KR20110082016A - Methods and apparatus for soft data generation for memory devices using reference cells - Google Patents

Abstract

A method and apparatus for generating soft data for a memory device using a reference cell are provided. Writing known data into one or more reference cells, reading one or more of the reference cells, obtaining read statistics based on the read one or more reference cells, and at least one based on the obtained read statistics. Acquiring a soft data value of at least one soft data value is generated in the memory device. Read statistics may optionally be obtained for one or more desired locations in the memory array or for a given pattern (PATT) in one or more attacker cells. The read statistics may optionally include asymmetric statistics obtained for a plurality of possible values.

Description

Soft data value generation method and soft data value generation system {METHODS AND APPARATUS FOR SOFT DATA GENERATION FOR MEMORY DEVICES USING REFERENCE CELLS}

Cross Reference of Related Application

This application is filed in US Patent Provisional Application Nos. 61 / 194,751 and 2009.6.30, filed Sep. 30, 2008 and entitled "Methods and Apparatus for Soft Demapping and Intercell Interference Mitigation in Flash Memories." Claims priority of 49333, each of which is incorporated herein by reference.

The present application is an international patent application named "Methods and Apparatus for Soft Data Generation for Memory Devices Based on Performance Factor Adjustment", an international patent application named "Methods and Apparatus for Soft Data Generation for Memory Devices", and "Methods and Apparatus for Soft Data Generation for Memory Devices Using Decoder Performance Feedback ", each of which is filed concurrently with the present application and incorporated herein by reference.

Field of invention

FIELD OF THE INVENTION The present invention relates generally to flash memory devices, and more particularly to improved intercell interference effects, back pattern dependencies, backlash and other distortions in such flash memory devices. Soft demapping and soft data generation techniques.

Many memory devices, such as flash memory devices, use analog memory cells to store data. Each memory cell stores an analog value, such as a charge or voltage, also referred to as a stored value. The stored value represents the information stored in the cell. In flash memory devices, for example, each analog memory cell typically stores a specific voltage. The range of possible analog values for each cell is typically divided into threshold regions, each region corresponding to one or more data bit values. Data is written to an analog memory cell by writing a nominal analog value corresponding to one or more desired bits.

For example, a single-level cell (SLC) flash memory device stores one bit per memory cell (or two possible memory states). Multi-level cell (MLC) flash memory devices, on the other hand, store two or more bits per memory cell (ie, each cell has four or more programmable states). For a more detailed discussion of MLC flash memory devices, see, for example, the method filed March 31, 2009, which is incorporated herein by reference, "Methods and Apparatus for Storing Data in a Multi-Level Cell Flash Memory Device with Cross." -International Patent Application PCT / US09 / 36810, "Page Sectors, Multi-Page Coding And Per-Page Coding".

For example, in a multilevel NAND flash memory device, a floating gate device is employed with a programmable threshold voltage in a range divided by a number of intervals each corresponding to a different multibit value. To program a given multi-bit value into a memory cell, the threshold voltage of the floating gate device in the memory cell is programmed to a threshold voltage interval corresponding to that value.

Analog values stored in memory cells are often distorted. Distortion is typically due to, for example, back pattern dependency (BPD), noise, and intercell interference (ICI). For a more detailed discussion of distortion in flash memory devices, see, for example, JD Lee et al., "Effects of Floating-Gate Interference on NAND Flash Memory Cell Operation," IEEE Electron Device Letters, 264-266 (March 2002) or Ki-Tae Park, et al., "A Zeroing Cell-to-Cell Interference Page Architecture With Temporary LSB Storing and Parallel MSB Program Scheme for MLC NAND Flash Memories," IEEE J. of Solid State Circuits, Vol. 43, No. 4, 919-928, (April 2008).

Many techniques have been proposed or suggested to reduce the effects of ICI and other disturbances. For example, Ki-Tae Park et al. Describe existing programming techniques such as even / odd programming, bottom up programming and multistage programming that reduce ICI. International Patent Application PCT / US09 / 49333, entitled "Methods and Apparatus for Soft Demapping and Intercell Interference Mitigation in Flash Memories," filed June 30, 2009, discloses a method and apparatus for soft demapping and disturbance reduction in flash memory. Initiate.

These existing methods have helped to improve the decoding performance of flash memory, but they have many limitations, and overcoming them will further improve the reliability of flash memory. For example, modern flash memories typically only provide hard data to the flash control system for decoding. However, it is well known that soft data can improve error rate performance in the decoding process. Therefore, there is a need for soft data generation techniques that use hard data from flash memory to estimate or enhance soft data and thereby improve decoding performance.

