The present disclosure relates generally to memory devices and in particular the present disclosure relates to memory devices configured to apply different weights to different strings of memory cells coupled to a data line and methods.
Memory devices are typically provided as internal, semiconductor, integrated circuits in computers or other electronic devices. There are many different types of memory including random-access memory (RAM), read only memory (ROM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), and flash memory.
Flash memory devices have developed into a popular source of non-volatile memory for a wide range of electronic applications. Flash memory devices typically use a one-transistor memory cell that allows for high memory densities, high reliability, and low power consumption. Changes in threshold voltage of the cells, through programming of charge storage structures, such as floating gates or trapping layers or other physical phenomena, determine the data state of each cell. Common uses for flash memory include personal computers, personal digital assistants (PDAs), digital cameras, digital media players, digital recorders, games, appliances, vehicles, wireless devices, cellular telephones, and removable memory modules, and the uses for flash memory continue to expand.
Flash memory may utilize architectures known as NOR flash and NAND flash. The designation is derived from the logic used to read the devices. In a NOR flash architecture, a column of memory cells are coupled in parallel with each memory cell coupled to a data line, such as a bit line. A “column” refers to a group of memory cells that are commonly coupled to a data line, such as a bit line. It does not require any particular orientation or linear relationship, but instead refers to the logical relationship between memory cell and data line.
Typically, the array of memory cells for NAND flash memory devices is arranged such that the control gate of each memory cell of a row of the array is connected together to form an access line, such as a word line. Columns of the array include strings (often termed NAND strings) of memory cells connected together in series, source to drain, between a pair of select lines, a source select line and a drain select line. The source select line includes a source select gate at each intersection between a NAND string and the source select line, and the drain select line includes a drain select gate at each intersection between a NAND string and the drain select line. Each source select gate is connected to a source line, while each drain select gate is connected to a data line, such as column bit line.
Content addressable memories (CAM) are memories that implement a lookup table function in a single clock cycle. They use dedicated comparison circuitry to perform the lookups. CAM applications are often used in network routers for packet forwarding and the like. Each individual memory cell in a CAM usually requires its own comparison circuit in order to allow the CAM to detect a match between a bit of input data, such as an input feature vector (e.g., sometimes referred to as a key or key data) with a bit of data, such as a data feature vector, stored in the CAM. However, not all bits are the same. For example, binary expressions may include a most significant bit (MSB), a least significant bit (LSB), and bits between the MSB and LSB.
For the reasons stated above, and for other reasons stated below which will become apparent to those skilled in the art upon reading and understanding the present specification, there is a need in the art for identifying and differentiating between MSBs, LSBs and bits between the MSB and LSB in comparisons between input data and data stored in memory.
In the following detailed description of the embodiments, reference is made to the accompanying drawings that form a part hereof. In the drawings, like numerals describe substantially similar components throughout the several views. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. Other embodiments may be utilized and structural, logical, and electrical changes may be made without departing from the scope of the present invention.
The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present disclosure is defined only by the appended claims, along with the full scope of equivalents to which such claims are entitled.
For example, the data feature vectors stored in columns 212 of memory array 104 may correspond to known entities, such as known patterns, and memory device 100 may be configured to determine whether an input feature vector at least partially matches a particular known entity represented by a data feature vector, thereby at least potentially identifying the input feature vector as being the particular known entity.
Memory device 100 may be configured to identify the input feature vector as being a particular known entity in response to determining that the input feature vector potentially matches the data feature vector representing the particular known entity. For example, memory device 100 may be configured to identify the data feature vectors most likely to match the input feature vector based on how closely the data feature vectors match the input feature vector.
Components of the input feature vector correspond to certain features (e.g., attributes) of the input feature vector, and thus the entity represented by the input feature vector. Similarly, components of the data feature vectors correspond to certain features (e.g., attributes) of the data feature vectors, and thus the entities represented by the data feature vectors. For example, each data feature vector and the input feature vector may be a pattern, such as of a person's face, fingerprint, etc., and the components may be the features of that pattern.
Components of the input feature vector may be compared to like components of the data feature vectors to determine which data feature vectors potentially match the input feature vector, thereby identifying the entity represented by the input feature vector. For some embodiments, an input feature vector may be determined to potentially match a data feature vector even though there might be a mismatch of certain components.
The h signals in
Memory cells 3151 and 3152 may be viewed as switches that selectively allow current to flow from a respective string 310 to bit line 215, or vice versa, when the remaining memory cells in the respective string are operating as pass-transistors. A mismatch between a component of a data feature vector stored in a string 310 and a like component of an input feature vector results in current flow in that string 310.
Resistors 3401 and 340N are respectively coupled in series with strings 3101 and 310N and respectively control the analog value of current through strings 3101 and 310N. For example, resistors 3401 and 340N may act as constant current sources of varying values.
The current value through a string can be set by setting the resistance of the respective resistor 340. For example, the resistance of a resistor 340 can be set by programming the threshold voltage of the resistor 340, setting the channel length of the resistor 340 to a specific value, setting the voltage VC applied to the control gate of the resistor 340 to a specific value, etc. Multiple resistors 340 may be in series for additional control (compensation) if appropriate.
The ratio of current may be binary. For example, for a unit of current through string 3101, the currents through strings 3102 to 310N might respectively be ½ to ½N-1 of a unit, where N is the block (e.g. string number), e.g., starting at 1. The string corresponding to the most significant bit, e.g., string 3101, may have a weight of 1 and a unit current in the event of a mismatch, and the string corresponding to the least significant bit, e.g., string 310N, may have the smallest weight, e.g., ½N-1, and a current of ½N-1 of a unit in the event of a mismatch.
In an example, a feature A1=(a1, a2, a3, a4, . . . , aN) of a data feature vector may be programmed as 1011, e.g., for N=4. A feature B1=(b1, b2, b3, b4, . . . , bN) of an input feature vector may be input as 1010, e.g., for N=4, for comparison to feature A1. Since a mismatch occurs in the fourth significant bit, corresponding to the least significant bit in this case, the value of the current flow will be ½(4-1)=⅛ of a unit. This may be referred to as the least significant bit value of current. Note that bits a1 through aN may be respectively stored in blocks 2101 through 210N (e.g., strings 3101 through 310N) for some embodiments.
A feature B1=(b1, b2, b3, b4, . . . , bN) of an input feature vector may input as 0011 for comparison to feature A1. Since a mismatch occurs in the first significant bit, corresponding to the most significant bit in this case, the value of the current flow will be the most significant bit value of current ½(1-1)=1 unit.
For some embodiments, the input feature vector can be made up of multiple features (e.g., components of the input feature vector), e.g., IF(B)=(B1, B2, B3, . . . , BN) and each data feature vector can be made up of multiple features (e.g., components of the data feature vectors), e.g., DF(A)=(A1, A2, A3, . . . , AN)
The page buffer 110 may measure the analog current flow in bit line 215. For example, zero current is due to an exact match, and the smallest amount of current is due to a single feature mismatch, e.g., a least significant bit mismatch. A current of weight one results from a most significant bit mismatch. All mismatches in a column may be summed in an analog fashion. The threshold value of what current constitutes a match can be programmable.
