WO2021002995A1 - Neural network memory - Google Patents
Neural network memory Download PDFInfo
- Publication number
- WO2021002995A1 WO2021002995A1 PCT/US2020/035816 US2020035816W WO2021002995A1 WO 2021002995 A1 WO2021002995 A1 WO 2021002995A1 US 2020035816 W US2020035816 W US 2020035816W WO 2021002995 A1 WO2021002995 A1 WO 2021002995A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- threshold voltage
- variable resistance
- memory cell
- sub
- resistance memory
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2178—Validation; Performance evaluation; Active pattern learning techniques based on feedback of a supervisor
- G06F18/2185—Validation; Performance evaluation; Active pattern learning techniques based on feedback of a supervisor the supervisor being an automated module, e.g. intelligent oracle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/778—Active pattern-learning, e.g. online learning of image or video features
- G06V10/7784—Active pattern-learning, e.g. online learning of image or video features based on feedback from supervisors
- G06V10/7792—Active pattern-learning, e.g. online learning of image or video features based on feedback from supervisors the supervisor being an automated module, e.g. "intelligent oracle"
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C11/00—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor
- G11C11/54—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor using elements simulating biological cells, e.g. neuron
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C11/00—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor
- G11C11/56—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor using storage elements with more than two stable states represented by steps, e.g. of voltage, current, phase, frequency
- G11C11/5621—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor using storage elements with more than two stable states represented by steps, e.g. of voltage, current, phase, frequency using charge storage in a floating gate
- G11C11/5628—Programming or writing circuits; Data input circuits
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C11/00—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor
- G11C11/56—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor using storage elements with more than two stable states represented by steps, e.g. of voltage, current, phase, frequency
- G11C11/5678—Digital stores characterised by the use of particular electric or magnetic storage elements; Storage elements therefor using storage elements with more than two stable states represented by steps, e.g. of voltage, current, phase, frequency using amorphous/crystalline phase transition storage elements
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C13/00—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G11C13/0002—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements
- G11C13/0004—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements comprising amorphous/crystalline phase transition cells
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C13/00—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G11C13/0002—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements
- G11C13/0021—Auxiliary circuits
- G11C13/003—Cell access
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C13/00—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G11C13/0002—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements
- G11C13/0021—Auxiliary circuits
- G11C13/0038—Power supply circuits
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C13/00—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G11C13/0002—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements
- G11C13/0021—Auxiliary circuits
- G11C13/004—Reading or sensing circuits or methods
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C13/00—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G11C13/0002—Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00 using resistive RAM [RRAM] elements
- G11C13/0021—Auxiliary circuits
- G11C13/0069—Writing or programming circuits or methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C2213/00—Indexing scheme relating to G11C13/00 for features not covered by this group
- G11C2213/70—Resistive array aspects
- G11C2213/77—Array wherein the memory element being directly connected to the bit lines and word lines without any access device being used
Definitions
- the present disclosure relates generally to operating apparatus, such as memories, and, more particularly, to neural network memory.
- Memory devices may typically be provided as internal, semiconductor, integrated circuits in computers or other electronic devices.
- memory There are various types of memory including volatile and non-volatile memory.
- Various memory arrays can be organized in a cross-point architecture with memory cells (e.g., two terminal cells) being located at intersections of a first and second signal lines used to access the cells (e.g., at intersections of access lines and sense lines).
- Some memory cells can be, for example, resistance variable memory cells whose logic state (e.g., stored data value) depends on the programmed resistance of the memory cell.
- Some variable resistance memory cells which may be referred to as self-selecting memory cells, comprise a single material which can serve as both a select element and a storage element for the memory cell.
- Figure 1 illustrates an example of an apparatus in the form of a memory array of a memory device in accordance with various embodiments of the present disclosure.
- Figure 2 illustrates an example of a memory array that supports neural network memory accordance with embodiments of the present disclosure.
- Figure 3 illustrates an example of a memory array that supports neural network memory in accordance with embodiments of the present disclosure.
- Figure 4 illustrates an example graph of threshold voltage distributions of memory cells in accordance with an embodiment of the present disclosure.
- Figure 5 illustrates an example graph of threshold voltage distributions of memory cells in accordance with an embodiment of the present disclosure.
- Figure 6 illustrates an example method for neural network memory using a memory device in accordance with embodiments of the present disclosure.
- Figure 7 illustrates an example method for neural network memory using a memory device in accordance with embodiments of the present disclosure.
- a synaptic weight may refer to the strength or amplitude of a connection between two nodes (e.g., a neuron).
- the nature and content of information transmitted through a neural network may be based in part on the properties of synapses formed between the nodes (e.g., synaptic weight).
- Memory arrays can be operated as neural network memory (e.g., neuromorphic systems and devices) and can be designed to achieve results that may not be possible with traditional computer architectures. For example, neuromorphic systems may be used to achieve results more commonly associated with biological systems such as learning, vision or visual processing, auditory processing, advanced computing, or other processes, or a combination thereof.
- an apparatus may include an array of variable resistance memory cells and a neural memory unit controller coupled to the array of variable resistance memory cells.
- the neural memory unit controller can be configured to apply a sub-threshold voltage pulse to a variable resistance memory cell of the array to change a threshold voltage of the variable resistance memory cell in an analog fashion from a voltage associated with a reset state to effectuate a first synaptic weight change; and apply additional sub-threshold voltage pulses to the variable resistance memory cell to effectuate each subsequent synaptic weight change.
- “a”,“an”, or“a number of’ can refer to one or more of something, and“a plurality of’ can refer to two or more such things.
- a memory device can refer to one or more memory devices, and a plurality of memory devices can refer to two or more memory devices.
- designators“M”,“P”,“R”,“B”,“S”, and“N”, as used herein, particularly with respect to reference numerals in the drawings, indicates that a number of the particular feature so designated can be included with a number of embodiments of the present disclosure. The number may be the same or different between designations.
- Figure 1 illustrates an example of an apparatus in the form of a memory array of a memory device in accordance with various embodiments of the present disclosure.
- an“apparatus” can refer to, but is not limited to, any of a variety of structures or combinations of structures, such as a circuit or circuitry, a die or dice, a module or modules, a device or devices, or a system or systems, for example.
- Figure 1 is an illustrative representation of various components and features of the memory device 100. As such, it should be appreciated that the components and features of the memory device 100 are shown to illustrate functional interrelationships, not their actual physical positions within the memory device 100.