Generally, a method and apparatus for generating soft data for a memory device using a reference cell is provided. According to one aspect of the present invention, there is provided a method of recording data known to at least one reference cell, reading at least one of the reference cells, acquiring read statistics based on the read at least one reference cell; At least one soft data value is generated in the memory device by obtaining the at least one soft data value based on the obtained read statistics. Known data can be written to a reference cell using one or more of voltage, current and resistance. Similarly, the read step reads one or more of data bits, voltage levels, current levels, and resistance levels. The known data may comprise a bit pattern or a symbol pattern. The generated soft data value may comprise (i) a soft read value used to generate one or more log likelihood ratios (LLRs) or (ii) one or more LLRs. The statistics may include one or more of bit based statistics, cell based and pattern dependent statistics.

In one embodiment, the read statistics are determined by comparing one or more bits read from the one or more reference cells with one or more bits in the known data. For example, the read statistics may be based on the number of faulty bits relative to the total number of reference bits in the known data. The read statistics or the soft data values (or both) may include endurance, number of program / erase cycles, number of read cycles, retention time, temperature, temperature change, process corners, intercell interference effects, It can be calculated as a function of one or more of a location in the memory device, a location in a wordline, a location in a memory array, or a pattern in an attacker cell.

According to one aspect, the read statistics can be obtained for one or more desired locations of a memory array, wherein the writing comprises writing the known data to the one or more desired locations. For example, the desired location may include one or more of a given bit, page location, wordline location, and bitline location within a multilevel cell.

According to another aspect, the read statistics may be obtained for a given pattern (PATT) in one or more attacker cells associated with at least one target cell, wherein the writing step comprises writing the known data to the at least one target cell. And the method comprises determining a number of erroneous target bits having the given pattern (PATT) in the one or more attacker cells and reading the for the given pattern based on the number of erroneous target bits. Obtaining statistics.

According to another aspect, the read statistics may comprise asymmetric read statistics obtained for a plurality of possible values, the method comprising determining a number of erroneous bits having a first value in the known data; Determining the number of erroneous bits having a second value in the known data, obtaining read statistics for the first value based on the number of erroneous bits having the first value; Obtaining read statistics for the second value based on the number of erroneous bits having the second value.

A more complete understanding of the present invention and further features and advantages of the present invention will be obtained with reference to the following detailed description and drawings.

1 is a schematic block diagram of a conventional flash memory system.
FIG. 2 illustrates an example threshold voltage distribution for the example flash memory of FIG. 1.
3 illustrates the architecture of an exemplary flash cell array in an MLC flash memory device.
4 illustrates an exemplary two-step MLC programming scheme for the voltage allocation scheme of FIG. 2.
5A and 5B together show another MLC programming scheme that reduces the ICI applied to neighboring cells.
6 illustrates an example flash cell array in more detail in an MLC flash memory device.
7 illustrates disturbances such as intercell interference, back pattern dependence, noise and other distortions present in the target cell due to many exemplary attacker cells.
8 is a schematic block diagram of an exemplary flash memory system including a controller-based soft de-mapping / soft data generation technique in accordance with the present invention.
9 is a schematic block diagram of an exemplary flash memory system that includes a memory based soft de-mapping / soft data generation technique in accordance with another embodiment of the present invention.
10 illustrates an exemplary flash read channel architecture with iterative demapping and decoding and selective interleaving.
11 illustrates an exemplary flash memory system with soft data generation in accordance with the present invention.
12A and 12B are flowcharts illustrating an exemplary soft demapping process and soft data generation process, respectively.
13 is an exemplary bipartite graph representation of a low-density parity-check (LDPC) code.
14 is a block diagram of an exemplary LDPC decoder architecture.
15 illustrates an exemplary flash memory system with soft data generation in accordance with one embodiment of the present invention.
16 is a trellis representing error probabilities p and q for an exemplary binary channel.
17A-17C are exemplary cell-based statistical tables that record statistics for reading data from flash memory.
18 is an exemplary pattern dependent cell based statistics table for recording pattern dependent statistics for reading data from flash memory.
FIG. 19 illustrates the example flash cell array of FIG. 3 in more detail for the reference cell embodiment of the present invention.
20 is a flow diagram illustrating an example implementation of a bit-based statistics generation process for a reference cell embodiment of the present invention.
21 is a flow diagram illustrating an example implementation of a cell based statistics generation process for a reference cell embodiment of the present invention.
22 is a flow diagram illustrating an example implementation of a bit-based statistics generation process for a decoding codeword embodiment of the present invention.
23 is a flow diagram illustrating an example implementation of a cell based statistics generation process for a decoding codeword embodiment of the present invention.
24 is a flow diagram illustrating an example bit-based location specific statistics generation process for calculating error probability statistics for a number of different locations in a memory array.
FIG. 25 is a flow diagram illustrating an example cell based location specific statistics generation process for calculating error probability statistics for a number of different locations in a memory array.
FIG. 26 shows a set of probability density functions that exhibit a pattern dependent disturbance effect on a given target cell, based on all possible values of each contributor cell.
FIG. 27 is a flow diagram illustrating an example bit-based pattern dependent statistics generation process for estimating error probability statistics that depends on a given data pattern in one or more attacker cells associated with at least one target cell.
FIG. 28 is a flow diagram illustrating an example cell-based pattern dependency statistics generation process for estimating statistics dependent on a given data pattern in one or more attacker cells associated with at least one target cell.
29 is a flow diagram illustrating an exemplary asymmetric statistics generation process for estimating error probability statistics for two possible binary values for a reference cell embodiment of the present invention.
30 is a flow diagram illustrating an exemplary asymmetric statistics generation process for estimating error probability statistics for two possible binary values for a decoding codeword embodiment of the present invention.
31 is a flowchart illustrating an example implementation of a statistics generation process using an unsatisfactory parity check.
FIG. 32 is a flow diagram illustrating an example location specific statistics generation process for estimating error probability statistics for a number of different locations in a memory array using an unsatisfactory parity check.
33 is a flow diagram illustrating an example asymmetric statistics generation process for estimating error probability statistics for two possible binary values using an unsatisfactory parity check.

