US20200286553A1 - In-memory computation device with inter-page and intra-page data circuits - Google Patents
In-memory computation device with inter-page and intra-page data circuits Download PDFInfo
- Publication number
- US20200286553A1 US20200286553A1 US16/297,504 US201916297504A US2020286553A1 US 20200286553 A1 US20200286553 A1 US 20200286553A1 US 201916297504 A US201916297504 A US 201916297504A US 2020286553 A1 US2020286553 A1 US 2020286553A1
- Authority
- US
- United States
- Prior art keywords
- page
- input
- circuits
- circuit
- block
- 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.)
- Granted
Links
- 239000013598 vector Substances 0.000 claims description 104
- 230000004044 response Effects 0.000 claims description 19
- 238000012546 transfer Methods 0.000 claims description 8
- 238000006243 chemical reaction Methods 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 abstract description 17
- 210000004027 cell Anatomy 0.000 description 43
- 239000010410 layer Substances 0.000 description 38
- 210000002569 neuron Anatomy 0.000 description 22
- 238000010586 diagram Methods 0.000 description 10
- 230000000946 synaptic effect Effects 0.000 description 8
- 238000003491 array Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000012549 training Methods 0.000 description 6
- 238000000034 method Methods 0.000 description 3
- 210000004205 output neuron Anatomy 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 239000003990 capacitor Substances 0.000 description 2
- 210000002364 input neuron Anatomy 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000004020 conductor Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000005055 memory storage Effects 0.000 description 1
- 229910044991 metal oxide Inorganic materials 0.000 description 1
- 150000004706 metal oxides Chemical class 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
Images
Classifications
-
- 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
- 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
- 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
- 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
- G11C13/0023—Address circuits or decoders
- G11C13/0026—Bit-line or column 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
- G11C13/0023—Address circuits or decoders
- G11C13/0028—Word-line or row circuits
-
- G—PHYSICS
- G11—INFORMATION STORAGE
- G11C—STATIC STORES
- G11C16/00—Erasable programmable read-only memories
- G11C16/02—Erasable programmable read-only memories electrically programmable
- G11C16/04—Erasable programmable read-only memories electrically programmable using variable threshold transistors, e.g. FAMOS
- G11C16/0408—Erasable programmable read-only memories electrically programmable using variable threshold transistors, e.g. FAMOS comprising cells containing floating gate transistors
- G11C16/0416—Erasable programmable read-only memories electrically programmable using variable threshold transistors, e.g. FAMOS comprising cells containing floating gate transistors comprising cells containing a single floating gate transistor and no select transistor, e.g. UV EPROM
-
- 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
Definitions
- the present invention relates to in-memory computing devices, and more particularly, to in-memory computing devices supporting efficient data sharing among multiple computational stages.
- a neural network is an information processing paradigm that is inspired by the way biological nervous systems process information. With the availability of large training datasets and sophisticated learning algorithms, neural networks have facilitated major advances in numerous domains such as computer vision, speech recognition, and natural language processing.
- FIG. 1 illustrates an example neural network 100 .
- the neural network 100 contains multiple neurons arranged in layers, which can be considered to be an example of a computational stage including many parallel operations.
- the neural network 100 includes an input layer 102 of input neurons (i.e., neurons that provide the input data), three hidden layers 106 , 108 and 110 of hidden neurons (i.e., neurons that perform computations and transfer information from the input neurons to the output neurons), and an output layer 104 of output neurons (i.e., neurons that provide the output data). Neurons in adjacent layers have synaptic layers of connections between them.
- the synaptic layer 112 connects neurons in the input layer 102 and the hidden layer 106
- the synaptic layer 114 connects neurons in the hidden layers 106 and 108
- the synaptic layer 116 connects neurons in the hidden layers 108 and 110
- the synaptic layer 118 connects the neurons in the hidden layer 110 and the output layer 104 . All these connections have weights associated with them.
- the neurons 122 , 124 and 126 in the hidden layer 106 are connected to a neuron 128 in the hidden layer 108 by connections with weights w 1 132 , w 2 134 and w 3 136 , respectively.
- the output for the neuron 128 in the hidden layer 108 can be calculated as a function of the inputs (x 1 , x 2 , and x 3 ) from the neurons 122 , 124 and 126 in the hidden layer 106 and the weights w 1 132 , w 2 134 and w 3 136 in the connections.
- the function can be expressed as follows:
- each product term is a product of a variable input x 1 and a weight w 1 .
- the weight w 1 can vary among the terms, corresponding for example to coefficients of the variable inputs x 1 .
- outputs from the other neurons in the hidden layer can also be calculated.
- the outputs of the two neurons in the hidden layer 110 act as inputs to the output neuron in the output layer 104 .
- Neural networks are used to learn patterns that best represent a large set of data.
- the hidden layers closer to the input layer in general learn high level generic patterns, and the hidden layers closer to the output layer in general learn more data-specific patterns.
- Training is a phase in which a neural network learns from training data. During training, the connections in the synaptic layers are assigned weights based on the results of the training session.
- Inference is a stage in which a trained neural network is used to infer/predict using input data and to produce output data based on the prediction.
- In-memory computing is an approach in which memory cells, organized in an in-memory computing device, can be used for both data processing and memory storage.
- a neural network can be implemented using an in-memory computing device for a number of synaptic layers.
- the weights for the sum-of-products function can be stored in memory cells of the in-memory computing device.
- the sum-of-products function can be realized as a circuit operation in the in-memory computing device in which the electrical characteristics of the memory cells of the array effectuate the function.
- An engineering issue associated with neural networks relates to movement of data among the synaptic layers.
- An in-memory computation device comprises a memory configured in a plurality of blocks useable for in-memory computations.
- Blocks B(n), n going from 0 to N ⁇ 1, in the plurality of blocks have corresponding page input circuits PI(n) and page output circuits PO(n) that are operatively coupled to sets of bit lines in the blocks.
- each block B(n) can include a set S(n) of bit lines coupled to its corresponding page output circuit and to its corresponding page input circuit.
- the device includes in some embodiments a data bus system for providing an external source of input data and a destination for output data.
- Data circuits are configurable to connect page input circuit PI(n) to one or more of page output circuit PO(n), page output circuit PO(n ⁇ 1), and the data bus system to source input data for a sensing cycle. This configuration can be done between each sensing cycle, or in longer intervals, in order to support a variety of neural network configurations and operations.
- a plurality of bit line bias circuits is connected to bit lines in the plurality of blocks.
- Bit line bias circuit Y(n) in the plurality of bit line bias circuits being operatively coupled to block B(n), and to page input circuit PI(n) in the plurality of page input circuits.
- the bit line biasing circuit can bias the bit lines in block B(n) in response to input voltages generated by page input circuit PI(n).
- FIG. 1 illustrates an example neural network
- FIG. 2 illustrates a NOR style memory array configured for in-memory computation of a sum-of-products suitable for use in a neural network.
- FIG. 3A is a simplified diagram of an in-memory computation device with inter-page and intra-page data circuits.
- FIG. 3B is a simplified diagram of an alternative in-memory computation device with inter-page and intra-page data circuits.
- FIG. 4 is a simplified diagram of an in-memory computation device using a NOR style memory array with drain side bias generated in response to input data.
- FIG. 5 is a diagram of an embodiment of page input circuitry usable in a device like that of FIG. 4 .
- FIG. 6 is a diagram of page output circuitry usable in a device like that of FIG. 4 .
- FIG. 7 is a diagram of a digital-to-analog converter usable to generate drain side bias voltages for use in an embodiment like that of FIG. 4 .
- FIG. 8 is a diagram of an alternative embodiment of page output circuitry usable in a device like that of FIG. 4 .
- FIGS. 2-8 A detailed description of embodiments of the present invention is provided with reference to the FIGS. 2-8 .
- FIG. 2 illustrates a memory array 210 including a plurality of memory cells (e.g. 211 , 212 ) configured in a NOR style.
- the memory cells in a given column of the array are connected between a bit line and a source line.
- a coefficient vector (a set of coefficients W(i)) is stored in the memory cells in a given row, such as on the row of cells connected to word line WL 0 .
- Input data (X 0 -X 3 ) is applied by biasing the bit lines BL 0 to BL 3 connecting a bias voltage to the drain side of the memory cells.
- a row of memory cells is selected by applying a word line signal to a selected word line (e.g. WL 0 ). Unselected word lines are biased in an off condition. Current is generated in the cells in the selected row that is a function of the coefficient W(i) stored in the accessed memory cell, and the input data X(i).
- the signals generated on the bit lines or source lines in response to the word line signal on the selected word line can be combined ( 215 ) to generate the sum-of-products output SUM by sensing a sum of the signals.
- the memory cells in the array of FIG. 2 can be nonvolatile memory cells, such as floating gate memory cells or dielectric charge trapping memory cells. Also, programmable resistance memory cells can be utilized, such as phase change memory, metal oxide memory and others. In some embodiments, the memory cells in the array of FIG. 2 can be volatile memory cells. The memory cells in the array of FIG. 2 can store one bit per cell in some examples. In other examples, the memory cells in the array of FIG. 2 can store multiple bits per cell. In yet other examples, the memory cells in the array of FIG. 2 can store analog values.
- the NOR style array can be implemented with many bit lines and word lines, storing thousands or millions of bits.
- the NOR style array of FIG. 2 can be implemented in a three-dimensional configuration, with many two-dimensional arrays stacked in many levels.
- In-memory computation circuits as described herein can use memory arrays of other styles in some examples, including for example AND style arrays or NAND style arrays.
- FIG. 3A is a simplified diagram of an in-memory computation device that includes a memory 300 that comprises an array of memory cells storing coefficients W(i), and including a plurality of word lines and a plurality of bit lines.
- the memory is configured in a plurality of blocks, B( 0 ), B( 1 ), and so on.
- Each block includes a plurality of bit lines and a plurality of word lines.
- the word lines can be shared among more than one block in some embodiments, so that he word lines for more than one block are driven and decoded together, or constitute a shared single word line conductor across more than one block. In other embodiments, the word lines are driven and decoded for individual blocks.
- the blocks B(n) in the plurality of blocks comprise corresponding YPASS circuits 301 , 302 , which combine the signals on the set S(n) of bit lines in the block into an output signal on a data line DL(n).
