the memory device being electrically coupled to the processing circuit.

9. The semiconductor circuit of claim 8, wherein the reference Memory array and the main Memory array comprise resistive-random-access Memory (ReRAM), phase-change Memory (phase-change Memory), or bridge-type Memory (bridge-access Memory), and the Memory device comprises flash Memory (flash Memory), Read-Only Memory (ROM), or one-time-programmable (OTP) Memory.

10. The semiconductor circuit of claim 8, wherein the reference memory array and the main memory array have the same memory structure.

Technical Field

The present invention relates to a semiconductor circuit and an operating method thereof, and more particularly, to a neural network and an operating method thereof, which belong to the technical field of integrated circuits.

Background

With the development of software technology, the deep learning of the neural network defined by software greatly improves the artificial intelligence capability, such as image recognition, voice recognition, natural language understanding and decision-making, through a universal learning process. The appearance of a Hardware Neural Network (HNN) further reduces the Hardware size, cost and power consumption of deep learning systems. HNNs, which are composed of a network of neurons interconnected by synapses, may have thousands of synapses, where the weights (weights) of the synapses (synapses) may be optimized during training.

Disclosure of Invention

The invention relates to a semiconductor circuit and an operation method thereof.

According to an aspect of the present invention, there is provided a method of operating a semiconductor circuit, comprising the steps of: operating the memory circuit at a first time to obtain a first memory state signal S1; operating the memory circuit at a second time after the first time to obtain a second memory state signal S2; calculating a difference of the first memory state signal S1 and the second memory state signal S2 to obtain a state difference signal SD; calculating to obtain an uncompensated output data signal OD, the uncompensated output data signal OD being related to the input data signal ID and the second memory state signal S2; and calculating the state difference signal SD and the uncompensated output data signal OD to obtain a compensated output data signal OD'.

According to another aspect of the invention, a semiconductor circuit is provided that includes a main memory array, a reference memory array, a memory device, and a processing circuit. The memory device is configured to store a first memory state signal S1 obtained by operating the reference memory array at a first time. The processing circuit is used for reading a second memory status signal S2 of the reference memory array at a second time after the first time, and is used for calculating an uncompensated output data signal OD. The uncompensated output data signal OD is related to the input data signal ID and to another second memory state signal S2 for the primary memory array at a second time. The memory device is electrically coupled to the processing circuit.

For a better understanding of the above and other aspects of the invention, reference should be made to the following detailed description of the embodiments, taken in conjunction with the accompanying drawings, in which:

drawings

Fig. 1 illustrates a semiconductor circuit according to an embodiment concept.

Fig. 2 illustrates a method of operating a semiconductor circuit according to an embodiment concept.

Fig. 3 and 4 illustrate a method of operating a semiconductor circuit according to an embodiment concept.

Fig. 5 retains conductance range curve results of a test at 150 ℃ using a semiconductor circuit of compensated output data according to an embodiment concept.

Fig. 6 is the results of a 150 ℃ retention test using a comparative example with uncompensated output data.

Fig. 7 illustrates a method of operating a semiconductor circuit according to another embodiment concept.

Fig. 8 shows curve results of a 150 ℃ retention test for a semiconductor circuit using compensated output data according to an embodiment concept.

Fig. 9 is the results of a 150 ℃ retention test using a comparative example with uncompensated output data.

[ notation ] to show

102: a processing circuit;

106: a memory device;

108: a primary memory array;

110: a reference memory array;

122: a data input;

124: a data output terminal;

s552, S554, S556, S558, S560: and (5) carrying out the following steps.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to specific embodiments and the accompanying drawings.

Fig. 1 illustrates a semiconductor circuit according to an embodiment concept. Fig. 2 illustrates a method of operating a semiconductor circuit according to an embodiment concept.

Referring to fig. 1, the semiconductor circuit includes a memory circuit and a processing circuit 102. The memory circuit includes a memory device 106, a main memory array 108, a reference memory array 110. The memory devices 106, the primary memory array 108, and the reference memory array 110 may be provided in the same chip product, such as formed on a semiconductor chip. The processing circuitry 102 may be electrically coupled between the memory device 106, the main memory array 108, and the reference memory array 110. In one embodiment, the semiconductor circuit of fig. 1 is a neural network node (neural network node), and a plurality of input data signals ID are processed by the processing circuit 102 from the data input terminal 122 to become output data signals and are sent to the data output terminal 124. The weights of the synapses to which the input data signals ID are to be multiplied, respectively, are stored in a plurality of memory cells of the primary memory array 108. The processing circuit 102 may be coupled between the data input 122 and the data output 124. The processing circuit 102 may be electrically coupled between the data input 122 and the data output 124.