Various aspects of the present invention relate to soft data generation techniques for improved decoding in memory devices such as single level cell or multi level cell (MLC) NAND flash memory devices. In the present specification, a multi-level cell flash memory includes a memory in which each memory cell stores two or more bits. Typically, multiple bits stored in one flash cell belong to different pages. Although the invention is described herein using memory cells that store analog values as voltages, the invention can be used with any storage scheme for a memory device, such as using voltage or current to represent stored data. This will be apparent to those skilled in the art.

1 is a schematic block diagram of a conventional flash memory system 100. As shown in FIG. 1, an exemplary flash memory system 100 includes a flash control system 110 and a flash memory block 160. Exemplary flash control system 110 includes flash controller 120, encoder / decoder block 140, and one or more buffers 145. In other embodiments, encoder / decoder block 140 and some buffers 145 may be internalized within flash controller 120. Encoder / decoder block 140 and buffer 145 may be implemented, for example, using known commercially available techniques and / or products.

Exemplary flash memory block 160 may include memory array 170 and one or more buffers 180, each of which may be implemented using known techniques and / or products. The memory array 170 may be a single-level or multi-level cell flash, such as NAND flash memory, phase-change memory (PCM), MRM memory, NOR flash memory, or other nonvolatile flash memory. It can be implemented in memory. Although the present invention is primarily described in the context of multi-level cell NAND flash memory, the present invention can also be applied to single-level cell flash memory and other non-volatile memories, which will be apparent to those skilled in the art.

Multi Level Cell Flash Memory

In multilevel cell NAND flash memories, a threshold detector is typically employed to convert voltage values associated with a particular cell into a predetermined memory state. 2 illustrates an example threshold voltage distribution for the example multi-level cell flash memory 170 of FIG. 1, based on the disclosure of US Pat. No. 6,522,580, which is incorporated herein by reference. In general, the threshold voltage of a cell is the voltage that the cell must apply to in order to conduct a certain amount of current. Threshold voltage is a measure of data stored in a cell.

In the exemplary embodiment shown in FIG. 2, each storage element uses four possible data states to store two bits of data in each memory cell. 2 shows four peaks 210-213, each peak corresponding to one state. In a multilevel cell flash device, different peaks 210-213 of the threshold voltage distribution graph 200 are used to store two bits in the cell.

Peaks 210-213 of threshold voltage distribution graph 200 are labeled with corresponding binary values. Therefore, when the cell is in the first state 210, it represents "1" for the lower bit (also known as the least significant bit) and "1" for the higher bit (also known as the most significant bit). . State 210 is generally the first unprogrammed or erased state of the cell. Similarly, if the cell is in the second state 211, it represents "0" for the lower bit and "1" for the higher bit. When the cell is in the third state 212, it represents "0" for the lower bit and "0" for the higher bit. Finally, if the cell is in the fourth state 213, it represents "1" for the lower bit and "0" for the higher bit.

Threshold voltage distribution 210 represents a distribution of threshold voltages V ₁ of cells in an array in an erased state (“11” data state) with a negative threshold voltage level below zero volts. Threshold voltage distributions 212 and 212 of memory cells that store "10" and "00" user data, respectively, are shown between 0 and 1 volt and between 1 and 2 volt, respectively. Threshold voltage distribution 213 shows the distribution of cells programmed to the "01" data state and has a threshold voltage level set between 2 and 4.5 volts of read pass voltage.