- the memory cells coupled to the bit lines in the set S(n) of bit line store a coefficient vector W(n) represented by the threshold voltages of the memory cells in a row or rows selected by word line signals.
- the YPASS circuits include a plurality of bit line bias circuits (not shown in FIG. 3A ) connected to the plurality of sets of bit lines.
- Bit line bias circuit Y(n) in the plurality of bit line bias circuits is operatively coupled to set S(n) of bit lines in the corresponding block B(n), and to page input circuit PI(n) in the plurality of page input circuits.
- the bit line bias circuit biases the bit lines in set S(n) in response to input voltages generated by page input circuit PI(n).
- the input signals X[1:z] on line 360 are input voltages in this example, applied to YPASS circuit 301 , and X[z+1:2z] on line 361 are applied to YPASS circuit 302 , to be applied to the corresponding set of bit lines having “z” members.
- YPASS circuits each of which includes N sets S(n) of bit lines, where a YPASS circuit in block (n) including set S(n) bit lines receives input signals, X[n*z+1:(n+1)*z)], for n going from 0 to N ⁇ 1.
- two blocks, B( 0 ) and B( 1 ), are illustrated.
- the number N can be 8 for example, or 16. In other embodiments, N can be in much higher.
- the memory 300 can comprise a single two-dimensional array, in which the blocks are arranged side-by-side in sequence.
- the memory 300 can comprise a three-dimensional array, including a plurality of stacked two-dimensional arrays.
- each two-dimensional array in the stack can comprise one block. Blocks arranged in sequence can reside on sequential levels of the stack.
- each two-dimensional array in the stack can comprise more than one block configured as illustrated.
- a data bus system represented by the input data block 320 and the input data block 321 , is provided on the device.
- the data bus system can be used as an input/output interface for data from an external data source, or from other circuitry on the device that generates data for use as the input vectors in some configurations.
- the input data block 320 and the input data block 321 can comprise, for examples, a data cache memory, a register or a latch coupled to input/output circuitry on the device or other data bus circuits.
- the output signals on lines DL 0 and DL 1 from the YPASS circuits 301 , 302 represent an in-memory computation result, combining the signals on the bit lines in the corresponding block B( 0 ) or B( 1 ), that are produced in response to weights stored in memory cells selected by a word line signal on a selected word line in the memory array, and to the input signals X[n*z+1:(n+1)*z)] from lines 360 , 361 for the corresponding block.
- Word line drivers 398 and decoders are included to provide the word line signals on selected word lines. In the simplified example shown in FIG.
- YPASS circuits 301 , 302 are coupled to corresponding page input circuits PI( 0 ), PI( 1 ) which provide input signals X[n*z+1:(n+1)*z)] on lines 360 and 361 , respectively. Also, YPASS circuits 301 , 302 are coupled to corresponding page output circuits PO( 0 ), PO( 1 ) to receive in-memory computation result on corresponding data lines DL 0 and DL 1 , respectively.
- the page input circuits include respective input registers 314 , 316 which store input data in the form of input vectors VI(n) received as input from data circuits on the device for the corresponding block B(n).
- the page input circuits can include circuits that convert the input vector VI(n), which have a number of bits equal to a multiple M of the number Z of bit lines in the block, to the input signals.
- the page output circuits (PO( 0 ) and PO( 1 )) include respective output registers 315 , 317 which store output vectors VO(n) generated in response to an in-memory computation using memory cells in the corresponding block B(n) for output to the data circuits on the device.
- the output vectors VO(n) can have the same number of bits as the input vectors VI(n), or a different number of bits.
- Data circuits on the device are configurable to interconnect the input register for a given page (e.g., PI( 1 )) singly or in any combination, with sources of input data, including the output register of a previous page (e.g., PO( 0 )), the output register of the same page (e.g., PO( 1 ) as feedback) and the data bus to source input data for a given sensing cycle.
- sources of input data including the output register of a previous page (e.g., PO( 0 )), the output register of the same page (e.g., PO( 1 ) as feedback) and the data bus to source input data for a given sensing cycle.
- An input vector including the input data applied to the input register in a given sensing cycle can be sourced by a single source, or a combination of multiple sources.
- the data circuits are configurable to transfer an output vector VO( 0 ) of page output circuit PO( 0 ) as all or part an input vector VI( 1 ) to the next page input circuit PI( 1 ) as represented by the line 340 . Also, data circuits are configurable to feed back the output vector VO( 0 ) of the page output circuit PO( 0 ) as all or part of an input vector VI( 0 ) for the page input circuit PI( 0 ), as represented by line 350 . In addition, the data circuits are configurable to connect the page input circuit PI( 0 ) to the data bus system including input data block 320 as represented by line 330 , to receive all or part of an input vector VI( 0 ) from another source in the bus system.
- the data circuits on the device are configurable to transfer an output vector VO( 1 ) of page output circuit PO( 1 ) as all or part of an input vector VI( 2 ) for the next page input circuit as represented by the line 341 .
- data circuits are configurable to feedback the output VO( 1 ) of the page output circuit PO( 1 ) as all or part of an input vector VI( 1 ) for the page input circuit PI( 1 ), as represented by line 351 .
- the data circuits are configurable to connect the page input circuit PI( 1 ) to the data bus system including input data block 321 as represented by line 331 , to receive all or part of an input vector VI( 1 ) from an another source on the bus system.
- Configuration circuits 399 are included on the device.
- the circuits 399 include logic, configuration parameters storage, or both.
- the configuration circuits 399 control the configuration of the data circuits for the routing of input vectors and output vectors among the page input and page output circuits.
- the configuration of the data circuits can be set dynamically for each sensing cycle as suits the needs of a particular implementation, using timed control signals delivered to switches in the data circuits.
- the configuration circuit can use volatile or nonvolatile configuration registers to set up the data circuits.
- Control circuits are coupled to the circuit of FIG. 3A , and include logic to control execution of in-memory computation operations, data circuit configuration operations in coordination with configuration circuits 399 , and memory operations like program and erase.
- the control circuits can comprise command decoders, configuration registers, state machines and timing circuits set up to control the operations.
- the control circuits can include voltage regulators, charge pumps and other bias circuits to provide appropriate bias voltages in support of the memory operations.
- FIG. 3B is a simplified diagram of an in-memory computation device that includes a memory 900 that comprises an array of memory cells storing coefficients W(i), and including a plurality of word lines and a plurality of bit lines.
- the memory is configured in a plurality of blocks, B( 0 ), B( 1 ), and so on. Each block includes a plurality of bit lines and a plurality of word lines. The word lines are driven and decoded for individual blocks in this example.
- the blocks B(n) in the plurality of blocks comprise corresponding YPASS circuits 901 , 902 , which combine the signals on the set S(n) of bit lines in the block into an output signal on a data line DL(n).
- the memory cells coupled to the bit lines in the set S(n) of bit line store a coefficient vector W(n) represented by the threshold voltages of the memory cells in a row or rows selected by one or more word line signals.
- the word line drivers include a plurality of bias circuits (not shown in FIG. 3B ) connected to the plurality of sets of word lines.
- Word line driver WD(n) in the plurality of word line drivers is operatively coupled to a set of word lines in the corresponding block B(n), and to page input circuit PI(n) in the plurality of page input circuits.
- the word line bias circuit biases the word lines in corresponding block in response to input voltages generated by page input circuit PI(n).
- the input signals X[1:z] on line 960 are input voltages in this example, applied to a word line driver 938 for block B( 0 ), and X[z+1:2z] on line 961 are applied to a word line driver for block B( 1 ), to be applied to the corresponding set of word lines having “z” members.
- two blocks, B( 0 ) and B( 1 ), are illustrated.
- the number N can be 8 for example, or 16. In other embodiments, N can be in much higher.
- the memory 900 can comprise a single two-dimensional array, in which the blocks are arranged side-by-side in sequence.
- the memory 900 can comprise a three-dimensional array, including a plurality of stacked two-dimensional arrays.
- each two-dimensional array in the stack can comprise one block. Blocks arranged in sequence can reside on sequential levels of the stack.
- each two-dimensional array in the stack can comprise more than one block configured as illustrated.
- a data bus system represented by the input data block 920 and the input data block 921 , is provided on the device.
- the data bus system can be used as an input/output interface for data from an external data source, or from other circuitry on the device that generates data for use as the input vectors in some configurations.
- the input data block 920 and the input data block 921 can comprise, for examples, a data cache memory, a register or a latch coupled to input/output circuitry on the device or other data bus circuits.
- the output signals on lines DL 0 and DL 1 from the YPASS circuits 901 , 902 represent an in-memory computation result, combining the signals on the bit lines in the corresponding block B( 0 ) or B( 1 ), that are produced in response to weights stored in memory cells selected by one or more word line signals on a selected word line in the memory array, and to the input signals X[n*z+1:(n+1)*z)] from lines 960 , 961 for the corresponding block.
- the word line drivers e.g.
- word line driver WD(n) in the plurality of word line drivers is operatively coupled to block B(n) in the plurality of blocks, and to page input circuit PI(n) in the plurality of page input circuits, and biases the word lines in block B(n) in response to input voltages generated by page input circuit PI(n).
- YPASS circuits 901 , 902 are coupled to corresponding page output circuits PO( 0 ), PO( 1 ) to receive in-memory computation result on corresponding data lines DL 0 and DL 1 , respectively.
- the page input circuits include respective input registers 914 , 916 which store input data in the form of input vectors VI(n) received as input from data circuits on the device for the corresponding block B(n).
- the page input circuits can include circuits that convert the input vector VI(n), which have a number of bits equal to a multiple M of the number Z of bit lines in the block, to the input signals.
- the page output circuits (PO( 0 ) and PO( 1 )) include respective output registers 915 , 917 which store output vectors VO(n) generated in response to an in-memory computation using memory cells in the corresponding block B(n) for output to the data circuits on the device.
- the output vectors VO(n) can have the same number of bits as the input vectors VI(n), or a different number of bits.
- Data circuits on the device are configurable to interconnect the input register for a given page (e.g., PI( 1 )) singly or in any combination, with sources of input data, including the output register of a previous page (e.g., PO( 0 )), the output register of the same page (e.g., PO( 1 ) as feedback) and the data bus to source input data for a given sensing cycle.