Referring to fig. 1 and fig. 2, the method for operating the semiconductor circuit includes step S552: the memory circuit is operated at a first time t1 to obtain a first memory state signal S1. In one embodiment, the reference memory array 110 may be programmed to a programmed memory state at a first time t1 to generate a first memory state signal S1 (e.g., conductance G) for the reference memory array 110_i，t1Or conductance G_ref，t1)。

A first memory state signal S1 (e.g., conductance G) from the reference memory array 110 may be read by the processing circuit 102_i，t1Or conductance G_ref，t1) And transmits the first memory state signal S1 to the memory device 106, such that the memory device 106 is utilized to store the first memory state signal S1. In one embodiment, the processing circuit 102 may generate a first memory state signal S1 from the reference memory array 110 (e.g., conductance values G for a plurality of memory cells of the reference memory array 110)_i，t1) Performs an operation on the first memory state signal S1 (e.g., conductance G)_ref，t1) To the memory device 106 for storage. In one embodiment, the first memory state signal S1 from the reference memory array 110 is a plurality of memory cells of the reference memory array 110The memory cell signal (e.g., conductance G) stored after programming_i，t1I 1, 2, 3 …) and the processing circuit 102 operates on the memory cell signals to obtain an operated first memory state signal S1 (e.g., conductance G)_ref，t1) Less than the memory cell state signal (e.g. conductance G)_i，t1I ═ 1, 2, 3 …) (e.g. conductance G)_i，t1Gmax) and is greater than the minimum of the memory cell state signals (e.g., conductance G)_i，t1Gmin) of the smallest. In one embodiment, the first memory state signal S1 (e.g., conductance G) is calculated_ref，t1) For storing cell state signals (e.g. conductance G)_i，t11, 2, 3 …), the arithmetic mean (mean), median (mean), or mode (mode).

In one embodiment, the reference memory array 110 and the main memory array 108 have the same memory structure, so the reference memory array 110 and the main memory array 108 after the visual programming have the same first memory state, i.e., the memory device 106 stores the first memory state signal S1 corresponding to the first memory state signal S1 derived from the main memory array 108.

In one embodiment, reference memory array 110 and primary memory array 108 may be programmed simultaneously at first time t1, and the programming steps are performed with the same programming parameters. In one embodiment, the programming step generates the same first memory state signal S1 for reference memory array 110 and primary memory array 108.

Then, step S554 is performed: the memory circuit is operated at a second time t2 after the first time t1 to obtain a second memory state signal S2.

A second memory state signal S2 (e.g., conductance G) from the reference memory array 110 at a second time t2 may be read utilizing the processing circuit 102_i，t2Or conductance G_ref，t2). In one embodiment, the processing circuit 102 may generate a second non-operated memory state signal S2 (e.g., conductance G) from the reference memory array 110_i，t2) Performing operation to obtain a second memory state signalS2 (e.g. conductance G)_ref，t2). In one embodiment, the second memory state signal S2 from the reference memory array 110 is a memory cell signal (e.g., conductance G) for a plurality of memory cells of the reference memory array 110_i，t2I 1, 2, 3 …) and the processing circuit 102 operates on the memory cell signals to obtain an operated second memory state signal S2 (e.g., conductance G)_ref，t2) Less than the memory cell state signal (e.g. conductance G)_i，t2I ═ 1, 2, 3 …) (e.g. conductance G)_i，t2Gmax) and is greater than the minimum of the memory cell state signals (e.g., conductance G)_i，t2Gmin) of the smallest. In one embodiment, the second memory state signal S2 (e.g., conductance G) is calculated_ref，t2) Memory cell state signal (e.g. conductance G)_i，t21, 2, 3 …), the arithmetic mean (mean), median (mean), or mode (mode). In one embodiment, the first memory state signal S1 (e.g., conductance G) is calculated_ref，t1) ANDing the second memory state signal S2 (e.g., conductance G)_ref，t2) Are obtained based on the same operation, e.g., all median.

In one embodiment, the reference memory array 110 and the main memory array 108 have the same memory structure, and thus the reference memory array 110 and the main memory array 108 at the second time t2 are considered to have the same second memory state, i.e., the second memory state signal S2 read or operated by the processing circuit 102 is equivalent to the second memory state signal S2 derived from the main memory array 108.

