WO2024048704A1 - 人工知能処理装置および人工知能処理装置の重み係数書き込み方法 - Google Patents
人工知能処理装置および人工知能処理装置の重み係数書き込み方法 Download PDFInfo
- Publication number
- WO2024048704A1 WO2024048704A1 PCT/JP2023/031711 JP2023031711W WO2024048704A1 WO 2024048704 A1 WO2024048704 A1 WO 2024048704A1 JP 2023031711 W JP2023031711 W JP 2023031711W WO 2024048704 A1 WO2024048704 A1 WO 2024048704A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- memory element
- nonvolatile memory
- variable resistance
- resistance nonvolatile
- voltage
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C27/00—Electric analogue stores, e.g. for storing instantaneous values
- G11C27/005—Electric analogue stores, e.g. for storing instantaneous values with non-volatile charge storage, e.g. on floating gate or MNOS
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
- G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
- G06F7/544—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices for evaluating functions by calculation
- G06F7/5443—Sum of products
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06G—ANALOGUE COMPUTERS
- G06G7/00—Devices in which the computing operation is performed by varying electric or magnetic quantities
- G06G7/12—Arrangements for performing computing operations, e.g. operational amplifiers
- G06G7/14—Arrangements for performing computing operations, e.g. operational amplifiers for addition or subtraction
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06G—ANALOGUE COMPUTERS
- G06G7/00—Devices in which the computing operation is performed by varying electric or magnetic quantities
- G06G7/12—Arrangements for performing computing operations, e.g. operational amplifiers
- G06G7/16—Arrangements for performing computing operations, e.g. operational amplifiers for multiplication or division
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06G—ANALOGUE COMPUTERS
- G06G7/00—Devices in which the computing operation is performed by varying electric or magnetic quantities
- G06G7/48—Analogue computers for specific processes, systems or devices, e.g. simulators
- G06G7/60—Analogue computers for specific processes, systems or devices, e.g. simulators for living beings, e.g. their nervous systems ; for problems in the medical field
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
- G06N3/065—Analogue means
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H10—SEMICONDUCTOR DEVICES; ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
- H10B—ELECTRONIC MEMORY DEVICES
- H10B63/00—Resistance change memory devices, e.g. resistive RAM [ReRAM] devices
-
- H—ELECTRICITY
- H10—SEMICONDUCTOR DEVICES; ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
- H10B—ELECTRONIC MEMORY DEVICES
- H10B63/00—Resistance change memory devices, e.g. resistive RAM [ReRAM] devices
- H10B63/80—Arrangements comprising multiple bistable or multi-stable switching components of the same type on a plane parallel to the substrate, e.g. cross-point arrays
- H10B63/82—Arrangements comprising multiple bistable or multi-stable switching components of the same type on a plane parallel to the substrate, e.g. cross-point arrays the switching components having a common active material layer
Definitions
- the present disclosure relates to an artificial intelligence processing device and a weighting coefficient writing method thereof, and particularly relates to an artificial intelligence processing device using a variable resistance nonvolatile memory element whose resistance value changes depending on an applied electrical signal.
- IoT Internet of Things
- AI artificial intelligence
- neural network technology which is an engineering imitation of human brain-type information processing, is used, and research and development of semiconductor integrated circuits that can perform neural network calculations at high speed and with low power consumption is being actively conducted. There is.
- Patent Document 1 discloses a conventional neural network calculation circuit.
- a neural network consists of basic elements called neurons (sometimes called perceptrons), in which multiple inputs are connected by connections called synapses, each with a different connection weighting coefficient, and by connecting multiple neurons to each other, It can perform advanced computational processing such as image recognition and voice recognition.
- the neuron performs a product-sum calculation operation in which the products of each input and each connection weighting coefficient are added together.
- the product-sum calculation circuit includes a memory circuit and a register circuit that store inputs and connection weighting coefficients, a multiplication circuit that multiplies the input and connection weighting coefficients, an accumulator circuit that cumulatively adds the multiplication results, and a control that controls the operation of these circuit blocks. Consists of circuits. All future circuit blocks will be composed of digital circuits.
- the neural network calculation circuit is constructed using a variable resistance nonvolatile memory that can set multi-gradation analog resistance values or conductance (hereinafter simply referred to as "conductance"), which is the reciprocal of the resistance value.
- conductance a conductance corresponding to a coupling weighting coefficient is stored in a memory element, a voltage value corresponding to an input is applied to the nonvolatile memory element, and at this time, an analog current value flowing through the nonvolatile memory element is utilized.
- the product-sum calculation operation performed in a neuron stores multiple connection weighting coefficients as conductance in multiple non-volatile memory elements, applies a voltage value corresponding to the input to multiple non-volatile memory elements, and stores multiple connection weight coefficients as conductance in multiple non-volatile memory elements. This is performed by obtaining an analog current value, which is the sum of the current values flowing through the elements, as a result of a sum-of-products operation.
- the conductance to be written to each nonvolatile memory is calculated from the coupling weight coefficient derived in advance, and the conductance is written to each nonvolatile memory.
- Non-Patent Document 1 Yet another example of a conventional neural network calculation circuit is disclosed in Non-Patent Document 1.
- the neural network calculation circuit is configured using a variable resistance nonvolatile memory whose conductance can be set, and the conductance corresponding to the coupling weighting coefficient is stored in the nonvolatile memory element, and the analog voltage corresponding to the input is stored in the nonvolatile memory element. They are the same in that they apply a value to a non-volatile memory element and utilize an analog current value flowing through the non-volatile memory element at this time.
- the amount of change between the conductance at the time before writing and the conductance set after writing is first derived, and the amount of change in conductance is calculated according to the amount of change in conductance. Writing is performed to the non-volatile memory element.
- the write operation is generally performed based on the conductance itself written to the non-volatile memory
- the neural network calculation circuit as in Non-Patent Document 1 The circuit differs in that a write operation is generally performed based on the amount of change in conductance written into a nonvolatile memory before and after writing.
- the neural network arithmetic circuits using nonvolatile memory elements disclosed in Patent Document 2 and Non-Patent Document 1 can both achieve lower power consumption than the neural network arithmetic circuits configured with digital circuits described above. Therefore, in recent years, process development, device development, and circuit development of variable resistance nonvolatile memory whose conductance can be set have been actively conducted.
- the neural network arithmetic circuit which performs a write operation based on the conductance itself written to the nonvolatile memory, writes using the conductance itself derived in advance, so it is not possible to accurately write the conductance to the nonvolatile memory element. Therefore, after the product is shipped (that is, after the artificial intelligence processing device such as the neural network calculation circuit is shipped), it is suitable for the "artificial intelligence processing device for inference” that mainly processes only the product-sum calculations in the neural network. .
- update of the connection weight coefficients that is, learning
- conductance is frequently updated after the product is shipped.
- a problem with the "artificial intelligence processing device for learning” is that learning cannot be performed efficiently when using a neural network calculation circuit that performs writing operations based on conductance itself.
- FIG. 1A shows a learning process in a neural network arithmetic circuit in which a write operation is performed based on conductance itself.
- inference is performed using the connection weight coefficient of each nonvolatile memory element (S10), the difference between the inference result and the teacher label is checked (S11), and the connection weight coefficient of each nonvolatile memory element is The amount of change at the time of update is calculated (S12), the current coupling weighting coefficient is read out for each nonvolatile memory element (S13), and the updated coupling weighting coefficient is calculated for each nonvolatile memory element based on the readout coupling weighting coefficient.
- the calculated connection weighting coefficients are calculated (S14) and written into each nonvolatile memory element (S15). In other words, six steps are required to write the connection weighting coefficients themselves.
- FIG. 1B shows a learning process in a neural network arithmetic circuit in which a writing operation is performed based on the amount of change in conductance before and after writing.
- inference is performed using the connection weight coefficient of each nonvolatile memory element (S10), the difference between the inference result and the teacher label is checked (S11), and the connection weight coefficient of each nonvolatile memory element is The amount of change at the time of updating is calculated (S12), and the connection weighting coefficient is updated for each nonvolatile memory element so that the current connection weighting coefficient changes by the calculated amount of change in the connection weighting coefficient (S20).
- S10 connection weight coefficient of each nonvolatile memory element
- S11 the difference between the inference result and the teacher label
- connection weight coefficient of each nonvolatile memory element is The amount of change at the time of updating is calculated (S12), and the connection weighting coefficient is updated for each nonvolatile memory element so that the current connection weighting coefficient changes by the calculated amount of change in the connection weighting coefficient (S20).
- the conductance can be updated in the learning process shown in Figure 1B, and the conductance can be updated after the product is shipped. It is suitable for "artificial intelligence processing equipment for learning" that updates frequently.
- the neural network calculation circuit in which the learning process shown in FIG. It is necessary to change the conductance to a value derived in advance by striking multiple times in succession. At that time, there is a problem that variations occur in the writing characteristics and initial settings cannot be performed with high efficiency and accuracy.
- An object of the present invention is to provide an artificial intelligence processing device using a variable resistance nonvolatile memory element capable of performing the following steps, and a method for writing weighting coefficients therein.
- An artificial intelligence processing device includes one substrate, a first variable resistance nonvolatile memory element and a second variable resistance nonvolatile memory element, each of which is mounted on the one substrate and has the same structure and maintains conductance. and a variable resistance nonvolatile memory element, and applying a first voltage pulse having a first voltage to the first variable resistance nonvolatile memory element.
- the second resistance change is achieved by rewriting the conductance of the nonvolatile memory element and applying a second voltage pulse having a second voltage different from the first voltage to the second variable resistance nonvolatile memory element. and a write circuit that rewrites the conductance of the nonvolatile memory element.
- a weighting coefficient writing method for an artificial intelligence processing device is a weighting coefficient writing method possessed by the artificial intelligence processing device, wherein the artificial intelligence processing device holds a weighting coefficient in a sum-of-products operation as a conductance. , comprising a first variable resistance nonvolatile memory element and a second variable resistance nonvolatile memory element having the same structure, and the writing method includes a first variable resistance nonvolatile memory element with respect to the first variable resistance nonvolatile memory element.
- the setting (initial setting) of the coupling weighting coefficient with high efficiency and high accuracy at the time of product shipment, etc. makes it possible to simultaneously update (learn) the connection weighting coefficients in a highly efficient manner later on.
- FIG. 1A is a diagram showing a learning process in a neural network arithmetic circuit in which a write operation is performed based on conductance itself.
- FIG. 1B is a diagram showing a learning process in a neural network calculation circuit in which a writing operation is performed based on the amount of change in conductance before and after writing.
- FIG. 2A is a schematic cross-sectional view of the first variable resistance nonvolatile memory element according to the embodiment.
- FIG. 2B is a schematic cross-sectional view of the second variable resistance nonvolatile memory element according to the embodiment.