- the memory device 100 includes a memory array 101.
- the memory array 101 includes memory cells 105 that may be programmable to different logic states. In some
- each memory cell 105 may be programmable to two states, denoted as a logic 0 and a logic 1. In some embodiments, a memory cell 105 may be programmable to more than two logic states.
- a memory cell 105 may, in some embodiments, include a variable resistance memory cell such as a self-selecting memory' cell.
- a self- selecting memory cell is a memory cell comprising a single elialeogenide material to operate as a selection component and a storage component.
- variable resistance memory cells can be self-selecting memory cells and can be arranged into an array (e.g., a neural memory unit) to mimic neuro biological functionality such as learning.
- the self-selecting memory cells include a chalcogenide material that can change a logic state (e.g., set or reset) of the chalcogenide material that is in an amorphous state responsive to a magnitude of voltage.
- a set state can be conductive (e.g., low-resistance to current flow) and a reset status can be less conductive (e.g., higher resistance to current flow).
- the change in state of the chal cogenide material between a set state and a reset state can change the threshold voltage value of the self- selecting memory cell in an analog fashion.
- the analog threshold voltage change of the chalcogenide material of the self- selecting memory cell can represent a synaptic weight in a neuromorphic memory system. Changes in synaptic weight can represent and/or be interpreted to represent learning and other biological functions.
- each row of memory cells 105 is connected to a first signal line 110- 1, 110-2, or 110-N (e.g., an access line, sometimes referred to as a word line), and each column of memory cells 105 is connected to a second signal line 115- 1 , 1 15-2, or 1 15- S (e.g., a sense line, sometimes referred to as a bit line).
- Both access lines 110 and sense lines 115 may also be genericalJy referred to as signal lines.
- Access lines 110 and sense lines 115 may be substantially perpendicular to one another and may support an array of memory cells.
- memory array 101 may include a s ubstrate (further described herein in connection with Figure 2).
- one memory cell 105 may be located at the intersection of two signal lines such as an access line 1 10 and a sense line 1 15.
- memory cell 105 is located at the intersection of access line 1 10- 1 and sense line 115-S. This intersection may be referred to as the address of the memory cell 105.
- a target memory ceil 105 may be a memory' cell 105 located at the intersection of an energized (e.g., activated) access line 1 10 and an energized (e.g. , activated) sense line 1 15; that is, an access line 1 10 and a sense line 1 15 may both be energized in order to read or write a memory cell 105 at their intersection.
- Other memory ceils 105 that are in electronic communication with (e.g., connected to) the same access line 1 10 or sense line 1 15 may be referred to as untargeted memory cells 105.
- an electrode may couple a memory cell 105 to an access line 1 10 or a sense line 1 15.
- the term “electrode” may refer to an electrical conductor, and in some cases, may be employed as an electrical contact to a memory ceil or other component of a memory array.
- An electrode may include a trace, wire, conductive line, conductive layer, or the like that provides a conductive path between elements or components of the memory array 101 of the memory' device 100.
- the term electrode may refer in some eases to an access line, such as an access line 1 10 or a sense line such as sense line 1 15, as well as in some cases to an additional conductive element employed as an electrical contact between an access line and a memory ceil 105.
- a memory cell 105 may comprise a chalcogenide material positioned between a first electrode and a second electrode.
- the first electrode may couple the chaicogenide material to an access line 1 10
- the second electrode couple the chalcogenide material to a sense line 1 15.
- the first electrode and the second electrode may be the same material (e.g., carbon) or different material.
- a memory cell 105 may be coupled directly with one or more access lines, and electrodes other than the access lines may be omitted.
- Chaicogenide materials may be materials or alloys that include at least one of the elements S, Se, and Te.
- Chaicogenide materials may include alloys of S, Se, Te, Ge, As, Al, Sb, Au, indium (In), gallium (Ga), tin (Sn), bismuth (Bi), palladium (Pd), cobalt (Co), oxygen (O), silver (Ag), nickel (Ni), platinum (Pt)
- Example chaicogenide materials and alloys may include, but are not limited to, Ge-Te, In- Se, Sb-Te, Ga- Sb, In- Sb, As-Te, Al ⁇ Te, Ge- Sb-Te, Te-Ge-As, In- Sb-Te, Te- Sn- Se, Ge- Se-Ga, Bi- Se- Sb, Ga- Se- Te, Sn- Sb-Te, In- Sb- Ge, Te-Ge- Sb- S, Te-Ge- Sn-O, Te- Ge- Sn-Au, Pd- Te-Ge- Sn, In- Se- Ti-Co
- the hyphenated chemical composition notation indicates the elements included in a particular compound or alloy and is intended to represent ail stoichiometries involving the indicated elements.
- Ge- Te may include Ge x Te y , where x and y may be any positive integer.
- Other examples of variable resistance materials may include binary metal oxide materials or mixed valence oxide including two or more metals (e.g., transition metals, alkaline earth metals), and/or rare earth metals.
- Embodiments are not limited to a particular variable resistance material or materials associated with the memory elements of the memory cells.
- variable resistance materials can be used to form memory elements and may include chaicogenide materials, colossal magnetoresistive materials, or polymer- based materials, among others.
- Operations such as reading, and writing may be performed on memory cells 105 by activating or selecting access line 110 and sense line 115.
- Activating or selecting an access line 110 or a sense line 115 may include applying a voltage to the respective line.
- Access lines 110 and sense lines 115 may be made of conductive materials such as metals (e.g., copper (Cu), aluminum (Al), gold (Au), tungsten (W), titanium (Ti)), metal alloys, carbon, conductively- doped semiconductors, or other conductive materials, alloys, compounds, or the like.
- the logic storing device of a cell may be electrically isolated from the sense line by a selection component.
- the access line 110 may be connected to and may control the selection component.
- the selection component may be a transistor and the access line 110 may be connected to the gate of the transistor.
- the selection component may be a variable resistance component of a variable resistance memory cell, which may comprise chalcogenide material.
- the variable resistance memory cell can be a self-selecting memory cell, comprising a single material (e.g., the chalcogenide material) which can serve as both a select element and a storage element for the memory cell.
- Activating the access line 110 may result in an electrical connection or closed circuit between the logic storing device of the memory cell 105 and its corresponding sense line 1 15. The sense line may then be accessed to either read or write the memory cell 105.