Therefore, in the exemplary embodiment of FIG. 2, 0 volts, 1 volt and 2 volts can be used as the voltage level threshold between each level or state. The voltage level threshold is used to determine the voltage level or state of a cell given by flash memory 160 (eg, sensing circuitry in flash memory 160). The flash memory 160 will allocate one or more bits to each cell based on the comparison of the measured voltage and the voltage level threshold, which is sent to the flash control system 110 as a hard decision. Additionally or alternatively, in implementations that use soft information, flash memory 160 may transmit the measured voltage or a quantized version of the measured voltage as soft information to flash control system 110, in which case the memory cell More bits than the number of bits stored in are used to represent the measured voltage.

It is also important that the cell is normally programmed using known program / verify techniques. In general, during a program / verify cycle, flash memory 160 applies a voltage that gradually increases until the minimum target threshold voltage is exceeded to store charge in the cell transistor. For example, when programming the '10' data state in the example of FIG. 2, flash memory 160 gradually increases the voltage to store charge in the cell transistor until it exceeds a minimum target threshold voltage of 0.4V. Can be authorized.

As detailed below, each of the 2 bits stored in a single memory cell is from a different page. In other words, each of the two bits stored in each memory cell carries a different page address. The bits on the right shown in FIG. 2 are accessed when the lower page address is entered. The left bit is accessed when the upper page address is entered.

3 illustrates the architecture of an example flash cell array 300 of a multi-level cell (MLC) flash memory device 160, where each example cell typically corresponds to a floating gate transistor that stores two bits. . In Figure 3, each cell is associated with two numbers for two pages to which two bits belong. Exemplary cell array section 300 shows wordlines n through n + 2 and four bitlines. Exemplary flash cell array 300 is divided into even and odd pages, where, for example, cells with even numbers (such as cells with numbers 0 and 2) correspond to even pages and (numbers 1 and 3). A cell with an odd number (such as a cell with) corresponds to an odd page. Wordline n stores, for example, even pages (0 and 2) of even bit lines and odd pages (1 and 3) of odd bit lines.

3 also shows an example program sequence in which even or odd bitline cells are selected and programmed in succession (bottom to top) in the displayed order. The numbers indicate the order in which the pages are programmed. For example, page 0 is programmed before page 1. For further discussion of the programming of even and odd pages, see, for example, K.-T. Park et al., "A Zeroing Cell-to-Cell Interference Page Architecture with Temporary LSB Storing and Parallel MSB Program Scheme for MLC NAND Flash Memories," (IEEE Journal of Solid-State Circuits, Vol. 43, No. 4, 919-928 (April 2008).

4 illustrates an exemplary two stage MLC programming scheme 400 for the voltage allocation scheme of FIG. As shown in FIG. 4, during the LSB program phase, the state of the selected cell in the erased state 410 moves to the lowest program state 411 if the LSB is zero. Therefore, in the LSB programming step, the memory cells are programmed from erase state '11' to '10'. Next, during the MSB program phase, two states, states '00' 412 and '01' 413 are sequentially formed according to the previous LSB data. Generally, during the MSB programming phase, the '10' state is programmed to '00' and the state '11' is programmed to '01'.

Note that the programming scheme 400 of FIG. 4 shows the maximum voltage shift associated with a state change from state 410 to state 413. Many programming schemes have been proposed or suggested to reduce the maximum voltage variation associated with a state change, thereby reducing the ICI caused by the voltage variation.

5A and 5B together show another MLC programming scheme 500 that reduces the ICI applied to neighboring cells. As shown in Figure 5A, during the LSB programming phase, the memory cells are programmed in a similar manner to SLC programming from state '11' to 'x0' as a temporary (middle) state. After adjacent cells in the same wordline are also LSB programmed, the distribution may be widened as shown by peak 510 in FIG. 5A by ICI. Then, in the MSB programming step shown in FIG. 5B, the 'x0' state is programmed to '00' or '10' as the final state corresponding to the input data, or the '11' state to the final '01' state. Is programmed. In general, all memory cells except the '11' cells can be reprogrammed from their temporary program state to the LSB to their final stage in the MSB programming phase, thereby greatly reducing the ICI caused by adjacent cells. The cell in the final state does not experience the ICI that it experienced when in the intermediate state because the cell was reprogrammed to the final state. The cell in the final state will only go through the ICIs that have been there since the final state. As discussed above, the multi-step programming sequence of FIGS. 5A and 5B using an intermediate program state reduces the maximum voltage changes, thus reducing the ICI caused by these voltage changes. For example, in FIG. 5B, it can be seen that during the MSB programming phase, the maximum eye shift is related to the transition from state '11' to '01' and the transition from state 'x0' to '10', respectively. These voltage shifts are significantly less than the maximum voltage shift from state '11' to '01' in FIG.