- FIG. 1A is a diagram showing a learning process in a neural network arithmetic circuit in which a write operation is performed based on conductance itself.
- FIG. 1B is a diagram showing a learning process in a neural network calculation circuit in which a writing operation is performed based on the amount of change in conductance before and after writing.
- FIG. 2A is a schematic cross-sectional view of the first variable resistance nonvol
- FIG. 3 is a diagram showing a change in conductance when the resistance of the first variable resistance nonvolatile memory element according to the embodiment is changed by continuously applying a plurality of voltage pulses of the same polarity and the same voltage.
- FIG. 4 shows a second variable resistance nonvolatile memory element according to an embodiment using multiple voltages of the same polarity and the same voltage, which are different in voltage magnitude from those of the first variable resistance nonvolatile memory element 10.
- FIG. 6 is a diagram showing changes in conductance when resistance is changed by continuous application of pulses.
- FIG. 5A is a diagram illustrating definitions of a first variable resistance nonvolatile memory element and a second variable resistance nonvolatile memory element.
- FIG. 5B is a diagram illustrating driving conditions of the first variable resistance nonvolatile memory element and the second variable resistance nonvolatile memory element.
- FIG. 6A is a circuit diagram of a memory cell in the prior art.
- FIG. 6B is a cross-sectional view of a memory cell in the prior art.
- FIG. 7A is a circuit diagram of a memory cell according to an embodiment.
- FIG. 7B is a cross-sectional view illustrating an example of a memory cell according to the embodiment.
- FIG. 7C is a cross-sectional view showing an example of the memory cell according to the embodiment, which is different from FIG. 7B.
- FIG. 8A is a circuit diagram of a memory cell according to another embodiment.
- FIG. 8B is a diagram showing an example of a cross-sectional view of a memory cell of an artificial intelligence processing device in which the first variable resistance nonvolatile memory element described in FIG. 8A is mounted.
- FIG. 8C is a diagram showing an example of a cross-sectional view of a memory cell of the artificial intelligence processing device in which the second variable resistance nonvolatile memory element described in FIG. 8A is mounted.
- FIG. 9A is a block diagram showing a model of the artificial intelligence processing device according to the embodiment.
- FIG. 9B is a diagram illustrating the function of the neuron shown in FIG. 9A.
- FIG. 10A is a diagram showing an example of a circuit that implements the neuron shown in FIG. 9B.
- FIG. 10B is a diagram illustrating another example circuit that implements the neuron shown in FIG. 9B.
- FIG. 11 is a block diagram showing the overall configuration of an artificial intelligence processing device made up of neurons shown in FIG. 10A.
- FIG. 12A is a diagram showing an example of voltages applied via the word line WL, bit lines BL1, BL2, and source line SL during writing and reading to the memory cell shown in FIG. 10A.
- FIG. 12B is a diagram showing an example of voltages applied via word lines WL1, WL2, bit line BL, and source line SL during writing and reading to the memory cell shown in FIG. 10B.
- FIG. 13 is a flowchart showing an example of the operation of the control circuit shown in FIG.
- constituent elements in the following embodiments constituent elements that are not described in the independent claims representing the top concept of the present disclosure may be used in forms that may be adopted although they are not necessarily necessary to achieve the objects of the present disclosure. It is explained as configuring.
- the present inventor has developed a system for setting high-accuracy coupling weighting coefficients (initial setting) at the time of product shipment, etc. in an artificial intelligence processing device using variable resistance non-volatile memory elements, and for setting highly efficient coupling weighting coefficients after product shipment, etc.
- high-accuracy coupling weighting coefficients initial setting
- variable resistance non-volatile memory elements variable resistance non-volatile memory elements
- the present inventors have developed an artificial intelligence processing device using variable resistance nonvolatile memory elements by mounting two variable resistance nonvolatile memory elements with the same structure on one substrate, and By applying continuous voltage pulses of the same polarity and voltage to each nonvolatile memory element under different driving conditions (specifically, voltages of different magnitudes), the conductance of one nonvolatile memory element changes little by little.
- the other non-volatile memory element has a structure in which the conductance changes greatly when the first voltage pulse is applied, and the amount of conductance change due to the second and subsequent voltage pulses is small, thereby solving the problem of resistance change. Operational efficiency of highly accurate connection weighting coefficient setting (initial setting) at the time of product shipment and highly efficient connection weighting coefficient updating (learning) after product shipment in an artificial intelligence processing device using non-volatile memory elements
- “same structure” means that it is made of substantially the same material and has substantially the same structure, and the content of minute amounts of material components such as impurities is different, or the same structure is used. Even if there are differences due to dimensional variations that occur during the manufacturing process, they are still included in the "same structure.” Typically, identical structures are formed by identical manufacturing processes.
- variable resistance nonvolatile memory element one of the variable resistance nonvolatile memory elements whose conductance is to be changed under two different driving conditions according to the embodiment will be referred to as a "first variable resistance nonvolatile memory element,” and the other will be referred to as a “second variable resistance nonvolatile memory element.”
- first variable resistance nonvolatile memory element one of the variable resistance nonvolatile memory elements whose conductance is to be changed under two different driving conditions according to the embodiment
- second variable resistance nonvolatile memory element An example of the configuration of the first variable resistance nonvolatile memory element and the second variable resistance nonvolatile memory element will be described. As described above, the first variable resistance nonvolatile memory element and the second variable resistance nonvolatile memory element are driven under different driving conditions, but have the same structure.
- variable resistance nonvolatile memory element and the second variable resistance nonvolatile memory element are driven under different driving conditions.
- FIG. 2A is a schematic diagram showing a configuration example of the first variable resistance nonvolatile memory element 10 according to the embodiment.
- the first variable resistance nonvolatile memory element 10 is a variable resistance nonvolatile memory element that is used for setting (initial setting) highly accurate coupling weighting coefficients at the time of product shipment, etc. It is a sexual memory element.
- the first variable resistance nonvolatile memory element 10 includes a substrate 1, a first electrode 2 formed on the substrate 1, and a metal oxide layer formed on the first electrode 2. It includes a variable resistance layer 3 formed and a second electrode 4 formed on the variable resistance layer 3. The first electrode 2 and the second electrode 4 are electrically connected to the variable resistance layer 3. That is, the first variable resistance nonvolatile memory element 10 includes a first electrode 2 , a second electrode 4 , and a variable resistance layer 3 interposed between the first electrode 2 and the second electrode 4 .
- first electrode 2 may be of the same size as the second electrode 4, and the first electrode 2, second electrode 4, and variable resistance layer 3 may be arranged upside down or sideways. It may be placed in
- the substrate 1 is composed of a silicon substrate on which circuit elements such as transistors are formed, for example.
- the first electrode 2 or the second electrode 4 is made of a noble metal, such as Au (gold), Pt (platinum), Ir (iridium), Pd (palladium), Ru (ruthenium), etc. Constructed using one material.
- the second electrode 4 in contact with the second tantalum oxide layer 3b is made of a noble metal
- the first electrode 2 is made of a noble metal or a non-noble metal. Due to such electrode characteristics, the first variable resistance nonvolatile memory element 10 has resistance variable characteristics shown in FIG. 3, which will be described later.
- the resistance (in other words, conductance) of the variable resistance layer 3 changes depending on the voltage pulse applied between the first electrode 2 and the second electrode 4.
- the resistance change layer 3 is made of a metal oxide, and is formed by laminating a first tantalum oxide layer 3a and a second tantalum oxide layer 3b.
- the oxygen content of the second tantalum oxide layer 3b is higher than the oxygen content of the first tantalum oxide layer 3a.
- composition of the first tantalum oxide layer 3a is TaOx, 0 ⁇ x ⁇ 2.5, and when the composition of the second tantalum oxide layer 3b is TaOy, if x ⁇ y good.
- FIG. 2B is a schematic diagram showing a configuration example of the second variable resistance nonvolatile memory element 20 according to the embodiment.
- the second variable resistance nonvolatile memory element 20 is a variable resistance nonvolatile memory element used for highly efficient updating (learning) of coupling weighting coefficients after product shipment, etc., whose use is determined by the driving conditions described later. It is a memory element.
- the second variable resistance nonvolatile memory element 20 has the same structure as the first variable resistance nonvolatile memory element 10, and includes a substrate 11 and a substrate 11 formed on the substrate 11. It includes a first electrode 12 , a variable resistance layer 13 formed as a metal oxide layer on the first electrode 12 , and a second electrode 14 formed on the variable resistance layer 13 .
- the resistance change layer 13 is made of metal oxide, and is formed by laminating a first tantalum oxide layer 13a and a second tantalum oxide layer 13b.
- the substrate 11, the first electrode 12, the variable resistance layer 13, the second electrode 14, the first tantalum oxide layer 13a, and the second tantalum oxide layer 13b correspond to the first variable resistance nonvolatile memory element 10, respectively. Constructed from the same material as the components.
- the first electrode 2 is formed on the substrate 1 by sputtering. Thereafter, a tantalum oxide layer is formed on the first electrode 2 by a so-called reactive sputtering method in which a Ta target is sputtered in argon gas and oxygen gas.
- the oxygen content in the tantalum oxide layer can be easily adjusted by changing the flow rate ratio of oxygen gas to argon gas. Note that the substrate temperature can be set to room temperature without being particularly heated.
- the outermost surface of the tantalum oxide layer formed as described above is oxidized to modify the surface.
- a layer with a higher oxygen content is formed by sputtering using a tantalum oxide (for example, Ta2O5) target with a high oxygen content.
- a region (second region) having a higher oxygen content than the unoxidized region (first region) of the tantalum oxide layer is formed on the surface of the previously formed tantalum oxide layer.
- variable resistance layer 3 is constituted by the layer 3b.
- the second electrode 4 is formed by sputtering on the variable resistance layer 3 formed as described above.
- the first electrode 2 the oxygen-deficient first tantalum oxide layer 3a, the second tantalum oxide layer 3b and the second electrode 4 to form a first variable resistance nonvolatile memory element 10 in which the variable resistance layer 3 is sandwiched between the first electrode 2 and the second electrode 4.
- patterning was performed at once using the same mask in this step, but patterning may be performed individually for each step.
- the sizes and shapes of the first electrode 2, the second electrode 4, and the variable resistance layer 3 can be adjusted using a photomask and photolithography.
- the size of the second electrode 4 and the variable resistance layer 3 is 0.1 ⁇ m x 0.1 ⁇ m (area 0.01 ⁇ m 2 ), and the size of the portion where the first electrode 2 and the variable resistance layer 3 are in contact is also
- the size is 0.1 ⁇ m ⁇ 0.1 ⁇ m (area 0.01 ⁇ m 2 )
- the size and shape are not limited to this, and can be changed as appropriate depending on the layout design.