- sources of input data including the output register of a previous page (e.g., PO( 0 )), the output register of the same page (e.g., PO( 1 ) as feedback) and the data bus to source input data for a given sensing cycle.
- An input vector including the input data applied to the input register in a given sensing cycle can be sourced by a single source, or a combination of multiple sources.
- the data circuits are configurable to transfer an output vector VO( 0 ) of page output circuit PO( 0 ) as all or part an input vector VI( 1 ) to the next page input circuit PI( 1 ) as represented by the line 940 . Also, data circuits are configurable to feed back the output vector VO( 0 ) of the page output circuit PO( 0 ) as all or part of an input vector VI( 0 ) for the page input circuit PI( 0 ), as represented by line 950 .
- the data circuits are configurable to connect the page input circuit PI( 0 ) to the data bus system including input data block 920 as represented by line 930 , to receive all or part of an input vector VI( 0 ) from another source in the bus system.
- the data circuits on the device are configurable to transfer an output vector VO( 1 ) of page output circuit PO( 1 ) as all or part of an input vector VI( 2 ) for the next page input circuit as represented by the line 941 .
- data circuits are configurable to feedback the output VO( 1 ) of the page output circuit PO( 1 ) as all or part of an input vector VI( 1 ) for the page input circuit PI( 1 ), as represented by line 951 .
- the data circuits are configurable to connect the page input circuit PI( 1 ) to the data bus system including input data block 921 as represented by line 931 , to receive all or part of an input vector VI( 1 ) from an another source on the bus system.
- Configuration circuits 999 are included on the device.
- the circuits 999 include logic, configuration parameters storage, or both.
- the configuration circuits 999 control the configuration of the data circuits for the routing of input vectors and output vectors among the page input and page output circuits.
- the configuration of the data circuits can be set dynamically for each sensing cycle as suits the needs of a particular implementation, using timed control signals delivered to switches in the data circuits.
- the configuration circuit can use volatile or nonvolatile configuration registers to set up the data circuits.
- Control circuits are coupled to the circuit of FIG. 3B , and include logic to control execution of in-memory computation operations, data circuit configuration operations in coordination with configuration circuits 999 , and memory operations like program and erase.
- the control circuits can comprise command decoders, configuration registers, state machines and timing circuits set up to control the operations.
- the control circuits can include voltage regulators, charge pumps and other bias circuits to provide appropriate bias voltages in support of the memory operations.
- FIG. 4 is a diagram of an in-memory computation device like that of FIG. 3A with information about embodiments of the YPASS circuits 401 , 402 , the page input circuits PI( 0 ), PI( 1 ) and the page output circuits PO( 0 ), PO( 1 ).
- the in-memory computation device includes a memory array 410 along with supporting circuitry as mentioned with respect to FIG. 3A .
- the memory array 410 includes a plurality of blocks B( 0 ), B( 1 ), . . .
- a set S(n) of bit lines in each block is coupled to a corresponding YPASS circuit 401 , 402 .
- the YPASS circuit 401 includes a bit line bias circuit.
- the bit line bias circuit includes a plurality of clamp transistors 490 - 493 having one source/drain terminal coupled to a bit line in the block B( 0 ) of the array 410 , and another source/drain terminal coupled to a summing node on the data line DL 0 .
- the gates of the clamp transistors 490 - 493 are connected to corresponding input signals in the set of input signals X[1:z] provided by the page input circuit 419 .
- the YPASS circuit 402 includes a bit line bias circuit.
- the bit line bias circuit includes a plurality of clamp transistors 494 - 497 , each having one source/drain terminal coupled to a bit line in the block B( 1 ) of the array 410 , and another source/drain terminal coupled to a summing node on the data line DL 1 .
- the gates of the clamp transistors 494 - 497 are connected to corresponding input signals in the set of input signals X[z+1:2z] provided by the page input circuit 459 .
- the page input circuit 419 (PI( 0 )) includes an input register 420 and supporting logic 421 , which converts the input vector stored in the input register 420 to the input signals X[1:z] applied to the gates of the clamp transistors.
- the page input circuit 459 (PI( 1 )) includes an input register 460 and supporting logic 461 , which converts the input vector stored in the input register 460 to the input signals X[z+1:2z] applied to the gates of the clamp transistors.
- the page output circuit 429 (PO( 0 )) includes an N-bit sense amplifier 430 , or other type of multilevel sensing circuit or analog-to-digital converter.
- the output of the sense amplifier 430 is coupled to an output register 431 or other type of register.
- the output register 431 is connected to a compute unit 432 which can accumulate or otherwise process data stored in the output register 431 .
- the circuits in the page output circuit 429 convert the signal on the DL 0 line into an output vector VO(n) for the computation based on a sum-of-signals generated by memory cells on a selected word line on the block B(n).
- the sum-of-signals can correspond to a sum-of-products.
- the page output circuit 469 (PO( 1 )) includes an N-bit sense amplifier 470 , or other type of multilevel sensing circuit or analog-to-digital converter.
- the output of the sense amplifier 470 is coupled to an output register 471 or other type of register.
- the output register 471 in this example is connected to a compute unit 472 which can accumulate or otherwise process data stored in the output register 471 .
- the circuits in the page output circuit 469 convert the signal on the DL 1 line into an output vector VO(n) for the computation based on a sum-of-signals generated by memory cells on a selected word line on the block B(n).
- the cell current in each memory cell on the selected word line coupled to a bit line in the set of bit lines can be represented in one example memory system by the equation:
- Data circuits on the device are configurable to transfer an output vector VO( 0 ) of page output circuit 429 as an input vector VI( 1 ) for the next page input circuit 459 as represented by the line 443 . Also, data circuits are configurable to feedback the output VO( 0 ) of the page output circuit 429 as an input vector VI( 0 ) for the page input circuit 419 as represented by line 442 . In addition, the data circuits are configurable to connect the page input circuit 419 to the data bus system 440 as represented by line 441 , to receive an input vector VI( 0 ) from an another source on the bus system.
- data circuits on the device are configurable to transfer an output vector VO( 1 ) of page output circuit 469 as an input vector VI( 1 ) for the next page input circuit as represented by the line 483 .
- data circuits are configurable to feedback the output VO( 1 ) of the page output circuit 469 as an input vector VI( 1 ) for the page input circuit 459 as represented by line 482 .
- the data circuits are configurable to connect the page input circuit 459 to the data bus system 480 as represented by line 481 to receive an input vector VI( 1 ) from an another source on the bus system.
- FIG. 5 illustrates one example of a page input circuit 500 , representing page input circuit PI(n), which is configured to deliver the input signals on bus 505 to the YPASS circuits of the block B(n) in the memory array. Also, portions of the data circuits which are configured to deliver the input vector VI(n) are illustrated.
- the page input circuit 500 includes an input register 501 that comprises a plurality of latches 520 - 523 .
- Data circuits include a set of switches 510 , 511 , 512 , 513 operable as a multiplexer in response to the signal SW.
- the switches 510 , 511 , 512 , 513 connect and disconnect the register 501 to the previous page output circuit which provides output vector VO(n ⁇ 1) on lines 508 of the data circuits for use as input vector VI(n).
- the data circuits include a multiplexer 551 and input bus register 550 .
- the multiplexer 551 is responsive to the signal SW_IN to connect and disconnect the input data from the input bus register 550 to deliver an input vector VI(n) to the input register 501 .
- the data circuits include a bus 509 which is coupled to a multiplexer (see, MUX 642 in FIG. 6 ), to receive the output vector VO(n) as feedback from the page output circuit PO(n) coupled to the same block B(n), for use as input vector VI(n).
- the signal SW and the signal SW_IN are provided by a configuration circuit such as configuration circuits 399 and configuration circuits 999 shown in FIGS. 3A and 3B .
- the signals SW and SW_IN are provided independently for the page input circuits and page output circuits operatively coupled with each block B(n).
- the input vector VI(n) includes Z chunks of M bits of data.
- Input register 501 applies the Z chunks of M bits of input vector VI(n) to conversion circuits 502 that provide the bias voltages X 5 to X 8 in the Figure to establish the drain level of the selected memory cells in the corresponding block B(n).
- the circuits 502 include one M bit digital-to-analog converter 530 , 531 , 532 , 533 for each of the Z chunks in the input vector VI(n) to generate Z analog bias voltages for corresponding bit lines in the block B(n).
- M is 2 and Z is 4 with an input vector including 8 bits.
- the input vector can include 16 bits, or any number of bits. Also, the number of chunks in the input vector used to generate a bias voltage for one of the corresponding bit lines in the set of bit lines can be determined by the number Z of bit lines in the set.
- FIG. 6 illustrates one example of a page output circuit 600 , representing page output circuit PO(n) which is coupled to the block B(n) and to the page input circuit PI(n).
- the page output circuit 600 is configured to receive the signal on data line DL(n) from the YPASS circuits coupled to the block B(n).
- the data line DL(n) is connected to a sense amplifier 601 or other type of sensing circuit or analog-to-digital converter.
- the sense amplifier 601 in this example includes an adjustable current source I 1 which is connected to node 610 , which is also connected to data line DL(n).
- a precharge transistor 612 is coupled to a sensing node SEN.
- a capacitor 611 is coupled to the sensing node SEN.
- Transistor M 2 is connected in series between node 610 and the sensing node SEN.
- the sensing node SEN is connected to the gate of transistor M 1 .
- Transistor M 1 is connected between ground and output node A.
- the output node A is coupled to a set of switches responsive to respective switch signals SWQ 1 to SWQ 8 in this example for an 8-bit output.
- the switches apply data signals Q 1 to Q 8 to corresponding latches in the output register 620 .
- the output register 620 stores the output vector VO(n) for the stage.
- the adjustable current source I 1 is operated in coordination with the switch signals SWQ 1 to SWQ 8 to sense a plurality of levels of the signal on DL(n) and store the resulting sensing results in the output register 620 .
- Data circuits are coupled to the output register 620 for transferring the output vector VO(n) on lines 630 to the page input circuit in the next stage as the input vector VI(n+1), or in feedback on lines 631 as the input vector VI(n) for the page input circuit PI(n) of the same stage through a multiplexer 642 , which is controlled by the signal SW_FB.