- the resulting signal may be used to determine the stored logic state.
- a first logic state may correspond to no current or a negligibly small current through the memory cell 105, whereas a second logic state may correspond to a finite current.
- a memory cell 105 may include a self-selecting memory cell having two terminals and a separate selection component may be omitted. As such, one terminal of the self- selecting memory cell may be electrically connected to an access line 110 and the other term inal of the self- selecting mem ory cell may be electrically connected to a sense line 1 1 5.
- Accessing memory cells 105 may be controlled through a row decoder 120 and a column decoder 130.
- a row decoder 120 may receive a row address from the memory controller 140 and activate the appropriate access line 1 10 based on the received row address.
- a column decoder 130 may receive a column address from the memory controller 140 and activate the appropriate sense line 115.
- memory array 101 may include multiple access lines 110- 1, 110-2, and 110-N and multiple sense lines 1 1 5- 1, 115-2, and 115- S, where S and N depend on the array size.
- an access line 1 10 and a sense line 115 e.g., 1 10- 1 and sense line 1 15- S
- the memory cell 105 at the intersection may be accessed.
- a memory cell 105 may be read, or sensed, by sense component 125 to determine the programmed state of the memory cell 105. For example, a voltage may be applied to a memory cell 105 (using the corresponding access line 110 and sense line 115) and the presence of a resulting current through the memory cell 105 may depend on the applied voltage and the threshold voltage of the memory ceil 105. In some cases, more than one voltage may be applied. Additionally, if an applied voltage does not result in current flow, other voltages may be applied until a current is detected by sense component 125. By assessing the voltage that resulted in current flow, the logic state of the memory cell 105 may be determined.
- the voltage may be ramped up in magnitude until a current flow is detected. In other cases, predetermined voltages may be applied sequentially until a current is detected. Likewise, a current may be applied to a memory cell 105 and the magnitude of the voltage to create the current may depend on the electrical resistance or the threshold voltage of the memory cell 105.
- the memory cell 105 may comprise a chalcogenide material.
- the chalcogenide material of the memory' cell may remain in an amorphous state during the access operation.
- operating the memory cell may include applying various shapes of programming pulses to the memory' cell to determine a particular threshold voltage of the memory' cell that is, a threshold voltage of a memory cell may be modified by changing a shape of a programming pulse.
- a particular threshold voltage of the memory cell 105 may be determined by applying various shapes of read pulses to the memory cell. For example, when an applied voltage of a read pulse exceeds the particular threshold voltage of the memory cell, a finite amount of current may flow through the memory cell. Similarly, when the applied voltage of a read pulse is less than the particular threshold voltage of the memory cell, no appreciable amount of current may flow through the memory cell.
- the memory cell is a variable resistance memory cell (e.g., a self- selecting memory cell)
- applying read pulses that are less than the threshold voltage of the memory cell can change the threshold voltage of the memory cell 105 in an analog fashion.
- an initial threshold voltage of a variable resistance memory cell may change incrementally (e.g., in an analog fashion), responsive to having been pulsed below the initial threshold voltage (pulsed with a sub-threshold voltage). This change in threshold voltage can be responsive to the chalcogenide material of the variable resistance memory cell being altered.
- sense component 125 may read information stored in a selected memory cell 105 by detecting the current flow or lack thereof through the memory cell 105.
- the m emory cell 105 e.g., variable resistance memory cell and/or a self-selecting memory cell
- the m emory cell 105 may store one bit of data based on threshold voltage levels (e.g., two threshold voltage levels) associated with the chalcogenide material, with the threshold voltage levels at which current flows through the memory cell 105 indicative of a logic state stored by the memory cell 105.
- the memory cell 105 may exhibit a certain number of different threshold voltage levels (e.g., three or more threshold voltage levels), thereby storing more than one bit of data.
- Sense component 125 may include various transistors or amplifiers in order to detect and amplify a difference in the signals associated with a sensed memory cell 105, which may be referred to as latching. The detected logic state of memory cell 105 may then be output through column decoder 130 as output 135. In some cases, sense component 125 may be part of a column decoder 130 or row decoder 120 Or, sense component 125 may be connected to or in electronic communication with column decoder 130 or row decoder 120. An ordinary person skilled in the art would appreciate that sense component 125 may be associated either with column decoder or row decoder without losing its functional purposes.
- a transistor or transistors discussed herein may represent a field- effect transistor (FET) and comprise a three terminal device including a source, drain, and gate.
- the terminals may be connected to other electronic elements through conductive materials, e.g., metals.
- the source and drain may be conductive and may comprise a heavily -doped, e.g., degenerate, semiconductor region.
- the source and drain may be separated by a lightly-doped semiconductor region or channel. If the channel is n-type (i.e., majority carriers are electrons), then the FET may be referred to as a n-type FET. If the channel is p-type (i.e., majority carriers are holes), then the FET may be referred to as a p-type FET.
- the channel may be capped by an insulating gate oxide.
- the channel conductivity may be controlled by applying a voltage to the gate. For example, applying a positive voltage or negative voltage to an n-type FET or a p-type FET, respectively, may result in the channel becoming conductive.
- a transistor may be "on” or “activated” when a voltage greater than or equal to the transistor's threshold voltage is applied to the transistor gate. The transistor may be “off or “deactivated” when a voltage less than the transistor's threshold voltage is applied to the transistor gate.
- a memory cell 105 may be set or written by similarly activating the relevant access line 110 and sense line 115, and at least one logic value may be stored in the memory cell 105.
- Column decoder 130 or row decoder 120 may accept data, for example input/output 135, to be written to the memory cells 105.
- accessing the memory cell 105 may degrade or destroy the logic state and re-write or refresh operations may be performed to return the original logic state to memory cell 105.
- the capacitor may be partially or completely
- activating an access line 1 10 may result in the discharge of all memory cells in the row (e.g., coupled with the access line 110); thus, several or all memory cells 105 in the row may need to be re written. But in non-volatile memory, such as variable resistance memory cells, self-selecting memory cells, and/or PCM memory, accessing the memory cell 105 may not destroy the logic state and, thus, the memory cell 105 may not require re-writing after accessing.
- non-volatile memory such as variable resistance memory cells, self-selecting memory cells, and/or PCM memory
- the memory controller 140 may control the operation (e.g., read, write, re-write, refresh, discharge) of memory cells 105 through the various components, for example, row decoder 120, column decoder 130, and sense component 125. In some cases, one or more of the row decoder 120, column decoder 130, and sense component 125 may be co-located with the memory controller 140.