- a plurality of voltage pulses of the same polarity and the same voltage with different magnitudes for each of the variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 for example, a first voltage pulse having a first voltage for the memory element 10, and a second voltage pulse having a second voltage different from the first voltage for the second variable resistance nonvolatile memory element 20).
- the conductances of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 change between them.
- a write voltage pulse of negative polarity is applied between the first electrode 2 and the second electrode 4 of the first variable resistance nonvolatile memory element 10 or the second variable resistance nonvolatile memory element 20.
- the process in which the conductance of the variable resistance layer 3 increases and the variable resistance layer 3 changes from a high resistance state to a low resistance state is called resistance lowering (or “set”).
- the conductance of the variable resistance layer 3 is reduced, and the process of changing the variable resistance layer 3 from a low resistance state to a high resistance state is increased (or (also called “reset”).
- lowering and increasing the resistance of the variable resistance nonvolatile memory element is also referred to as "writing” or "rewriting” the conductance or resistance value.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 operate as nonvolatile memory elements.
- the initial process will be explained.
- the initial process is normally executed only once before the first writing.
- the initial process is a preparatory process for realizing a stable resistance change operation in the subsequent lowering and higher resistance, and is also called "breaking" or "forming.”
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 immediately after manufacture have an initial resistance value that is higher than the high resistance state when the resistance changes. Even if a low resistance voltage pulse or a high resistance voltage pulse during normal operation is applied in this state, no resistance change occurs.
- an initial voltage pulse is applied between the first electrode 2 and the second electrode 4.
- the magnitude of the voltage applied to each of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 (that is, the driving conditions) will be different;
- the conductance of the first variable resistance non-volatile memory element 10 and the second variable resistance non-volatile memory element 20 changes between a high resistance state and a low resistance state. It will change between.
- the initial process is a first resistance change in an initial state to which no voltage has been applied after manufacturing the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20. This is a process performed on the type nonvolatile memory element 10 and the second variable resistance type nonvolatile memory element 20.
- a local region called a filament having a degree of oxygen deficiency greater than the degree of oxygen deficiency in the surrounding area is formed in the resistance change layer 3.
- the filament is formed through the initial process, but it is not necessarily necessary to form the filament through the initial process, and it is not necessary to form the filament through the initial process.
- a large oxide layer may be provided.
- FIG. 3 shows that in the first variable resistance non-volatile memory element 10 according to the present embodiment, the first variable resistance non-volatile memory element 10 according to the present embodiment is This is a resistance change characteristic when the conductance of the nonvolatile memory element 10 is changed to a high resistance state or a low resistance state.
- the horizontal axis shows the number of voltage pulse applications, and the vertical axis shows conductance.
- FIG. 3 when the resistance is lowered by continuously applying multiple voltage pulses of the same polarity and the same voltage to the first variable resistance nonvolatile memory element 10, a black circle on the upper side of the rectangular waveform in FIG.
- the amount of change in conductance with the first voltage pulse is large, and the conductance increases from a high resistance state to a state close to a low resistance state due to the first voltage pulse. Thereafter, even if the second and third pulses are applied in succession, the amount of change in conductance remains very small compared to the amount of change in conductance caused by the first voltage pulse.
- a plurality of voltage pulses of the same polarity and the same voltage are continuously applied to the first variable resistance nonvolatile memory element 10 to increase the resistance, as shown in the black circle plot on the lower side of the rectangular waveform in FIG.
- the amount of change in conductance with the first voltage pulse is large, and the conductance decreases from a low resistance state to a state close to a high resistance state due to the first voltage pulse. Thereafter, even if the second and third pulses are applied in succession, the amount of change in conductance remains very small compared to the amount of change in conductance caused by the first voltage pulse. In other words, the conductance of the first variable resistance nonvolatile memory element 10 is reduced by continuous application of the first voltage pulse having a voltage different from that of the second variable resistance nonvolatile memory element 20. It has the characteristic of changing continuously (or non-linearly).
- the conductance itself in a low resistance state (in other words, the multi-gradation analog resistance value) is adjusted by a current limiting circuit connected to the element. It is possible to do so.
- the first variable resistance nonvolatile memory element 10 has a characteristic that the conductance changes greatly with the first voltage pulse in the application of continuous voltage pulses, and the amount of change is small with subsequent voltage pulses. Therefore, the first variable resistance nonvolatile memory element 10 can be said to be a variable resistance nonvolatile memory element suitable for setting (initial setting) highly accurate coupling weighting coefficients at the time of product shipment.
- firmware update and learning can be performed before and after product shipment in an artificial intelligence processing device using a variable resistance nonvolatile memory element.
- the coupling weighting coefficient is insufficient due to model updates, regular maintenance, and changes in the conductance of the second variable resistance nonvolatile memory element 20, the conductance setting value of each element is acquired in advance and written to each element. becomes possible.
- FIG. 4 shows that, in the second variable resistance non-volatile memory element 20 according to the present embodiment, the magnitude of the voltage is different from that of the first variable resistance non-volatile memory element 10, and the same polarity and the same voltage are used.
- This is a resistance change characteristic when the conductance of the second variable resistance nonvolatile memory element 20 is changed to a high resistance state or a low resistance state by continuous application of a plurality of voltage pulses (that is, second voltage pulses).
- the horizontal axis shows the number of voltage pulse applications, and the vertical axis shows conductance.
- the rate of change in conductance at the first voltage pulse is smaller than the rate of change in conductance at the first voltage pulse in the first variable resistance nonvolatile memory element 10.
- the second variable resistance nonvolatile memory element 20 exhibiting such resistance change characteristics, a voltage pulse with a polarity toward lower resistance or a polarity toward higher resistance is generated regardless of the conductance before the pulse is applied.
- a voltage pulse By applying a voltage pulse, it is possible to cause a certain amount of increase in conductance (lower resistance) or decrease in conductance (higher resistance).
- the second variable resistance nonvolatile memory element 20 has a characteristic that the conductance gradually changes with successive voltage pulses. Therefore, the second variable resistance nonvolatile memory element 20 can be said to be a variable resistance nonvolatile memory element suitable for highly efficient updating (learning) of the coupling weighting coefficients after product shipment.
- the second variable resistance non-volatile memory element 20 in an artificial intelligence processing device using a variable resistance non-volatile memory element, each It becomes possible to write directly to the element to increase or decrease the conductance by a certain amount.
- the second voltage pulse changes with respect to the amount of conductance change in the first voltage pulse in the second variable resistance nonvolatile memory element 20.
- the ratio of the amount of change in conductance at the second voltage pulse to the amount of change in conductance at the first voltage pulse in the first variable resistance nonvolatile memory element 10 is It's also big.
- the second variable resistance nonvolatile memory element 20 has a continuous conductance due to the continuous application of the second voltage pulse having a voltage different from that of the first variable resistance nonvolatile memory element 10. It has the characteristic of changing linearly (or linearly).
- the first variable resistance nonvolatile memory element 10 almost reaches the set conductance with a single voltage pulse, although the difference is caused by the difference in driving conditions (that is, the magnitude of the applied voltage pulse). Comparing the second variable resistance nonvolatile memory element 20 whose conductance gradually changes by continuous application of multiple voltage pulses of the same polarity and the same voltage, the first variable resistance nonvolatile memory element 20 whose conductance can be set by a single voltage pulse.
- the variable resistance nonvolatile memory element 10 is considered to maintain a specific resistance state more firmly, and is therefore considered to have high retention characteristics after writing (that is, storage durability).
- the second variable resistance nonvolatile memory element 20 whose conductance gradually changes due to the continuous application of multiple voltage pulses of the same polarity and the same voltage changes the conductance gradually, so that the degree of element deterioration due to writing itself is reduced. It is considered to be small and have high endurance characteristics (that is, write characteristics).
- FIG. 5A is a diagram illustrating the definitions of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20.
- a voltage pulse of the same polarity and the same voltage is applied 20 times in a row in the process of lowering the resistance, and then a voltage pulse of the opposite polarity and the same voltage is applied 20 times in a row in the process of increasing the resistance.
- -Cycle (“n th _Cycle” in FIG. 5A) is repeated 50 times, and the relationship between the number of voltage pulses (horizontal axis) and conductance (vertical axis) is shown.
- 0 th _LR, 1 th _LR, and Max_LR are the conductance after the 0th voltage pulse application (that is, as an initial value) and the 1st voltage pulse application in the resistance reduction process in 1 Super-Cycle, respectively.
- the later conductance indicates the maximum conductance.
- 0 th _HR, 1 th _HR, and Min_HR are the conductance after the 0th voltage pulse application (that is, as an initial value) and the conductance after the 1st voltage pulse application in the high resistance process in 1 Super-Cycle, respectively. Conductance, indicating the minimum conductance.
- Ratio_LR the rate of change in conductance due to the first voltage pulse in the resistance reduction process.
- Ratio_LR (1 th _LR-0 th _LR)/(Max_LR-0 th _LR) (Formula 1)
- Ratio_HR rate of change in conductance due to the first voltage pulse in the process of increasing resistance
- Ratio_HR (1 th _HR-0 th _HR)/(Min_HR-0 th _HR) (Formula 2)
- Ratio_LR and Ratio_HR are collectively simply referred to as Ratio.
- variable resistance nonvolatile memory element driven under the driving condition of Ratio>0.7 is defined as a "first variable resistance nonvolatile memory element” or a digital RAND ( It is called a Resistive Analog Neuro Device.
- first variable resistance nonvolatile memory element or a digital RAND ( It is called a Resistive Analog Neuro Device.
- second variable resistance nonvolatile memory element or analog RAND.
- the "second variable resistance nonvolatile memory element" or analog RAND does not necessarily have to be driven under the same driving conditions (that is, Ratio_LR ⁇ 0.4 and Ratio_HR ⁇ 0.4), and may be driven under a drive condition that satisfies at least one of Ratio_LR ⁇ 0.4 and Ratio_HR ⁇ 0.4.
- FIG. 5B is a diagram illustrating the driving conditions of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20.
- the voltage of the voltage pulse to increase the resistance of the variable resistance non-volatile memory element as digital RAND is 1.0
- the voltage ratio of the actually applied voltage pulse (HR-pulse_Voltage_Ratio) is plotted on the horizontal axis.
- HR-pulse_Voltage_Ratio the voltage ratio of the actually applied voltage pulse
- It is a diagram in which the results obtained in an experiment are plotted with the vertical axis representing the Ratio obtained with a voltage pulse.
- the circle plots show the results obtained in the process of lowering the resistance
- the square plots show the results obtained in the process of increasing the resistance.
