Method and apparatus for data refresh of simulated non-volatile memory in deep learning neural network

By combining CMOS technology and nonvolatile memory arrays in artificial neural networks, the data drift in vector-matrix multiplication arrays is solved, and efficient data refresh and precision are achieved , suitable for high-performance information processing.

CN111837189BActive Publication Date: 2024-11-08SILICON STORAGE TECHNOLOGY INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201980018504.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2018-05-25
Filing Date
2019-01-16
Publication Date
2024-11-08
Estimated Expiration
2039-01-16

AI Technical Summary

Technical Problem

In artificial neural networks, the flash memory cells in the vector-matrix multiplication array drift due to temperature changes and process variations, causing data to become incorrect over time, making it difficult for the prior art to effectively detect and refresh the data to its original state.

Method used

Using the combination of CMOS technology and nonvolatile memory arrays, the charge amount of memory cells is maintained by detecting data drift and performing data refresh, and the specific voltage operation of floating gates and word line terminals is used to maintain the charge amount of memory cells, realizing independent programming and erasing of memory cells. Discrete and read ensure that each unit in the memory array can be tuned and fine-tuned accurately.

Benefits of technology

Effectively detect and correct data drift, ensure that the data in the memory array can still maintain accuracy after a long time, improve the computing efficiency and energy efficiency of artificial neural networks, and is suitable for high-performance information processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111837189B_ABST
    Figure CN111837189B_ABST
Patent Text Reader

Abstract

A number of embodiments of a data refresh method and apparatus for a vector-matrix multiplication (VMM) array in an artificial neural network are disclosed. Various embodiments of a data drift detector suitable for detecting data drift in flash memory cells within a VMM array are disclosed.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Priority declaration

[0002] This application claims priority to U.S. Provisional Patent Application No. 62 / 642,867, filed on March 14, 2018, entitled “Method and Apparatus for Data Refresh for Analog Neuromorphic Memory in Artificial Neural Network,” and U.S. Patent Application No. 15 / 990,220, filed on May 25, 2018, entitled “Method and Apparatus for Data Refresh for Analog Non-volatile Memory in Deep Learning Neural Network.” Technical Field

[0003] Many embodiments of a method and apparatus for refreshing data of flash memory cells in a vector-matrix multiplication (VMM) array for use in an artificial neural network are disclosed. Background Art

[0004] Artificial neural networks simulate biological neural networks (the central nervous system of animals, especially the brain), and are used to estimate or approximate functions that may depend on a large number of inputs and are usually unknown. Artificial neural networks typically consist of layers of interconnected "neurons" that exchange messages with each other.

[0005] Figure 1 An artificial neural network is shown, where circles represent inputs or layers of neurons. Connections (called synapses) are represented by arrows and have numerical weights that can be adjusted based on experience. This allows the neural network to adapt to the input and be able to learn. Typically, a neural network includes multiple layers of inputs. There are typically one or more intermediate layers of neurons, and an output layer of neurons that provide the output of the neural network. Neurons at each level make decisions based on data received from the synapses, either individually or collectively.

[0006] One of the main challenges in developing artificial neural networks for high-performance information processing is the lack of adequate hardware technology. In fact, practical neural networks rely on a large number of synapses to achieve high connectivity between neurons, i.e. very high computational parallelism. In principle, such complexity can be achieved by digital supercomputers or clusters of dedicated graphics processing units. However, in addition to being high cost, these approaches are also mediocre in energy efficiency compared to biological networks, which consume less energy mainly due to the low-precision analog calculations they perform. CMOS analog circuits have been used for artificial neural networks, but given the large number of neurons and synapses, the synapses of most CMOS implementations are too large.

[0007] Applicant previously disclosed an artificial (simulated) neural network utilizing one or more non-volatile memory arrays as synapses in U.S. Patent Application 15 / 594,439, which is incorporated herein by reference. The non-volatile memory array operates as an simulated neuromorphic memory. The neural network device includes a first plurality of synapses configured to receive a first plurality of inputs and generate a first plurality of outputs therefrom, and a first plurality of neurons configured to receive the first plurality of outputs. The first plurality of synapses includes a plurality of memory cells, wherein each of the memory cells includes: a spaced-apart source region and a drain region formed in a semiconductor substrate, wherein a channel region extends between the source region and the drain region; a floating gate disposed over a first portion of the channel region and insulated from the first portion; and a non-floating gate disposed over a second portion of the channel region and insulated from the second portion. Each of the plurality of memory cells is configured to store a weight value corresponding to a plurality of electrons on the floating gate. The plurality of memory cells are configured to multiply the first plurality of inputs by the stored weight values ​​to generate a first plurality of outputs.

[0008] Each nonvolatile memory cell used in an analog neuromorphic memory system must be erased and programmed to hold a very specific and precise amount of charge in the floating gate. For example, each floating gate must hold one of N different values, where N is the number of different weights that can be indicated by each cell. Examples of N include 16, 32, and 64.