- Memory controller 140 may generate row and column address signals in order to activate the desired access line 110 and sense line 115. Memory controller 140 may also generate and control various voltages or currents used during the operation of memory device 100.
- the amplitude, shape, polarity, and/or duration of an applied voltage or current discussed herein may be adjusted or varied and may be different for the various operations discussed in operating the memory device 100.
- one, multiple, or all memory cells 105 within memory array 101 may be accessed simultaneously; for example, multiple or all cells of memory array 101 may be accessed simultaneously during a reset operation in which all memory cells 105, or a group of memory cells 105, are set to a single logic state.
- Various memory cells 105 of the memory device 100 may be grouped into memory units that are configured to store an analog value (e.g., a neural memory unit).
- the memory controller 140 may be coupled to the neural memory unit and referred to as a neural memory unit controller.
- the neural memory units may be configured to mimic neuro-biological architectures.
- the neural memory units may leverage properties of chalcogenide material within a memory cell to change the properties of the chalcogenide material.
- the changed properties of the chalcogenide material may alter the threshold voltage of the memory cell, this operation may be referred to as the memory cell‘storing’ an analog value, as one example value.
- the analog value of a neural memory unit and/or multiple analog values of memory cells of a neural memory unit may be interpreted as a result of learning in a neural memory unit.
- An external host and/or a portion or the entire memory device 100 can generate and/or receive a learning algorithm.
- the learning algorithm is an algorithm that can be used in machine learning to help the neural memory unit imitate (e.g., mimic, emulate, etc.) neuro-biological architectures.
- a learning algorithm can include variables of a learning event.
- a learning event can be a quantity of pulses of voltage values, a magnitude of voltage values (e.g., above or sub-threshold voltages), and/or a length of time a pulse is applied to be applied to memory cells (e.g., variable resistance memory cells) of a neural memory unit.
- a neural memory unit controller e.g., the memory controller 140
- a neural memory unit controller (e.g., memory controller 140) can be configured to apply a learning algorithm including learning events.
- the neural memory unit controller can apply a quantity of sub-threshold voltage pulses to alter a single chalcogenide material of a self-selecting memory cell (e.g., memory cell 105) that is in an amorphous condition associated with a reset status to a condition associated with a set status.
- the chalcogenide material of the self-selecting memory cell can move between two electrical states responsive to a sub-threshold voltage pulse.
- the neural memory unit controller (e.g., memory controller 140) can apply an iteration of a learning algorithm by applying sub-threshold voltage pulses to the variable resistance memory cells.
- the threshold voltage of the cells may change in an analog fashion responsive to the sub-threshold voltage pulses applied to the variable resistance memory cell.
- Each of the sub-threshold voltage pulses can change the synaptic weight (e.g., increase or decrease) of the variable resistance memory cell.
- the increase and/or decrease in synaptic weight can indicate whether learning has occurred in a neural memory unit.
- the neural memory unit controller can monitor the neural memory unit to determine when additional sub threshold pulses can be applied based on the synaptic weights. For example, the neural memory unit controller can refrain from applying an additional sub-threshold voltage pulse in response to the learning algorithm indicating that additional learning has not occurred (e.g., the threshold voltage did not change or increase).
- the neural memory unit controller can change the sub-threshold voltage pulses responsive to a degree of learning. For example, the neural memory unit controller can apply an additional sub threshold voltage pulse as a relatively longer pulse to the variable resistance memory cell when an indication of relatively strong learning occurred (e.g., a decrease in the threshold voltage of the variable resistance memory cell of the array). Separately or concurrently, the neural memory unit controller can increase (or decrease) the quantity of sub-threshold voltage pulses responsive to the degree of learning occurring.
- the neural memory unit can be configured to monitor the degree of learning (e.g., the changing synaptic weights of the variable resistance memory cells of the array) and adjust variables (e.g., learning events) of a learning algorithm.
- FIG. 2 illustrates an example of a memory array 201 that supports neural network memory in accordance with embodiments of the present disclosure.
- Memory array 201 may be an example of portions of memory array 101 described with reference to F igure 1.
- Memory array 201 may include memory cells 205 that are positioned above a substrate 204.
- Memory array 201 may also include access lines 210- 1 and 210-2, and sense lines 215- 1 and 215- 2, which may be examples of access lines 110 and sense lines 1 15, as described with reference to Figure 1.
- memory cell 205 may be a self- selecting memory cell.
- some elements included in Figure 2 are labeled with a numeric indicator, other corresponding elements are not labeled, though they are the same or would be understood to be similar, in an effort to increase the visibility and clarity of the depicted features.
- the memory' cell 205 can be a self-selecting memory cell and can include a first electrode 211, a chalcogemde material 213, and a second electrode 217.
- signal lines e.g., access line 1 10 and sense line 1 15 of Figure 1
- the electrode layer of the signal lines may interface with a memory material (e.g., chalcogenide material 213).
- signal lines e.g., access line 1 10, sense line 115
- the architecture of memory array 201 may in some cases be referred to as an example of a cross-point architecture, as a memory cell 205 may be formed at a topological cross-point between an access line 210 and a sense line 215 as illustrated in Figure 2.
- a cross-point architecture may offer relatively high-density data storage with lower production costs compared to some other memory architectures.
- a memory array with a cross-point architecture may have memory cells with a reduced area and, resultantly, may support an increased memory cell density compared to some other architectures.
- a cross-point architecture may have a 4F 2 memory cell area, where F is the smallest feature size (e.g., a minimum feature size), compared to other architectures with a 6F ⁇ memory cell area, such as those with a three-terminal selection component.
- a DRAM memory array may use a transistor, which is a three-terminal device, as the selection component for each memory cell, and thus a DRAM memory array comprising a given number of memory cells may have a larger memory cell area compared to a memory array with a cross-point architecture comprising the same number of memory cells.
- Figure 2 shows 1 level of a memory array (e.g., a memory deck), other configurations may include any number of decks.
- one or more of the memory' decks may include self-selecting memory cells that include chaleogenide material 213.
- Memory cells e.g , memory cell 205, may be incorporated as part of memory units that may be configured to store values, which may in some cases be or include analog values.
- applying an electronic pulse to a chaleogenide material 213 may cause the chaleogenide material 213 to be affected, which may in some cases include changing physical form.