In one embodiment, the reference memory array 110 and the main memory array 108 include a resistive random access memory (ReRAM), a phase change memory (phase change memory), or a bridge-type memory (bridge-access memory), the data retention of which may vary with time, i.e., the memory status signals from the reference memory array 110 and/or the main memory array 108 transition from the first memory status signal S1 at the first time t1 to the second memory status signal S2 at the second time t 2. And the difference between the first memory state signal S1 and the second memory state signal S2 may be larger as the retention time (retention time) interval is larger. Therefore, the memory state signals (or weights) of primary memory array 108 at second time t2 are not accurate. In one embodiment, the reference memory array 110 and the main memory array 108 have the same one-transistor-one-resistance random access memory (1-resistor-1-ReRAM, 1T1R) structure, and use 1T1R of the main memory array 108 as synapses.

Then, step S556 is performed: the difference of the first memory state signal S1 and the second memory state signal S2 is calculated to obtain a state difference signal SD. The processing circuit 102 may be utilized to read a first memory state signal S1 (e.g., conductance G) present in the memory device 106_ref，t1) And in response to a first memory state signal S1 (e.g., conductance G)_ref，t1) And a second memory state signal S2 (e.g., conductance G)_ref，t2) A calculation is performed to obtain a state difference signal SD (e.g., Δ G)_ref，t2-G_ref，t1). In one embodiment, the Memory device 106 includes a Flash Memory (Flash Memory), a Read-Only Memory (ROM), or a One Time Program (OTP) Memory, and the Memory state of the Memory device is substantially not changed with time, so that the first Memory state signal S1 stored in the Memory device 106 is not changed with time, and the Memory state at the first time t1 can be substantially maintained even at the second time t 2. In one embodiment, the state difference signal SD is the second memory state signal S2 minus the first memory state signal S1, which may be expressed as Δ G in one embodiment_ref，t2-G_ref，t1. That is, the state difference signal SD (e.g., difference Δ) may be viewed as the degree of difference in the memory state signals of the primary memory array 108 between the first time t1 and the second time t 2.

Step S558 is performed after step S552: to obtain the uncompensated output data signal OD. The uncompensated output data signal OD is related to the input data signal ID at the data input 122 and to the second memory state signal S2 at the second time t2 for the primary memory array 108. The data can be read from the data by the processing circuit 102The input data signal ID at the input terminal 122 is read, the second memory state signal S2 of the primary memory array 108 is read, and the uncompensated output data signal OD is calculated based on the input data signal ID and the second memory state signal S2. In one embodiment, the input data signal ID is a voltage V_iThe second memory state signal S2 is conductance G_i，t2The uncompensated output data signal OD being the current I_out，t2＝∑V_i*G_i，t2I is 1, 2, 3 …. Since the uncompensated output data signal OD is related to the second memory state signal S2 at the second time t2 of the primary memory array 108, the difference between the second memory state signal S2 and the first memory state signal S1 at the first time t1 causes the uncompensated output data signal OD to decrease in accuracy.

Step S560: the state difference signal SD and the uncompensated output data signal OD are calculated to obtain a compensated output data signal OD'. In one embodiment, the processing circuit 102 may be utilized to perform this step. In an embodiment, the accuracy of the calculated compensated output data signal OD 'is higher than the accuracy of the uncompensated output data signal OD, i.e. the compensated output data signal OD' is closer to the output data signal obtained based on the memory state of the primary memory array 108 at the first time t 1. In one embodiment, the compensated output data signal OD' is the current I_out，t2m＝I_otut，t2-Δ*∑V_i。

The first and second memory state signals S1 and S2 have a first electrical measurement unit, the input data signal ID has a second electrical measurement unit, the uncompensated output data signal OD and the compensated output data signal OD 'have a third electrical measurement unit, and the first, second and third electrical measurement units are different electrical measurement units and conform to ohm' S law. In one embodiment, the first electrical property measurement unit is a conductance (electrical conductivity) unit, the second electrical property measurement unit is a voltage unit, and the third electrical property measurement unit is a current unit.

In one embodiment, SD is S2-S1, OD is ID S2, and OD is OD- (SD).