[0009] Conventional flash memory devices do not need to refresh their data because each flash memory cell can hold its value indefinitely (usually limited to two possible levels, or sometimes limited to four possible levels, i.e., N=2 or 4). However, in the flash memory cells used in VNM, since N is significantly larger, there is an increased likelihood that the data will become incorrect (i.e., corrupted) over time due to data drift that naturally occurs over time based on temperature changes and process variations.

[0010] What is needed is a mechanism for detecting data drift in a flash memory cell in a VNM and refreshing the data in the cell to its original desired state. Summary of the invention

[0011] Many embodiments of a method and apparatus for refreshing data of flash memory cells in a vector-matrix multiplication (VMM) array for use in an artificial neural network are disclosed. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram showing an artificial neural network.

[0013] Figure 2 is a cross-sectional side view of a conventional 2-gate non-volatile memory cell.

[0014] Figure 3 is a cross-sectional side view of a conventional 4-gate non-volatile memory cell.

[0015] Figure 4 is a cross-sectional side view of a conventional 3-gate non-volatile memory cell.

[0016] Figure 5 FIG. 4 is a cross-sectional side view of another conventional 2-gate non-volatile memory cell.

[0017] Figure 6 A schematic diagram showing different stages of an exemplary artificial neural network utilizing a non-volatile memory array.

[0018] Figure 7 is a block diagram showing a vector multiplier matrix.

[0019] Figure 8 is a block diagram showing the various stages of the vector multiplier matrix.

[0020] Fig. 9 An implementation of a vector multiplier matrix is ​​depicted.

[0021] Fig.10 Depicted Fig. 9 The operating voltage of the vector multiplier matrix.

[0022] Fig.11 Another implementation of a vector multiplier matrix is ​​depicted.

[0023] Fig.12 Depicted Fig.11 The operating voltage of the vector multiplier matrix.

[0024] Fig.13 Another implementation of a vector multiplier matrix is ​​depicted.

[0025] Fig.14 Depicted Fig.13 The operating voltage of the vector multiplier matrix.

[0026] Fig.15 Another implementation of a vector multiplier matrix is ​​depicted.

[0027] Fig.16 Describes Fig.15 The operating voltage of the vector multiplier matrix.

[0028] Fig.17 Another implementation of a vector multiplier matrix is ​​depicted.

[0029] Fig.18 A system for detecting data drift in a vector multiplier matrix is ​​depicted.

[0030] Fig.19 Describes the data refresh method.

[0031] Fig. 20 Another data refresh method is described.

[0032] Fig.21 Characterizes the fast unit in the vector multiplier matrix.

[0033] Fig. 22 Characterizes the fast unit in the vector multiplier matrix.

[0034] Fig.23 The mechanism used to assess data drift is described.

[0035] Fig.24 An implementation of a data drift detector is depicted.

[0036] Fig.25 Another implementation of a data drift detector is depicted.

[0037] Fig.26 Depicted Fig.25 An exemplary sampling period for an implementation scheme of .

[0038] Fig. 27 Another implementation of a data drift detector is depicted. DETAILED DESCRIPTION

[0039] The artificial neural network of the present invention utilizes a combination of CMOS technology and non-volatile memory arrays.

[0040] Non-volatile memory cell

[0041] Digital non-volatile memory is well known. For example, U.S. Patent No. 5,029,130 ​​("the '130 patent") discloses a split-gate non-volatile memory cell array and is incorporated herein by reference for all purposes. Such memory cells are Figure 2 . Each memory cell 210 includes a source region 14 and a drain region 16 formed in a semiconductor substrate 12, with a channel region 18 therebetween. A floating gate 20 is formed over and insulated from (and controls the conductivity of) a first portion of the channel region 18, and is formed over a portion of the source region 16. A word line terminal 22 (which is typically coupled to a word line) has a first portion disposed over and insulated from (and controls the conductivity of) a second portion of the channel region 18, and a second portion extending upward and located over the floating gate 20. The floating gate 20 and the word line terminal 22 are insulated from the substrate 12 by a gate oxide. A bit line 24 is coupled to the drain region 16.

[0042] The storage cell 210 is erased (where electrons are removed from the floating gate) by placing a high positive voltage on the word line terminal 22, which causes the electrons on the floating gate 20 to tunnel through the intermediate insulator via Fowler-Nordheim tunneling from the floating gate 20 to the word line terminal 22.

[0043] The memory cell 210 is programmed (where electrons are placed on the floating gate) by placing a positive voltage on the word line terminal 22 and a positive voltage on the source 16. Electron current will flow from the source 16 to the drain 14. When the electrons reach the gap between the word line terminal 22 and the floating gate 20, the electrons will accelerate and become hot. Due to the electrostatic attraction from the floating gate 20, some of the heated electrons will be injected through the gate oxide 26 onto the floating gate 20.

[0044] The memory cell 210 is read by placing a positive read voltage on the drain 14 and the word line terminal 22 (which turns on the channel region below the word line terminal). If the floating gate 20 is positively charged (i.e., electrons are erased and positively coupled to the drain 16), the portion of the channel region below the floating gate 20 is also turned on, and current will flow through the channel region 18, which is sensed as the erased state or the "1" state. If the floating gate 20 is negatively charged (i.e., programmed by electrons), the portion of the channel region below the floating gate 20 is mostly or completely turned off, and no current (or very little current) will flow through the channel region 18, which is sensed as the programmed state or the "0" state.