- the physical forms for some chaleogenide materials 213 include an amorphous state and a crystalline state. The resistance of these physical forms is different, thereby allowing the chaleogenide material 213 to maintain a physical state which may be referred to as storing a logic (e.g , sense a logic and/or an analog value).
- applying an electronic pulse to a chaleogenide material 213 may not change a phase of the chaleogenide material 213 and the chaleogenide material 213 can remain amorphous.
- set and reset states may be obtained by applying pulses of different (e.g., opposite) polarity, in some embodiments moreover, the threshold voltage can be modified or tuned by- applying voltage pulses of appropriate amplitude and/or polarity.
- Figure 3 illustrates an example of a memory array 301 that supports neural network memory in accordance with embodiments of the present disclosure.
- the memory array 301 may be similar to the example memory arrays 101 and 201 of Figures 1 and 2.
- the memory array 301 may include a neural memory unit 323 which can include a portion or all of the variable resistance memory cells 305 in the array 301.
- the memory array 301 may include multiple variable resistance memory cells 305- 1, 305-2, 305- 3, 305- 4, 305- 5, 305- 6, 305-7, 305- 8, 305-9, 305- 10, 305- 11, and 305-P (may be collectively referred to as the variable resistance memory cells 305).
- the memory array 301 may include a neural memory unit 323 which can include a portion or all of the variable resistance memory cells 305 in the array 301.
- the variable resistance memory cells 305 of the neural memory unit 323 may be coupled to a plurality of sense lines 315- 1, 315-2, 315- S (may be collectively referred to as sense lines 315) and a plurality of access lines 310- 1, 310-2, 310- 3, 310-N (may be collectively referred to as access lines 310).
- variable resistance memory cells of the memory array 301 may not be included in the neural memory unit 323.
- the neural memory unit 323 includes variable resistance memory cells 305- 1, 305-2, 305-3, 305-4, 305- 5, 305- 6, 305-7, 305- 8, and 305-9 and the variable resistance memory cells 305- 10, 305- 11, and 305-P may be excluded from the neural memory unit 323.
- variable resistance memory cells 305 coupled to access line 310-N may be included in the memory array 301 but not necessarily be a part of the neural memory unit 323. In such examples, the variable resistance memory cells may be excluded from the total analog value stored by the neural memory unit 323.
- a neural memory unit controller may select a neural memory unit 323 for a read operation.
- the neural memory unit controller may select one or more variable resistance memory cells 305 of the neural memory unit 323 for a read operation.
- the neural memory unit controller may identify and/or select one or more sense lines 315 and/or access lines 310 associated with the neural memory unit 323.
- the neural memory unit controller may provide an input 341 to the variable resistance memory cells 305 of the neural memory unit 323.
- the input 341 may comprise a plurality of voltage values (e.g., sub-threshold voltage values) applied to a plurality of access lines 310 (e.g., VI, V2, V3).
- the neural memory unit controller may bias the access lines 310 to one or more voltage values (e.g., a read voltage value) included in the input 341. In some cases, the access lines 310 are all biased to the same read voltage. In some cases, the neural memory unit controller may bias one or more access lines to a different voltage than the other access lines.
- the neural memory unit controller may also bias an unselected access line 310-N (e.g., an access lines not included in the neural memory unit 323) to a read voltage value.
- the read voltage values applied to the one or more unselected access lines 310-N are the same as the voltage values applied to the selected access lines 310- 1, 310-2, 310- 3.
- the read voltage values applied to the unselected access lines 310- N are different than a voltage value applied to one of the selected access lines 310- 1, 310-2, 310- 3.
- the neural memory unit controller may detect an output 343 that includes one or more signals generated on one or more sense lines 315 coupled with the neural memory unit 323.
- the output 343 on the sense lines 315 may be generated based on applying the input 341 to the access lines 310 coupled with the neural memory unit 323.
- the signals of the output 343 may comprise current signals (e.g., II, 12, 13).
- An individual signal or an individual threshold voltage may be detected on each sense line 315 coupled with the neural memory unit 323.
- Each signal or weight may have different variable resistance memory cells contributing to the signal.
- variable resistance memory cells 305- 1, 305-4, and 305-7 may contribute to the signal on the first sense line 315- 1.
- the variable resistance memory cells 305-2, 305- 5, and 305- 8 may contribute to the signal on the second sense line 315-2.
- the variable resistance memory cells 305- 3, 305-6, and 305-9 may contribute to the signal on the third sense line 315- S.
- the neural memory unit controller may determine the analog value stored in the neural memory unit 323 based on the detecting the signals generated on the sense lines 315 coupled with the neural memory unit 323.
- the neural memory unit controller may combine signals and/or the threshold voltage (e.g., synaptic weights) on each sense line 315 to generate a total weight (may be referred to as a total analog value).
- the analog value may be proportional to and/or based on the total weight.
- the neural memory unit controller may sum the signals of the sense lines 315 to generate the total weight.
- the neural memory unit controller may generate a product by applying a weight matrix to an input vector of voltages that represent the one or more voltages on one or more access lines (e.g. VI, V2, V3).
- the neural memory unit controller may apply a vector by matrix multiplication operation, where the vector is the input 341 (VI, V2, and V3), and the matrix is the threshold voltage of the respective variable resistance memory cells 305 included in the neural memory unit 323.
- the neural memory unit controller can be configured to apply a respective sub-threshold voltage pulse via each of the access lines 310 to change a threshold voltage of a respective variable resistance memory cell 305 coupled to each of the access lines 310 and to a particular sense line 315- 1 in an analog fashion toward a voltage associated with a set state to effectuate synaptic weight changes.
- Each of the variable resistance memory cells 305- 1, 305-4, and 305-7 can receive the sub-threshold voltage in this instance.
- the neural memory unit controller can be further configured to, apply another respective sub-threshold voltage pulse via each of the plurality of access lines 310 to change a threshold voltage of a respective variable resistance memory cell 305 coupled to each access lines 310 and to a different sense line 315-2 in an analog fashion toward a voltage associated with a set state to effectuate synaptic weight changes.
- each of the variable resistance memory cells 305-2, 305-5, and 305- 8 can receive the sub-threshold voltage.
- This example method can continue for any number of sense lines 315 to effectuate synaptic weight changes of the variable resistance memory cells 305.