- the signal SW_FB is provided by a configuration logic/store such as configuration circuits 399 and configuration circuits 999 shown in FIGS. 3A and 3B , for page input circuits and page output circuits operatively coupled with each block B(n).
- FIG. 7 illustrates an example digital-to-analog converter 700 which can be utilized in a page input circuit like that shown in FIG. 5 .
- a chunk of 2 bits D 1 and D 2 of the input vector register 705 are applied as input to a 4-to-1 multiplexer 704 .
- a reference voltage generating circuit includes an operational amplifier 701 which receives a reference voltage on its “+” input and a feedback signal on its “ ⁇ ” input.
- the output of the operational amplifier 701 drives the gate of a PMOS transistor 702 , controlling current flow through a string of resistors 703 . Nodes between the resistors provide output voltage levels O 1 to O 4 as inputs to the multiplexer 704 .
- the reference voltage is applied to the “ ⁇ ” input of the operational amplifier 710 , the output of which drives the gate of an NMOS transistor 711 connected between the supply potential and a resistor 712 .
- the resistor is connected to ground.
- the node on the resistor 712 is applied as feedback to the “+” input of the operational amplifier 710 .
- the operational amplifier 710 in this example is configured in a unity gain mode, and its output is applied as one of the signals X[i] used to bias a bit line in the block B(n).
- FIG. 8 illustrates an alternative embodiment of a page output circuit, representing page output circuit PO(n) which is coupled to the block B(n) and to the page input circuit PI(n).
- the page output circuit PO(n) is configured to receive the signal on data line DL(n) from the YPASS circuits coupled to the block B(n).
- the data line DL(n) is connected to a sense amplifier 820 or other type of sensing circuit or analog-to-digital converter.
- the sense amplifier 820 in this example includes an adjustable current source I 1 which is connected to node 810 , which is also connected to data line DL(n).
- a precharge transistor 812 is coupled to a sensing node SEN.
- a capacitor 811 is coupled to the sensing node SEN.
- Transistor M 2 is connected in series between node 810 and the sensing node SEN.
- the sensing node SEN is connected to the gate of transistor M 1 .
- Transistor M 1 is connected between ground and output node A.
- the output node A is coupled to a set of switches including two members in this example, responsive to respective switch signals SWQ 1 to SWQ 2 in this example for an 2-bit output.
- the switches apply data signals Q 1 to Q 2 to corresponding latches (e.g. 821 ) in the output register 802 .
- the output register 802 stores a chunk of output vector VO(n) for the stage.
- the adjustable current source I 1 is operated in coordination with the switch signals SWQ 1 to SWQ 2 to sense a plurality of levels of the signal on DL(n) and store the resulting sensing results in the output register 802 in each sensing cycle. For example, in each sensing cycle, the sum of the signals from the set of memory cells on one selected word line in the selected set of bit lines can be converted to a 2-bit chunk.
- the output of the output register 802 is applied to a layer interface block 803 .
- the output vector VO(n) can be generated using bit lines on a plurality of layers of a 3D memory.
- the interface block 803 provides the chunks of data generated by the sense amplifier 800 , which are combined with corresponding chunks generated by interface blocks on other layers of the 3D memory to form an output vector VO(n) having Z chunks of M bits.
- the page output circuit PO(n) is described as comprising chunk-wide portions on a plurality of layers of a 3D memory.
- the page input circuit PI(n) can be configured in a similar way for connection to a block B(n) coupled to memory cells on the plurality of layers of the 3D memory.
- the interface block 803 can be used to accumulate additional chunks of the output vector in a sequence of sensing cycles using memory cells on a set of word lines in a single layer or single 2D array.
- Data circuits are coupled to the interface block 803 for transferring the output vector VO(n) on lines 830 to the page input circuit in the next stage as the input vector VI(n+1), or in feedback on lines 831 as the input vector VI(n) for the page input circuit PI(n) of the same stage through a multiplexer 842 , which is controlled by the signal SW_FB.
- the signal SW_FB is provided by a configuration circuit such as configuration circuits 399 and configuration circuits 999 shown in FIGS. 3A and 3B , for page input circuits and page output circuits operatively coupled with each block B(n).
- An in-memory computing structure in which a sum-of-products result latched in a page output circuit can feedback as an input vector for the page input circuit in the same block, or pass as an input vector to the page input circuit in the next page.
- the page input vector in each block can be coupled to a bus system that provides input vectors from other sources on the bus.
- the input data can be latched from outside the in-memory computing device, from a feedback loop of the previous state output from the same block, or from the output data in a previous block.
- a large bandwidth with very little routing capacitance can be achieved. Connection between two blocks can be very short, so that the previous layer or previous page output can be delivered to the next layer, or next block, fast and with low power.
- the structure can be applied to artificial intelligence processors use for both training and inference steps in deep learning.
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Neurology (AREA)
- Theoretical Computer Science (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Computer Hardware Design (AREA)
- Read Only Memory (AREA)
Abstract
Description
- The present invention relates to in-memory computing devices, and more particularly, to in-memory computing devices supporting efficient data sharing among multiple computational stages.
- A neural network is an information processing paradigm that is inspired by the way biological nervous systems process information. With the availability of large training datasets and sophisticated learning algorithms, neural networks have facilitated major advances in numerous domains such as computer vision, speech recognition, and natural language processing.
- The basic unit of computation in a neural network is often referred to as a neuron. A neuron receives inputs from other neurons, or from an external source, performs an operation, and provides an output.
FIG. 1 illustrates an exampleneural network 100. Theneural network 100 contains multiple neurons arranged in layers, which can be considered to be an example of a computational stage including many parallel operations. Theneural network 100 includes aninput layer 102 of input neurons (i.e., neurons that provide the input data), threehidden layers synaptic layer 112 connects neurons in theinput layer 102 and thehidden layer 106, thesynaptic layer 114 connects neurons in thehidden layers synaptic layer 116 connects neurons in thehidden layers synaptic layer 118 connects the neurons in thehidden layer 110 and the output layer 104. All these connections have weights associated with them. For example, theneurons hidden layer 106 are connected to aneuron 128 in thehidden layer 108 by connections withweights w 1 132, w2 134 and w3 136, respectively. The output for theneuron 128 in thehidden layer 108 can be calculated as a function of the inputs (x1, x2, and x3) from theneurons hidden layer 106 and theweights w 1 132, w2 134 and w3 136 in the connections. The function can be expressed as follows: -
- In the sum-of-products expression above, each product term is a product of a variable input x1 and a weight w1. The weight w1 can vary among the terms, corresponding for example to coefficients of the variable inputs x1. Similarly, outputs from the other neurons in the hidden layer can also be calculated. The outputs of the two neurons in the
hidden layer 110 act as inputs to the output neuron in the output layer 104. - Neural networks are used to learn patterns that best represent a large set of data. The hidden layers closer to the input layer in general learn high level generic patterns, and the hidden layers closer to the output layer in general learn more data-specific patterns. Training is a phase in which a neural network learns from training data. During training, the connections in the synaptic layers are assigned weights based on the results of the training session. Inference is a stage in which a trained neural network is used to infer/predict using input data and to produce output data based on the prediction.
- In-memory computing is an approach in which memory cells, organized in an in-memory computing device, can be used for both data processing and memory storage. A neural network can be implemented using an in-memory computing device for a number of synaptic layers. The weights for the sum-of-products function can be stored in memory cells of the in-memory computing device. The sum-of-products function can be realized as a circuit operation in the in-memory computing device in which the electrical characteristics of the memory cells of the array effectuate the function.
- An engineering issue associated with neural networks relates to movement of data among the synaptic layers. In some embodiments, there can be thousands of neurons in each layer, and the routing of the outputs of the neurons to the inputs of other neurons can be a time consuming aspect of the execution of the neural network.
- It is desirable to provide in-memory neural network technology that supports efficient movement of data among the computational components of the system.
- An in-memory computation device is described that comprises a memory configured in a plurality of blocks useable for in-memory computations. Blocks B(n), n going from 0 to N−1, in the plurality of blocks have corresponding page input circuits PI(n) and page output circuits PO(n) that are operatively coupled to sets of bit lines in the blocks. For example, each block B(n) can include a set S(n) of bit lines coupled to its corresponding page output circuit and to its corresponding page input circuit. The device includes in some embodiments a data bus system for providing an external source of input data and a destination for output data. Data circuits are configurable to connect page input circuit PI(n) to one or more of page output circuit PO(n), page output circuit PO(n−1), and the data bus system to source input data for a sensing cycle. This configuration can be done between each sensing cycle, or in longer intervals, in order to support a variety of neural network configurations and operations.
- In a device described herein, a plurality of bit line bias circuits is connected to bit lines in the plurality of blocks. Bit line bias circuit Y(n) in the plurality of bit line bias circuits being operatively coupled to block B(n), and to page input circuit PI(n) in the plurality of page input circuits. The bit line biasing circuit can bias the bit lines in block B(n) in response to input voltages generated by page input circuit PI(n).
- Other aspects and advantages of the present invention can be seen on review of the drawings, the detailed description and the claims, which follow.