[0045] Table 1 depicts typical voltage ranges that may be applied to the terminals of the memory cell 210 for performing read, erase, and program operations:

[0046] Table 1: Figure 2 Operation of the flash memory unit 210

[0047] WL BL SL Read 2-3V 0.6V-2V 0V Erase About 11-13V 0V 0V programming 1-2V 1-3μA 9-10V

[0048] Other split-gate memory cell configurations are known. For example, Figure 3 A quad-gate memory cell 310 is depicted, which includes a source region 14, a drain region 16, a floating gate 20 over a first portion of a channel region 18, a select gate 28 (typically coupled to a word line) over a second portion of the channel region 18, a control gate 22 over the floating gate 20, and an erase gate 30 over the source region 14. This configuration is described in U.S. Pat. No. 6,747,310, which is incorporated herein by reference for all purposes. Here, except for the floating gate 20, all gates are non-floating, meaning that they are electrically connected or capable of being electrically connected to a voltage source. Programming is shown by heated electrons from the channel region 18 injecting themselves onto the floating gate 20. Erasing is shown by electrons tunneling from the floating gate 20 to the erase gate 30.

[0049] Table 2 depicts typical voltage ranges that may be applied to the terminals of the memory cell 310 for performing read, erase, and program operations:

[0050] Table 2: Figure 3 Operation of the flash memory unit 310

[0051] WL / SG BL CG EG SL Read 1.0V-2V 0.6V-2V 0-2.6V 0-2.6V 0V Erase -0.5V / 0V 0V 0V / -8V 8-12V 0V programming 1V 1μA 8-11V 4.5-9V 4.5-5V

[0052] Figure 4 A split-gate tri-gate memory cell 410 is depicted. The memory cell 410 is Figure 3 The memory cell 310 is the same as the memory cell 410 except that the memory cell 410 does not have a separate control gate. The erase operation (erasing through the erase gate) and the read operation are the same as Figure 3 The operation is similar to that of , except that there is no control gate bias. The programming operation is also done without a control gate bias, so the programming voltage on the source line is higher to compensate for the lack of control gate bias.

[0053] Table 3 depicts typical voltage ranges that may be applied to the terminals of the memory cell 410 for performing read, erase, and program operations:

[0054] Table 3: Figure 4 Operation of the flash memory unit 410

[0055] WL / SG BL EG SL Read 0.7V-2.2V 0.6V-2V 0-2.6V 0V Erase -0.5V / 0V 0V 11.5V 0V programming 1V 2-3μA 4.5V 7-9V

[0056] Figure 5 A stacked gate memory cell 510 is depicted. The memory cell 510 is similar to Figure 2 Memory cell 210 of the embodiment of the present invention is shown, except that floating gate 20 extends over the entire channel region 18, and control gate 22 extends over floating gate 20, separated by an insulating layer. Erase, program and read operations operate in a similar manner as previously described for memory cell 210.

[0057] Table 4 depicts typical voltage ranges that may be applied to the terminals of the memory cell 510 for performing read, erase, and program operations:

[0058] Table 4: Figure 5 Operation of the flash memory unit 510

[0059] CG BL SL P-sub Read 2-5V 0.6–2V 0V 0V Erase -8 to -10V / 0V FLT FLT 8-10V / 15-20V programming 8-12V 3-5V 0V 0V

[0060] In order to utilize a memory array comprising one of the above-described types of non-volatile memory cells in an artificial neural network, two modifications are made. First, the circuitry is configured so that each memory cell can be individually programmed, erased, and read without adversely affecting the memory states of other memory cells in the array, as further explained below. Second, continuous (analog) programming of the memory cells is provided.

[0061] Specifically, the memory state (i.e., charge on the floating gate) of each memory cell in the array can be changed continuously from a fully erased state to a fully programmed state, and vice versa, independently and with minimal disturbance to other memory cells. In another embodiment, the memory state (i.e., charge on the floating gate) of each memory cell in the array can be changed continuously from a fully programmed state to a fully erased state, and vice versa, independently and with minimal disturbance to other memory cells. This means that the cell storage device is analog, or at least can store one of many discrete values ​​(such as 16 or 64 different values), which allows very precise and individual tuning of all cells in the memory array, and this makes the memory array ideal for storing and fine-tuning the synaptic weights of a neural network.

[0062] Neural Network Using Non-Volatile Memory Cell Array

[0063] Figure 6 A non-limiting example of a neural network utilizing a non-volatile memory array is conceptually illustrated. This example uses the non-volatile memory array neural network for a facial recognition application, but any other suitable application may also be implemented using a non-volatile memory array-based neural network.