In
For example, if a corresponding digit of data of a received input feature vector has a binary bit value of b=0, a voltage of about 2V, corresponding to a bit value of d1=0, may be applied to the control gate of memory cell 3152 of string 3101, and a voltage of about 4V, corresponding to a bit value of d1*=1, may be applied to the control gate of memory cell 3151 of string 3101. That is, d1*=1 and d1=0 may correspond to a binary bit value of b=0 of a digit of data of the input feature vector, as shown in
Memory cells 3151 and 3152 of string 3101 might be respectively programmed to threshold voltages Vt1 and Vt1* to store at least a portion of a value of a feature of the data feature vector stored in string 3101. For example, memory cells 3151 and 3152 of string 3101 might be respectively programmed to threshold voltages Vt1 and Vt1* of about 1V and about 3V to store a binary value of a=0 for a digit of data of the stored data feature vector. For example, when memory cell 3151 is programmed to a threshold voltage Vt1 of about 1V, memory cell 3151 may store a bit value of z=0, and when memory cell 3152 is programmed to a threshold voltage Vt1* of about 3V, memory cell 3152 may store a bit value of z*=1. This means that when memory cells 3151 and 3152 respectively store bit values of z=0 and z*=1, memory-cell pair 3151, 3152 stores a binary value of a=0, as shown in
Memory cells 3151 and 3152 of string 3101 might be respectively programmed to threshold voltages Vt1 and Vt1* of about 3V and about 1V to store a binary value of a=1 for the digit of data of the data feature vector. For example, when memory cell 3151 is programmed to a threshold voltage Vt1 of about 3V, memory cell 3151 may store a bit value of z=1, and when memory cell 3152 is programmed to a threshold voltage Vt1* of about 1V, memory cell 3152 may store a bit value of z*=0. This means that when memory cells 3151 and 3152 respectively store bit values of z=1 and z*=0, memory-cell pair 3151, 3152 stores a binary value of a=1, as shown in
The logic truth table of
If the voltage of the signal corresponding to the respective d*/d bit value is greater than the programmed threshold voltage (e.g., Vt* or Vt) of the respective memory cell (e.g., 3151 and 3152), the respective memory cell will conduct, e.g., and will be ON. If the voltage of the signal corresponding to the respective d*/d bit value is less than the programmed threshold voltage (e.g., Vt* or Vt) of the respective memory cell (e.g., 3151 and 3152), the respective memory cell will not conduct, e.g., and will be OFF. According to an embodiment implementing the truth table of
Referring to
Continuing with such an example, if the digit b has a value of binary 1, a signal having a voltage level of about 2V, e.g., corresponding to a bit value d1*=0, might be applied to the control gate of the first memory cell (e.g., memory cell 3151) of the pair, and a signal having a voltage level of about 4V, e.g., corresponding to a bit value d1=1, might be applied to the control gate of a second memory cell of the pair (e.g., memory cell 3152).
Further continuing with the example, if do not care data X has been inserted into the digit d, a signal having a voltage level of about 0V might be applied to the control gate of the first memory cell (e.g., memory cell 3151) of the pair, and a signal having a voltage level of about 0V might also be applied to the control gate of the second memory cell of the pair (e.g., memory cell memory cell 3152), thus ensuring neither cell of the pair conducts, assuring a match no matter what actual data is stored by the pair of cells.
The F entry in the first column of
Each row of the second column of
For example, if the digit a has a value of binary 0, a first memory cell (e.g., memory cell 3151 of string 3101) of the pair might be programmed to a threshold voltage (e.g., Vt1) of about 1V, while the second memory cell of the pair (e.g., memory cell 3152 of string 3101) might be programmed to a threshold voltage (e.g., Vt1*) of about 3V. Continuing with such an example, if the digit a has a value of binary 1, a first memory cell (e.g., memory cell 3151 of string 3101) of the pair might be programmed to a threshold voltage (e.g., Vt1) of about 3V, while the second memory cell of the pair (e.g., memory cell 3152 of string 3101) might be programmed to a threshold voltage (e.g., Vt1*) of about 1V.
Further continuing with the example, to store do not care data X for digit a, a first memory cell (e.g., memory cell 3151 of string 3101) of the pair might be programmed to a threshold voltage (e.g., Vt1) of about 3V, and the second memory cell of the pair (e.g., memory cell 3152 of string 3101) might also be programmed to a threshold voltage (e.g., Vt1*) of about 3V, thus ensuring that at least one cell of the pair does not conduct regardless of the selected the voltage of the signal corresponding to the respective d*/d bit value, and thereby assuring a match regardless of the value of digit b. Still further, in a another variation, the memory cells have not been programmed to store either a binary 0, a binary 1, or do not care data X (e.g., they both have a threshold voltage of about 1V), in which case a match with digit b is only detected if do not care data has been inserted into digit b.
Each row of the third column of
Each column 212 may include memory cells from each of blocks 210. Each block 210 may include access lines (e.g., word lines 220) respectively coupled to rows of memory cells in each block 210 so that each block 210 is organized in rows and columns of memory cells.
Bit lines 2151 to 215L may be respectively coupled, one-to-one, to current sense amplifiers, such as sense amps 2251 to 225L that may have analog-to-digital converter functionality. Sense amps 2251 to 225L may be respectively coupled, one-to-one, to comparators 2281 to 228L. A single register 230 (e.g., that may be referred to as difference register), in some embodiments, may be coupled to comparators 2281 to 228L. Sense amps 225, comparators 228, and register 230 may be part of page buffer 110 for some embodiments. For other embodiments, there may be a plurality of individual difference registers, where the individual difference registers are coupled one-to-one to comparators 2281 to 228L.
Word lines 220 may be coupled to respective ones of drivers 232 for receiving voltages therefrom. Mask registers 240 may be coupled to drivers 232, and may be used to mask certain blocks 210 or certain cells of blocks 210 when reading from or writing to array 204. For example, mask registers 240 may be coupled one-to-one to word lines 220. Drivers 232 and mask register 240 may be part of input buffer 120 for some embodiments. In other words, mask registers 240 may mask memory cells coupled thereto. For example, a mask register 240 may store a mask bit that causes a memory cell coupled thereto to be masked.
Input (e.g., key) registers 245 (e.g., which may also be referred to as input feature registers) may be coupled one-to-one to mask registers 240, and thus to word lines 220, and may be part of input buffer 120 for some embodiments. For other embodiments, the input feature registers 245 and mask registers 240 may be interchanged so that input feature registers 245 are between mask registers 240 and drivers 232.
Note that the set of input feature registers receive input data, such as input feature vectors, for comparison to data, such as data feature vectors, stored in columns 212 of memory array 204. For example, an input feature vector may be programmed into resisters 245.
Each component of a data feature vector may be stored in one or more blocks 210. For some embodiments, multiple components of a data feature vector may be stored in a single block. For other embodiments, components for multiple data feature vectors may be stored in a single block.
Each feature (e.g., component) of a feature vector may be expressed by a binary expression of data having a most significant bit, a least significant bit, and bits of differing significance between the most significant bit and the least significant bit, where bits of the input feature vector are compared to bits of each data feature vector having like significance. For example, the most significant bit of the input feature vector may be compared to the most significant bit of a data feature vector, the least significant bit of the input feature vector to the least significant bit of the data feature vector, etc. The bits of a feature (e.g., component) of a data feature vector may be respectively stored in different blocks 210 of a group of blocks 210; the bits of another feature (e.g., component) of the data feature vector may be respectively stored in different blocks 210 of another group of blocks 210; etc.
For some embodiments, more than one bit of a component of a data feature vector can be stored in a single block. For example, multiple bits for a single component of a data feature vector can be stored in a single block. In addition, bits for multiple components of a data feature vector can be stored in a single block.
For some embodiments, each sense amp 225 may sense a level of a current IA on a respective one of bit lines 215 and may convert the sensed current level IA to a digital representation ID (e.g., a voltage level representing a sensed current level) of the level of the sensed current IA. Each sense amp 225 may send the representation ID to a respective comparator 228.
Register 230 may store a particular reference Iref (e.g., a voltage level) to be compared to the representation ID at the respective comparator 228. That is, the respective comparator 228 may receive the particular reference Iref from register 230 and compare the particular reference Iref to the representation ID. For example, comparator 2281 may receive the particular reference Iref from register 230 and compare the particular reference Iref to representation Iii that is a digital representation of the level of the sensed current IA1 on bit line 2151; comparator 2282 may receive the particular reference Iref from register 230 and compare the particular reference Iref to representation ID2 that is a digital representation of the level of the sensed current IA2 on bit line 2152; and so on, until comparator 228L receives the particular reference Iref from register 230 and compares the particular reference Iref to representation IDL that is a digital representation of the level of the sensed current IAL on bit line 215L.