6 illustrates in more detail an exemplary flash cell array 600 in a multi-level cell (MLC) flash memory device 130. As shown in FIG. 6, the flash cell array 600 stores 3 bits per flash cell, c _i . 6 shows a flash cell array architecture for one block, where each exemplary cell corresponds to a floating gate transistor that typically stores three bits. The exemplary cell array 600 is composed of m word lines and n bit lines. Typically, in current multi-page cell flash memory, the bits in a single cell belong to different pages. In the example of FIG. 6, three bits for each cell correspond to three different pages, and each wordline stores three pages. In the following discussion, pages 0, 1, and 2 are referred to as lower, middle, and upper page levels within a wordline.

As described above, the flash cell array can be further divided into even and odd pages, in which case, for example, cells with even numbers (such as cells 2 and 4 in FIG. 6) correspond to even pages and (FIG. Cells with odd numbers (such as cells 1 and 3 of 6) correspond to odd pages. In this case, the page (such as page 0) will contain even pages in even cells (even page 0) and odd pages in odd cells (odd page 0).

Intercell Interference and Other Disturbance

7 illustrates disturbances such as intercell interference, back pattern dependence, noise and other distortions present in the target cell due to many exemplary attacker cells. In Figure 7, the following notation is used.

WL: word line;

BL: bit line;

BLo: odd bitline;

BLe: Even Bitline; And

C: capacitance.

For example, ICI is caused by an attacker cell 720 that is programmed after the target cell 710 is programmed. ICI changes the voltage V _t of the target cell 710. In an exemplary embodiment, a "bottom up" programming scheme is assumed, with adjacent attacker cells on wordlines i and i + 1 causing ICI for target cell 710. With this upward programming of the block, the ICI from the lower wordline i-1 is removed and up to five neighboring cells contribute to the ICI as the attacker cell 720 as shown in FIG. However, the techniques disclosed herein may be generalized even when attacker cells from other wordlines, such as wordline i-1, contribute to ICI, which will be apparent to those skilled in the art. If attacker cells from wordlines i-1, i and i + 1 contribute to ICI, up to eight nearest neighbor cells should be considered. Other cells further away from the target cell may be ignored if their contribution to the ICI is negligible. In general, attacker cell 720 is identified by analyzing programming sequence schemes (such as upward or odd / even techniques, etc.) to identify attacker cells 720 that are programmed after a given target cell 710.

In general, V _t is the voltage representing the data stored on the cell and is obtained during a read operation. V _t is the same resolution by a read operation, for example, as a soft voltage value with a higher accuracy than the number of bits stored per cell, or equal to the number of bits stored per cell (eg 3 bits per cell flash). Can be obtained as a value quantized to a hard voltage level.

For further discussion of ICI reduction techniques, see, for example, International Patent Application PCT / US09 / 49326, entitled “Methods and Apparatus for Read-Side Intercell Interference Mitigation in Flash Memories,” each of which is incorporated herein by reference. Or International Patent Application PCT / US09 / 49327 entitled "Methods and Apparatus for Write-Side Intercell Interference Mitigation in Flash Memories."

Soft data generation

The present invention provides soft demapping and soft data generation techniques for flash memories. In one exemplary embodiment described below with respect to FIG. 12A, enhanced soft data is generated from soft data allocated by flash memory using probability statistics such as probability density function, approximation, bit-based probability, or cell-based probability. do. In another exemplary embodiment described below with respect to FIG. 12B, soft data is generated from hard data allocated by the flash memory using probability statistics such as a probability density function, its approximation, bit-based probability, or cell-based probability. In general, data allocated by flash memory is initially obtained. The invention then generates or enhances soft information such as probability or reliability information based on the data from the flash memory. The generated soft information can optionally be used for soft decision decoding. As used herein, the term "probability density function" includes probability density functions, and their approximations, such as histogram and Gaussian approximation.

8 is a schematic block diagram of an exemplary flash memory system 800 incorporating a controller-based soft data generation technique in accordance with the present invention. As shown in FIG. 8, the exemplary flash memory system 800 includes a flash control system 810 and a flash memory block 860 connected by an interface 850. Exemplary flash control system 810 typically includes flash controller 820 and read channel 825 on one or more integrated circuits.

Exemplary read channel 825 includes signal processing unit 830, encoder / detoder block 840, and one or more buffers 845. Note that the term "read channel" may also encompass recording channels. In other embodiments, encoder / decoder block 840 and some buffers 845 may be implemented inside flash controller 820. Encoder / decoder block 840 and buffer 845 may be implemented by, for example, modifying, as used herein, known, commercially available techniques and / or products to provide the features and functionality of the present invention. .