-
FIG. 1 illustrates an example neural network. -
FIG. 2 illustrates a NOR style memory array configured for in-memory computation of a sum-of-products suitable for use in a neural network. -
FIG. 3A is a simplified diagram of an in-memory computation device with inter-page and intra-page data circuits. -
FIG. 3B is a simplified diagram of an alternative in-memory computation device with inter-page and intra-page data circuits. -
FIG. 4 is a simplified diagram of an in-memory computation device using a NOR style memory array with drain side bias generated in response to input data. -
FIG. 5 is a diagram of an embodiment of page input circuitry usable in a device like that ofFIG. 4 . -
FIG. 6 is a diagram of page output circuitry usable in a device like that ofFIG. 4 . -
FIG. 7 is a diagram of a digital-to-analog converter usable to generate drain side bias voltages for use in an embodiment like that ofFIG. 4 . -
FIG. 8 is a diagram of an alternative embodiment of page output circuitry usable in a device like that ofFIG. 4 . - A detailed description of embodiments of the present invention is provided with reference to the
FIGS. 2-8 . -
FIG. 2 illustrates amemory array 210 including a plurality of memory cells (e.g. 211, 212) configured in a NOR style. The memory cells in a given column of the array are connected between a bit line and a source line. In the illustrated example, there are columns of memory cells between bit line BL0 and source line SL0, between bit line BL1 and source line SL1, between bit line BL2 and source line SL2, and between bit line BL3 and source line SL3. To execute a sum-of-products operation in-memory, a coefficient vector (a set of coefficients W(i)) is stored in the memory cells in a given row, such as on the row of cells connected to word line WL0. Input data (X0-X3) is applied by biasing the bit lines BL0 to BL3 connecting a bias voltage to the drain side of the memory cells. A row of memory cells is selected by applying a word line signal to a selected word line (e.g. WL0). Unselected word lines are biased in an off condition. Current is generated in the cells in the selected row that is a function of the coefficient W(i) stored in the accessed memory cell, and the input data X(i). The signals generated on the bit lines or source lines in response to the word line signal on the selected word line can be combined (215) to generate the sum-of-products output SUM by sensing a sum of the signals. - The memory cells in the array of
FIG. 2 can be nonvolatile memory cells, such as floating gate memory cells or dielectric charge trapping memory cells. Also, programmable resistance memory cells can be utilized, such as phase change memory, metal oxide memory and others. In some embodiments, the memory cells in the array ofFIG. 2 can be volatile memory cells. The memory cells in the array ofFIG. 2 can store one bit per cell in some examples. In other examples, the memory cells in the array ofFIG. 2 can store multiple bits per cell. In yet other examples, the memory cells in the array ofFIG. 2 can store analog values. - The NOR style array can be implemented with many bit lines and word lines, storing thousands or millions of bits. The NOR style array of
FIG. 2 can be implemented in a three-dimensional configuration, with many two-dimensional arrays stacked in many levels. - In-memory computation circuits as described herein can use memory arrays of other styles in some examples, including for example AND style arrays or NAND style arrays.
-
FIG. 3A is a simplified diagram of an in-memory computation device that includes amemory 300 that comprises an array of memory cells storing coefficients W(i), and including a plurality of word lines and a plurality of bit lines. The memory is configured in a plurality of blocks, B(0), B(1), and so on. Each block includes a plurality of bit lines and a plurality of word lines. The word lines can be shared among more than one block in some embodiments, so that he word lines for more than one block are driven and decoded together, or constitute a shared single word line conductor across more than one block. In other embodiments, the word lines are driven and decoded for individual blocks. - In this example, the blocks B(n) in the plurality of blocks comprise corresponding
YPASS circuits - Also, the YPASS circuits include a plurality of bit line bias circuits (not shown in
FIG. 3A ) connected to the plurality of sets of bit lines. Bit line bias circuit Y(n) in the plurality of bit line bias circuits is operatively coupled to set S(n) of bit lines in the corresponding block B(n), and to page input circuit PI(n) in the plurality of page input circuits. The bit line bias circuit biases the bit lines in set S(n) in response to input voltages generated by page input circuit PI(n). The input signals X[1:z] online 360 are input voltages in this example, applied toYPASS circuit 301, and X[z+1:2z] online 361 are applied toYPASS circuit 302, to be applied to the corresponding set of bit lines having “z” members. In general, there can be a large number of blocks including YPASS circuits, each of which includes N sets S(n) of bit lines, where a YPASS circuit in block (n) including set S(n) bit lines receives input signals, X[n*z+1:(n+1)*z)], for n going from 0 toN− 1. - In the illustrated embodiment, two blocks, B(0) and B(1), are illustrated. In general, there can be a plurality of blocks B(n), where “n” ranges from 0 to N−1, and N can be any positive integer more than 1. In some embodiments, the number N can be 8 for example, or 16. In other embodiments, N can be in much higher.
- The
memory 300 can comprise a single two-dimensional array, in which the blocks are arranged side-by-side in sequence. In other embodiments, thememory 300 can comprise a three-dimensional array, including a plurality of stacked two-dimensional arrays. In this case, each two-dimensional array in the stack can comprise one block. Blocks arranged in sequence can reside on sequential levels of the stack. In other arrangements, each two-dimensional array in the stack can comprise more than one block configured as illustrated. - A data bus system, represented by the input data block 320 and the input data block 321, is provided on the device. The data bus system can be used as an input/output interface for data from an external data source, or from other circuitry on the device that generates data for use as the input vectors in some configurations. The input data block 320 and the input data block 321 can comprise, for examples, a data cache memory, a register or a latch coupled to input/output circuitry on the device or other data bus circuits.
- The output signals on lines DL0 and DL1 from the
YPASS circuits lines Word line drivers 398 and decoders are included to provide the word line signals on selected word lines. In the simplified example shown inFIG. 3A ,YPASS circuits lines YPASS circuits - In this illustration, the page input circuits (PI(0) and PI(1)) include respective input registers 314, 316 which store input data in the form of input vectors VI(n) received as input from data circuits on the device for the corresponding block B(n). The page input circuits can include circuits that convert the input vector VI(n), which have a number of bits equal to a multiple M of the number Z of bit lines in the block, to the input signals. Also, the page output circuits (PO(0) and PO(1)) include respective output registers 315, 317 which store output vectors VO(n) generated in response to an in-memory computation using memory cells in the corresponding block B(n) for output to the data circuits on the device. The output vectors VO(n) can have the same number of bits as the input vectors VI(n), or a different number of bits.
- Data circuits on the device are configurable to interconnect the input register for a given page (e.g., PI(1)) singly or in any combination, with sources of input data, including the output register of a previous page (e.g., PO(0)), the output register of the same page (e.g., PO(1) as feedback) and the data bus to source input data for a given sensing cycle. An input vector including the input data applied to the input register in a given sensing cycle can be sourced by a single source, or a combination of multiple sources. The data circuits are configurable to transfer an output vector VO(0) of page output circuit PO(0) as all or part an input vector VI(1) to the next page input circuit PI(1) as represented by the
line 340. Also, data circuits are configurable to feed back the output vector VO(0) of the page output circuit PO(0) as all or part of an input vector VI(0) for the page input circuit PI(0), as represented byline 350. In addition, the data circuits are configurable to connect the page input circuit PI(0) to the data bus system including input data block 320 as represented byline 330, to receive all or part of an input vector VI(0) from another source in the bus system. - Likewise, the data circuits on the device are configurable to transfer an output vector VO(1) of page output circuit PO(1) as all or part of an input vector VI(2) for the next page input circuit as represented by the
line 341. Also, data circuits are configurable to feedback the output VO(1) of the page output circuit PO(1) as all or part of an input vector VI(1) for the page input circuit PI(1), as represented byline 351. In addition, the data circuits are configurable to connect the page input circuit PI(1) to the data bus system including input data block 321 as represented byline 331, to receive all or part of an input vector VI(1) from an another source on the bus system. -
Configuration circuits 399 are included on the device. Thecircuits 399 include logic, configuration parameters storage, or both. Theconfiguration circuits 399 control the configuration of the data circuits for the routing of input vectors and output vectors among the page input and page output circuits. The configuration of the data circuits can be set dynamically for each sensing cycle as suits the needs of a particular implementation, using timed control signals delivered to switches in the data circuits. Alternatively, the configuration circuit can use volatile or nonvolatile configuration registers to set up the data circuits. - Control circuits, not shown, are coupled to the circuit of
FIG. 3A , and include logic to control execution of in-memory computation operations, data circuit configuration operations in coordination withconfiguration circuits 399, and memory operations like program and erase. The control circuits can comprise command decoders, configuration registers, state machines and timing circuits set up to control the operations. The control circuits can include voltage regulators, charge pumps and other bias circuits to provide appropriate bias voltages in support of the memory operations. -
FIG. 3B is a simplified diagram of an in-memory computation device that includes amemory 900 that comprises an array of memory cells storing coefficients W(i), and including a plurality of word lines and a plurality of bit lines. The memory is configured in a plurality of blocks, B(0), B(1), and so on. Each block includes a plurality of bit lines and a plurality of word lines. The word lines are driven and decoded for individual blocks in this example. - In this example, the blocks B(n) in the plurality of blocks comprise corresponding
YPASS circuits - Also, the word line drivers include a plurality of bias circuits (not shown in
FIG. 3B ) connected to the plurality of sets of word lines. Word line driver WD(n) in the plurality of word line drivers is operatively coupled to a set of word lines in the corresponding block B(n), and to page input circuit PI(n) in the plurality of page input circuits. The word line bias circuit biases the word lines in corresponding block in response to input voltages generated by page input circuit PI(n). The input signals X[1:z] online 960 are input voltages in this example, applied to aword line driver 938 for block B(0), and X[z+1:2z] online 961 are applied to a word line driver for block B(1), to be applied to the corresponding set of word lines having “z” members. - In the illustrated embodiment, two blocks, B(0) and B(1), are illustrated. In general, there can be a plurality of blocks B(n), where “n” ranges from 0 to N−1, and N can be any positive integer more than 1. In some embodiments, the number N can be 8 for example, or 16. In other embodiments, N can be in much higher.
- The
memory 900 can comprise a single two-dimensional array, in which the blocks are arranged side-by-side in sequence. In other embodiments, thememory 900 can comprise a three-dimensional array, including a plurality of stacked two-dimensional arrays. In this case, each two-dimensional array in the stack can comprise one block. Blocks arranged in sequence can reside on sequential levels of the stack. In other arrangements, each two-dimensional array in the stack can comprise more than one block configured as illustrated. - A data bus system, represented by the input data block 920 and the input data block 921, is provided on the device. The data bus system can be used as an input/output interface for data from an external data source, or from other circuitry on the device that generates data for use as the input vectors in some configurations. The input data block 920 and the input data block 921 can comprise, for examples, a data cache memory, a register or a latch coupled to input/output circuitry on the device or other data bus circuits.