For some embodiments, a representation ID that is less than or equal to the particular reference Iref may be indicative of a potential match, between data (e.g., a data feature vector) stored in a respective column 212 and input data (e.g., an input feature vector) stored in registers 245, and a representation ID that is greater than the particular reference Iref may be indicative of no match between the data stored in the respective column 212 and the data stored in registers 245. For example, memory device 100 may be configured to determine that there is a potential match in response to the representation ID being less than or equal to the particular reference Iref and no match in response to the representation ID being greater than the particular reference Iref. For example, if the representation ID1 is less than or equal to the particular reference Iref, the input feature vector is potentially an exact match to the data feature vector stored in column 2121, and if the number ID2 is less than or equal to the particular reference Iref, the input feature vector is potentially an exact match to the data feature vector stored in column 2122, whereas if the representation IDL is greater than the particular reference Iref, the input feature vector is not a potential match to the data feature vector stored in column 212L.
For other embodiments, to vary the number of potential matches and/or to identify data feature vectors that may have a greater or lesser potential of matching the input feature vector, register 230 may be programmable so that a user, for example, can program different particular references Iref into register 230. For example, larger particular references Iref, can potentially identify a larger number of data feature vectors that potentially match the input feature vector than smaller particular references Iref, in that the larger particular reference Iref allows for data feature vectors that are further away from matching the input feature vector than the smaller particular references Iref. For example, larger particular references Iref may be used when a smaller reference Iref yields little or no data feature vectors potentially matching the input feature vector. Conversely, if there are too many data feature vectors potentially matching the input feature vector, the particular reference Iref may be reduced, in that the smaller particular reference Iref requires data feature vectors potentially matching the input feature vector to be closer to the input feature vector than a larger value of the particular reference Iref.
For some embodiments, progressively reducing the particular reference Iref can act to identify the data feature vectors that are closest to the input feature vector. For other embodiments, setting the particular reference Iref to zero identifies data feature vectors that are an exact match to the input feature vector.
The level of a current IA on a respective one of bit lines 215 might not be zero, in that there may be some current leakage through a memory cell that is turned off, and thus though the string containing that memory cell and the bit line 215 coupled to that string. Therefore, the digital representation ID that is compared to Iref might not be zero. Therefore, to compensate for such leakage, Iref may be set to a certain value, e.g., corresponding to a predetermined current leakage that may occur, for some embodiments. Then, for some embodiments, if digital representation ID is less than or equal to the certain value, digital representation ID is taken to be zero. As such, register 230 is configured to store a value of Iref that compensates for current leakage.
Typical construction of a memory cell 315 may include a source 325, a drain 328, a charge-storage structure 330 (e.g., a floating gate, charge trap, etc.) that can store a charge that determines a data value of the cell, and a control gate 332, as shown in
A resistor (e.g., one or more resistors) 340 may be coupled in series with each string 310 of memory cells 315. For example, one or more resistors 340 may be between and coupled in series with a source select gate 320 and a memory cell 3151 located at an end of a respective string 310 that is opposite the end coupled to a drain select gate 322. Each source select gate 320 may selectively couple a resistor 340, and thus a respective string 310, to a source line (not shown). Although two series-coupled resistors 340 are shown in
For embodiments where more than one resistor 340 is coupled to a string 310, the resistances of the respective resistors may be same as each other or different than each other. In the example of
The overall resistance of the one or more resistors 340 coupled to each string 3101 in block 2101 may be different than the overall resistance of the one or more resistors 340 coupled to each string 3102 in block 2102, and the overall resistance of the one or more resistors 340 coupled to each string 3102 in block 2102 may be different than the overall resistance of the one or more resistors 340 coupled to each string 310N in block 210N. For example, for some embodiments, the overall resistance of the one or more resistors 340 coupled to each string in each block 210 may be different than the overall resistance of the one or more resistors 340 coupled to each string 310 in every other block 210.
The one or more resistors 340 respectively coupled in series with strings 3101, 3102, and . . . 310N respectively cause different levels of current to flow through strings 3101, 3102, and . . . 310N, e.g., when strings 3101, 3102, and . . . 310N are discharged from a common voltage level to ground. For example, the strings may be discharged through an activated source select gate 320 to a source line that may be coupled to ground.
The one or more resistors 340 coupled to each string 310 act to weight the respective string 310. Accordingly, differently weighted strings 3101 to 310N make different contributions to the overall level of the current flowing through a precharged bit line 215 coupled to and discharging through the differently weighted strings 3101 to 310N. This allows the bits of data respectively stored in the memory cells 315 of strings 3101 to 310N to be weighted differently. For example, the bits of data respectively stored in the memory cells 315 of strings 3101 to 310N can be weighted according to their significance.
For example, for some embodiments, strings 3101 to 310N may be respectively coupled to increasing resistances, and the memory cells of strings 3101 to 310N may respectively store most significant bits to least significant bits. In other words, the lower the resistance coupled to a given string, the higher the potential current flow through that string and the higher the significance of the bits of data stored in the memory cells of that string. For some embodiments, each string 3101 may be coupled to the lowest resistance and may store the most significant bits of data, and each string 310N may be coupled to the highest resistance and may store the least significant bits of data.
For some embodiments, each resistor 340 may be a transistor configured to act as a resistor. For example, the resistance of a transistor may be related to the channel length of the transistor, where the greater the channel length the greater the resistance. Therefore, the resistances on each string may be preset, e.g., during fabrication of the memory array, by fabricating the transistor, acting as a resistor, to have a predetermined channel length.
For example, strings 3101 to 310N may be respectively coupled to increasing resistances by fabricating the transistors, acting as resistors, respectively coupled in series with strings 3101 to 310N to respectively have increasing channel lengths. For embodiments, where resistors with different resistances are coupled to a particular string, the different resistances may be respectively preset by respectively coupling transistors with different channel lengths to the particular string.
For other embodiments, each resistor 340 may be a programmable resistor. For example, each resistor 340 may be configured in a manner similar to a memory cell 315. In one example, each resistor 340 may be a charge-storage cell having a charge storage structure 342 and a control gate 344 coupled to (and in some cases forming) a control line 345. For example, the control gates of resistors 3401,1, 3402,1, 3401,2, 3402,2, 3401,N, and 3402,N may be respectively coupled to control lines 3451,1, 3452,1, 3451,2, 3452,2, 3451,N, and 3452,N.
The resistance of a charge-storage cell is related to a difference between the threshold voltage programmed into the charge-storage cell and a voltage VC applied to the control gate 344 (e.g., the voltage placed on a corresponding control line 345). For example, a small difference, e.g., corresponding to charge-storage cell being partially ON, may produce a large resistance, and progressively increasing the difference, may progressively decrease the resistance until the difference is large enough that the charge-storage cell is fully ON, e.g., is placed in a read mode. For other embodiments, the each resistor 340 may be programmed to the same threshold voltage, and the voltages VC applied to the control gates may be adjusted to adjust the resistance. Control lines 345 may be coupled to drivers, such as drivers 232, that apply the voltages VC.
For some embodiments, the voltage VC that is to be applied to the control gate of a resistor 340 to effect a certain reduction in the current level may be determined from an iterative process. For example, the resistor is programmed to a certain threshold voltage and the voltage VC is adjusted until the desired current level is obtained. The voltage VC thus obtained may be subsequently used in conjunction with the threshold voltage to cause the resistor 340 to set the current flow through a corresponding string 310.
In one example, the voltages VC1,1 and VC2,1 respectively applied to control lines 3451,1 and 3452,1 and the threshold voltages of the resistors 3401,1 and 3402,1 respectively coupled to control lines 3451,1 and 3452,1 may cause the voltages VC1,1 and VC2,1 to act as pass voltages so that resistors 3401,1 and 3402,1 provide little resistance to any current flowing through the respective string 3101. As such, the resistors 3401,1 and 3402,1 may be set to reduce the current level by substantially a factor of one, e.g., substantially no reduction, and thus the strings 3101 coupled in series with resistors 3401,1 and 3402,1 may be said to have a weight factor of substantially one. For some embodiments, the strings 3101 in columns 212 may store the most significant bits of data (e.g., data feature vectors) stored in those columns.