[0064] For this example, S0 is the input, which is a 32×32 pixel RGB image with 5-bit precision (i.e., three 32×32 pixel arrays, one for each color R, G, and B, with 5-bit precision per pixel). The synapse CB1 from S0 to C1 has both different sets of weights and shared weights, and scans the input image with a 3×3 pixel overlapping filter (kernel), shifting the filter by 1 pixel (or more than 1 pixel as dictated by the model). Specifically, the values ​​of 9 pixels in the 3×3 portion of the image (i.e., called the filter or kernel) are provided to the synapse CB1, whereby these 9 input values ​​are multiplied by the appropriate weights, and after summing the outputs of this multiplication, a single output value is determined and provided by the first neuron of CB1 for the pixels of one of the layers C1 that generates the feature map. The 3×3 filter is then shifted to the right by one pixel (i.e., a column of three pixels on the right is added, and a column of three pixels on the left is released), whereby the 9 pixel values ​​in this newly positioned filter are provided to the synapse CB1, whereby they are multiplied by the same weights and a second single output value is determined by the associated neuron. This process continues until the 3×3 filter scans all three colors and all bits (precision values) over the entire 32×32 pixel image. The process is then repeated using different sets of weights to generate different feature maps for C1 until all feature maps for layer C1 are calculated.

[0065] At C1, in this example, there are 16 feature maps, each with 30×30 pixels. Each pixel is a new feature pixel extracted from the product of the input and the kernel, so each feature map is a two-dimensional array, so in this example, the synapse CB1 is composed of a two-dimensional array of 16 layers (remember that the neuron layers and arrays referred to in this article are logical relationships, not necessarily physical relationships, that is, the arrays do not have to be oriented to physical two-dimensional arrays). Each of the 16 feature maps is generated by one of sixteen different sets of synaptic weights applied to the filter scan. The C1 feature maps can all relate to different aspects of the same image features, such as boundary recognition. For example, a first map (generated using a first set of weights, shared for all scans used to generate the first map) can identify circular edges, a second map (generated using a second set of weights different from the first set of weights) can identify rectangular edges, or the aspect ratio of certain features, and so on.

[0066] Before going from C1 to S1, an activation function P1 (pooling) is applied, which pools the values ​​from consecutive non-overlapping 2×2 regions in each feature map. The purpose of the pooling stage is to average neighboring positions (or a max function can also be used) to, for example, reduce the dependency of edge positions and reduce the data size before entering the next stage. At S1, there are 16 15×15 feature maps (i.e., 16 different arrays of 15×15 pixels each). The synapses and associated neurons from S1 to C2 in CB2 scan the maps in S1 using a 4×4 filter, where the filter is shifted by 1 pixel. At C2, there are 22 12×12 feature maps. Before going from C2 to S2, an activation function P2 (pooling) is applied, which pools the values ​​from consecutive non-overlapping 2×2 regions in each feature map. At S2, there are 22 6x6 feature maps. The activation function is applied to the synapse CB3 from S2 to C3, where each neuron in C3 is connected to each map in S2. At C3, there are 64 neurons. Synapse CB4 from C3 to output S3 fully connects S3 to C3. The output at S3 includes 10 neurons, where the highest output neuron determines the class. For example, this output may indicate recognition or classification of the content of the original image.

[0067] Each level of synapses is implemented using an array or portion of an array of non-volatile memory cells. Figure 7 A block diagram of a vector-matrix multiplication (VMM) array that includes nonvolatile memory cells and is used as a synapse between an input layer and the next layer. Specifically, VMM 32 includes a nonvolatile memory cell array 33, an erase gate and word line gate decoder 34, a control gate decoder 35, a bit line decoder 36, and a source line decoder 37, which decode the input of the memory array 33. In this example, the source line decoder 37 also decodes the output of the memory cell array. Alternatively, the bit line decoder 36 can decode the output of the memory array. The memory array has two purposes. First, it stores the weights to be used by the VMM. Second, the memory array effectively multiplies the input with the weights stored in the memory array and then adds them one by one on the output line (source line or bit line) to produce an output that will serve as the input of the next layer or the input of the final layer. By performing multiplication and addition functions, the memory array eliminates the need for separate multiplication and addition logic circuits, and is also highly power-efficient due to in-situ memory calculations.

[0068] The output of the memory array is provided to a differential adder (such as an adding operational amplifier) ​​38, which sums the output of the memory cell array to create a single value for the convolution. The differential adder is to achieve the sum of positive weights and negative weights with the positive input. The summed output value is then provided to an activation function circuit 39, which corrects the output. The activation function may include a sigmoid, tanh, or ReLU function. The corrected output value becomes an element of the feature map of the next layer (e.g., C1 in the above description) and is then applied to the next synapse to produce the next feature map layer or the final layer. Therefore, in this example, the memory array constitutes a plurality of synapses (which receive their inputs from an existing neuron layer or from an input layer such as an image database), and the summing operational amplifier 38 and the activation function circuit 39 constitute a plurality of neurons.