- a region satisfying "Ratio>0.7” is described as “Digital operation”, and the variable resistance nonvolatile memory element operates as the first variable resistance nonvolatile memory element 10 or digital RAND. It is an area. Further, the region satisfying "Ratio ⁇ 0.4" is described as “Analog-like operation”, and is a region where the variable resistance nonvolatile memory element operates as the second variable resistance nonvolatile memory element 20 or analog RAND. be.
- the driving conditions of Equation 3 below are met, that is, the voltage pulse with a voltage smaller than 0.71 times the voltage of the voltage pulse that increases the resistance of the variable resistance nonvolatile memory element as digital RAND is used to change the resistance.
- the variable resistance nonvolatile memory element can be operated as the second variable resistance nonvolatile memory element 20 or analog RAND.
- FIG. 6A and 6B are a circuit diagram and a cross-sectional view of a resistance variable nonvolatile memory element in the prior art.
- FIG. 6A shows a circuit diagram of a memory cell of an artificial intelligence processing device equipped with a conventional variable resistance nonvolatile memory element.
- the memory cell MC is composed of a resistance variable non-volatile memory element RP and a cell transistor T0 connected in series, and is a "1T1R" type composed of one cell transistor T0 and one resistance variable non-volatile memory element RP. memory cell.
- the word line WL of the memory cell MC is connected to the gate terminal of the cell transistor T0
- the bit line BL is connected to the variable resistance nonvolatile memory element RP
- the source line SL is connected to the source terminal of the cell transistor T0.
- FIG. 6B shows a cross-sectional view of a memory cell of an artificial intelligence processing device equipped with a conventional variable resistance nonvolatile memory element RP.
- Diffusion regions 61a and 61b are formed on the substrate 60, and the diffusion region 61a serves as the source terminal of the cell transistor T0, and the diffusion region 61b serves as the drain terminal of the cell transistor.
- the area between the diffusion regions 61a and 61b acts as a channel region of the cell transistor T0, and an oxide film 62 and a gate electrode 63 made of polysilicon are formed on this channel region, and operates as the cell transistor T0.
- Diffusion region 61a which is the source terminal of cell transistor T0, is connected to source line SL, which is first wiring layer 65a, via via 64a.
- Diffusion region 61b which is the drain terminal of cell transistor T0, is connected to first wiring layer 65b via via 64b.
- the first wiring layer 65b is connected to a second wiring layer 67 via a via 66, and the second wiring layer 67 is connected to a variable resistance nonvolatile memory element RP via a via 68a.
- the variable resistance nonvolatile memory element RP is composed of a first electrode 2, a variable resistance layer 3, and a second electrode 4.
- the variable resistance nonvolatile memory element RP is connected to the bit line BL, which is the third wiring layer 69, via the via 68b.
- variable resistance nonvolatile memory element RP only the first variable resistance nonvolatile memory element 10 according to the embodiment, which approximately reaches the set conductance with a single voltage pulse, is used in the artificial intelligence processing device. It will be installed. Therefore, it is difficult to achieve both highly efficient and accurate connection weighting coefficient setting (initial setting) at the time of product shipment, and highly efficient connection weighting coefficient update (learning) after product shipment.
- FIGS. 7A to 7C are diagrams showing examples of a circuit diagram and a cross-sectional view of a variable resistance nonvolatile memory element according to an embodiment.
- FIG. 7A shows a circuit diagram of a memory cell of an artificial intelligence processing device in which both the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are mounted.
- the memory cell MC is a first variable resistance nonvolatile memory element 10 whose conductance can be set by a single voltage pulse, and whose conductance gradually changes by continuously applying multiple voltage pulses of the same polarity and voltage that satisfy the above formula 3.
- a "1T2R" type memory is composed of a second variable resistance nonvolatile memory element 20 connected to a cell transistor T0, and is a "1T2R” type memory composed of one cell transistor T0 and two variable resistance nonvolatile memory elements. It is a cell.
- the word line WL of the memory cell MC is connected to the gate terminal of the cell transistor T0, the bit line BL1 is connected to the first variable resistance nonvolatile memory element 10, and the bit line BL2 is connected to the second variable resistance nonvolatile memory element 10.
- the source line SL is connected to the source terminal of the cell transistor T0.
- FIG. 7B is an example of a cross-sectional view of a memory cell of an artificial intelligence processing device in which both the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 shown in FIG. 7A are mounted. This is what is shown.
- a first variable resistance nonvolatile memory element 10 and a second variable resistance nonvolatile memory element 20 are respectively arranged in a first layer sandwiched between two wiring layers and a first layer sandwiched between two wiring layers.
- An example structure is illustrated in which the structure is arranged in a second layer different from the first layer.
- Diffusion regions 71a and 71b are formed on a semiconductor substrate 70, and the diffusion region 71a acts as the source terminal of the cell transistor T0, and the diffusion region 71b acts as the drain terminal of the cell transistor.
- the area between the diffusion regions 71a and 71b acts as a channel region of the cell transistor T0, and an oxide film 72 and a gate electrode 73 made of polysilicon are formed on this channel region, and operates as the cell transistor T0.
- Diffusion region 71a which is the source terminal of cell transistor T0, is connected to source line SL, which is first wiring layer 75a, via via 74a.
- Diffusion region 71b which is the drain terminal of cell transistor T0, is connected to first wiring layer 75b via via 74b.
- first wiring layer 75b is connected to a second wiring layer 77 via a via 76a
- second wiring layer 77 is connected to the first variable resistance nonvolatile memory element 10 via a via 78a.
- the first variable resistance nonvolatile memory element 10 includes a first electrode 2, a variable resistance layer 3, and a second electrode 4.
- the first variable resistance nonvolatile memory element 10 is connected to the bit line BL1, which is the third wiring layer 79, via the via 78b.
- the second wiring layer 77 is connected to the second variable resistance nonvolatile memory element 20 via the via 76c.
- the second variable resistance nonvolatile memory element 20 includes a first electrode 12 , a variable resistance layer 13 , and a second electrode 14 .
- the second variable resistance nonvolatile memory element 20 is connected to the bit line BL2, which is the first wiring layer 75c, via the via 76b.
- the center of the first variable resistance nonvolatile memory element 10 and the center of the second variable resistance nonvolatile memory element 20 coincide with each other when viewed in a direction perpendicular to the substrate plane.
- FIG. 7C is a cross-sectional view of a memory cell of an artificial intelligence processing device in which both the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 shown in FIG. 7A are mounted.
- a structural example is illustrated in which a first variable resistance nonvolatile memory element 10 and a second variable resistance nonvolatile memory element 20 are arranged in the same layer sandwiched between two wiring layers. .
- Diffusion regions 81a and 81b are formed on a semiconductor substrate 80, and the diffusion region 81a acts as the source terminal of the cell transistor T0, and the diffusion region 81b acts as the drain terminal of the cell transistor.
- the area between the diffusion regions 81a and 81b acts as a channel region of the cell transistor T0, and an oxide film 82 and a gate electrode 83 made of polysilicon are formed on this channel region to operate as the cell transistor T0.
- Diffusion region 81a which is the source terminal of cell transistor T0, is connected to source line SL, which is first wiring layer 85a, via via 84a.
- Diffusion region 81b which is the drain terminal of cell transistor T0, is connected to first wiring layer 85b via via 84b.
- the first wiring layer 85b is connected to the second wiring layer 87b via a via 86c, and the second wiring layer 87b is connected to the first variable resistance nonvolatile memory element 10 via a via 88a.
- the first variable resistance nonvolatile memory element 10 includes a first electrode 2, a variable resistance layer 3, and a second electrode 4.
- the first variable resistance nonvolatile memory element 10 is connected to the bit line BL1, which is the third wiring layer 89b, via the via 88b.
- the first wiring layer 85b is connected to the second wiring layer 87a via the via 86a, and the second wiring layer 87a is connected to the second variable resistance nonvolatile memory element 20 via the via 88c.
- the second variable resistance nonvolatile memory element 20 includes a first electrode 12 , a variable resistance layer 13 , and a second electrode 14 .
- the second variable resistance nonvolatile memory element 20 is connected to the bit line BL2, which is the third wiring layer 89a, via the via 88d.
- the center of the first variable resistance nonvolatile memory element 10 and the center of the second variable resistance nonvolatile memory element 20 do not coincide when viewed in a direction perpendicular to the substrate plane.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are arranged between different wiring layers between the first wiring layer and the third wiring layer. , the same effect can be obtained even if it is placed between other wiring layers, for example between different wiring layers, such as between the second wiring layer and the fourth wiring layer.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are arranged in a layer sandwiched between the second wiring layer and the third wiring layer.
- the same effect can be obtained even if it is placed between other wiring layers, for example, in a layer sandwiched between a first wiring layer and a second wiring layer.
- the current flowing through the bit line BL or the source line SL is defined as the current flowing through the memory cell MC.
- the current flowing through the memory cell MC is the sum of the currents flowing through the bit line BL1 and the bit line BL2, or the current flowing through the source line SL, that is, the sum of the current flowing through the first variable resistance nonvolatile memory element 10 and the first variable resistance nonvolatile memory element 10. It is defined as the sum of the currents flowing through the two variable resistance nonvolatile memory elements 20.
- the signal on the word line WL corresponds to an input signal input to one neuron
- the signal on the word line WL corresponds to the input signal input to one neuron
- the total conductance of the non-volatile memory elements 20 corresponds to one coupling weighting coefficient corresponding to the input signal
- the first variable resistance non-volatile memory element 10 and the second variable resistance non-volatile memory element 20 are connected to each other.
- the sum of flowing currents (that is, the sum of currents flowing through bit line BL1 and bit line BL2, or the current flowing through source line SL) corresponds to the product of the input signal and the coupling weighting coefficient.
- FIGS. 8A to 8C are diagrams showing examples of a circuit diagram and a cross-sectional view of a variable resistance nonvolatile memory element according to another embodiment.
- FIG. 8A shows a circuit diagram of a memory cell of an artificial intelligence processing device in which both the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are mounted.
- the memory cell MC1 is composed of a first variable resistance nonvolatile memory element 10 whose conductance can be set by a single voltage pulse and a cell transistor T1, and the memory cell MC2 has the same polarity that satisfies the above formula 3.
- a cell transistor T2 is connected to a second variable resistance nonvolatile memory element 20 whose conductance gradually changes due to the continuous application of a plurality of voltage pulses of the same voltage, and two cell transistors T1 and T2 are connected to each other.
- the word line WL1 of the memory cell MC1 is connected to the gate terminal of the cell transistor T1
- the word line WL2 of the memory cell MC2 is connected to the gate terminal of the cell transistor T2
- the bit line BL is connected to the first variable resistance nonvolatile memory element.
- 10 and a second variable resistance nonvolatile memory element 20 is connected to the source terminals of cell transistors T1 and T2.
- FIG. 8B shows an example of a cross-sectional view of the memory cell MC1 of the artificial intelligence processing device in which the first variable resistance nonvolatile memory element 10 shown in FIG. 8A is mounted.