Continuing with the example, the voltages VC1,2 and VC2,2 respectively applied to control lines 3451,2 and 3452,2 and the threshold voltages of the resistors 3401,2 and 3402,2 respectively coupled to control lines 3451,2 and 3452,2 may cause the overall (e.g., combined) resistance of the resistors 3401,2 and 3402,2 to be, for example, substantially two times the combined resistance of resistors 3401,1 and 3402,1. As such, resistors 3401,2 and 3402,2 may act to reduce the current level by substantially a factor of 2, and thus the strings 3102 coupled in series with resistors 3401,2 and 3402,2 may be said to have a weight factor of substantially ½. That is, the current flow through resistors 3401,2 and 3402,2 may be substantially a factor of two less than through resistors 3401,1 and 3402,1. For some embodiments, the strings 3102 in columns 212 may store the next-most (e.g., the second-most) significant bits of data (e.g., data feature vectors) stored in those columns.
Continuing further with the example, the voltages VC1,N and VC2,N respectively applied to control lines 3451,N and 3452,N and the threshold voltages of the resistors 3401,N and 3402,N respectively coupled to control lines 3451,N and 3452,N may cause the combined resistance of the resistors 3401,N and 3402,N coupled thereto to be, for example, substantially 2N-1 times the combined resistance of resistors 3401,1 and 3402,1. As such, resistors 3401,N and 3402,N may act to reduce the current level by substantially a factor of 2N-1, and thus the strings 310N coupled in series with resistors 3401,N and 3402,N may be said to have a weight factor of substantially ½″ That is, the current flow through resistors 3401,N and 3402,N may be substantially a factor of 2N-1 less than through resistors 3401,1 and 3402,1. For some embodiments, the strings 310N in columns 212 may store the least significant bits of data (e.g., data feature vectors) stored in those columns.
For some embodiments, bits of data of a data feature vector to be compared to like bits of data of an input feature vector may be stored in pairs of memory cells 315 so that each string 310 may store M/2 bits of data. For example, a memory-cell pair 3151, 3152 of string 3101 may store the most significant bit of a component of a data feature vector to be compared to the most significant bit of the same component of the input feature vector; a memory-cell pair 3151, 3152 of string 310N may store the least significant bit of the component of the data feature vector to be compared to the least significant bit of the component of the input feature vector; and the memory-cell pair 3151, 3152, of string 3102 may store a bit (e.g., the second-most significant bit) of the component of the data feature vector between the most significant bit and the least significant bit of the component of the data feature vector to be compared to a like bit (e.g., the second-most significant bit) of the component of the input feature vector between the most significant bit and the least significant bit of the component of the input feature vector.
Another memory-cell pair of string 3101 may store the most significant bit of another component of the data feature vector stored in the respective column 212 to be compared to the most significant bit of another component of the input feature vector; another memory-cell pair of string 3102 may store the second-most significant bit of the other component of the data feature vector stored in the respective column 212 to be compared to the second-most significant bit of the other component of the input feature vector; and another memory-cell pair of string 310N may store the least significant bit of the other component of the data feature vector stored in the respective column 212 to be compared to the least significant bit of the other component of the input feature vector. Each bit of a input feature vector may correspond to the values (e.g., bit values d and d*) stored in two registers 245 (
In the example of
Component B1 may be the most important feature in a comparison of input feature vector 400 to data feature vectors stored in memory array 204, component B2 the next-most (e.g., the second-most) important feature of input feature vector 400, and component BN the least important feature of input feature vector 400. Therefore, the bit of component B1 may be the most significant bit of a binary expression of input feature vector 400; the bit of component B2 may be the second-most significant bit of the binary expression of input feature vector 400; and the bit of component BN may be the least significant bit of the binary expression of input feature vector 400. As such, input feature vector 400 may be represented by the binary expression 011, e.g., IF(B)=011.
Note that each component (e.g., bit in this example) of input feature vector 400 may use two register bits d and d*. For example, a register bit 0 may cause a voltage VWL=2V to be applied to the word line 220 corresponding to the register containing bit 0, and a register bit 1 may cause a voltage VWL=4V to be applied to the word line 220 corresponding to the register containing bit 1.
For some embodiments, a 0 value (bit) of input feature vector 400 may be coded as a first voltage (e.g., 2V) in a first register bit and a second voltage (e.g., 4V) in a second register bit in a pair of registers 245, and a 1 value (bit) of input feature vector 400 may be coded as the second voltage (e.g., 4V) in a first register bit and the first voltage (e.g., 2V) in the second register bit in a pair of registers 245, as shown in
In the example of
For example, a data feature vector DF1(A)=(A1, A2, AN)1 may be stored in column 2121; a data feature vector DF2(A)=(A1, A2, AN)2 may be stored in column 2122; and a data feature vector DFL(A)=(A1, A2, AN)L may be stored in column 212L. Data feature vector DF1(A) may be equal to the binary expression 000; data feature vector DF2(A) may be equal to the binary expression 011; and data feature vector DFN(A) may be equal to the binary expression 111. Therefore, there is an exact match between input feature vector 400 (e.g., input feature vector IF(B)=011) and data feature vector DF2(A), a mismatch between the most significant bits of input feature vector IF(B) and data feature vector DFN(A), and a mismatch between the second-most significant bits of input feature vector IF(B) and data feature vector DF1(A) and the least significant bits of input feature vector IF(B) and data feature vector DF1(A). As discussed below, the most, second-most, and least significant bits may respectively have the weight factors of 1, ½, and ½N-1.
For some embodiments, a plurality of bits may be stored in a string of memory cells. For example, each of a plurality of memory-cell pairs in a string may store a bit. The bits stored in a string of memory cells may be bits of different components of a data feature vector. Alternatively, the bits stored in a string of memory cells may be different bits of a single component of a data feature vector.
The data stored in columns 2121, 2122, and 212L in block 2101 may represent a component (e.g., a feature) A1 of the data feature vectors stored in columns 2121, 2122, and 212L and may be in the same category (e.g., of the same type) as component B1 of input feature vector 400; the data stored in columns 2121, 2122, and 212L in block 2102 may represent a component (e.g., a feature) A2 of the data feature vectors stored in columns 2121, 2122, and 212L and may be in the same category (e.g., of the same type) as component B2 of input feature vector 400; and the data stored in columns 2121, 2122, and 212L in block 210N may represent a component (e.g., a feature) AN of the data feature vectors stored in columns 2121, 2122, and 212L and may be in the same category (e.g., of the same type) as component BN of input feature vector 400. For example input feature vector 400 and the data feature vectors stored in columns 2121, 2122, and 212L may represent people, and like components A1 and B1 may be race, like components A2 and B2 may be height, and like components AN and BN may be eye color, i.e., race, height, and eye color may be different categories.
Therefore, component B1 of input feature vector 400 (e.g., input feature vector IF(B)) will be compared to component A1 of data feature vectors DF1(A), DF2(A), and DFL(A). Component B2 of the input feature vector IF(B) will be compared to component A2 of data feature vectors DF1(A), DF2(A), and DFL(A). Component BN of the input feature vector IF(B) will be compared to component AN of data feature vectors DF1(A), DF2(A), and DFL(A).
Component A1 of the data feature vector stored in each of columns 2121, 2122, and 212L may be stored in the memory-cell pair 3151, 3152 of string 3101 in each of columns 2121, 2122, and 212L; component A2 of the data feature vector stored in each of columns 2121, 2122, and 212L may be stored in the memory-cell pair 3151, 3152 of string 3102 in each of columns 2121, 2122, and 212L; and component AN of the data feature vector stored in each of columns 2121, 2122, and 212L may be stored in memory-cell pair 3151, 3152 of string 310N in each of columns 2121, 2122, and 212L. Note that strings 3101, 3102, and 310N are respectively in blocks 2101, 2102, and 210N (
For some embodiments, the components A1, A2, and A3 of the data feature vector stored in each of columns 2121, 2122, and 212L may respectively be the most, second-most, and least important features in the comparison to input data feature vector 400, and thus may respectively be the most, second-most, and least significant bits of the data feature vector stored in each of columns 2121, 2122, and 212L. Component B1 of input data feature vector 400 is to be compared to data stored in the memory-cell pair 3151, 3152 of string 3101 in each of columns 2121, 2122, and 212L. Component B2 of input data feature vector 400 is to be compared to the data stored in the memory-cell pair 3151, 3152 of string 3102 in each of columns 2121, 2122, and 212L. Component AN of input data feature vector 400 is to be compared to the data stored in the memory-cell pair 3151, 3152 of string 310N in each of columns 2121, 2122, and 212L. That is, the most significant bit of the input data (e.g., corresponding to input feature vector 400) is compared to the most significant bit of the data stored in each of columns 2121, 2122, and 212L; the second-most significant bit of the input data is compared to the second-most significant bit of the data stored in each of columns 2121, 2122, and 212L; and the least significant bit of the input data is compared to the least significant bit of the data stored in each of columns 2121, 2122, and 212L.