[0069] Figure 8 32a. The input is converted from digital to analog by a digital-to-analog converter 31 and provided to an input VMM 32a. The output generated by the input VMM 32a is provided as an input to the next VMM (hidden level 1) 32b, which in turn generates an output provided as an input to the next VMM (hidden level 2) 32b, and so on. The layers of the VMM 32 serve as different layers of synapses and neurons of a convolutional neural network (CNN). Each VMM can be an independent non-volatile memory array, or multiple VMMs can utilize different portions of the same non-volatile memory array, or multiple VMMs can utilize overlapping portions of the same non-volatile memory array.

[0070] Vector-Matrix Multiplication (VMM) Array

[0071] Fig. 9 Depicted is a Neuron VMM 900, which is particularly suitable for Figure 2 Memory cells of the type shown, and used as synapses and components of neurons between the input layer and the next layer. VMM 900 includes a memory array 903 of non-volatile memory cells, a reference array 901 and a reference array 902. Reference arrays 901 and 902 are used to convert current inputs flowing into terminals BLR0-3 into voltage inputs WL0-3. As shown, reference arrays 901 and 902 are in the column direction. Generally speaking, the reference array direction is orthogonal to the input line. In practice, the reference memory cell is a diode connected to the current input flowing therein through a multiplexer (multiplexer 914, which includes a multiplexer and a cascade transistor VBLR for biasing the reference bit line). The reference cell is tuned to a target reference level.

[0072] The memory array 903 serves two purposes. First, it stores weights to be used by the VMM 900. Second, the memory array 903 effectively multiplies the inputs (current inputs provided in terminals BLR0-3; reference arrays 901 and 902 convert these current inputs into input voltages to be provided to word lines WL0-3) by the weights stored in the memory array to produce an output, which will be the input to the next layer or the input to the final layer. By performing a multiplication function, the memory array eliminates the need for a separate multiplication logic circuit and is also highly power efficient. Here, the voltage input is provided on the word line and the output appears on the bit line during a read (inference) operation. The current placed on the bit line performs a summing function of all currents from the memory cells connected to the bit line.

[0073] Fig.10 Depicted are operating voltages for VMM 900. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for unselected cells, the bit line for a selected cell, the bit line for unselected cells, the source line for a selected cell, and the source line for unselected cells. The rows indicate read, erase, and program operations.

[0074] Fig.11 Depicted is a neuron VMM 1100, which is particularly suitable for Figure 2 1 and 11. The VMM 1100 includes a memory array 1103 of nonvolatile memory cells, a reference array 1101, and a reference array 1102. The VMM 1100 is similar to the VMM 900, except that in the VMM 1100, the word lines extend in the vertical direction. There are two reference arrays 1101 (at the top, which provides a reference for converting the input current into a voltage for even rows) and 1102 (at the bottom, which provides a reference for converting the input current into a voltage for odd rows). Here, the input is provided on the word line, and the output appears on the source line during a read operation. The current placed on the source line performs a summing function of all currents from the memory cells connected to the source line.

[0075] Fig.12 Depicted are operating voltages for the VMM 1100. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for unselected cells, the bit line for a selected cell, the bit line for unselected cells, the source line for a selected cell, and the source line for unselected cells. The rows indicate read, erase, and program operations.

[0076] Fig.13 Depicted is a neuron VMM 1300, which is particularly suitable for Figure 313. The VMM 1300 includes a memory array 1301 of nonvolatile memory cells, a reference array 1302 (providing a reference for converting input current into a voltage for even rows), and a reference array 1303 (providing a reference for converting input current into a voltage for odd rows). The VMM 1300 is similar to the VMM 900, except that the VMM 1300 also includes a control line 1306 coupled to the control gates of a row of memory cells and a control line 1307 coupled to the erase gates of the memory cells of an adjacent row. Here, the directions of the word lines, control gate lines, and erase gate lines are the same. The VMM also includes a reference bit line select transistor 1304 (part of multiplexer 1314) that selectively couples the reference bit line to the bit line contact of a selected reference memory cell, and a switch 1305 (part of multiplexer 1314) that selectively couples the reference bit line to a control line 1306 of a particular selected reference memory cell. Here, during a read operation, an input is provided on a word line (of memory array 1301) and an output appears on a bit line such as bit line 1309. The current placed on the bit line performs a summing function of all the currents from the memory cells connected to the bit line.

[0077] Fig.14 Depicted are operating voltages for the VMM 1300. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for an unselected cell, the bit line for a selected cell, the bit line for an unselected cell, the control gate for a selected cell, the control gate for an unselected cell in the same sector as the selected cell, the control gate for an unselected cell in a different sector than the selected cell, the erase gate for a selected cell, the erase gate for an unselected cell, the source line for a selected cell, the source line for an unselected cell. The rows indicate read, erase, and program operations.

[0078] Fig.15 Depicted is a Neuron VMM 1500, which is particularly suitable for Figure 3 1500 is similar to VMM 1100, except that in VMM 1500, the erase gate line (such as erase gate line 1501) extends in the vertical direction. Here, the input is provided on the bit line, and the output appears on the source line. The current placed on the bit line performs the summing function of all currents from the memory cells connected to the bit line.