- The output signals on lines DL0 and DL1 from the
YPASS circuits lines FIG. 3B , the word line drivers (e.g. 938) for the blocks B(0) and B(1) are coupled to corresponding page input circuits PI(0), PI(1) which provide input signals X[n*z+1:(n+1)*z)] onlines YPASS circuits - In this illustration, the page input circuits (PI(0) and PI(1)) include respective input registers 914, 916 which store input data in the form of input vectors VI(n) received as input from data circuits on the device for the corresponding block B(n). The page input circuits can include circuits that convert the input vector VI(n), which have a number of bits equal to a multiple M of the number Z of bit lines in the block, to the input signals. Also, the page output circuits (PO(0) and PO(1)) include respective output registers 915, 917 which store output vectors VO(n) generated in response to an in-memory computation using memory cells in the corresponding block B(n) for output to the data circuits on the device. The output vectors VO(n) can have the same number of bits as the input vectors VI(n), or a different number of bits.
- Data circuits on the device are configurable to interconnect the input register for a given page (e.g., PI(1)) singly or in any combination, with sources of input data, including the output register of a previous page (e.g., PO(0)), the output register of the same page (e.g., PO(1) as feedback) and the data bus to source input data for a given sensing cycle. An input vector including the input data applied to the input register in a given sensing cycle can be sourced by a single source, or a combination of multiple sources. The data circuits are configurable to transfer an output vector VO(0) of page output circuit PO(0) as all or part an input vector VI(1) to the next page input circuit PI(1) as represented by the
line 940. Also, data circuits are configurable to feed back the output vector VO(0) of the page output circuit PO(0) as all or part of an input vector VI(0) for the page input circuit PI(0), as represented byline 950. In addition, the data circuits are configurable to connect the page input circuit PI(0) to the data bus system including input data block 920 as represented byline 930, to receive all or part of an input vector VI(0) from another source in the bus system. - Likewise, the data circuits on the device are configurable to transfer an output vector VO(1) of page output circuit PO(1) as all or part of an input vector VI(2) for the next page input circuit as represented by the
line 941. Also, data circuits are configurable to feedback the output VO(1) of the page output circuit PO(1) as all or part of an input vector VI(1) for the page input circuit PI(1), as represented byline 951. In addition, the data circuits are configurable to connect the page input circuit PI(1) to the data bus system including input data block 921 as represented byline 931, to receive all or part of an input vector VI(1) from an another source on the bus system. -
Configuration circuits 999 are included on the device. Thecircuits 999 include logic, configuration parameters storage, or both. Theconfiguration circuits 999 control the configuration of the data circuits for the routing of input vectors and output vectors among the page input and page output circuits. The configuration of the data circuits can be set dynamically for each sensing cycle as suits the needs of a particular implementation, using timed control signals delivered to switches in the data circuits. Alternatively, the configuration circuit can use volatile or nonvolatile configuration registers to set up the data circuits. - Control circuits, not shown, are coupled to the circuit of
FIG. 3B , and include logic to control execution of in-memory computation operations, data circuit configuration operations in coordination withconfiguration circuits 999, and memory operations like program and erase. The control circuits can comprise command decoders, configuration registers, state machines and timing circuits set up to control the operations. The control circuits can include voltage regulators, charge pumps and other bias circuits to provide appropriate bias voltages in support of the memory operations. -
FIG. 4 is a diagram of an in-memory computation device like that ofFIG. 3A with information about embodiments of theYPASS circuits memory array 410 along with supporting circuitry as mentioned with respect toFIG. 3A . Thememory array 410 includes a plurality of blocks B(0), B(1), . . . A set S(n) of bit lines in each block is coupled to acorresponding YPASS circuit - The
YPASS circuit 401 includes a bit line bias circuit. In this embodiment, the bit line bias circuit includes a plurality of clamp transistors 490-493 having one source/drain terminal coupled to a bit line in the block B(0) of thearray 410, and another source/drain terminal coupled to a summing node on the data line DL0. The gates of the clamp transistors 490-493 are connected to corresponding input signals in the set of input signals X[1:z] provided by thepage input circuit 419. - The
YPASS circuit 402 includes a bit line bias circuit. In this embodiment, the bit line bias circuit includes a plurality of clamp transistors 494-497, each having one source/drain terminal coupled to a bit line in the block B(1) of thearray 410, and another source/drain terminal coupled to a summing node on the data line DL1. The gates of the clamp transistors 494-497 are connected to corresponding input signals in the set of input signals X[z+1:2z] provided by thepage input circuit 459. - In the example shown in
FIG. 4 , the page input circuit 419 (PI(0)) includes aninput register 420 and supportinglogic 421, which converts the input vector stored in theinput register 420 to the input signals X[1:z] applied to the gates of the clamp transistors. Likewise, the page input circuit 459 (PI(1)) includes aninput register 460 and supportinglogic 461, which converts the input vector stored in theinput register 460 to the input signals X[z+1:2z] applied to the gates of the clamp transistors. - In the example shown in
FIG. 4 , the page output circuit 429 (PO(0)) includes an N-bit sense amplifier 430, or other type of multilevel sensing circuit or analog-to-digital converter. The output of thesense amplifier 430 is coupled to anoutput register 431 or other type of register. Theoutput register 431 is connected to acompute unit 432 which can accumulate or otherwise process data stored in theoutput register 431. In combination, the circuits in thepage output circuit 429 convert the signal on the DL0 line into an output vector VO(n) for the computation based on a sum-of-signals generated by memory cells on a selected word line on the block B(n). The sum-of-signals can correspond to a sum-of-products. - Likewise, the page output circuit 469 (PO(1)) includes an N-
bit sense amplifier 470, or other type of multilevel sensing circuit or analog-to-digital converter. The output of thesense amplifier 470 is coupled to anoutput register 471 or other type of register. Theoutput register 471 in this example is connected to acompute unit 472 which can accumulate or otherwise process data stored in theoutput register 471. In combination, the circuits in thepage output circuit 469 convert the signal on the DL1 line into an output vector VO(n) for the computation based on a sum-of-signals generated by memory cells on a selected word line on the block B(n). - In this example, the cell current in each memory cell on the selected word line coupled to a bit line in the set of bit lines can be represented in one example memory system by the equation:
-
- Data circuits on the device are configurable to transfer an output vector VO(0) of
page output circuit 429 as an input vector VI(1) for the nextpage input circuit 459 as represented by theline 443. Also, data circuits are configurable to feedback the output VO(0) of thepage output circuit 429 as an input vector VI(0) for thepage input circuit 419 as represented byline 442. In addition, the data circuits are configurable to connect thepage input circuit 419 to thedata bus system 440 as represented byline 441, to receive an input vector VI(0) from an another source on the bus system. - Likewise, data circuits on the device are configurable to transfer an output vector VO(1) of
page output circuit 469 as an input vector VI(1) for the next page input circuit as represented by theline 483. Also, data circuits are configurable to feedback the output VO(1) of thepage output circuit 469 as an input vector VI(1) for thepage input circuit 459 as represented byline 482. In addition, the data circuits are configurable to connect thepage input circuit 459 to thedata bus system 480 as represented byline 481 to receive an input vector VI(1) from an another source on the bus system. -
FIG. 5 illustrates one example of apage input circuit 500, representing page input circuit PI(n), which is configured to deliver the input signals onbus 505 to the YPASS circuits of the block B(n) in the memory array. Also, portions of the data circuits which are configured to deliver the input vector VI(n) are illustrated. - In this example, the
page input circuit 500 includes aninput register 501 that comprises a plurality of latches 520-523. - Data circuits include a set of
switches switches register 501 to the previous page output circuit which provides output vector VO(n−1) onlines 508 of the data circuits for use as input vector VI(n). Also, the data circuits include amultiplexer 551 andinput bus register 550. Themultiplexer 551 is responsive to the signal SW_IN to connect and disconnect the input data from theinput bus register 550 to deliver an input vector VI(n) to theinput register 501. Also the data circuits include abus 509 which is coupled to a multiplexer (see,MUX 642 inFIG. 6 ), to receive the output vector VO(n) as feedback from the page output circuit PO(n) coupled to the same block B(n), for use as input vector VI(n). The signal SW and the signal SW_IN are provided by a configuration circuit such asconfiguration circuits 399 andconfiguration circuits 999 shown inFIGS. 3A and 3B . The signals SW and SW_IN are provided independently for the page input circuits and page output circuits operatively coupled with each block B(n). - In this example, for an embodiment in which block B(n) includes a number Z of bit lines, the input vector VI(n) includes Z chunks of M bits of data.