Each memory cell 3151 is programmed to a threshold voltage Vt, and each memory cell 3152 is programmed to a threshold voltage Vt*. The values of threshold voltages Vt* and Vt are shown in
To compare component B1 of input feature vector 400 to the data stored in the memory-cell pair 3151, 3152 of string 3101 in each of columns 2121, 2122, and 212L, the register bits d1* and d1 corresponding to component B1 of input feature vector 400 may cause a voltage VWL1=4V (e.g., corresponding to a bit value of d1*=1) to be applied to the word line 2201 (
The memory cells 3152 in the strings 3101 of columns 2121 and 2122 are OFF, in that the voltage VWL2=2V applied to the word line 2202 in block 2121 is less than the threshold voltage (e.g., Vt1*=3V) of these memory cells 3152, and is thus insufficient to turn these memory cells 3152 ON. However, the memory cell 3152 in the string 3101 in column 212L is turned ON, in that the voltage VWL2=2V applied to the word line 2202 is greater than the threshold voltage (e.g., Vt1*=1V) on that memory cell 3152, and is sufficient to turn that memory cell 3152 ON. The memory cells 3151 in the strings 3101 of columns 2121, 2122, and 212L are all ON, in that the voltage VWL1=4V applied to the word line 2201 in block 2121 is greater than the threshold voltage on those memory cells 3151 (e.g., Vt1=1V for the memory cells 3151 in columns 2121 and 2122 and Vt1=3V for the memory cell 3151 in column 212L), and is sufficient to turn those memory cells 3151 ON.
For some embodiments, a memory cell 3152 having a threshold voltage of Vt1*=3V stores a bit value of 1, and a memory cell 3151 having a threshold voltage of Vt1=1V stores a bit value of 0. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 0 when memory cell 3151 stores a bit value of 0 and memory cell 3152 stores a bit value of 1. A memory cell 3152 having a threshold voltage of Vt1*=1V stores a bit value of 0, and a memory cell 3151 having a threshold voltage of Vt1=3V stores a bit value of 1. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 1 when memory cell 3152 stores a bit value of 0 and memory cell 3151 stores a bit value of 1. As such, component A1 of the data feature vector stored in column 2121 has a binary value of 0, which matches component B1 of the input feature vector. Component A1 of the data feature vector stored in column 2122 has a binary value of 0, which matches component B1 of the input feature vector. Component A1 of the data feature vector stored in column 212L has a binary value of 1, which does not match (e.g., mismatches) component B1 of the input feature vector.
Therefore, the ON, OFF status of the memory-cell pairs 3151, 3152 in the strings 3101 of columns 2121 and 2122 represents a match between component B1 of input feature vector 400 and the data stored in the memory-cell pairs 3151, 3152 in the strings 3101 of columns 2121 and 2122, i.e., the match between component B1 of input feature vector 400 and the component A1 of the data feature vectors stored in columns 2121 and 2122. The ON, ON status of the memory-cell pair 3151, 3152 in the string 3101 of column 212L represents a mismatch between component B1 of input feature vector 400 and the data stored in the memory-cell pair 3151, 3152 in the string 3101 of column 212L, i.e., the mismatch between component B1 of input feature vector 400 and the component A1 of the data feature vector stored in column 212L.
For some embodiments, the bit lines 2151, 2152, and 215L (
Since component A1 of the data feature vector stored in each of columns 2121, 2122, and 212L is the most significant bit of the data stored in each of columns 2121, 2122, and 212L and is stored in the memory-cell pair 3151, 3152 of string 3101 in each of columns 2121, 2122, and 212L, the string 3101 in each of columns 2121, 2122, and 212L, and thus block 2101 containing string 3101, may be weighted with the largest weight factor of the strings 3101, 3102, 310N and thus of the blocks 2101, 2102, 210N, e.g., a weight factor of 1. The weight factor of 1 may signify that component A1 of the data feature vector stored in each of columns 2121, 2122, and 212L is the most significant bit of the data stored in each of columns 2121, 2122, and 212L.
To set the weight factor of 1, the one or more resistors 340 coupled in series with the string 3101 in each of columns 2121, 2122, and 212L may be adjusted so that there is substantially no added resistance due to one or more resistors 340. Therefore, the level of the current on charged bit line 215L that discharges through the string 3101 in column 212L is reduced by a factor of substantially one, e.g., substantially no reduction.
For embodiments where the one or more resistors 340 are configured as charge-storage cells, the voltage applied to the control gates thereof may cause the one or more resistors 340 to operate as pass transistors to produce the weight factor of 1. For example, the series-coupled resistors 3401,1 and 3401,2 (
Continuing with the example of
The memory cells 3152 in the strings 3102 of columns 2121, 2122, and 212L are all ON, in that the voltage VWL2=4V applied to the word line 2202 in block 2122 is greater than the threshold voltage on those memory cells 3152 (e.g., Vt2*=1V for the memory cells 3152 in columns 2122 and 212L and Vt2*=3V for the memory cell 3152 in column 2121), and is sufficient to turn those memory cells 3152 ON. The memory cells 3151 in the strings 3102 of columns 2122 and 212L are OFF, in that the voltage VWL1=2V applied to the word line 2201 in block 2122 is less than the threshold voltage of these memory cells 3152 (e.g., Vt2=3V), and is thus insufficient to turn these memory cells 3151 ON. However, the memory cell 3151 in the string 3102 in column 2121 is turned ON, in that the voltage VWL1=2V applied to the word line 2201 in block 2122 is greater than the threshold voltage on that memory cell 3151 (e.g., Vt2=1V), and is sufficient to turn that memory cell 3151 ON.
For some embodiments, a memory cell 3152 having a threshold voltage of Vt2*=3V stores a bit value of 1, and a memory cell 3151 having a threshold voltage of Vt2=1V stores a bit value of 0. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 0 when memory cell 3152 stores a bit value of 1 and memory cell 3151 stores a bit value of 0. A memory cell 3152 having a threshold voltage of Vt2*=1V stores a bit value of 0, and a memory cell 3151 having a threshold voltage of Vt2=3V stores a bit value of 1. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 1 when memory cell 3152 stores a bit value of 0 and memory cell 3151 stores a bit value of 1. As such, components A2 of the data feature vectors stored in columns 2122 and 212L have binary values of 1, which match component B2 of the input feature vector 400. Component A2 of the data feature vector stored in column 2121 has a binary value of 0, which does not match (e.g., mismatches) component B2 of the input feature vector 400.
Therefore, the OFF, ON status of the memory-cell pairs 3151, 3152 in the strings 3102 of columns 2122 and 212L represents a match between component B2 of the input feature vector 400 and the data stored in the memory-cell pairs 3151, 3152 in the strings 3102 of columns 2122 and 212L, i.e., the match between component B2 of input feature vector 400 and the component A2 of the data feature vectors stored in columns 2122 and 212L. The ON, ON status of the memory-cell pair 3151, 3152 in the string 3102 of column 2121 represents a mismatch between component B2 of the input feature vector 400 and the data stored in the memory-cell pair 3151, 3152 in the string 3102 of column 2121, i.e., the mismatch between component B2 of the input feature vector 400 and the component A2 of the data feature vector stored in column 2121.