[0079] Fig.16Depicted are operating voltages for VMM 1500. The columns in the table indicate the voltages placed on the word line for selected cells, word line for unselected cells, bit line for selected cells, bit line for unselected cells, control gate for selected cells, control gate for unselected cells in the same sector as the selected cells, control gate for unselected cells in a different sector than the selected cells, erase gate for selected cells, erase gate for unselected cells, source line for selected cells, source line for unselected cells. The rows indicate read, erase, and program operations.

[0080] Fig.17 Depicted is a Neuron VMM 1700, which is particularly suitable for Figure 3 1700 and serves as a synapse and component of neurons between the input layer and the next layer. VMM 1700 includes a memory array 1701 of non-volatile memory cells and a reference array 1702 (at the top of the array). Alternatively, another reference array may be provided at the bottom, similar to Fig.10 Reference array. In other respects, VMM 1700 is similar to the previously described VMM, except that in VMM 1700, control gate lines (such as control gate lines 1703) extend in the vertical direction (therefore reference array 1702 is in the row direction orthogonal to the input control gate lines), and erase gate lines (such as erase gate lines 1704) extend in the horizontal direction. Here, input is provided on the control gate lines, and output occurs on the source lines. In one embodiment, only even-numbered rows are used, and in another embodiment, only odd-numbered rows are used. The current placed on the source lines performs the summing function of all currents of the memory cells connected to the source lines.

[0081] As described herein for neural networks, flash memory cells are preferably configured to operate in the subthreshold region.

[0082] The memory cell described in this article is weakly reverse biased:

[0083] Ids=Io*e (Vg-Vth) / kVt =w*Io*e (Vg) / kVt

[0084] w=e (-Vth) / kVt

[0085] For an I to V logarithmic converter that uses a memory cell to convert input current to input voltage:

[0086] Vg=k*Vt*log[Ids / wp*Io]

[0087] For a memory array used as a vector matrix multiplier VMM, the output current is:

[0088] Iout=wa*Io*e (Vg) / kVt ,Right now

[0089] Iout=(wa / wp)*Iin=W*Iin

[0090] W=e (Vthp-Vtha) / kVt

[0091] The word line or control gate may be used as the input of the memory cell for the input voltage.

[0092] Alternatively, the flash memory cell can be configured to operate in the linear region:

[0093] Ids=beta*(Vgs-Vth)*Vds;beta=u*Cox*W / L

[0094] Wα(Vgs-Vth)

[0095] For an I to V linear converter, a memory cell operating in the linear region can be used to linearly convert the input / output current to the input / output voltage.

[0096] Other implementations of the ESF vector-matrix multiplier are described in US Patent Application 15 / 826,345, which is incorporated herein by reference. A source line or a bit line can be used as a neuron output (current summing output).

[0097] Fig.18 Depicted is an emulated neuromorphic memory system 1800 that includes a system controller 1801, an emulated neuromorphic memory engine 1802, and an emulated neuromorphic memory engine 1803. The emulated neuromorphic memory engine 1802 and the emulated neuromorphic memory engine 1803 each contain a VNM array programmed with weights. A portion of the array in the VNM in the emulated neuromorphic memory engine 1803 is dedicated to the process of determining whether data has drifted.

[0098] Fig.19A method 1900 for refreshing data in a VNM is depicted. A set of flash memory cells are sampled (sample selection in step 1901). The samples are analyzed for each possible level (N value) stored by the flash memory cells (L0, ..., LN-1), and the difference (DRx) is calculated between each pair of adjacent levels (step 1902). For each level, the difference (DRx) is compared to a target (ideal) (DRx_target) (step 1903). If any difference (DRx) exceeds the target, a refresh flag is set (step 1904). A restore data process is then performed (step 1905). This step can be accomplished by an on-chip controller or an off-chip controller. In step 1905, a re-tuning step (e.g., re-programming) is customized for each level. The cell is first read back, and then, depending on the level, an incremental re-tuning is applied to the level. A lookup table (Icell vs. Ituning_target) can be used for this purpose. In this step, a tuning algorithm (iterative verification and procedure) is used to reach the target (such as by using the algorithm described in U.S. Patent Application 15 / 826,345, which is incorporated herein by reference). The sampled data retention cells are also restored. Once the current of each cell equals the target current of the level stored in the cell (step 1906), the process is complete (step 1908). If it is never reached, the cell is considered bad and the process is complete (step 1907).

[0099] Fig. 20Another method 2000 for refreshing data in a VNM is depicted. A set of exemplary flash memory cells are sampled (step 2001). The samples are analyzed for each possible level (N value) stored by the flash memory cells (L0, ..., LN-1), and the difference (DRx) is calculated between each pair of adjacent levels (step 2002). For each level, the difference (DRx) is compared to the target (ideal) (DRx_target) (step 2003). If any difference (DRx) exceeds the target, a refresh flag is set (step 2004). A restore data process is then performed (step 2005). The restore can be done by an on-chip or off-chip controller. In step 2005, a re-tuning step (e.g., re-programming) is customized for each level. The cell (array cell and sampled data retention cell) is first read back, and then, depending on the level, an incremental re-tuning is applied to the level by a fine incremental restore algorithm. A lookup table (Icell vs. Ituning target) can be used for this purpose. In this step, a fine tuning algorithm (fine step iterative verification and programming) is used to reach the target (such as by using the algorithm described in U.S. Patent Application 15 / 826,345, which is incorporated herein by reference). Alternatively, data recovery can be accomplished by reprogramming the entire VMM array by transferring all weights from training to the VMM array. Once the current of each cell is equal to the target current for the level stored in the cell (step 2006), the process is complete (step 2007). If it has never been reached, the cell is considered bad (step 2008). A redundant array sector is then used to replace the bad array sector (step 2009). If the number of redundant cells used in this manner exceeds a threshold (step 2010), the entire sector is considered bad (step 2011).