Exemplary signal processing unit 830 includes one or more processors that each implement one or more soft demapper and / or soft data generation processes 835, described below with respect to FIGS. 12A and 12B, respectively. Exemplary flash memory block 860 includes a memory array 870 and one or more buffers 880, each of which may be implemented using known commercial techniques and / or products.

In various embodiments of the disclosed soft data generation techniques, example interface 850 may have to carry additional information related to conventional flash memory systems, such as values representing information associated with an attacker cell. Therefore, interface 850 may need to have a greater capacity or faster rate than the interface of a conventional flash memory system. Interface 850 is optionally an international patent, for example, entitled "Methods and Apparatus for Interfacing Between a Flash Memory Controller and a Flash Memory Array", filed June 30, 2009, and incorporated herein by reference. It can be implemented in accordance with the teachings of application PCT / US09 / 49328 (Agent Reg. No. 08-0769), which, for example, increases the information carrying capacity of the interface 850 using double data rate (DDR) technology. Let's do it. During a write operation, interface 850 delivers the program value to be stored in the target cell, typically using page or wordline level access techniques. For a more detailed discussion of exemplary page or wordline level access techniques, see, for example, the methods "Methods and Apparatus for Storing Data in a Multi-Level Cell Flash Memory Device with Cross-", which is incorporated herein by reference. See, for example, International Patent Application PCT / US09 / 36810, filed March 11, 2009, entitled " Page Sectors, Multi-Page Coding and Per-Page Coding. &Quot;

During the read operation, the interface 850 delivers the hard and / or soft read values obtained from the memory array 870 for the target cell and the attacker cell. For example, in addition to the read value for a page into a target cell, the read value for one or more adjacent pages of the upper / lower wordline or adjacent even or odd bit lines is passed through the interface bus. In the embodiment of Figure 8, the disclosed soft data generation techniques are implemented outside of flash memory, typically in processing techniques optimized for logic circuits to achieve the lowest area. However, this sacrifices additional attack cell data that may be delivered on interface 850.

9 is a schematic block diagram of an exemplary flash memory system 900 that includes a memory-based soft data generation technique in accordance with another embodiment of the present invention. As shown in FIG. 9, an exemplary flash memory system 900 includes a flash control system 910 and a flash memory block 960, which are connected by an interface 950.

Exemplary flash control system 910 includes flash controller 920 and optional read channel 925, typically on one or more integrated circuits. In other embodiments, encoder / decoder block 940 and some buffers 945 may be implemented within flash controller 920. Exemplary flash controller 920 may be implemented, for example, using known commercially available technologies and / or products, as modified herein to support the features and functionality of the present invention. Exemplary read channel 925 includes an encoder / decoder block 940 and one or more buffers 945. Encoder / decoder block 940 and buffer 945 may be implemented using known commercially available techniques and / or products.

Exemplary flash memory block 960 includes a memory array 970 and one or more buffers 980, each of which may be implemented using known commercially available techniques and / or products. In addition, example flash memory block 960 includes one or more processors, each implementing one or more soft demapping and / or soft data generation processes 990 described below in connection with, for example, FIGS. 12A and 12B. An example signal processing unit 985 is included.

In various embodiments of the disclosed soft data generation techniques, example interface 950 may have to carry additional information about a conventional flash memory system, such as a value representing information associated with an attacker cell. Therefore, the interface 950 may need to have a greater capacity or faster rate than the interface of a conventional flash memory system. Interface 950 is optionally an international patent, for example, entitled "Methods and Apparatus for Interfacing Between a Flash Memory Controller and a Flash Memory Array", filed June 30, 2009, and incorporated herein by reference. It can be implemented in accordance with the teachings of application PCT / US09 / 49328 (Agent No. 08-0769), which, for example, increases the information carrying capacity of the interface 950 using double data rate (DDR) technology. Let's do it.

During a write operation, the interface 950 delivers program data to be stored in the target cell and the attack cell. During a read operation, interface 950 delivers new hard or soft read values or data for the target cell (s) and optionally the attacker cell. Typically, the information carried for a single read access is one page or wordline of data. Transferring data for a target cell is not costly to implement a soft data generation process in memory using memory processing techniques typically used to fabricate flash memory optimized for memory rather than logic circuitry. Note that it reduces the bandwidth requirement for.

 FIG. 10 is in accordance with the teachings of International Patent Application PCT / US09 / 49333, entitled "Methods and Apparatus for Soft Demapping and Intercell Interference Mitigation in Flash Memories", incorporated by reference herein, filed June 30, 2009; An example flash read architecture 1000 with recursive demapping and decoding, and optional interleaving is shown. As shown in FIG. 10, an exemplary write path includes an encoder 1010, an optional interleaver 1020, a serial-to-parallel converter 1030, and a mapper 1040. . Data is written to and read from memory 1050 in a known manner. Exemplary read paths include a soft demapper or soft data generator 1060, a parallel-to-serial converter 1070, a deinterleaver 1080, a decoder 1090 and an interleaver 1095. In general, as described below, the soft demapper or soft data generator 1060 is processed by the decoder 1090 and generated in a recursive manner to generate new soft information until the recursive process converges to the last decision. This generates soft information which will be described in detail later, which is fed back to the soft demapper.