Fig. 3 and 4 illustrate a method of operating a semiconductor circuit according to an embodiment concept. Referring to FIG. 3, a plurality of input data V1, V2, V3 … (i.e., voltage V)_iI 1, 2, 3 …) are multiplied by the weight G stored in the memory cell of the primary memory array, respectively₁、G₂、G₃… (i.e. conductance G)_iI 1, 2, 3 …, the conductance at the first time t1 may be denoted as G_i，t1The conductance at the second time t1 may be labeled G_i，t2) Each product of (1)₁、I₂、I₃(i.e. current I)_iI1, 2, 3 …, the current at the first time t1 may be denoted as I_i，t1The current at the second time t2 can be labeled as I_i，t2) Is the sum of the output data I_out＝∑V_i*G_i. Referring to FIG. 4, in this embodiment, the memory state G stored from the memory cells of the reference memory array 110₁、G₂、G₃… a reference memory state G is obtained by calculating a conductance range G (range) between a maximum memory state Gmax and a minimum memory state Gmin_refI.e. Gmin. ltoreq.G_refLess than or equal to Gmax. Gmin, Gmax gradually shift from the first time t1 to the second time t2, so the reference memory state G at the first time t1 and the second time t2_refIs not the same (i.e. G)_ref，t1≠G_ref，t2). Reading a reference memory state G stored by a memory device at a first time t1_ref，t1And calculating a reference memory state G_ref，t1Reference memory state G of a memory cell of reference memory array 110 at a second time t2_ref，t2The difference between the two is expressed by the formula G_ref，t2＝G_ref，t1+ Δ denotes, i.e. Δ ═ G_ref，t2-G_ref，t1. The output data associated with the primary memory array at a first time t1 is I_out，t1＝∑V_i*G_i，t1. The uncompensated output data associated with the primary memory array at the second time t2 is I_out，t2＝∑V_i*G_i，t2＝∑V_i*(G_i，t1+Δ)＝∑V_i*G_i，t1+Δ∑V_i. According to the data relation, compensated output dataI_out，t2m＝I_out，t2-Δ*∑V_i. Compared with the uncompensated output data I_out，t2Compensated output data I_out，t2mAnd is more accurate. Fig. 5 shows the result of the G range curve of the 150 ℃ retention test for a semiconductor circuit using compensated output data according to an embodiment concept. Fig. 6 is the results of a 150 ℃ retention test using a comparative example with uncompensated output data. The vertical axis is based on the cumulative probability (cumulative probability) of 100 memory arrays. As can be seen from the results of fig. 5 and 6, the output data using compensation according to the embodiment concept has better accuracy (accuracy) and reliability.

FIG. 7 illustrates a method of operating a semiconductor circuit according to another concept of an embodiment in which two reference memory states are taken between a maximum memory state and a minimum memory state stored by memory cells of a reference memory array 110, i.e., a first reference memory state G at a first time t1_ref1，t1And a second reference memory state G_ref2，t1And a first reference memory state G at a second time t2_ref1，t2And a second reference memory state G_ref2，t2. The first reference memory state at the first time t1 and the second time t2 are different in formula G_ref1，t2＝a*G_ref1，t1+ b represents; the second reference memory state at the first time t1 and the second time t2 are different in formula G_ref2，t2＝a*G_ref2，t1+ b represents; from the above relationship, the expression G can be derived_t2＝(d_t2/d_t1)*(G_t1-G_tefl，t2)+G_ref1，t1Wherein d is_t1＝G_ref2，t1-G_ref1，t1And d is_t2＝G_ref2，t2-G_ref1，t2. The output data of the primary memory array at a first time t1 is ∑ V_i*G_i，t1＝I_out，t1(the concept can refer to fig. 3). The uncompensated output data of the primary memory array at a second time t2 is ∑ V_i*G_i，t2＝∑V_i*[A(G_i，t1-B)+C]＝A∑V_i*G_i，t1-AB∑V_i+C∑V_i. Can pass the sum of input data ∑ V_iAnd a reference storeThe state difference compensates the uncompensated output data to obtain compensated output data that is more accurate than the uncompensated output data. Fig. 8 is a G range curve result of a 150 ℃ retention test for a semiconductor circuit using compensated output data according to an embodiment concept. FIG. 9 shows the results of a 150 ℃ retention test on a comparative example with uncompensated output data. As can be seen from the results of fig. 8 and 9, the output data using compensation according to the concept of the embodiment has better accuracy and reliability.

While the present invention has been described with reference to the above embodiments, it is not intended to be limited thereto. Those skilled in the art to which the invention pertains will readily appreciate that various modifications and adaptations can be made without departing from the spirit and scope of the invention. Therefore, the protection scope of the present invention is defined by the claims of the application.

The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.