- the neural memory unit controller can be configured to read a voltage from each of the sense lines 315- 1, and 315-2 to determine a total analog value of the memory array 301 (e.g., a total analog value of the neural memory unit 323) which may be the output 343. [0057]
- the neural memory unit controller can be configured to apply subsequent voltage pulses as input 341 based on the detected total analog value of the sense lines 315- 1 and 315-2.
- the neural memory unit controller can be configured to apply a vector (e.g., the input 341-V1, V2, and V3) responsive to a learning algorithm.
- the learning algorithm and/or the neural memory unit controller can be configured to perform matrix
- vectors e.g., input 341 each comprised of the respective threshold voltage of each variable resistance memory cell in each of the plurality of sense lines 215- 1, and 215-2, and/or 215- S.
- FIG. 4 illustrates an example graph of threshold voltage distributions of neural network memory in accordance with an embodiment of the present disclosure.
- the example graph 484 illustrates an x-axis of threshold voltage (Vth) 469 of variable resistance memory cells (e.g., self-selecting memory cells) where the Vth increases from the left side of the page to the right side of the page.
- Graph 484 illustrates y-axis 467 representing a statistical normal quantile of bit (e.g., standard deviation in gaussian distribution).
- the graph 484 illustrates arrays of variable resistance memory cells during different programming and read combinations.
- the legend below the x-axis includes respective line notation for different combinations of polarities of programming pulses and read values.
- a notation for negative programming with a negative read 461 a notation for negative programming and a positive read 462, a notation for positive programming and a negative read 464, and a notation for a positive programming with a positive read 466.
- Each variable resistance memory cell of a memory array can be programmed to include a threshold voltage Vth 469.
- Each variable resistance memory cell can spread the Vth values within a range between a high Vth state (e.g., a reset pulse) and a low Vth state (e.g., a set pulse) when voltage is pulsed below the threshold voltage Vth 469.
- the programming pulse and/or other voltage pulses applied across the variable resistance memory cells can be referred to herein as being of a polarity.
- Polarity refers to a bias in the voltages of write and read operations.
- a sense line and an access line can be biased at different voltages for write and read operations.
- a programming pulse can be of a positive or negative polarity and can depend on read polarity.
- the behavior of the variable resistance memory cell may determine if the read pulse determined in response to the programming pulse has a positive or negative voltage value.
- variable resistance memory cells such as the variable resistance memory ceils in the array can be configured by a neural memory unit controller to change a logic state by manipulating their respective chalcogenide material in an amorphous state. This state change can increase or decrease the resistive value of the self-selecting memory cell which can be represented as a synaptic weight.
- variable resistance memory cells may include memory cells configured to store multiple levels and/or that may have wide sense windows.
- types of memory may be configured to perform training operations by pulse (e.g., spike) control.
- training operations may include spike-timing- dependent plasticity (STOP).
- STOP may be form of Hebbian learning that is induced by correlation between spikes transmitted between nodes (e.g., neurons).
- STDP may be an example of a process that adjusts the strength of connections between nodes (e.g., neurons).
- Figure 4 illustrates a large analog value spread (e.g., wide sense window/a digital state with bit to bit variability) of Vth values in instances where the read value is negative (e.g., neg-neg 461 and pos-neg 464) compared to instances where the read value is positive (e.g., neg-pos 462 and pos-pos 466), which may have a narrower sense window- and a smaller Vth spread.
- Vth values e.g., wide sense window/a digital state with bit to bit variability
- Variable resistance memory cells can be used in neuromorphic applications where the array of variable resistance memory cells can be trained by external pulses to obtain certain analog Vth values.
- Current along signal lines e.g., sense lines 315 and access lines 310 of Figure 3
- can represent the final output e.g., output 343, a total analog value of a neural memory unit 323 of Figure 3
- This may be obtained by using the analog voltage inputs in a matrix by a vector multiplication function.
- FIG. 5 illustrates an example graph of threshold voltage distributions of neural network memory in accordance with an embodiment of the present disclosure.
- the example graph 578 illustrates an x-axis of threshold voltages (Vth) 569 of variable resistance memory cells (e.g., self-selecting memory cells) where the Vth increases from the left side of the page (a set state 581) to the right side of the page (a reset state 582).
- Graph 578 illustrates y-axis 567 representing a statistical normal quantile of bit (e.g., standard deviation in gaussian distribution).
- the legend of the graph 578 illustrates a solid line 579 representing a memory array of variable resistance memory cells at 10,000 reads at a sub-threshold voltage pulse.
- the legend of the graph 578 illustrates a broken line 580 of variable resistance memory cells at 65,000 reads at sub-threshold voltage pulses. While quantities of pulses are recited here as examples, any number of pulses could be used.
- the graph 578 illustrates threshold voltage (Vth) 569 which can be a range of analog values representing an incremental change in logic state of the chalcogenide material of a variable resistance memory cell (e.g., a self- selecting memory cell).
- Vth threshold voltage
- the chalcogenide material of the variable resistance memory cell can change from a reset state 582 corresponding to a higher Vth (e.g., low conductance, high resistance to current flow) toward a set state 581 corresponding to a lower Vth (e.g., high conductance, low resistance to current flow).
- the chalcogenide material of a self-selecting memory cell can be in an amorphous state and have different threshold voltage values while maintaining the amorphous state, based at least in part, on a number of sub-threshold pulses, a magnitude of sub-threshold pulses, and/or a duration of the sub-threshold pulses.
- variable resistance memory cells of 579 and/or 580 can change threshold voltages in an analog manner from a reset state 582 to a set state toward 581.
- the threshold voltage 569 in a range of analog values between a voltage associated with a reset state 582 and the voltage associated with the set state 581 can change responsive to sub-threshold voltage pulses, where each of the range of analog values corresponds to a synaptic weight.
- the quantity of pulses and/or the magnitude of sub-threshold voltage pulses can be associated with strong learning event and/or a weak learning event.
- the learning events can be set by a learning algorithm, where a weak learning event may be a relatively low quantity of pulses (e.g., less than 1000) and/or pulse that is not far enough below the Vth of the variable resistance memory cell.
- a strong learning event may be a relatively high quantity of pulses (e.g., greater than 1000) and/or a pulse that is a larger magnitude below the Vth of the variable resistance memory cell. Stronger learning events may correspond to a greater change in threshold voltages of the variable resistance memory cells.