The OFF status of the memory cells 3151 in the strings 3102 of columns 2122 and 212L acts to prevent the charged bit lines 2152 and 215L from discharging through the strings 3102 in columns 2122 and 212L, signaling the match. However, the ON status of both memory cell 3151 and memory cell 3152 in the string 3102 of column 2121 allows the charged bit line 2151 to discharge through the string 3102 in column 212L, signaling the mismatch.
Since component A2 of the data feature vector stored in each of columns 2121, 2122, and 212L is the second-most significant bit of the data stored in each of columns 2121, 2122, and 212L and is stored in the memory-cell pair 3151, 3152 of string 3102 in each of columns 2121, 2122, and 212L, the string 3102 in each of columns 2121, 2122, and 212L, and thus block 2102, may be weighted with second largest weight factor of the strings 3101, 3102, 310N, and thus of the blocks 2101, 2102, 210N, e.g., a weight factor of ½. The weight factor of ½ may signify that component A2 of the data feature vector stored in each of columns 2121, 2122, and 212L is the second-most significant bit of the data stored in each of columns 2121, 2122, and 212.
To set the weight factor of ½, the one or more resistors 340 coupled in series with the string 3102 in each of columns 2121, 2122, and 212L may be adjusted so that the level of any current discharging through strings 3102 in columns 2121, 2122, and 212L is reduced by substantially a factor of 2. Therefore, the level of the current on the charged bit line 2151 that discharges through the string 3102 in column 2121 is reduced by a factor of substantially 2.
For embodiments where the one or more resistors 340 are configured as charge-storage cells, the voltage applied to the control gates thereof in conjunction with the threshold voltage programmed in the one or more resistors 340 may cause the one or more resistors 340 to turn partially on so as to reduce the current flow by substantially the factor of 2 and thus produce the weight factor of ½. For example, the difference between the voltage applied to the control gate of a resistor coupled in series with the string 3102 in each of columns 2121, 2122, and 212L and the threshold voltage of that resistor may be different (e.g., less) than the difference between the voltage applied to the control gate of a resistor coupled in series with the string 3101 in each of columns 2121, 2122, and 212L and the threshold voltage of that resistor.
In embodiments where the series-coupled resistors 3401,2 and 3402,2 (
Continuing with the example of
Note that the blocks 2103 to 210N-1 between blocks 2102 and 210N may be masked, e.g., by mask bits stored in mask registers 240 (
The memory cells 3152 in the strings 310N of columns 2121, 2122, and 212L are all ON, in that the voltage VWL2=4V applied to the word line 2202 in block 212N is greater than the threshold voltage on those memory cells 3152 (e.g., VtN*=1 Von the memory cells 3152 in columns 2122 and 212L and VtN*=3V on the memory cell 3152 in column 2121), and is sufficient to turn those memory cells 3152 ON. The memory cells 3151 in the strings 310N of columns 2122 and 212L are OFF, in that the voltage VWL1=2V applied to the word line 2201 in block 212N is less than the threshold voltage of these memory cells 3151 (e.g., VtN=3V), and is thus insufficient to turn these memory cells 3152 ON. However, the memory cell 3151 in the string 310N in column 2121 is turned ON, in that the voltage VWL1=2V applied to the word line 2201 in block 212N is greater than the threshold voltage on that memory cell 3151 (e.g., VtN=1V), and is sufficient to turn that memory cell 3151 ON.
For some embodiments, a memory cell 3152 having a threshold voltage of VtN*=3V stores a bit value of 1, and a memory cell 3151 having a threshold voltage of VtN=1V stores a bit value of 0. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 0 when memory cell 3151 stores a bit value of 1 and memory cell 3152 stores a bit value of 0. A memory cell 3152 having a threshold voltage of VtN*=1V stores a bit value of 0, and a memory cell 3151 having a threshold voltage of VtN=3V stores a bit value of 1. Therefore, the memory-cell pair 3151, 3152 stores a value of binary 1 when memory cell 3152 stores a bit value of 0 and memory cell 3151 stores a bit value of 1. As such, components AN of the data feature vectors stored in columns 2122 and 212L have binary values of 1, which match component BN of the input feature vector 400. Component AN of the data feature vector stored in column 2121 has a binary value of 0, which does not match (e.g., mismatches) component BN of the input feature vector 400.
The OFF, ON status of the memory-cell pairs 3151, 3152 in the strings 310N of columns 2122 and 212L represents a match between component AN of input feature vector 400 and the data stored in the memory-cell pairs 3151, 3152 in the strings 310N of columns 2122 and 212L, i.e., the match between component BN of input feature vector 400 and the component AN of the data feature vectors stored in columns 2122 and 212L. The ON, ON status of the memory-cell pair 3151, 3152 in the string 310N of column 2121 represents a mismatch between component BN of input feature vector 400 and the data stored in the memory-cell pair 3151, 3152 in the string 310N of column 2121, i.e., a mismatch between component BN of input feature vector 400 and the component AN of the data feature vector stored in column 2121.
The OFF status of the memory cells 3151 in the strings 310N of columns 2122 and 212L acts to prevent the charged bit lines 2152 and 215L from discharging through the strings 310N in columns 2122 and 212L, signaling the match. However, the ON status of both memory cell 3151 and memory cell 3152 in the string 310N of column 2121 allows the charged bit line 2151 to discharge through the string 310N in column 2121L, signaling the mismatch.
Since component AN of the data feature vector stored in each of columns 2121, 2122, and 212L is the least significant bit of the data stored in each of columns 2121, 2122, and 212L and is stored in the memory-cell pair 3151, 3152 of string 310N in each of columns 2121, 2122, and 212L, the string 310N in each of columns 2121, 2122, and 212L, and thus block 210N, may be weighted with smallest weight factor of the strings 3101, 3102, 310N, and thus of the blocks 2101, 2102, 210N, e.g., a weight factor of ½N-1, where N is greater than 2. The weight factor of ½N-1 may signify that component AN of the data stored in each of columns 2121, 2122, and 212L is the least significant bit of the data stored in each of columns 2121, 2122, and 212.
To set the weight factor of ½N-1, the one or more resistors 340 coupled in series with the string 310N in each of columns 2121, 2122, and 212L may be adjusted so that the level of any current discharging through strings 310N in columns 2121, 2122, and 212L is reduced by substantially a factor of 2N-1. Therefore, the level of the current on charged bit line 2151 that discharges through the string 310N in column 2121 is reduced by a factor of substantially 2N-1.
For embodiments where the one or more resistors 340 are configured as charge-storage cells, the voltage applied to the control gates thereof in conjunction with the threshold voltage programmed in the one or more resistors 340 may cause the one or more resistors 340 to turn partially on so as to reduce the current flow by substantially the factor of 2N-1 and thus produce the weight factor of ½N-1. For example, the difference between the voltage applied to the control gate of a resistor coupled in series with the string 310N in each of columns 2121, 2122, and 212L and the threshold voltage of that resistor may be different (e.g., less) than the difference between the voltage applied to the control gate of a resistor coupled in series with the string 3101 in each of columns 2121, 2122, and 212L and the threshold voltage of that resistor and different (e.g., less) than the difference between the voltage applied to the control gate of a resistor coupled in series with the string 3102 in each of columns 2121, 2122, and 212L and the threshold voltage of that resistor.
In embodiments where the series-coupled resistors 3401,N and 3402,N (
For some embodiments, the voltages VCL1,N and VC2,N may be respectively applied to control lines 3451,N and 3452,N substantially concurrently (e.g., concurrently) with applying the voltage VWL1=2V the word line 2201 in block 212N and applying the voltage VWL2=4V the word line 2202 in block 212N, as well as substantially concurrently (e.g., concurrently) with applying the respective the voltages VC1,2 and VC2,2 to control lines 3451,2 and 3452,2 in block 2102 and with respectively applying the respective the voltages VC1,1 and VC2,1 to control lines 3451,1 and 3452,1 in block 2101, and thus substantially concurrently (e.g., concurrently) with applying the voltage VWL1=2V to the word line 2201 in block 2102 and applying the voltage VWL2=4V to the word line 2202 in block 2102 and with applying the voltage VWL1=4V to the word line 2201 in block 2101 and applying the voltage VWL2=2V to the word line 2202 in block 2101. Therefore, the weight factors may applied concurrently to the blocks 2101, 2102, and 210N and thus to strings 3101, 3102, and 310N.