- Diffusion regions 91a and 91b are formed on a semiconductor substrate 90, and the diffusion region 91a serves as the source terminal of the cell transistor T1, and the diffusion region 91b serves as the drain terminal of the cell transistor T1.
- the area between the diffusion regions 91a and 91b acts as a channel region of the cell transistor T1, and an oxide film 92 and a gate electrode 93 made of polysilicon are formed on this channel region to operate as the cell transistor T1.
- Diffusion region 91a which is the source terminal of cell transistor T1 is connected to source line SL, which is first wiring layer 95a, via via 94a.
- Diffusion region 91b which is the drain terminal of cell transistor T1 is connected to first wiring layer 95b via via 94b.
- first wiring layer 95b is connected to a second wiring layer 97 via a via 96
- second wiring layer 97 is connected to the first variable resistance nonvolatile memory element 10 via a via 98a.
- the first variable resistance nonvolatile memory element 10 includes a first electrode 2, a variable resistance layer 3, and a second electrode 4.
- the first variable resistance nonvolatile memory element 10 is connected to the bit line BL, which is the third wiring layer 99, via the via 98b.
- FIG. 8C shows an example of a cross-sectional view of the memory cell MC2 of the artificial intelligence processing device in which the second variable resistance nonvolatile memory element 20 described in FIG. 8A is mounted.
- Diffusion regions 101a and 101b are formed on the same substrate 90 as in FIG. 8B, and the diffusion region 101a acts as the source terminal of the cell transistor T2, and the diffusion region 101b acts as the drain terminal of the cell transistor T2.
- the area between the diffusion regions 101a and 101b acts as a channel region of the cell transistor T2, and an oxide film 102 and a gate electrode 103 made of polysilicon are formed on this channel region, and operates as the cell transistor T2.
- Diffusion region 101a which is the source terminal of cell transistor T2 is connected via via 104a to source line SL, which is the first wiring layer 95a common to FIG. 8B.
- Diffusion region 101b which is the drain terminal of cell transistor T2, is connected to first wiring layer 105 via via 104b.
- first wiring layer 105 is connected to a second wiring layer 107 via a via 106
- second wiring layer 107 is connected to a second variable resistance nonvolatile memory element 20 via a via 108a.
- the second variable resistance nonvolatile memory element 20 includes a first electrode 12 , a variable resistance layer 13 , and a second electrode 14 .
- the second variable resistance nonvolatile memory element 20 is connected to the bit line BL, which is the third wiring layer 99 common to FIG. 8B, via the via 108b.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are placed in a layer sandwiched between the second wiring layer and the third wiring layer.
- the same effect can be obtained even if it is placed between other wiring layers, for example, between different wiring layers such as a first wiring layer and a second wiring layer.
- variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are arranged between different wiring layers.
- the current flowing through the bit line BL or the source line SL that is, the current flowing through the variable resistance nonvolatile memory element RP
- the current flowing through the memory cell MC is defined as the current flowing through the memory cell MC.
- the current flowing through the bit line BL or the source line SL is the sum of the current flowing through the memory cell MC1 and the current flowing through the memory cell MC2, that is, the sum of the current flowing through the first variable resistance nonvolatile memory element 10 and the second current flowing through the memory cell MC2. is defined as the sum of the currents flowing through the variable resistance nonvolatile memory element 20.
- a signal flowing in common to word lines WL1 and WL2 corresponds to an input signal input to one neuron.
- the sum of the conductances of the first resistance variable nonvolatile memory element 10 and the second resistance variable nonvolatile memory element 20 corresponds to one coupling weighting coefficient corresponding to the input signal, and the first resistance variable nonvolatile memory element 20 corresponds to one coupling weighting coefficient corresponding to the input signal.
- the sum of the currents flowing through the variable resistance storage element 10 and the second variable resistance nonvolatile storage element 20 is the product of the input signal and the coupling weighting coefficient. Equivalent to.
- the first variable resistance nonvolatile memory element 10 approximately reaches the set conductance with a single voltage pulse, and the same polarity that satisfies the above formula 3,
- both elements can be used.
- the first variable resistance non-volatile memory element 10 is subject to firmware updates, learning model updates, periodic maintenance, and changes in the conductance of the second variable resistance non-volatile memory element 20 before and after product shipment.
- the conductance is changed when the coupling weighting coefficient is insufficiently updated, and the second variable resistance nonvolatile memory element 20 is used in such a way that the conductance is changed when the coupling weighting coefficient is updated for learning after product shipment.
- This makes it possible to achieve both highly efficient and highly accurate connection weighting coefficient setting (initial setting) at the time of product shipment, and highly efficient connection weighting coefficient updating (learning) after product shipment.
- one of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 is selected depending on the application at that time.
- a first neural network region that uses the connection weight coefficient settings of the existing neural network as is by applying the transfer learning or reinforcement learning using the connection weight coefficient settings of the existing neural network; If there is a second neural network area to perform new learning, the conductance of the first variable resistance nonvolatile memory element 10 is changed to set the coupling weight coefficient of the first neural network area. Transfer learning or reinforcement learning can be performed efficiently by changing the conductance of the second variable resistance non-volatile memory element 20 to set the connection weighting coefficient of the second neural network region. can.
- FIG. 9A is a block diagram showing a model of the artificial intelligence processing device 200 according to the embodiment.
- the artificial intelligence processing device 200 is a neural network composed of an input layer 201, a plurality of hidden layers 202, and an output layer 203.
- Each layer (input layer 201, hidden layer 202, output layer 203) is composed of a plurality of neurons 210.
- Output data from the neurons 210 forming the previous layer is input to each neuron 210 via a synapse 211 .
- FIG. 9B is a diagram illustrating the function of neuron 210 shown in FIG. 9A.
- the neuron 210 receives the output data from the neuron 210 forming the previous layer as input data x i via the synapse 211, and multiplies the received input data x i by the connection weighting coefficient w i corresponding to the synapse 211.
- a sum-of-products operation ( ⁇ wi ⁇ x i ) is performed to add the product ( w i ⁇ x i ) for all input data x i .
- the neuron 210 adds an internal bias b to the result of the product-sum operation, and applies the obtained result ( ⁇ w i ⁇ x i +b) to an internal activation function f such as a step function.
- an internal activation function f such as a step function.
- FIG. 10A is a diagram showing an example of a circuit that implements neuron 210 shown in FIG. 9B.
- the neuron 210 includes a product-sum calculation circuit 215, a word line selection circuit 230, a determination circuit 250, and a column gate (transistors YTi1, YTi2, and transistor DTi).
- the product-sum calculation circuit 215 is a circuit that performs a product-sum calculation in the neuron 210, and includes a plurality of 1T2R type memory cells shown in FIG. are connected in such a way that they are common to each other.
- the current output from each memory cell corresponds to the product (w i ⁇ x i ) of the input data x i and the coupling weighting coefficient w i , and when these currents are combined, the current flowing to the source line SLi ( Alternatively, the sum of the currents flowing through the bit lines BLi1 and BLi2) corresponds to the product-sum calculation result ⁇ w i ⁇ x i .
- the input data x i is "1" or "0”
- the coupling weighting coefficient w i is between the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile This corresponds to the total conductance of the magnetic memory element 20.
- a plurality of product-sum calculation circuits 215 corresponding to all the neurons 210 shown in FIG. 9A are arranged in rows, and among them, one product-sum calculation circuit 215 is , performs a sum-of-products operation in one neuron 210.
- the word line selection circuit 230 selects or de-selects memory cells on a row-by-row basis via word lines WL0 to WLn to the gate terminals of transistors Ti included in memory cells MCi0 to MCin constituting the product-sum calculation circuit 215. This is a circuit that supplies input data x0 to xn for .
- the determination circuit 250 is a circuit that executes the activation function f possessed by the neuron 210, and is a circuit that executes the activation function f that the neuron 210 has.
- a value ( ⁇ wi ⁇ x i +b) obtained by adding the internal bias b to the total current ( ⁇ w i ⁇ x i ) flowing through the circuit is compared with a predetermined threshold value and output.
- the determination circuit 250 can perform processing in parallel on a plurality of product-sum calculation circuits 215 arranged in the row direction.
- Transistors YTi1 and YTi2 constituting the column gate connect bit lines BLi1 and BLi2 of memory cells MCi0 to MCin to a predetermined level, respectively, in response to a signal input to the gate terminal during writing and reading to and from memory cells MCi0 to MCin. Connect or disconnect from the power supply voltage. Further, the transistor DTi connects or disconnects the source line SLi to a predetermined power supply voltage according to a signal input to the gate terminal during writing and reading to and from the memory cells MCi0 to MCin.
- FIG. 10B is a diagram showing another circuit example (neuron 210a) that implements neuron 210 shown in FIG. 9B.
- the neuron 210a includes a product-sum calculation circuit 215a, a word line selection circuit 230a, a determination circuit 250a, and a column gate (transistor YTi, transistor DTi).
- This neuron 210a has the same basic function as the neuron 210 shown in FIG. 10A, but unlike the neuron 210 in FIG. 10A which is composed of 1T2R type memory cells MCij, it corresponds to 2T2R type memory cells MCij.
- the connection configuration is as follows.
- Word line selection circuit 230a outputs two word lines WLj1 and WLj2 to each of memory cells MCi0 to MCin.
- the determination circuit 250a adds an internal bias b to the current ( ⁇ wi ⁇ x i ) flowing through the bit line BLi or source line SLi indicating the result of the product-sum calculation output from the product-sum calculation circuit 215a.
- the value ( ⁇ wi ⁇ x i +b) is compared with a predetermined threshold value and output.
- one transistor YTi is provided for each bit line BLi to switch connection and disconnection to a predetermined power supply voltage.
- FIG. 11 is a block diagram showing the overall configuration of the artificial intelligence processing device 200 consisting of neurons 210 shown in FIG. 10A.
- the artificial intelligence processing device 200 includes a memory cell array 220, a word line selection circuit 230, a column gate 240, a determination circuit 250, a write circuit 260, and a control circuit 270.
- Word line selection circuit 230, column gate 240, and determination circuit 250 are the same as those described in FIG. 10A.
- the write circuit 260 is a circuit that supplies the predetermined power supply voltage explained in FIG. 10A, and is a current limiting circuit for writing a desired conductance (in other words, a multi-gradation analog resistance value) into the memory cell MCij. has.
- the write circuit 260 applies a first voltage pulse having a first voltage to the first variable resistance nonvolatile memory element 10, thereby increasing the first resistance.
- the second resistance is increased.
- the conductance of the variable nonvolatile memory element 20 is rewritten.