- Figure 6 illustrates an example method 660 for neural network memory in accord ance with embodiments of the present disclosure.
- the operations of method 660 may he implemented by a neural memory unit controller (e.g., die controller 140 of Figure 1) or its components as described herein.
- the operations of method 660 may be performed within a neural memory unit as described with reference to Figures 1- 5.
- a neural memory unit controller may execute a set of codes to control the functional elements of a memory device to perform the functions described below. Additionally, or alternatively, the neural memory unit controller may perform aspects of the functions described below using special-purpose hardware.
- method 660 includes applying, by a neural memory unit controller coupled to an array of variable resistance memory cells (e.g., the variable resistance memory cells 105 of Figure 1), a weak sub-threshold voltage pulse to a variable resistance memory cell of the array (e.g., the array 301 of Figure 3) in response to a first weak learning event.
- a neural memory unit controller coupled to an array of variable resistance memory cells (e.g., the variable resistance memory cells 105 of Figure 1), a weak sub-threshold voltage pulse to a variable resistance memory cell of the array (e.g., the array 301 of Figure 3) in response to a first weak learning event.
- the weak sub-threshold voltage pulse causes a weak change in a threshold voltage (e.g., Vth 569 of Figure 5) of the variable resistance memory cell in an analog fashion toward a voltage associated with a set state (e.g., the set state 581 of Figure 5) to effectuate a weak synaptic weight change.
- the weak sub-threshold voltage may be a voltage that is below the threshold voltage Vth of the variable resistance memory cell, but the difference in magnitude between the sub-voltage pulse and the Vth of the variable resistance memory cell not enough to change the Vth to a degree to represent learning.
- method 660 includes applying, by the neural memory unit controller coupled to the array of variable resistance memory cells, a strong sub-threshold voltage pulse to the variable resistance memory cell of the array.
- the strong sub-threshold voltage pulse may have a greater difference in magnitude between the voltage of the sub-threshold voltage pulse and the Vth of the variable resistance memory cell.
- the strong sub-threshold voltage pulse causes a strong change in the threshold voltage of the variable resistance memory cell in an analog fashion toward the voltage associated with the set state to effectuate a strong synaptic weight change.
- the application of the strong sub-threshold voltage pulse can include the difference in magnitude of the second pulse being larger than the first.
- the duration of the sub-threshold pulse can be manipulated to elicit a change in the Vth of the variable resistance memory cell.
- the application of the strong sub-threshold voltage pulse can include the difference in duration of the second pulse being longer than the first.
- FIG. 7 illustrates an example method 777 for neural network memory in accordance with embodiments of the present disclosure.
- the operations of method 777 may be implemented by a neural memory unit controller (eg , the controller 140 of Figure i) or its components as described herein.
- the operations of method 777 may be performed within a neural memory unit as described with reference to Figures 1 -6.
- a neural memory unit controller may execute a set of codes to control the functional elements of a memory' device to perform the functions described below'. Additionally, or alternatively, the neural memory unit controller may perform aspects of the functions described below using special-purpose hardware.
- the method 777 includes applying, by a neural memory unit controller coupled to an array of variable resistance memory cells (e.g., the variable resistance memory cells 105 of Figure 1), a first quantity of sub threshold voltage pulses to a variable resistance memory cell of the array in response to a weak learning event.
- the first quantity of sub-threshold voltage pulses can cause a weak change in a threshold voltage Vth (e.g., Vth 569 of Figure 5) of the variable resistance memory cell in an analog fashion toward a voltage associated with a set state (e.g., the set state 581 of Figure 5) to effectuate a weak synaptic weight change.
- the weak synaptic weight change can be an ineffective change in the voltage threshold Vth of the variable resistance memory cell.
- the learning algorithm may increase the quantity of pulses applied to the variable number resistance cell (e.g., may change the learning event).
- the method 777 includes applying, by the neural memory unit controller coupled to the array of variable resistance memory cells, a second quantity of sub-threshold voltage pulses to the variable resistance memory cell of the array in response to a strong learning event where the second quantity is greater than the first quantity.
- the neural network unit controller can increase the quantity of pulses from 1000 to 65000 to effectuate a change in Vth toward the set state.
- the second quantity of sub-threshold voltage pulses can cause a strong change in the threshold voltage of the variable resistance memory cell in an analog fashion toward the voltage associated with the set state to effectuate a strong synaptic weight change.
- the neural memory unit controller can apply subsequent sub-threshold voltage pulses to the variable resistance memory cell in response to subsequent learning events until a sub-threshold voltage pulse causes the variable resistance memory cell to reach the set state (e.g., a crystalline, highly conductive, low resistance to current flow state).