Input register 501 applies the Z chunks of M bits of input vector VI(n) toconversion circuits 502 that provide the bias voltages X5 to X8 in the Figure to establish the drain level of the selected memory cells in the corresponding block B(n). In this example, thecircuits 502 include one M bit digital-to-analog converter -
FIG. 6 illustrates one example of apage output circuit 600, representing page output circuit PO(n) which is coupled to the block B(n) and to the page input circuit PI(n). Thepage output circuit 600 is configured to receive the signal on data line DL(n) from the YPASS circuits coupled to the block B(n). - The data line DL(n) is connected to a
sense amplifier 601 or other type of sensing circuit or analog-to-digital converter. Thesense amplifier 601 in this example includes an adjustable current source I1 which is connected tonode 610, which is also connected to data line DL(n). Aprecharge transistor 612 is coupled to a sensing node SEN. Also acapacitor 611 is coupled to the sensing node SEN. Transistor M2 is connected in series betweennode 610 and the sensing node SEN. The sensing node SEN is connected to the gate of transistor M1. Transistor M1 is connected between ground and output node A. The output node A is coupled to a set of switches responsive to respective switch signals SWQ1 to SWQ8 in this example for an 8-bit output. The switches apply data signals Q1 to Q8 to corresponding latches in theoutput register 620. The output register 620 stores the output vector VO(n) for the stage. In operation, the adjustable current source I1 is operated in coordination with the switch signals SWQ1 to SWQ8 to sense a plurality of levels of the signal on DL(n) and store the resulting sensing results in theoutput register 620. - Data circuits are coupled to the
output register 620 for transferring the output vector VO(n) onlines 630 to the page input circuit in the next stage as the input vector VI(n+1), or in feedback onlines 631 as the input vector VI(n) for the page input circuit PI(n) of the same stage through amultiplexer 642, which is controlled by the signal SW_FB. The signal SW_FB is provided by a configuration logic/store such asconfiguration circuits 399 andconfiguration circuits 999 shown inFIGS. 3A and 3B , for page input circuits and page output circuits operatively coupled with each block B(n). -
FIG. 7 illustrates an example digital-to-analog converter 700 which can be utilized in a page input circuit like that shown inFIG. 5 . In this example, a chunk of 2 bits D1 and D2 of theinput vector register 705 are applied as input to a 4-to-1multiplexer 704. A reference voltage generating circuit includes anoperational amplifier 701 which receives a reference voltage on its “+” input and a feedback signal on its “−” input. The output of theoperational amplifier 701 drives the gate of aPMOS transistor 702, controlling current flow through a string ofresistors 703. Nodes between the resistors provide output voltage levels O1 to O4 as inputs to themultiplexer 704. In response to the chunk of input data, the reference voltage is applied to the “−” input of theoperational amplifier 710, the output of which drives the gate of anNMOS transistor 711 connected between the supply potential and aresistor 712. The resistor is connected to ground. The node on theresistor 712 is applied as feedback to the “+” input of theoperational amplifier 710. Theoperational amplifier 710 in this example is configured in a unity gain mode, and its output is applied as one of the signals X[i] used to bias a bit line in the block B(n). -
FIG. 8 illustrates an alternative embodiment of a page output circuit, representing page output circuit PO(n) which is coupled to the block B(n) and to the page input circuit PI(n). The page output circuit PO(n) is configured to receive the signal on data line DL(n) from the YPASS circuits coupled to the block B(n). - The data line DL(n) is connected to a
sense amplifier 820 or other type of sensing circuit or analog-to-digital converter. Thesense amplifier 820 in this example includes an adjustable current source I1 which is connected tonode 810, which is also connected to data line DL(n). Aprecharge transistor 812 is coupled to a sensing node SEN. Also acapacitor 811 is coupled to the sensing node SEN. Transistor M2 is connected in series betweennode 810 and the sensing node SEN. The sensing node SEN is connected to the gate of transistor M1. Transistor M1 is connected between ground and output node A. The output node A is coupled to a set of switches including two members in this example, responsive to respective switch signals SWQ1 to SWQ2 in this example for an 2-bit output. The switches apply data signals Q1 to Q2 to corresponding latches (e.g. 821) in theoutput register 802. The output register 802 stores a chunk of output vector VO(n) for the stage. In operation, the adjustable current source I1 is operated in coordination with the switch signals SWQ1 to SWQ2 to sense a plurality of levels of the signal on DL(n) and store the resulting sensing results in theoutput register 802 in each sensing cycle. For example, in each sensing cycle, the sum of the signals from the set of memory cells on one selected word line in the selected set of bit lines can be converted to a 2-bit chunk. - The output of the
output register 802 is applied to alayer interface block 803. In this example, the output vector VO(n) can be generated using bit lines on a plurality of layers of a 3D memory. In this example, theinterface block 803 provides the chunks of data generated by thesense amplifier 800, which are combined with corresponding chunks generated by interface blocks on other layers of the 3D memory to form an output vector VO(n) having Z chunks of M bits. - In this example, the page output circuit PO(n) is described as comprising chunk-wide portions on a plurality of layers of a 3D memory. The page input circuit PI(n) can be configured in a similar way for connection to a block B(n) coupled to memory cells on the plurality of layers of the 3D memory.
- In other embodiments, the
interface block 803 can be used to accumulate additional chunks of the output vector in a sequence of sensing cycles using memory cells on a set of word lines in a single layer or single 2D array. - Data circuits are coupled to the
interface block 803 for transferring the output vector VO(n) onlines 830 to the page input circuit in the next stage as the input vector VI(n+1), or in feedback onlines 831 as the input vector VI(n) for the page input circuit PI(n) of the same stage through amultiplexer 842, which is controlled by the signal SW_FB. The signal SW_FB is provided by a configuration circuit such asconfiguration circuits 399 andconfiguration circuits 999 shown inFIGS. 3A and 3B , for page input circuits and page output circuits operatively coupled with each block B(n). - An in-memory computing structure is described in which a sum-of-products result latched in a page output circuit can feedback as an input vector for the page input circuit in the same block, or pass as an input vector to the page input circuit in the next page. Also, the page input vector in each block can be coupled to a bus system that provides input vectors from other sources on the bus. Thus, the input data can be latched from outside the in-memory computing device, from a feedback loop of the previous state output from the same block, or from the output data in a previous block. With this scheme, a large bandwidth with very little routing capacitance can be achieved. Connection between two blocks can be very short, so that the previous layer or previous page output can be delivered to the next layer, or next block, fast and with low power. The structure can be applied to artificial intelligence processors use for both training and inference steps in deep learning.
- While the present invention is disclosed by reference to the preferred embodiments and examples detailed above, it is to be understood that these examples are intended in an illustrative rather than in a limiting sense. It is contemplated that modifications and combinations will readily occur to those skilled in the art, which modifications and combinations will be within the spirit of the invention and the scope of the following claims.
Claims (20)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/297,504 US10783963B1 (en) | 2019-03-08 | 2019-03-08 | In-memory computation device with inter-page and intra-page data circuits |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/297,504 US10783963B1 (en) | 2019-03-08 | 2019-03-08 | In-memory computation device with inter-page and intra-page data circuits |
Publications (2)
Publication Number | Publication Date |
---|---|
US20200286553A1 true US20200286553A1 (en) | 2020-09-10 |
US10783963B1 US10783963B1 (en) | 2020-09-22 |
Family
ID=72335450
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US16/297,504 Active US10783963B1 (en) | 2019-03-08 | 2019-03-08 | In-memory computation device with inter-page and intra-page data circuits |
Country Status (1)
Country | Link |
---|---|
US (1) | US10783963B1 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022232055A1 (en) * | 2021-04-25 | 2022-11-03 | University Of Southern California | Embedded matrix-vector multiplication exploiting passive gain via mosfet capacitor for machine learning application |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11636325B2 (en) | 2018-10-24 | 2023-04-25 | Macronix International Co., Ltd. | In-memory data pooling for machine learning |
US11562229B2 (en) | 2018-11-30 | 2023-01-24 | Macronix International Co., Ltd. | Convolution accelerator using in-memory computation |
US11934480B2 (en) | 2018-12-18 | 2024-03-19 | Macronix International Co., Ltd. | NAND block architecture for in-memory multiply-and-accumulate operations |
US11132176B2 (en) | 2019-03-20 | 2021-09-28 | Macronix International Co., Ltd. | Non-volatile computing method in flash memory |
TWI782288B (en) * | 2020-06-29 | 2022-11-01 | 大陸商珠海興芯存儲科技有限公司 | Pseudo-analog memory computing circuit |
Family Cites Families (70)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE2619663C3 (en) | 1976-05-04 | 1982-07-22 | Siemens AG, 1000 Berlin und 8000 München | Field effect transistor, method of its operation and use as a high-speed switch and in an integrated circuit |
US4987090A (en) | 1987-07-02 | 1991-01-22 | Integrated Device Technology, Inc. | Static ram cell with trench pull-down transistors and buried-layer ground plate |
JP3073645B2 (en) | 1993-12-27 | 2000-08-07 | 株式会社東芝 | Nonvolatile semiconductor memory device and method of operating the same |
US6107882A (en) | 1997-12-11 | 2000-08-22 | Lucent Technologies Inc. | Amplifier having improved common mode voltage range |
US6960499B2 (en) | 1998-02-24 | 2005-11-01 | Texas Instruments Incorporated | Dual-counterdoped channel field effect transistor and method |
US6313486B1 (en) | 2000-06-15 | 2001-11-06 | Board Of Regents, The University Of Texas System | Floating gate transistor having buried strained silicon germanium channel layer |
US6829598B2 (en) | 2000-10-02 | 2004-12-07 | Texas Instruments Incorporated | Method and apparatus for modeling a neural synapse function by utilizing a single conventional MOSFET |
US6703661B2 (en) | 2001-12-27 | 2004-03-09 | Ching-Yuan Wu | Contactless NOR-type memory array and its fabrication methods |