The data stored in the memory-cell pairs 3151, 3152 in the strings 3101, 3102, and 310N in column 2122 respectively matches components B1, B2, and BN of input feature vector 400, i.e., the components A1, A2, and AN of the data feature vector DF2(A) stored in column 2122 respectively match the components B1, B2, and BN of input feature vector 400. Therefore, there is an exact match between input feature vector 400 and the data feature vector DF2(A) stored in the memory-cell pairs 3151, 3152 in column 2122. This is reflected by the fact that a memory cell of the pairs 3151, 3152 in each of the strings 3101, 3102, and 310N in column 2122 is OFF (i.e., memory cell 3152 in string 3101, memory cell 3151 in string 3102, and memory cell 3151 in string 310N are OFF). This means that charged bit line 2152 coupled to the strings 3101, 3102, and 310N in column 2122 is prevented from being discharged through the strings 3101, 3102, and 310N in column 2122, meaning there is no current flow through bit line 2152, and the level of the current IA2 (
Therefore, in the event of an exact match between an input feature vector and a data feature vector stored in the memory cells of a column, there is no current flow on the charged bit line coupled to those memory cells. That is, an exact match can be determined by detecting zero current flow through a charged bit line coupled memory cells whose data are compared with an input feature vector.
The components A2 and AN of the data feature vector DFL(A) stored in column 212L respectively match the components B2 and BN of input feature vector 400. That is, the data stored in the memory-cell pairs 3151, 3152 in the strings 3102 and 310N in column 212L respectively match components B2 and BN of input feature vector 400, and current is prevented from flowing from charged bit line 215L though the strings 3102 and 310N in column 212L by the turned-off memory cells 3151 in the strings 3102 and 310N in column 212L.
However, component A1 of the data feature vector DFL(A) stored in column 212L does not match component B1 of input feature vector 400. That is, the data stored in the memory-cell pair 3151, 3152 in the string 3101 in column 212L does not match component B1 of input feature vector 400, and current IAL (
Component A1 of the data feature vector DF1(A) stored in column 2121 matches component B1 of input feature vector 400. That is, the data stored in the memory-cell pairs 3151, 3152 in the string 3101 in column 2121 matches component B1 of input feature vector 400, and current is prevented from flowing from charged bit line 2151 though the string 3101 in column 2121 by the turned-off memory cell 3152 in the string 3101 in column 2121.
Component A2 of the data feature vector DF1(A) stored in column 2121 does not match component B2 of input feature vector 400. That is, the data stored in the memory-cell pair 3151, 3152 in the string 3102 in column 2121 does not match component B2 of input feature vector 400, and current flows from charged bit line 2151 through the string 3102 in column 2121. Component AN of the data feature vector DF1(A) stored in column 2121 does not match component BN of input feature vector 400. That is, the data stored in the memory-cell pair 3151, 3152 in the string 310N in column 2121 does not match component BN of input feature vector 400, and current also flows from charged bit line 2151 through the string 310N in column 2121. The total current IA1 (
The level of the current IAL on bit line 215L is greater than the level of the current IA1 on bit line 2151 when N is greater than 2. The current IAL on bit line 215L is the current through string 3101 in column 212L, and the level of this current is substantially unreduced due to the weight factor of 1 for block 2101 and string 3101. Due to the weight factor of ½ for block 2102, and thus string 3102, the level of the current flowing through string 3102 is about ½ the current on bit line 2151. Due to the weight factor of ½N-1 for block 210N, and thus string 310N, the level of the current flowing through string 310N is about ½N-1 the current on bit line 2151. Therefore, the level of the total current IA1 on bit line 2151 (the sum of the current flowing through strings 3102 and 310N) is about (½+½N-1) of the level of the current on bit line 215L.
The level of the current on a bit line is indicative of the degree (e.g., closeness) of a match between the data (e.g., data feature vectors) stored in memory cells coupled to the bit line and input data (e.g., a data feature vector), e.g., the lower the level the better (e.g., closer) the match. The level of the current IA2 on bit line 2152 is zero, in that there is an exact match between the data stored in the memory-cell pairs coupled to bit line 2152 and input feature vector 400.
The mismatch between the data feature vector DFL(A) stored in memory-cell pairs coupled to bit line 215L and input feature vector 400 occurred between the component B1 of input feature vector 400, i.e., the most important component of input feature vector 400 to the comparison, and the component A1 of the data feature vector DFL(A) stored in a memory-cell pair in block 2101 with a weight factor of 1, i.e., the component A1 of the data feature vector DFL(A) in column 212L that is most important to the comparison.
The mismatch between the data feature vector DF1(A) stored in memory-cell pairs coupled to bit line 2151 and input feature vector 400 occurred between component B2 of input feature vector 400, i.e., the second most important component of input feature vector 400 to the comparison, and the component A2 of the data feature vector DF1(A) stored in a memory-cell pair in block 2102 with a weight factor of ½, i.e., the component A2 of the data feature vector DF1(A) in column 2121 that is second most important to the comparison.
The mismatch between the data feature vector DF1(A) stored in memory-cell pairs coupled to bit line 2151 and input feature vector 400 further occurred between the component BN of input feature vector 400, i.e., the least important component of input feature vector 400 to the comparison, and the component AN of the data feature vector DF1(A) stored in a memory-cell pair in block 210N with a weight factor of ½N-1, i.e., the component AN of the data feature vector DF1(N) stored in column 2121 least important to the comparison.
Therefore, the data feature vector DF1(A) stored in the memory-cell pairs in column 2121 coupled to bit line 2151 is closer to matching (e.g., more likely to match) input feature vector 400 than the data feature vector DFL(A) stored in the memory-cell pairs in column 212L coupled to bit line 215L. This is evident by the fact that the current level in bit line 2151 is lower than the current level in bit line 215L.
The current levels IA1, IA2, and IAL, respectively in bit lines 2151, 2152, and 215L may be respectively converted to the digital representations ID1, ID2, and IDL respectively at sense amps 2251, 2252, and 225L as described above in conjunction with
If it is deemed that a mismatch between input feature vector 400 and the data feature vectors stored in columns 212 occurs when component B1 of input feature vector 400 mismatches the component A1 of the data feature vectors stored in in strings 3101 of columns 212, the particular reference Iref may be selected to be less than a digital representation of the level of the unweighted current in strings 3101 of columns 212. For example, the particular reference Iref may be selected to be less than a digital representation of the level of the current through string 3101 in column 212L, e.g., less than the representation IDL of the level of the current IAL on bit line 215L, due to both of memory cells 3151 and 3152 in string 3101 of column 212L being ON and block 2101 and string 3101 of column 212L having a weight factor of one. That is, comparator 228L will indicate that the representation IDL exceeds the particular reference Iref, thereby indicating a mismatch, as indicated above in conjunction with
Selecting the particular reference Iref to be less than representation IDL and greater than the representation hi, representing the level of the current of IA1, will cause the data feature vector DFL(A) stored in column 212L to be identified as mismatching input feature vector 400 and the data feature vectors DFL(A) and DFL(A) stored in columns 2121 and 2122 to be identified as potentially matching input feature vector 400. This differentiates the data feature vector stored in column 212L from the data feature vectors stored in columns 2121 and 2122. The differentiation of the data feature vector stored in column 212L from the data feature vector stored in column 2121 is enabled by the weighting of the strings in the respective columns.
Note that the data feature vectors stored in both of columns 2121 and 212L mismatches the data stored in input feature vector 400. However, the mismatch between the data feature vector stored in column 212L and input feature vector 400 is more critical to the comparison between the data feature vectors stored in columns 212 and input feature vector 400 than the mismatch between the data feature vector stored in column 2121 and input feature vector 400. The mismatch between the data feature vector stored in column 212L and input feature vector 400 is due to a mismatch between component B1 of input feature vector 400 and the component A1 of the data feature vector stored in column 212L that has a weight factor of 1. The mismatch between the data feature vector stored in column 2121 and input feature vector 400 is due to a mismatch between component B2 of input feature vector 400 and the component A2 of the data feature vector stored in column 2121 that has a weight factor of ½ and a mismatch between component BN of input feature vector 400 and the component AN of the data feature vector stored in column 2121 that has a weight factor of ½N-1.