The equation employed by the soft demapper 1060 to generate soft information (LLR) in accordance with the present invention is discussed in the section "Calculating Soft Data (LLR) Using Reading Statistics" below. As shown in FIG. 10, the soft information generated by the demapper 1060 is recursive between the soft demapper 1060, the deinterleaver 1080, the decoder 1090, and the interleaver 1095 of the feedback path. Can be used for demapping and decoding.

Soft data generation based on data from flash

The present invention recognizes that current flash memories 860 and 960 typically provide only hard data to the flash control system 810 and 910. However, it is known that soft data can improve error rate performance in the decoding process. Therefore, in accordance with one aspect of the present invention, hard data from flash memory 860, 960 is used to estimate soft data and thereby improve decoding performance in flash control system 810, 910. For example, as discussed below, the statistical properties of hard data can be used to estimate or enhance soft data. The generated soft data may then be used for decoding, such as trust propagation decoding of the LDPC code, to improve error rate performance.

According to another aspect of the present invention, flash memory 860, 960 provides soft data or soft information to flash control system 810, 910. The enhanced soft data is generated from the soft data provided by the flash memories 860 and 960 to improve decoding performance in the flash control system 810 and 910. In an implementation using soft information, the flash memory system 860, 960 sends the measured voltage or a quantized version of the measured voltage as soft information to the flash control system 810, 910, where the bit stored in the memory cell is stored. More than the number of bits are used to represent the measured voltage.

11 illustrates an example flash memory system 1100 with controller based soft data generation in accordance with an embodiment of the present invention. As shown in FIG. 11, an exemplary flash memory system 1100 includes a flash memory block 1100 and a flash control system 1120, which are connected by an interface 1115. As described below, soft or hard data values (or both) may be assigned by flash memory block 1110 and communicated through interface 1115 to flash control system 1120 for further decoding and processing. . Exemplary flash control system 1120 includes a soft demapper / soft data generator 1200 described below with respect to FIGS. 12A and 12B and a decoder 1400 described below with respect to FIGS. 13-14. The decoder 1400 may use an LDPC decoding algorithm such as, for example, trust propagation, message passing, sum-product, or min-sum algorithm. Can be implemented.

As shown in FIG. 11, soft information generated by the soft demapper / soft data generator 1200 is optionally recursive demapping and decoding between the soft demapper / soft data generator 1200 and the decoder 1400. Can be used for In general, as shown in FIG. 11, the soft demapper / soft data generator 1200 generates soft information in the form of LLR ( L _e ), which is described below in "Calculating Soft Data (LLR) Using Reading Statistics." Discussed in the section. Initially, the LLR L _e calculated by the soft demapper / soft data generator 1200 is based on soft or hard reads (or both) and corresponding statistics from the flash memory 1110. LLR L _e by decoder 1400 to generate new soft information L _a fed back to soft demapper / soft data generator 1200 in a recursive manner until the recursive process converges to a final decision. This is processed.

soft Demapper / Soft data generator (1200)

12A is a flow diagram illustrating an exemplary soft demapping process 1200 that incorporates features of the present invention for generating enhanced soft data from soft data provided by flash memories 810, 910. As shown in FIG. 12A, exemplary soft demapping process 1200 initially obtains soft data r from flash memory 810, 910 for a target cell during step 1210, and optionally, with the target cell. Obtain one or more values h representing data stored in the associated attacker cell (s).

The soft demapping process 1200 then obtains statistics (or probabilities), such as one or more probability density functions, based on r and optionally h during step 1220. Statistics are discussed further in the "Gathering Statistics" section below.

The obtained statistics are then used to calculate the LLR during step 1230. LLR (s) are discussed below in the section "Calculation of Soft Data (LLR) Using Reading Statistics". The calculated LLR is then provided to the decoder 1400 during step 1240, or optionally to a deinterleaver. The calculated LLR can optionally be used to make a final decision on the read data, for example based on the sign of the LLR.

)를 획득하고, 선택적으로, 타겟 셀과 관련된 공격자 셀(들)에 저장된 데이터를 나타내는 하나 이상의 값(

)을 획득한다. 하드 데이터(

)는 예를 들어 플래시 메모리(810, 910)에 의해 각 셀에 할당된 이진 비트 또는 수준일 수 있다.12B is a flow diagram illustrating an exemplary soft data generation process 1250 incorporating features of the present invention for generating soft data from hard data provided by flash memories 810 and 910. As shown in FIG. 12B, the exemplary soft data generation process 1250 may initially include hard data (from hard memory (810, 910) for the target cell during step 1260.