JP4620943B2 (en) | 2003-10-16 | 2011-01-26 | キヤノン株式会社 | Product-sum operation circuit and method thereof |
US7057216B2 (en) | 2003-10-31 | 2006-06-06 | International Business Machines Corporation | High mobility heterojunction complementary field effect transistors and methods thereof |
US6906940B1 (en) | 2004-02-12 | 2005-06-14 | Macronix International Co., Ltd. | Plane decoding method and device for three dimensional memories |
US20050287793A1 (en) | 2004-06-29 | 2005-12-29 | Micron Technology, Inc. | Diffusion barrier process for routing polysilicon contacts to a metallization layer |
US8058636B2 (en) | 2007-03-29 | 2011-11-15 | Panasonic Corporation | Variable resistance nonvolatile memory apparatus |
KR20090037690A (en) | 2007-10-12 | 2009-04-16 | 삼성전자주식회사 | Non-volatile memory device, method of operating the same and method of fabricating the same |
US8860124B2 (en) | 2009-01-15 | 2014-10-14 | Macronix International Co., Ltd. | Depletion-mode charge-trapping flash device |
JP5462490B2 (en) | 2009-01-19 | 2014-04-02 | 株式会社日立製作所 | Semiconductor memory device |
JP5317742B2 (en) | 2009-02-06 | 2013-10-16 | 株式会社東芝 | Semiconductor device |
US8203187B2 (en) | 2009-03-03 | 2012-06-19 | Macronix International Co., Ltd. | 3D memory array arranged for FN tunneling program and erase |
US9099181B2 (en) | 2009-08-19 | 2015-08-04 | Grandis, Inc. | Non-volatile static ram cell circuit and timing method |
JP2011065693A (en) | 2009-09-16 | 2011-03-31 | Toshiba Corp | Non-volatile semiconductor memory device |
US8275728B2 (en) | 2009-11-05 | 2012-09-25 | The United States Of America As Represented By The Secretary Of The Air Force | Neuromorphic computer |
US8311965B2 (en) | 2009-11-18 | 2012-11-13 | International Business Machines Corporation | Area efficient neuromorphic circuits using field effect transistors (FET) and variable resistance material |
CN107293322B (en) | 2010-02-07 | 2021-09-21 | 芝诺半导体有限公司 | Semiconductor memory device having permanent and non-permanent functions and including conductive floating body transistor, and method of operating the same |
US8331127B2 (en) | 2010-05-24 | 2012-12-11 | Macronix International Co., Ltd. | Nonvolatile memory device having a transistor connected in parallel with a resistance switching device |
US20110297912A1 (en) | 2010-06-08 | 2011-12-08 | George Samachisa | Non-Volatile Memory Having 3d Array of Read/Write Elements with Vertical Bit Lines and Laterally Aligned Active Elements and Methods Thereof |
WO2012015450A1 (en) | 2010-07-30 | 2012-02-02 | Hewlett-Packard Development Company, L.P. | Systems and methods for modeling binary synapses |
US20120044742A1 (en) | 2010-08-20 | 2012-02-23 | Micron Technology, Inc. | Variable resistance memory array architecture |
US8432719B2 (en) | 2011-01-18 | 2013-04-30 | Macronix International Co., Ltd. | Three-dimensional stacked and-type flash memory structure and methods of manufacturing and operating the same hydride |
US8630114B2 (en) | 2011-01-19 | 2014-01-14 | Macronix International Co., Ltd. | Memory architecture of 3D NOR array |
US8759895B2 (en) | 2011-02-25 | 2014-06-24 | Micron Technology, Inc. | Semiconductor charge storage apparatus and methods |
JP5722180B2 (en) | 2011-09-26 | 2015-05-20 | 株式会社日立製作所 | Nonvolatile memory device |
US9698185B2 (en) | 2011-10-13 | 2017-07-04 | Omnivision Technologies, Inc. | Partial buried channel transfer device for image sensors |
US9430735B1 (en) | 2012-02-23 | 2016-08-30 | Micron Technology, Inc. | Neural network in a memory device |
JP5998521B2 (en) | 2012-02-28 | 2016-09-28 | セイコーエプソン株式会社 | Nonvolatile semiconductor memory and method for manufacturing nonvolatile semiconductor memory |
US8981445B2 (en) | 2012-02-28 | 2015-03-17 | Texas Instruments Incorporated | Analog floating-gate memory with N-channel and P-channel MOS transistors |
JP2014053056A (en) | 2012-09-06 | 2014-03-20 | Toshiba Corp | Semiconductor storage device |
US9019771B2 (en) | 2012-10-26 | 2015-04-28 | Macronix International Co., Ltd. | Dielectric charge trapping memory cells with redundancy |
US20140149773A1 (en) | 2012-11-29 | 2014-05-29 | Agency For Science, Technology And Research | Latch circuit and data processing system |
KR20140113024A (en) | 2013-03-15 | 2014-09-24 | 에스케이하이닉스 주식회사 | Resistance variable Memory Device And Method of Driving The Same |
KR102179899B1 (en) | 2013-08-05 | 2020-11-18 | 삼성전자주식회사 | Neuromophic system and configuration method thereof |
CN105612617B (en) | 2013-10-11 | 2018-08-24 | 夏普株式会社 | Semiconductor device |
KR102139944B1 (en) | 2013-11-26 | 2020-08-03 | 삼성전자주식회사 | Three dimensional semiconductor device |
US9698156B2 (en) | 2015-03-03 | 2017-07-04 | Macronix International Co., Ltd. | Vertical thin-channel memory |
US9536969B2 (en) | 2014-09-23 | 2017-01-03 | Taiwan Semiconductor Manufacturing Co., Ltd. | Self-aligned split gate flash memory |
SG11201702791TA (en) | 2014-10-15 | 2017-05-30 | Agency Science Tech & Res | Flip-flop circuit, method of controlling a flip-flop circuit and memory device |
US9431099B2 (en) | 2014-11-11 | 2016-08-30 | Snu R&Db Foundation | Neuromorphic device with excitatory and inhibitory functionalities |
US9356074B1 (en) | 2014-11-17 | 2016-05-31 | Sandisk Technologies Inc. | Memory array having divided apart bit lines and partially divided bit line selector switches |
KR20160073847A (en) | 2014-12-17 | 2016-06-27 | 에스케이하이닉스 주식회사 | Electronic device and method for fabricating the same |
US9524980B2 (en) | 2015-03-03 | 2016-12-20 | Macronix International Co., Ltd. | U-shaped vertical thin-channel memory |
US9379129B1 (en) | 2015-04-13 | 2016-06-28 | Macronix International Co., Ltd. | Assist gate structures for three-dimensional (3D) vertical gate array memory structure |
KR20160122531A (en) | 2015-04-14 | 2016-10-24 | 에스케이하이닉스 주식회사 | Electronic device |
TWI580087B (en) | 2015-04-28 | 2017-04-21 | 旺宏電子股份有限公司 | Memory device and manufacturing method of the same |
US9934463B2 (en) | 2015-05-15 | 2018-04-03 | Arizona Board Of Regents On Behalf Of Arizona State University | Neuromorphic computational system(s) using resistive synaptic devices |
CN105718994B (en) | 2015-07-29 | 2019-02-19 | 上海磁宇信息科技有限公司 | Cellular array computing system |
KR101701250B1 (en) | 2015-08-03 | 2017-02-01 | 서울대학교산학협력단 | Multi-layered neuron array for deep belief network and neuron array operating method |
US9589982B1 (en) | 2015-09-15 | 2017-03-07 | Macronix International Co., Ltd. | Structure and method of operation for improved gate capacity for 3D NOR flash memory |
US9892800B2 (en) | 2015-09-30 | 2018-02-13 | Sunrise Memory Corporation | Multi-gate NOR flash thin-film transistor strings arranged in stacked horizontal active strips with vertical control gates |
US9842651B2 (en) | 2015-11-25 | 2017-12-12 | Sunrise Memory Corporation | Three-dimensional vertical NOR flash thin film transistor strings |
WO2017091338A1 (en) | 2015-11-25 | 2017-06-01 | Eli Harari | Three-dimensional vertical nor flash thin film transistor strings |
KR102382727B1 (en) | 2016-03-18 | 2022-04-08 | 가부시키가이샤 한도오따이 에네루기 켄큐쇼 | Semiconductor device and system using the same |
CN105789139B (en) | 2016-03-31 | 2018-08-28 | 上海新储集成电路有限公司 | A kind of preparation method of neural network chip |
JP2017195275A (en) | 2016-04-20 | 2017-10-26 | 東芝メモリ株式会社 | Semiconductor storage device and manufacturing method of the same |
KR102508532B1 (en) | 2016-05-02 | 2023-03-09 | 삼성전자주식회사 | Sense amplifier and a memory device comprising thereof |
CN106530210B (en) | 2016-10-31 | 2019-09-06 | 北京大学 | The device and method that parallel-convolution calculates are realized based on resistive memory array |
US10777566B2 (en) | 2017-11-10 | 2020-09-15 | Macronix International Co., Ltd. | 3D array arranged for memory and in-memory sum-of-products operations |
US10719296B2 (en) | 2018-01-17 | 2020-07-21 | Macronix International Co., Ltd. | Sum-of-products accelerator array |
US20190244662A1 (en) | 2018-02-02 | 2019-08-08 | Macronix International Co., Ltd. | Sum-of-products array for neuromorphic computing system |
US10242737B1 (en) * | 2018-02-13 | 2019-03-26 | Macronix International Co., Ltd. | Device structure for neuromorphic computing system |
US10635398B2 (en) | 2018-03-15 | 2020-04-28 | Macronix International Co., Ltd. | Voltage sensing type of matrix multiplication method for neuromorphic computing system |
JP7070190B2 (en) | 2018-07-18 | 2022-05-18 | 株式会社デンソー | Neural network circuit |
-
2019
- 2019-03-08 US US16/297,504 patent/US10783963B1/en active Active
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022232055A1 (en) * | 2021-04-25 | 2022-11-03 | University Of Southern California | Embedded matrix-vector multiplication exploiting passive gain via mosfet capacitor for machine learning application |
Also Published As
Publication number | Publication date |
---|---|
US10783963B1 (en) | 2020-09-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10783963B1 (en) | In-memory computation device with inter-page and intra-page data circuits | |
CN111722830B (en) | Nonvolatile computing method of flash memory | |
US10860682B2 (en) | Binary, ternary and bit serial compute-in-memory circuits | |
CN108009640B (en) | Training device and training method of neural network based on memristor | |
US11119674B2 (en) | Memory devices and methods for operating the same | |
US11157810B2 (en) | Resistive processing unit architecture with separate weight update and inference circuitry | |
KR102672586B1 (en) | Artificial neural network training method and device | |
US10453527B1 (en) | In-cell differential read-out circuitry for reading signed weight values in resistive processing unit architecture | |
US11138497B2 (en) | In-memory computing devices for neural networks | |
US11797833B2 (en) | Competitive machine learning accuracy on neuromorphic arrays with non-ideal non-volatile memory devices | |
CN112734019A (en) | Neuromorphic packaging device and neuromorphic computing system | |
US20210064974A1 (en) | Formation failure resilient neuromorphic device | |
CN111859261A (en) | Computing circuit and operating method thereof | |
Doevenspeck et al. | Noise tolerant ternary weight deep neural networks for analog in-memory inference | |
KR102511526B1 (en) | Hardware-based artificial neural network device | |
US11120864B2 (en) | Capacitive processing unit | |
CN111243648A (en) | Flash memory unit, flash memory module and flash memory chip | |
JP7480391B2 (en) | Storage for in-memory computing | |
US20230053948A1 (en) | Apparatus and method with in-memory computing | |
CN117409830A (en) | In-memory computing memory device and solid state drive module | |
JPH0683622A (en) | Information processor | |
CN117559980A (en) | Delay buffer unit, delay buffer array, electronic device and operation method | |
CN118138026A (en) | Delay buffer unit, electronic device, delay buffer array and operation method thereof |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: MACRONIX INTERNATIONAL CO., LTD., TAIWAN Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:HUNG, CHUN-HSIUNG;YANG, SHANG-CHI;REEL/FRAME:048552/0935 Effective date: 20190305 |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
CC | Certificate of correction | ||
MAFP | Maintenance fee payment |
Free format text: PAYMENT OF MAINTENANCE FEE, 4TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1551); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY Year of fee payment: 4 |