- the second voltage in the high resistance process is 0.71 times or less of the first voltage in the high resistance process, for example, 0.7 times the voltage, 0.5 times the voltage, 0.3 times the voltage, Alternatively, it is one voltage specified from a plurality of voltages that are 0.71 times or less.
- the control circuit 270 is a circuit that controls writing and reading to and from the memory cell MCij by controlling the entire artificial intelligence processing device 200, and includes, for example, a memory storing a program, a processor, and the like. More specifically, when changing (that is, writing) the connection weighting coefficients of the artificial intelligence processing device 200, the control circuit 270 determines whether the high-precision connection weighting coefficients are set (initial setting) at the time of product shipment or the like.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element constituting each memory cell MCij may be used for highly efficient updating (learning) of coupling weighting coefficients later on.
- the artificial intelligence processing device 200 is controlled to change the conductance of only one of the two.
- control circuit 270 controls the first variable resistance nonvolatile memory element constituting each memory cell MCij.
- the artificial intelligence processing device 200 is controlled to use the total value of the currents flowing through the variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20.
- FIG. 11 is a block diagram showing the entire configuration of the artificial intelligence processing device 200 made up of the neurons 210 shown in FIG. 10A, the entire configuration of the artificial intelligence processing device 200 made of the neurons 210a shown in FIG.
- the block diagram illustrating the block diagram is the same as that in FIG. 11 except for the above-mentioned connection wiring, so illustration and explanation thereof will be omitted.
- FIG. 12A is a diagram showing an example of voltages applied via the word line WL, bit lines BL1, BL2, and source line SL during writing and reading to the memory cell MCij shown in FIG. 10A.
- setting increasing resistance
- Examples of applied voltages are shown when reading data from the memory cell MCij and when reading from the memory cell MCij.
- FIG. 12A shows an example of voltages applied via the word line WL, bit lines BL1, BL2, and source line SL during writing and reading to the memory cell MCij shown in FIG. 10A.
- mode 1 indicates writing of conductance to the first variable resistance nonvolatile memory element 10 (first rewriting step)
- mode 2 indicates writing of conductance to the first variable resistance nonvolatile memory element 10
- mode 2 indicates writing of conductance to the first variable resistance nonvolatile memory element 10
- 3 shows writing of conductance to the memory element 20 (second rewriting step).
- writing to memory cell MCij means setting or changing a coupling weighting coefficient in memory cell MCij
- “reading to memory cell MCij” means measuring the current flowing through memory cell MCij. means.
- a pulse voltage V g_on (for example, 2 V) is supplied to the gate terminal, a reset voltage V H1 (for example, 2 V) is applied to the bit line BL1, a reference voltage V SS (for example, 0 V) is applied to the bit line BL2, and a reference voltage V H1 (for example, 0 V) is applied to the source line SL.
- a voltage V ss (eg, 0V) is applied.
- a pulse voltage V g_on (for example, 2 V) is supplied to the gate terminal, a reference voltage V SS (for example, 0 V) is applied to the bit line BL1, and a reset voltage V H2 (for example, 1.4 V (that is, 0.4 V of V H1) is applied to the bit line BL2. 7 times)) is applied, and a reference voltage V ss (for example, 0 V) is applied to the source line SL.
- writing is performed under different driving conditions. (In other words, writing is performed under the same driving conditions).
- a pulse voltage V g_on (for example, 2 V) is supplied to the gate terminal, a reference voltage V ss (for example, 0 V) is applied to the bit line BL1, a set voltage V set (for example, 2 V) is applied to the bit line BL2, and a set voltage V set (for example, 2 V) is applied to the source line SL.
- a voltage V set (eg, 2V) is applied.
- a pulse voltage V g_on (for example, 2V) is supplied to the gate terminal, a set voltage V set (for example, 2V) is applied to the bit line BL1, a reference voltage V ss (for example, 0V) is applied to the bit line BL2, and a set voltage is applied to the source line SL.
- a voltage V set (eg, 2V) is applied.
- the read voltage V read is applied to the first resistance variable nonvolatile memory element 10 and the second resistance variable nonvolatile memory element 20, and the read voltage V read is applied from the memory cell MCij to the first resistance variable nonvolatile memory element 10 and the second resistance variable nonvolatile memory element 20.
- the sum of the currents flowing through the variable resistance storage element 10 and the second variable resistance nonvolatile storage element 20 (that is, one product (w i x i ) is output.
- the current output from the memory cell MCij flows to the source line SL (also the sum of the currents flowing to the bit lines BL1 and BL2), and is measured by the determination circuit 250 as the result of the sum-of-products operation ( ⁇ wi ⁇ x i ). Ru.
- FIG. 12B is a diagram showing an example of voltages applied via the word lines WL1, WL2, bit line BL, and source line SL during writing and reading to the memory cell MCij shown in FIG. 10B.
- setting increasing resistance
- Examples of applied voltages are shown when reading data from the memory cell MCij and when reading from the memory cell MCij.
- the pulse voltage V g_on For example, 2V
- a pulse voltage V g_off for example, 0V
- a reset voltage V H1 for example, 2V
- a reference voltage Vss for example, 0V
- a pulse voltage V g_off (for example, 0V) is supplied to the gate terminal
- a pulse voltage V g_on (for example, 2V) that turns on the transistor Ti2 via the word line WL2 is supplied to the gate terminal
- a reset voltage V H2 (for example, 1.4V (that is, 0.7 times VH1 )) is applied
- a reference voltage Vss (for example, 0V) is applied to the source line SL.
- writing is performed under different driving conditions. (In other words, writing is performed under the same driving conditions).
- a pulse voltage V g_on (for example, 2V) is supplied to the gate terminal
- a pulse voltage V g_off (for example, 0V) that turns off the transistor Ti2 is supplied to the gate terminal via the word line WL2
- a reference voltage V ss (for example, 0V) is supplied to the bit line BL.
- a set voltage V set (for example, 2V) is applied to the source line SL.
- a pulse voltage V g_off (for example, 0V) is supplied to the gate terminal
- a pulse voltage V g_on (for example, 2V) that turns on the transistor Ti2 via the word line WL2 is supplied to the gate terminal
- a reference voltage V ss (for example, 0V) is supplied to the bit line BL.
- a set voltage V set (for example, 2V) is applied to the source line SL.
- the read voltage V read is applied to the first resistance variable nonvolatile memory element 10 and the second resistance variable nonvolatile memory element 20, and the read voltage V read is applied from the memory cell MCij to the first resistance variable nonvolatile memory element 10 and the second resistance variable nonvolatile memory element 20.
- Currents flowing through the variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are output.
- the current output from all the memory cells MCij constituting the product-sum calculation circuit 215 flows to the source line SL and the bit line BL, and is measured by the determination circuit 250 as the result of the product-sum calculation ( ⁇ w i ⁇ x i ). be done.
- FIG. 13 is a flowchart showing an example of the operation of the control circuit 270 shown in FIG. 11.
- the control circuit 270 determines whether the connection weighting coefficient setting process to be performed now is the first case in which the connection weighting coefficients are changed for initial setting, or the second case in which the connection weighting coefficients are changed in learning. (S30).
- the first case includes firmware updates before and after shipment of the artificial intelligence processing device 200, learning model updates, periodic maintenance, and changes in the conductance of the second variable resistance nonvolatile memory element 20. At least one of the cases in which the weighting coefficients are insufficiently updated is included, and the second case includes a case in which the connection weighting coefficients are updated for learning after the artificial intelligence processing device 200 is shipped.
- the control circuit 270 controls the column gate 240 or the word line selection circuit. 230 and the write circuit 260, the first variable resistance nonvolatile memory element 10 is selected for each of the memory cells MCij (S31), and the selected first variable resistance nonvolatile memory element 10 is On the other hand, by writing according to the drive conditions corresponding to mode 1 shown in FIG. 12A or 12B, the connection weighting coefficient derived in advance is set (S32). In other words, the control circuit 270 writes the conductance derived in advance to the first variable resistance nonvolatile memory element 10 (that is, in "mode 1") for each of the memory cells MCij (first rewriting). step).
- step S30 determines whether the connection weighting coefficient setting process is updating the connection weighting coefficient by learning (that is, the second case) ("updating the connection weighting coefficient by learning" in S30).
- the control circuit 270 performs inference using the connection weighting coefficients held in the current memory cell array 220 (S35), and after confirming the difference between the inference result and the teacher label (S36), each of the memory cells MCij , the amount of change in the connection weighting coefficient at the time of updating is calculated (S37).
- the control circuit 270 selects the second variable resistance nonvolatile memory element 20 for each of the memory cells MCij by controlling the column gate 240 or the word line selection circuit 230 (S38), and By writing into the variable resistance non-volatile memory element 20 of No. 2 under the drive conditions corresponding to mode 2 shown in FIG. 12A or 12B, the current connection weight is changed by the amount of change in the calculated connection weight coefficient.
- the connection weighting coefficients are updated so that the coefficients change (S39).
- the control circuit 270 updates the conductance of the second variable resistance nonvolatile memory element 20 (that is, in "mode 2") by the amount of change for each of the memory cells MCij (the second rewrite step).
- connection weighting coefficient setting initial setting
- learning connection weighting coefficient updating
- each step is executed by the control circuit 270, but some or all of these steps may be executed by a control circuit such as another processor placed outside the artificial intelligence processing device 200. It's okay.
- the artificial intelligence processing device 200 includes one substrate 60, etc., and a first variable resistance non-volatile non-volatile device having the same structure that is mounted on one substrate 60, etc., and each retains conductance.
- An arithmetic circuit such as a product-sum arithmetic circuit 215 including the variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 and a first voltage having a first voltage applied to the first variable resistance nonvolatile memory element 10.
- the write circuit 260 rewrites the conductance of the second variable resistance nonvolatile memory element 20 by applying two voltage pulses.
- the second voltage during the resistance increasing process is 0.71 times or less of the first voltage during the resistance increasing process.
- the first variable resistance nonvolatile memory element 10 which has the characteristic that the conductance changes greatly by the first voltage pulse in the application of continuous voltage pulses, and the amount of change is small in the subsequent voltage pulses.
- the second variable resistance nonvolatile memory element 20 which has a characteristic that the conductance gradually changes with successive voltage pulses, can be used for updating (learning) the coupling weighting coefficients.
- a variable resistance non-volatile memory element that is capable of both highly accurate setting of coupling weighting coefficients (initial setting) at the time of product shipment, and highly efficient updating (learning) of coupling weighting coefficients after product shipment. An artificial intelligence processing device using this technology is realized.
- the first variable resistance nonvolatile memory element 10 is different from the second variable resistance nonvolatile memory element 10.
- the second variable resistance nonvolatile memory element 20 has better storage durability than the first variable resistance nonvolatile memory element 10. It will be held.