[0100] about Fig.21 and Fig. 22 , it may be desirable to identify "fast cells," which are cells that can lose charge from their floating gates relatively quickly during programming operations compared to average cells, because such fast cells are more susceptible to data drift over time. Fig.21 Referring to graph 2100, fast cells can be identified as cells Ircell that achieve high current levels with relatively few programming attempts (counts). Fig. 22 22. Referring to graph 2200, it will be appreciated that a fast cell may have a higher Ir operating range than a normal cell. Fig. 22 An exemplary cell for data retention (drift) monitoring is shown. Cells outside of the minimum and maximum ranges are used for monitoring purposes. This is used to create monitoring cells with stress conditions that are worse than normal cells.

[0101] Fig.23A data drift detection scheme is depicted. In this example, a certain current level is expected to be between Target 1 and Target 2. If the sense current is between these targets, no refresh is required, which means that the drift is acceptable for system performance. However, if the sense current is outside the range between Target 1 and Target 2, a data refresh (recovery) process will be performed.

[0102] Fig.24 An embodiment of a data drift detector is depicted. The data drift detector 2400 receives a current 2402 (which may be a sensed current level through a selected cell or a difference between such currents) and a reference current 2401 in a comparator 2403, which outputs a value indicating whether the current 2402 exceeds the reference current 2401. The data drift detector 2400 compares the current drawn by the selected cell to a target current based on a value expected to have been stored in the selected cell.

[0103] Fig.25 Another embodiment of a data drift detector is depicted. The data drift detector 2500 includes sample-and-hold capacitors 2501 and 2504, switches 2502 and 2505, a current source 2506 (which may be a sensed current level through a selected cell or a difference between such currents), a reference current 2503, and a comparator 2507. In this embodiment, the two current sources are sampled by the capacitors through a determined ramp period and held, and then those sampled values ​​are compared by the comparator 2507. Thus, the data drift detector 2500 compares the sampled current drawn by the selected cell with a sampled target current based on the value expected to have been stored in the selected cell.

[0104] refer to Fig.26 As shown in graph 2600, the sampling period of data drift detector 2500 can be selected based on an understanding of how much time may pass before significant data drift occurs. This can reduce power consumption compared to data drift detector 2400.

[0105] Fig. 27 Another embodiment of a data drift detector is depicted. The data drift detector 2700 includes a selected memory cell 2701, a current source 2702, a reference voltage 2703, and a comparator 2704. In this embodiment, a direct comparison is made between the voltage stored on the floating gate of the selected memory cell 2701 (the voltage on node 2705 = approximately VFG - VtFG) and the reference voltage 2703 (the voltage that the selected memory cell 2701 is intended to store). The memory cell 2701 is configured in a source follower mode of operation. The reference voltage can be generated from a reference memory cell.

[0106] It should be noted that, as used herein, the terms "above" and "on" both inclusively include "directly on" (no intermediate material, element, or space is provided between) and "indirectly on" (intermediate material, element, or space is provided between). Similarly, the term "adjacent" includes "directly adjacent" (no intermediate material, element, or space is provided between) and "indirectly adjacent" (intermediate material, element, or space is provided between), "mounted to" includes "directly mounted to" (no intermediate material, element, or space is provided between) and "indirectly mounted to" (intermediate material, element, or space is provided between), and "electrically coupled to" includes "directly electrically coupled to" (no intermediate material or element electrically connecting the elements together between) and "indirectly electrically coupled to" (intermediate material or element electrically connecting the elements together between). For example, forming an element "above a substrate" may include directly forming the element on the substrate without intermediate materials / elements between the two, and indirectly forming the element on the substrate with one or more intermediate materials / elements between the two.

Claims

1. A method for detecting data drift in selected flash memory cells and refreshing data in an analog neuromorphic memory system, the method comprising: sensing a first voltage stored in a floating gate of a selected flash memory cell; comparing the first voltage with a second voltage provided by a reference circuit; If the first voltage exceeds the second voltage, refreshing the data stored in the floating gate by storing a third voltage on the floating gate, wherein the third voltage is closer to the second voltage than the first voltage, wherein the refreshing is accomplished by re-transferring training weights to selected memory cells. The method of claim 1 , wherein the reference circuit comprises a voltage source. The method of claim 1 , wherein the reference circuit comprises a reference memory cell. The method of claim 1 , wherein the selected memory cell is a split-gate memory cell. The method of claim 1 , wherein the selected memory cell is a stacked-gate memory cell. The method of claim 1 , wherein the selected memory cell is configured in a source follower mode.