- the set state e.g., a crystalline, highly conductive, low resistance to current flow state.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Computer Hardware Design (AREA)
- Crystallography & Structural Chemistry (AREA)
- Chemical & Material Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Multimedia (AREA)
- Semiconductor Memories (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
- Reduction Or Emphasis Of Bandwidth Of Signals (AREA)
- Feedback Control In General (AREA)
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202080047870.8A CN114051620A (en) | 2019-07-03 | 2020-06-03 | Neural network memory |
JP2021577224A JP2022539751A (en) | 2019-07-03 | 2020-06-03 | neural network memory |
KR1020227003266A KR20220028051A (en) | 2019-07-03 | 2020-06-03 | neural network memory |
EP20835010.8A EP3994624A4 (en) | 2019-07-03 | 2020-06-03 | Neural network memory |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/502,978 US11587612B2 (en) | 2019-07-03 | 2019-07-03 | Neural network memory with an array of variable resistance memory cells |
US16/502,978 | 2019-07-03 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021002995A1 true WO2021002995A1 (en) | 2021-01-07 |
Family
ID=74066809
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2020/035816 WO2021002995A1 (en) | 2019-07-03 | 2020-06-03 | Neural network memory |
Country Status (7)
Country | Link |
---|---|
US (1) | US11587612B2 (en) |
EP (1) | EP3994624A4 (en) |
JP (1) | JP2022539751A (en) |
KR (1) | KR20220028051A (en) |
CN (1) | CN114051620A (en) |
TW (1) | TWI778374B (en) |
WO (1) | WO2021002995A1 (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11544061B2 (en) * | 2020-12-22 | 2023-01-03 | International Business Machines Corporation | Analog hardware matrix computation |
US20230049032A1 (en) * | 2021-08-02 | 2023-02-16 | Silicon Storage Technology, Inc. | Output circuitry for analog neural memory in a deep learning artificial neural network |
US12033695B2 (en) * | 2022-05-09 | 2024-07-09 | Micron Technology, Inc. | Techniques for multi-level chalcogenide memory cell programming |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4876668A (en) * | 1985-07-31 | 1989-10-24 | California Institute Of Technology | Thin film memory matrix using amorphous and high resistive layers |
US20160358661A1 (en) * | 2012-02-23 | 2016-12-08 | Micron Technology, Inc. | Methods of operating memory |
US20170083811A1 (en) * | 2015-09-18 | 2017-03-23 | Samsung Electronics Co., Ltd. | Weighting device, neural network, and operating method of the weighting device |
US20170277628A1 (en) * | 2016-03-24 | 2017-09-28 | Somnath Paul | Technologies for memory management of neural networks with sparse connectivity |
Family Cites Families (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4248187B2 (en) * | 2002-03-27 | 2009-04-02 | シャープ株式会社 | Integrated circuit device and neuro element |
US20040257848A1 (en) * | 2003-06-18 | 2004-12-23 | Macronix International Co., Ltd. | Method for adjusting the threshold voltage of a memory cell |
JP5155843B2 (en) * | 2008-12-22 | 2013-03-06 | シャープ株式会社 | Information processing apparatus and neural network circuit using the same |
US8184469B2 (en) * | 2009-11-30 | 2012-05-22 | Micron Technology, Inc. | Stored multi-bit data characterized by multiple-dimensional memory states |
US10332004B2 (en) * | 2015-07-13 | 2019-06-25 | Denso Corporation | Memristive neuromorphic circuit and method for training the memristive neuromorphic circuit |
US10268949B2 (en) * | 2016-03-21 | 2019-04-23 | International Business Machines Corporation | Artificial neuron apparatus |
TWI625681B (en) | 2017-05-11 | 2018-06-01 | 國立交通大學 | Neural network processing system |
JP6724870B2 (en) * | 2017-06-19 | 2020-07-15 | 株式会社デンソー | Artificial neural network circuit training method, training program, and training device |
US10127494B1 (en) | 2017-08-02 | 2018-11-13 | Google Llc | Neural network crossbar stack |
WO2019100036A1 (en) * | 2017-11-20 | 2019-05-23 | The Regents Of The University Of California | Memristive neural network computing engine using cmos-compatible charge-trap-transistor (ctt) |
US10366747B2 (en) | 2017-11-30 | 2019-07-30 | Micron Technology, Inc. | Comparing input data to stored data |
US10424376B2 (en) | 2017-12-24 | 2019-09-24 | Micron Technology, Inc. | Material implication operations in memory |
FR3078814B1 (en) * | 2018-03-09 | 2020-10-02 | Commissariat Energie Atomique | ANALOGUE PROGRAMMING PROCESS OF A PHASE CHANGE MEMORY CELL BY MEANS OF IDENTICAL ELECTRIC PULSES |
US10622063B2 (en) * | 2018-06-27 | 2020-04-14 | Sandisk Technologies Llc | Phase change memory device with reduced read disturb and method of making the same |
US10643119B2 (en) * | 2018-07-24 | 2020-05-05 | Sandisk Technologies Llc | Differential non-volatile memory cell for artificial neural network |
-
2019
- 2019-07-03 US US16/502,978 patent/US11587612B2/en active Active
-
2020
- 2020-06-03 WO PCT/US2020/035816 patent/WO2021002995A1/en unknown
- 2020-06-03 CN CN202080047870.8A patent/CN114051620A/en active Pending
- 2020-06-03 KR KR1020227003266A patent/KR20220028051A/en not_active Application Discontinuation
- 2020-06-03 JP JP2021577224A patent/JP2022539751A/en not_active Ceased
- 2020-06-03 EP EP20835010.8A patent/EP3994624A4/en not_active Withdrawn
- 2020-06-17 TW TW109120423A patent/TWI778374B/en active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4876668A (en) * | 1985-07-31 | 1989-10-24 | California Institute Of Technology | Thin film memory matrix using amorphous and high resistive layers |
US20160358661A1 (en) * | 2012-02-23 | 2016-12-08 | Micron Technology, Inc. | Methods of operating memory |
US20170083811A1 (en) * | 2015-09-18 | 2017-03-23 | Samsung Electronics Co., Ltd. | Weighting device, neural network, and operating method of the weighting device |
US20170277628A1 (en) * | 2016-03-24 | 2017-09-28 | Somnath Paul | Technologies for memory management of neural networks with sparse connectivity |
Non-Patent Citations (2)
Title |
---|
ALEXANDER ORORBIA; AHMED AHMED ELSAID; TRAVIS DESELL: "Investigating RNN Memory using Neuro-Evolution: Investigating Recurrent Neural Network Memory Structures using Neuro-Evolution", ARXIV.ORG, 2 June 2019 (2019-06-02), pages 1 - 11, XP081026204 * |
See also references of EP3994624A4 * |
Also Published As
Publication number | Publication date |
---|---|
JP2022539751A (en) | 2022-09-13 |
US20210005250A1 (en) | 2021-01-07 |
US11587612B2 (en) | 2023-02-21 |
KR20220028051A (en) | 2022-03-08 |
TW202105391A (en) | 2021-02-01 |
TWI778374B (en) | 2022-09-21 |
CN114051620A (en) | 2022-02-15 |
EP3994624A4 (en) | 2023-07-19 |
EP3994624A1 (en) | 2022-05-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11062767B2 (en) | Techniques for programming neural memory unit using cell conditioning | |
TWI778374B (en) | Neural network memory | |
US11295811B2 (en) | Increase of a sense current in memory | |
US11211122B1 (en) | Increase of a sense current in memory | |
US12045712B2 (en) | Self select memory cell based artificial synapse | |
US11380391B2 (en) | Neural network memory with mechanism to change synaptic weight | |
US20240087619A1 (en) | Matrix formation for performing computational operations in memory | |
US20240242771A1 (en) | Using a subthreshold voltage for mapping in memory | |
US20240038322A1 (en) | Performing sense operations in memory | |
US20240249777A1 (en) | Storing one data value by programming a first memory cell and a second memory cell |
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: 20835010 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2021577224 Country of ref document: JP Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 20227003266 Country of ref document: KR Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 2020835010 Country of ref document: EP Effective date: 20220203 |