- the arithmetic circuit performs a product-sum operation using a composite value of the conductance of the first variable resistance nonvolatile memory element 10 and the conductance of the second variable resistance nonvolatile memory element 20 as one weighting coefficient.
- This is a product-sum calculation circuit 215. More specifically, the arithmetic circuit performs a product-sum operation on the sum of the current flowing through the first variable resistance nonvolatile memory element 10 and the current flowing through the second variable resistance nonvolatile memory element 20 as one product. I do.
- one coupling weighting coefficient is configured by the conductance of the first variable resistance nonvolatile memory element 10 and the conductance of the second variable resistance nonvolatile memory element 20, and the conductance of the first variable resistance nonvolatile memory element 20 constitutes one coupling weighting coefficient. It becomes possible to make a memory cell including the element 10 and the second variable resistance nonvolatile memory element 20 correspond to one neuron.
- first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 may be arranged in the same layer sandwiched between two wiring layers.
- first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are structurally arranged in parallel, so the manufacturing process is faster than when they are arranged in different layers. is simplified.
- first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are respectively arranged in a first layer sandwiched between two wiring layers and a first layer sandwiched between two wiring layers. Alternatively, it may be arranged in a second layer different from the first layer. As a result, the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 can be arranged to overlap when viewed from above with respect to the substrate, so that they are in the same layer.
- the chip size of the artificial intelligence processing device 200 is reduced compared to the case where the artificial intelligence processing device 200 is arranged.
- each of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 includes first electrodes 2 and 12, second electrodes 4 and 14, and first electrode 2 and 12 and the second electrodes 4 and 14 may include variable resistance layers 3 and 13 sandwiched between the electrodes 12 and the second electrodes 4 and 14.
- the first electrodes 2 and 12 of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 may be noble metal electrodes.
- the first electrodes 2 and 12 of the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 may contain any one of Ir and Pt. .
- the weighting coefficient writing method of the artificial intelligence processing device 200 is a weighting coefficient writing method possessed by the artificial intelligence processing device 200, in which the artificial intelligence processing device 200 holds the weighting coefficient in the product-sum calculation as a conductance.
- a first variable resistance nonvolatile memory element 10 and a second variable resistance nonvolatile memory element 20 having the same structure are provided, and a writing method is performed on the first variable resistance nonvolatile memory element 10.
- a first rewriting step (S31 to S32) of rewriting the conductance of the first variable resistance nonvolatile memory element 10 by applying a first voltage pulse having a first voltage
- a rewriting step (S38 to S39).
- the second voltage during the resistance increasing process is 0.71 times or less of the first voltage during the resistance increasing process.
- the first variable resistance nonvolatile memory element 10 which has the characteristic that the conductance changes greatly by the first voltage pulse in the application of continuous voltage pulses, and the amount of change is small in the subsequent voltage pulses.
- the second variable resistance nonvolatile memory element 20 which has a characteristic that the conductance gradually changes with successive voltage pulses, can be used for updating (learning) the coupling weighting coefficients.
- the first rewriting step is executed when the weighting coefficients are initially set for the artificial intelligence processing device 200, and the second rewriting step is made to be learned by the artificial intelligence processing device 200. This can be done when updating the weighting coefficients.
- the first variable resistance nonvolatile memory element 10 constitutes a first artificial intelligence area for using existing weighting coefficients of artificial intelligence processing as they are in transfer learning or reinforcement learning
- the non-volatile memory element 20 constitutes a second artificial intelligence area for performing new learning and updating weighting coefficients in transfer learning or reinforcement learning, thereby changing the first rewriting step to the first rewriting step. It can be executed when writing a weighting coefficient for one artificial intelligence area, and the second rewriting step can be executed when updating a weighting coefficient for a second artificial intelligence area.
- the present disclosure is not limited to this embodiment. Unless departing from the gist of the present disclosure, various modifications to the present embodiment that those skilled in the art can think of, and other forms constructed by combining some of the components of the embodiments are also within the scope of the present disclosure. included.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 have a variable resistance layer made of tantalum oxide, but the variable resistance layer is not limited to such materials, and may be composed of a transition metal such as hafnium or an oxide of aluminum.
- the resistance change layer of the first resistance change nonvolatile memory element 10 and the second resistance change nonvolatile memory element 20 is composed of the first tantalum oxide layer and the second tantalum oxide layer.
- it has a laminated structure, it is not limited to such a laminated structure, and may be composed of a single layer such as a tantalum oxide layer.
- one neuron is composed of one first variable resistance nonvolatile memory element 10 and one second variable resistance nonvolatile memory element 20, but at least one If the first variable resistance nonvolatile memory element 10 and at least one second variable resistance nonvolatile memory element 20 are provided, two or more first variable resistance nonvolatile memory elements 10 or Two or more second variable resistance nonvolatile memory elements 20 may be provided.
- the artificial intelligence processing device 200 is a neural network having the structure shown in FIG. It may also be a neural network composed of neurons.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are used as weighting coefficients in the product-sum calculation circuit 215, but their uses are as follows. It is not limited to the product-sum operation circuit.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 hold variable values corresponding to conductance, and perform various operations such as four arithmetic operations using the variable values. It may also be used in an arithmetic circuit.
- the first variable resistance nonvolatile memory element 10 and the second variable resistance nonvolatile memory element 20 are reset (“increased resistance”). Writing under different driving conditions was performed only when Alternatively, writing may be performed under different driving conditions.
- the artificial intelligence processing device using the variable resistance non-volatile memory element of the present disclosure is capable of setting highly efficient and highly accurate coupling weight coefficients (initial setting) at the time of product shipment, etc., and setting highly efficient coupling weights after product shipment, etc. This makes it possible to update (learn) coefficients at the same time, and is particularly useful as an edge AI processing device for IoT.
- Second electrode 10 1 resistance variable nonvolatile memory element 20 2nd resistance variable nonvolatile memory element 61a, 61b, 71a, 71b, 81a, 81b, 91a, 91b, 101a, 101b, diffusion region 62, 72, 82, 92, 102 Oxide film 63, 73, 83, 93, 103 Gate electrode (word line) 64a, 64b, 66, 68a, 68b, 74a, 74b, 76a, 76b, 76c, 78a, 78b, 84a, 84b, 86a, 86c, 88a, 88b, 88c, 88d, 94a, 94b, 96, 98a, 98b, 104a, 104b, 106, 108a
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Neurology (AREA)
- Neurosurgery (AREA)
- Physiology (AREA)
- Power Engineering (AREA)
- Semiconductor Memories (AREA)
Abstract
Description
[抵抗変化型不揮発性記憶素子の構成]
まず、実施形態に係る2種類の異なる駆動条件でコンダクタンスを変化させる対象となる抵抗変化型不揮発性記憶素子の一方を「第1の抵抗変化型不揮発性記憶素子」と呼び、他方を「第2の抵抗変化型不揮発性記憶素子」と呼び、それら第1の抵抗変化型不揮発性記憶素子及び第2の抵抗変化型不揮発性記憶素子の構成の一例について説明する。第1の抵抗変化型不揮発性記憶素子及び第2の抵抗変化型不揮発性記憶素子は、上述したように、異なる駆動条件で駆動されるが、同一構造を有する。