7. A method for detecting data drift in selected flash memory cells and refreshing data in an analog neuromorphic memory system, the method comprising: sensing a first current drawn by a selected flash memory cell during a read operation; comparing the first current to a second current drawn by a reference circuit; If the first current exceeds the second current, refreshing the data stored in the floating gate of the selected memory cell so that the selected flash memory cell draws a third current during a read operation, wherein the third current is closer to the second current than the first current, wherein the refreshing is accomplished by re-transferring the training weights to the selected memory cell. The method of claim 7 , wherein the reference circuit comprises a current source.

9. The method of claim 7, wherein the selected memory cell is a split-gate memory cell.

10. The method of claim 7, wherein the selected memory cell is a stacked gate memory cell.

11. A method for detecting data drift in selected flash memory cells and refreshing data in an analog neuromorphic memory system, the method comprising: sampling a first current drawn by a selected flash memory cell during a read operation to generate a first sampled value; sampling a second current drawn by the reference circuit to generate a second sampled value; comparing the first sampled value and the second sampled value; If the first sampled value exceeds the second sampled value, refreshing the data stored in the floating gate of the selected memory cell so that the selected flash memory cell draws a third current during a read operation, wherein the third current is closer to the second current than the first current, wherein the refreshing is accomplished by re-transferring the training weights to the selected memory cell. The method of claim 11 , wherein the reference circuit comprises a current source.

13. The method of claim 11, wherein the selected memory cell is a split-gate memory cell. The method of claim 11 , wherein the selected memory cell is a stacked-gate memory cell.

15. A method for detecting data drift in selected flash memory cells and refreshing data in an analog neuromorphic memory system, the method comprising: sensing a first current drawn by a selected flash memory cell during a read operation; comparing the first current to a second current drawn by a reference circuit; comparing the first current to a third current drawn by a reference circuit; refreshing the data stored in the floating gate of the selected memory cell if the first current exceeds the third current such that the selected flash memory cell draws a current during a read operation, wherein the drawn current is between the second current and the third current, wherein the refreshing is accomplished by re-transferring training weights to the selected memory cell; or If the first current exceeds the second current and does not exceed the third current, refreshing the data is suppressed. The method of claim 15 , wherein the reference circuit comprises a current source.

17. The method of claim 15, wherein the selected memory cell is a split-gate memory cell. The method of claim 15 , wherein the selected memory cell is a stacked gate memory cell.

19. A method for detecting data drift in selected flash memory cells and refreshing data in an analog neuromorphic memory system, the method comprising: identifying a fast cell among the flash memory cells, the fast cell being a cell that loses charge from its floating gate relatively quickly during a programming operation compared to an average cell of the flash memory cells; Sampling the fast unit to perform data drift detection; programming the flash cells for data drift detection at an exemplary level selected from N levels, N being the number of different weights that can be indicated by each flash memory cell; and The sampled fast cells are monitored for data drift detection, wherein the data is refreshed if data drift is detected, and wherein the refreshing is accomplished by re-transferring training weights to the sampled fast cells.

20. The method of claim 19, wherein the monitoring comprises: sensing a first current drawn by the sampled flash cell during a read operation; comparing the first current to a second current drawn by a reference circuit; comparing the first current to a third current drawn by a reference circuit; refreshing the data stored in the floating gate of the memory cell if the first current exceeds the third current so that the selected flash memory cell draws a current during a read operation, wherein the drawn current is between the second current and the third current; or If the first current exceeds the second current and does not exceed the third current, refreshing the data is suppressed. The method of claim 20 , wherein the reference circuit comprises a current source.

22. The method of claim 19, wherein the sampled flash cells are split-gate memory cells.

23. The method of claim 19, wherein the sampled flash cell is a stacked gate memory cell.

24. A method for detecting data drift in selected flash memory cells and recovering data in an analog neuromorphic memory system, the method comprising: selecting a sample of memory cells; Testing the selected samples for data drift detection; Programming the selected samples under test for data drift detection at an exemplary level selected from N levels, N being the number of different weights that can be indicated by each flash memory cell; as well as The selected samples under test are monitored for data drift detection, wherein when any current of the selected samples under test exceeds an incremental target, data recovery is performed, wherein the data recovery includes re-transferring training weights to the memory cells.

25. The method of claim 24, wherein any current of the selected sample being tested exceeds an increment target, then the data recovery flag is set.

26. The method of claim 24, wherein when any current of the memory cell being restored fails to meet a target, the memory cell being restored is identified as a bad cell.

27. The method of claim 26, further comprising replacing bad cells with redundant cells.

28. The method of claim 24, wherein the data recovery comprises cell tuning.

29. The method of claim 24, wherein the data recovery comprises tuning the unit by a target amount corresponding to the exemplary level.

30. The method of claim 24, wherein the memory cell is a split-gate memory cell.

31. The method of claim 24, wherein the memory cell is a stacked gate memory cell.