Selecting the particular reference Iref to be less than representation hi and greater than the representation ID2, representing the level of the current of IA2 (=0), will cause the data feature vectors stored in columns 2121 and 212L to be identified as mismatching input feature vector 400 and the data feature vector stored in column 2122 to be identified as potentially matching input feature vector 400. Note that the data feature vector stored in column 2122 matches input feature vector 400 exactly, but an exact match can only be identified by selecting the particular reference Iref to be zero and the representation ID2 being equal thereto.
Note that there may be some current in a string even though a memory cell in that string is turned off, owing to current leakage through the turned-off memory cell. As such, the current of IA2 may be non-zero in spite of the memory cell in the string being turned off. Therefore, an exact match will not be indicated by comparing the representation ID2, representing the level of the current of IA2, to Iref=zero. Therefore, register 230 may configured to store a certain value of Iref that compensates for the leakage, and memory device 100 may be configured to determine that the representation ID2 is zero when ID2 is less than or equal to the certain value.
To further illustrate that lower bit-line currents indicate closer matches, it is worthwhile to consider an example where the memory cell 3152 in string 3102 in column 2121 has a threshold voltage of 1V instead 3V, as shown in brackets in
In this example, memory cell 3152 remains ON in response to applying the voltage VWL2=4V to the word line 2202 in block 2102, but memory cell 3151 is OFF in response to applying the voltage VWL1=2V to the word line 2201 in block 2102.
The data feature vector DF1′(A)=(A1, A2′, AN)1=010 matches input feature vector 400 (IF(B)=(B1, B2, BN)=011) more closely than data feature vector DF1 (A)=(A1, A2, AN)1=000 that was previously stored in column 2121 and was discussed above. This is because the mismatch between data feature vector DF1′(A) and input feature vector IF(B) occurs only in the least significant bit, whereas the mismatch between data feature vector DF1 (A) and input feature vector IF(B) occurs both in the second-most significant bit and the in the least significant bit.
In the case of data feature vector DF1′(A), the current from charged bit line 2151 is prevented from flowing through string 3101 of column 2121, in that memory cell 3152 in string 3101 is OFF. Current from charged bit line 2151 is also prevented from flowing through string 3102 of column 2121, in that memory cell 3151 in string 3102 is OFF. Therefore, the current from charged bit line 2151 only flows through string 310N, in that both memory cell 3151 and memory cell 3152 in string 310N of column 2121 are ON.
However, in the case of data feature vector DF1′(A), the current from charged bit line 2151 that flows through string 310N is reduced by a factor of 2N-1, owing to the weight factor of ½″ on block 210N and string 310N. This current is now the only contribution to the current I′A1 (
Current level I′A1 may be subsequently converted to a representation I′D1, representing the level of the current I′A1, at sense amp 2251 and compared to the particular reference Iref at comparator 2281, as shown in
Bits b11, b12, b13, b21, b22, b23, b31, b32, and b23 may be stored in input buffer 120, as shown in
Input feature vector IF(B)=(B1, B2, B3) is to be compared to data feature vector DF(A)=(A1, A2, A3) that is stored in a column of blocks of memory cells in array 204. Component A1 may have bits a11, a12, and a13, where the first number in the subscript denotes the component number and the second number in the subscript denotes the bit number, e.g., j. Bits a11, a12, and a13 may be respectively the most significant bit (e.g., bit number j=1), the second-most significant bit (e.g., bit number j=2), and the least significant bit (e.g., bit number j=3) of component A1. Component A2 may have bits a11, a22, and a23. Bits a21, a22, and a23 may be respectively the most significant bit (e.g., bit number j=1), the second-most significant bit (e.g., bit number j=2), and the least significant bit (e.g., bit number j=3) of component A2. Component A3 may have bits a31, a32, and a323. Bits a31, a32, and a33 may be respectively the most significant bit (e.g., bit number j=1), the second-most significant bit (e.g., bit number j=2), and the least significant bit (e.g., bit number j=3) of component A3.
Bits a11, a12, a13, a21, a22, a23, a31, a32, and a23 are respectively stored on memory blocks 1 to 9 of the column, as shown in
Bits b11, b12, b13, b21, b22, b23, b31, b32, and b23 are to be compared one-to-one to bits a11, a12, a13, a21, a22, a23, a31, a32, and a23. The bits a may be weighted according to their position within a specific component, e.g., according to their bit number j. For example, the weight factor may be ½j-1, starting with j=1 for the most significant bit. Therefore, bits a11, a12, and an of component A1 may respectively have weight factors of 1, ½, and ¼; bits a11, a22, and a23 of component A2 may respectively have weight factors of 1, ½, and ¼; and bits a31, a32, and a33 of component A3 may respectively have weight factors of 1, ½, and ¼, as shown in
Alternatively, the bits a may be weighted according to the block number in which they are stored, e.g., the weight factor may be ½N-1, starting with N=1. Therefore, bits a11, a12, a13, a21, a22, a23, a31, a32, and a23 respectively have weight factors of 1, ½, ¼, ⅛, 1/16, 1/32, 1/64, 1/128, and 1/256, as shown in
Although the examples of
Memory device 602 includes an array of memory cells 204, row access circuitry 608 (e.g., that may include an input buffer, such as input buffer 120 in
The memory device 602 receives control signals (which represent commands) from the processor 620 over a control link 622. Memory device 602 receives data signals (which represent data) over a data (DQ) link 624. Address signals (which represent addresses) are received via an address link 626 that are decoded to access the memory array 204. Address buffer circuit 616 latches the address signals. The memory cells in array 204 are accessed in response to the control signals and the address signals.
For some embodiments, control circuitry 612 may be configured to determine whether a input data at least partially matches a data stored in memory array 204, as described above in conjunction with
It will be appreciated by those skilled in the art that additional circuitry and signals can be provided, and that the memory device of
Although specific embodiments have been illustrated and described herein, it will be appreciated by those of ordinary skill in the art that any arrangement that is calculated to achieve the same purpose may be substituted for the specific embodiments shown. Many adaptations of the embodiments will be apparent to those of ordinary skill in the art. Accordingly, this application is intended to cover any adaptations or variations of the embodiments.
This application is a continuation of U.S. patent application Ser. No. 14/798,845, titled “MEMORY DEVICES CONFIGURED TO APPLY DIFFERENT WEIGHTS TO DIFFERENT STRINGS OF MEMORY CELLS COUPLED TO A DATA LINE AND METHODS,” filed Jul. 14, 2015 (allowed), which is divisional of U.S. patent application Ser. No. 13/864,659, titled “MEMORY DEVICES CONFIGURED TO APPLY DIFFERENT WEIGHTS TO DIFFERENT STRINGS OF MEMORY CELLS COUPLED TO A DATA LINE AND METHODS,” filed Apr. 17, 2013, now U.S. Pat. No. 9,105,330 issued Aug. 11, 2015, which is commonly assigned and incorporated herein by reference in its entirety and which claims priority to U.S. Provisional Patent Application Ser. No. 61/625,286, filed Apr. 17, 2012, titled “MEMORY DEVICES CONFIGURED TO APPLY DIFFERENT WEIGHTS TO DIFFERENT STRINGS OF MEMORY CELLS COUPLED TO A DATA LINE AND METHODS,” which is incorporated herein by reference in its entirety, and is related to U.S. patent application Ser. No. 13/449,082, titled “METHODS AND APPARATUS FOR PATTERN MATCHING,” filed Apr. 17, 2012, which claims priority to U.S. Provisional Patent Application Ser. No. 61/476,574, titled “METHODS AND APPARATUS FOR PATTERN MATCHING,” filed Apr. 18, 2011.
Number | Date | Country | |
---|---|---|---|
61625286 | Apr 2012 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 13864659 | Apr 2013 | US |
Child | 14798845 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 14798845 | Jul 2015 | US |
Child | 15645009 | US |