次に、第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20の製造方法の一例について、これらの製造方法は同じであるので、第1の抵抗変化型不揮発性記憶素子10の製造方法の場合を用いて説明する。
次に、上述した製造方法により得られた第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20の動作について説明する。
次に、第1の抵抗変化型不揮発性記憶素子10及び第2の抵抗変化型不揮発性記憶素子20それぞれに対する駆動条件について、図5A及び図5Bを用いて、詳細に説明する。
図6Aおよび図6Bは、従来技術における抵抗変化型不揮発性記憶素子の回路図および断面図である。
次に、本開示に係る人工知能処理装置の実施例について、説明する。
本実施の形態では、第1の抵抗変化型不揮発性記憶素子10及び第2の抵抗変化型不揮発性記憶素子20をリセット(「高抵抗化」)する場合に、異なる駆動条件による書き込みを行っている。
第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20を流れる電流の合計を計測する場合には(「読み出し」)、ワード線WLを介してトランジスタTiをオンにする読み出し電圧Vg_read(例えば、1V)がゲート端子に供給され、ビット線BL1およびBL2に読み出し電圧Vread(例えば、0.4V)が印加され、ソース線SLに基準電圧Vss(例えば、0V)が印加される。これにより、第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20に対して、読み出し電圧Vreadが印加され、メモリセルMCijから、第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20を流れる電流の合計(つまり、一つの積(wi・xi)が出力される。よって、積和演算回路215を構成する全てのメモリセルMCijから出力される電流がソース線SLに流れ(ビット線BL1およびBL2に流れる電流の合計でもある)、積和演算の結果(Σwi・xi)として、判定回路250で計測される。
本実施の形態では、第1の抵抗変化型不揮発性記憶素子10及び第2の抵抗変化型不揮発性記憶素子20をリセット(「高抵抗化」)する場合に、異なる駆動条件による書き込みを行っている。
第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20を流れる電流の合計を計測する場合には(「読み出し」)、ワード線WL1およびWL2を介してトランジスタTi1およびTi2をオンにする読み出し電圧Vg_read(例えば、1V)がゲート端子に供給され、ビット線BLに読み出し電圧Vread(例えば、0.4V)が印加され、ソース線SLに基準電圧Vss(例えば、0V)が印加される。これにより、第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20に対して、読み出し電圧Vreadが印加され、メモリセルMCijから、第1の抵抗変化型不揮発性記憶素子10および第2の抵抗変化型不揮発性記憶素子20を流れる電流(その合計が一つの積(wi・xi)に相当する)が出力される。よって、積和演算回路215を構成する全てのメモリセルMCijから出力される電流がソース線SLおよびビット線BLに流れ、積和演算の結果(Σwi・xi)として、判定回路250で計測される。
2、12 第1電極
3、13 抵抗変化層
3a、13a 第1タンタル酸化物層
3b、13b 第2タンタル酸化物層
4、14 第2電極
10 第1の抵抗変化型不揮発性記憶素子
20 第2の抵抗変化型不揮発性記憶素子
61a、61b、71a、71b、81a、81b、91a、91b、101a、101b、 拡散領域
62、72、82、92、102 酸化膜
63、73、83、93、103 ゲート電極(ワード線)
64a、64b、66、68a、68b、74a、74b、76a、76b、76c、78a、78b、84a、84b、86a、86c、88a、88b、88c、88d、94a、94b、96、98a、98b、104a、104b、106、108a、108b ビア
65a、65b、75a、75b、75c、85a、85b、95a、95b、105 第1配線層
67、77、87a、87b、97、107 第2配線層
69、79、89a、89b、99 第3配線層
200 人工知能処理装置
201 入力層
202 隠れ層
203 出力層
210、210a ニューロン
211 シナプス
215、215a 積和演算回路
220 メモリセルアレイ
230、230a ワード線選択回路
240 カラムゲート
250、250a 判定回路
260 書き込み回路
270 制御回路
MC、MC1、MC2、MCij メモリセル
T0、T1、T2 セルトランジスタ
WL、WL1、WL2 ワード線
BL、BL1、BL2、BLi1、BLi2 ビット線
SL、SLi ソース線
Claims (15)
- 1つの基板と、
前記1つの基板に搭載され、それぞれがコンダクタンスを保持する、同一構造を有する第1の抵抗変化型不揮発性記憶素子と第2の抵抗変化型不揮発性記憶素子とを含む演算回路と、
前記第1の抵抗変化型不揮発性記憶素子に対して第1電圧を有する第1電圧パルスを印加することで前記第1の抵抗変化型不揮発性記憶素子のコンダクタンスを書き換え、前記第2の抵抗変化型不揮発性記憶素子に対して前記第1電圧とは異なる第2電圧を有する第2電圧パルスを印加することで前記第2の抵抗変化型不揮発性記憶素子のコンダクタンスを書き換える書き込み回路とを備える、
人工知能処理装置。 - 前記演算回路は、前記第1の抵抗変化型不揮発性記憶素子のコンダクタンスと前記第2の抵抗変化型不揮発性記憶素子のコンダクタンスとの合成値を一つの重み係数として用いる積和演算を行う積和演算回路である、
請求項1記載の人工知能処理装置。 - 前記演算回路は、前記第1の抵抗変化型不揮発性記憶素子を流れる電流と前記第2の抵抗変化型不揮発性記憶素子を流れる電流との合計を、一つの積として、前記積和演算を行う、
請求項2記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子のコンダクタンスは、前記第1電圧パルスが繰り返し印加された場合に、非連続的に変化し、
前記第2の抵抗変化型不揮発性記憶素子のコンダクタンスは、前記第2電圧パルスが繰り返し印加された場合に、連続的に変化する、
請求項1~3のいずれか1項に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子は、前記第2の抵抗変化型不揮発性記憶素子よりも優れた保管耐性を有し、
前記第2の抵抗変化型不揮発性記憶素子は、前記第1の抵抗変化型不揮発性記憶素子よりも優れた書き換え特性を有する、
請求項1~4のいずれか1項に記載の人工知能処理装置。 - 高抵抗化過程における前記第2電圧は、高抵抗化過程における前記第1電圧の0.71倍以下である、
請求項1~5のいずれか1項に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子及び前記第2の抵抗変化型不揮発性記憶素子は、2つの配線層で挟まれた同一の層に配置されている、
請求項1~6のいずれか1項に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子及び前記第2の抵抗変化型不揮発性記憶素子は、それぞれ、2つの配線層で挟まれた第1層、及び、2つの配線層で挟まれた、第1層とは異なる第2層に配置されている、
請求項1~6のいずれか1項に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子及び前記第2の抵抗変化型不揮発性記憶素子のそれぞれは、第1電極、第2電極、及び前記第1電極と前記第2電極との間に挟まれた抵抗変化層を含む、
請求項1~8のいずれか1項に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子及び前記第2の抵抗変化型不揮発性記憶素子の前記第1電極は、貴金属電極である、
請求項9に記載の人工知能処理装置。 - 前記第1の抵抗変化型不揮発性記憶素子及び前記第2の抵抗変化型不揮発性記憶素子の前記第1電極は、Ir、Ptのうちいずれか1つを含む、
請求項10に記載の人工知能処理装置。 - 人工知能処理装置が有する重み係数書き込み方法であって、
前記人工知能処理装置は、積和演算における重み係数をコンダクタンスとして保持する、同一構造を有する第1の抵抗変化型不揮発性記憶素子と第2の抵抗変化型不揮発性記憶素子とを備え、
前記書き込み方法は、
前記第1の抵抗変化型不揮発性記憶素子に対して第1電圧を有する第1電圧パルスを印加することで前記第1の抵抗変化型不揮発性記憶素子のコンダクタンスを書き換える第1の書き換えステップと、
前記第2の抵抗変化型不揮発性記憶素子に対して、少なくとも高抵抗化過程において前記第1電圧とは異なる第2電圧を有する第2電圧パルスを印加することで前記第2の抵抗変化型不揮発性記憶素子のコンダクタンスを書き換える第2の書き換えステップとを含む、
重み係数書き込み方法。 - 前記第1の書き換えステップは、前記人工知能処理装置に対して前記重み係数を初期設定する場合に、実行され、
前記第2の書き換えステップは、前記人工知能処理装置に対して学習させることで前記重み係数を更新する場合に、実行される、
請求項12記載の重み係数書き込み方法。 - 前記第1の抵抗変化型不揮発性記憶素子は、転移学習又は強化学習において、既にある人工知能処理の重み係数をそのまま用いるための第1人工知能領域を構成し、
前記第2の抵抗変化型不揮発性記憶素子は、転移学習又は強化学習において、新たに学習を行って重み係数を更新していくための第2人工知能領域を構成し、
前記第1の書き換えステップは、前記第1人工知能領域に対して前記重み係数を書き込む場合に、実行され、
前記第2の書き換えステップは、前記第2人工知能領域に対して前記重み係数を更新していく場合に、実行される、
請求項12記載の重み係数書き込み方法。 - 高抵抗化過程における前記第2電圧は、高抵抗化過程における前記第1電圧の0.71倍以下である、
請求項12~14のいずれか1項に記載の重み係数書き込み方法。
Priority Applications (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2024544553A JPWO2024048704A1 (ja) | 2022-09-02 | 2023-08-31 | |
| CN202380062520.2A CN119790401A (zh) | 2022-09-02 | 2023-08-31 | 人工智能处理装置及人工智能处理装置的权重系数写入方法 |
| US19/050,821 US20250182837A1 (en) | 2022-09-02 | 2025-02-11 | Artificial intelligence processing device and weight coefficient writing method for artificial intelligence processing device |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2022140283 | 2022-09-02 | ||
| JP2022-140283 | 2022-09-02 |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| US19/050,821 Continuation US20250182837A1 (en) | 2022-09-02 | 2025-02-11 | Artificial intelligence processing device and weight coefficient writing method for artificial intelligence processing device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2024048704A1 true WO2024048704A1 (ja) | 2024-03-07 |
Family
ID=90099798
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2023/031711 Ceased WO2024048704A1 (ja) | 2022-09-02 | 2023-08-31 | 人工知能処理装置および人工知能処理装置の重み係数書き込み方法 |
Country Status (4)
| Country | Link |
|---|---|
| US (1) | US20250182837A1 (ja) |
| JP (1) | JPWO2024048704A1 (ja) |
| CN (1) | CN119790401A (ja) |
| WO (1) | WO2024048704A1 (ja) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011135843A1 (ja) * | 2010-04-28 | 2011-11-03 | パナソニック株式会社 | 抵抗変化型不揮発性記憶装置及びその製造方法 |
| JP2020068048A (ja) * | 2018-10-18 | 2020-04-30 | 株式会社デンソー | 人工ニューラルネットワーク回路及び人工ニューラルネットワーク回路における学習値切替方法 |
| WO2023112674A1 (ja) * | 2021-12-13 | 2023-06-22 | ヌヴォトンテクノロジージャパン株式会社 | 人工知能処理装置および人工知能処理装置の学習推論方法 |
-
2023
- 2023-08-31 CN CN202380062520.2A patent/CN119790401A/zh active Pending
- 2023-08-31 JP JP2024544553A patent/JPWO2024048704A1/ja active Pending
- 2023-08-31 WO PCT/JP2023/031711 patent/WO2024048704A1/ja not_active Ceased
-
2025
- 2025-02-11 US US19/050,821 patent/US20250182837A1/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011135843A1 (ja) * | 2010-04-28 | 2011-11-03 | パナソニック株式会社 | 抵抗変化型不揮発性記憶装置及びその製造方法 |
| JP2020068048A (ja) * | 2018-10-18 | 2020-04-30 | 株式会社デンソー | 人工ニューラルネットワーク回路及び人工ニューラルネットワーク回路における学習値切替方法 |
| WO2023112674A1 (ja) * | 2021-12-13 | 2023-06-22 | ヌヴォトンテクノロジージャパン株式会社 | 人工知能処理装置および人工知能処理装置の学習推論方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN119790401A (zh) | 2025-04-08 |
| JPWO2024048704A1 (ja) | 2024-03-07 |
| US20250182837A1 (en) | 2025-06-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| TWI674534B (zh) | 使用非揮發性半導體記憶元件之類神經網路運算電路 | |
| US20220261624A1 (en) | Neural network circuits having non-volatile synapse arrays | |
| US11615299B2 (en) | Neural network computation circuit including non-volatile semiconductor memory element | |
| CN110892480B (zh) | 采用nvm矩阵电路以执行矩阵计算的非易失性(nv)存储器(nvm)矩阵电路 | |
| KR102567160B1 (ko) | 비휘발성의 시냅스 배열을 가지는 신경망 회로 | |
| TWI754567B (zh) | 神經形態裝置及其操作方法 | |
| CN111433792A (zh) | 可编程可重置人工神经网络的基于计数器的电阻处理单元 | |
| TWI699711B (zh) | 記憶體裝置及其製造方法 | |
| US20240304267A1 (en) | Artificial intelligence processing device and training inference method for artificial intelligence processing device | |
| WO2024021365A1 (zh) | 存储器单元、阵列电路结构及数据处理方法 | |
| WO2024048704A1 (ja) | 人工知能処理装置および人工知能処理装置の重み係数書き込み方法 | |
| KR20200024419A (ko) | 3차원 크로스바 메모리 구조를 이용한 뉴로 모픽 소자 | |
| CN113592081A (zh) | 数据处理装置及数据处理方法 | |
| KR102656337B1 (ko) | 삼항 정밀도 xor 논리 연산을 기반으로 하는 반도체 소자 및 이를 포함하는 뉴로모픽 컴퓨팅 시스템 | |
| WO2023171683A1 (ja) | ニューラルネットワーク演算回路 | |
| Fernando | Low Power, Dense Circuit Architectures and System Designs for Neural Networks using Emerging Memristors | |
| TW202514438A (zh) | 用於人工神經網路學習之三端突觸裝置、使用其之突觸陣列及操作其之方法 | |
| CN117808062A (zh) | 计算装置、电子装置以及用于计算装置的操作方法 | |
| JPWO2024048704A5 (ja) |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23860466 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2024544553 Country of ref document: JP |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202380062520.2 Country of ref document: CN |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| WWP | Wipo information: published in national office |
Ref document number: 202380062520.2 Country of ref document: CN |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 23860466 Country of ref document: EP Kind code of ref document: A1 |