) And optionally, one or more values representing data stored in the attacker cell (s) associated with the target cell (

). Hard data (

) May be, for example, a binary bit or level assigned to each cell by flash memory 810, 910.

)으로 매핑될 수 있다. 워드라인 액세스 기술로 전체 워드라인이 판독되어 셀 내의 모든 비트가 획득될 수 있고, 그로부터 하드 데이터 수준(

)이 획득된다. To calculate the LLR for one bit in a cell, the other bits are not available and the other bits in the cell are read using page and wordline access techniques, for example. For a more detailed discussion of an exemplary page or wordline level access technique, see, for example, the method and Apparatus for Storing Data in a Multi-Level Cell Flash Memory Device with Cross-Page, incorporated herein by reference. Sectors, Multi-Page Coding and Per-Page Coding ", and filed March 11, 2009, filed in International Patent Application PCT / US09 / 36810. With page access technology, the page on which the LLR is being computed can be read and, optionally, other pages on the same wordline can also be read so that hard data can be read at the level of the cell.

) Can be mapped. Wordline access technology allows the entire wordline to be read so that every bit in the cell can be obtained, from which the hard data level (

) Is obtained.

은, 예를 들어, 공격자 셀(720)(또는 공격자 셀(720)이 저장된 페이지 또는 워드라인)로부터 비트를 판독함으로써 획득된다. 공격자 셀을 판독하는 기술에 대한 더 상세한 논의는 예를 들어, 본 명세서에 참조로 포함되는 명칭이 "Methods and Apparatus for Read-Side Intercell Interference Mitigation in Flash Memories"인 국제 특허 출원 PCT/US09/49326호를 참조하라.pattern

Is obtained, for example, by reading the bits from the attacker cell 720 (or the page or wordline on which the attacker cell 720 is stored). A more detailed discussion of techniques for reading attacker cells is described, for example, in International Patent Application PCT / US09 / 49326, entitled "Methods and Apparatus for Read-Side Intercell Interference Mitigation in Flash Memories", which is incorporated herein by reference. See also.

및, 선택적으로,

에 기초하여 단계 1270 동안 하나 이상의 확률 밀도 함수와 같은 통계(또는 확률)을 획득한다. 통계는 비트 기반 또는 셀 기반 확률일 수도 있으며, 이는 아래의 "통계의 수집" 섹션에서 논의한다. 소프트 판독 값의 분포에 대한 가우스 근사가 채용되는 경우, 통계는 분포의 평균 값 또는 분산을 포함하며, 이는 다양한 가우스 근사식에 대해 "판독 통계를 사용한 소프트 데이터(LLR)의 계산" 섹션에서 논의한다. 평균 값 및 분산은, 예를 들어, 프로그램/소거 사이클, 판독 사이킬 및 온도와 같은 상이한 성능 인자에 대한 플래시 메모리 칩의 특성화(characterization) 동안 사전 계산되어 표에 저장될 수 있다. 평균 값 및 분산은 성능 인자에 기초하여, 또한 선택적으로 공격자 셀(720)에 저장된 패턴(

)에도 기초하여 표로부터 획득될 수 있다.The soft data generation process 1250 may then

And, optionally,

Obtain a statistic (or probability) such as one or more probability density functions during step 1270. Statistics may be bit-based or cell-based probabilities, which are discussed in the section "Collection of Statistics" below. If a Gaussian approximation for the distribution of soft readings is employed, the statistics include the mean value or variance of the distribution, which is discussed in the section "Calculation of Soft Data (LLR) Using Reading Statistics" for various Gaussian approximations. . The mean value and variance can be precomputed and stored in a table during the characterization of the flash memory chip for different performance factors such as, for example, program / erase cycles, read cycles, and temperature. The mean value and the variance are based on the performance factors, and optionally also the pattern (stored in the attacker cell 720).

Can also be obtained from the table.

The obtained statistics are then used to calculate the LLR (s) during step 1280. LLR (s) are discussed in the section "Calculation of Soft Data (LLR) Using Reading Statistics" below. As described in the section "Calculation of Soft Data (LLR) Using Reading Statistics", in addition to or instead of statistics, an a-priori LLR (L _a ) provided by a decoder calculates the LLR. Note that it can be used optionally. When an LLR is calculated for a bit in a cell, a priori LLR (L _a ) is used for at least one bit (optionally all other bits) in the cell. This requires that these other bits in the cell have been read and an a priori LLR L _a for them has been calculated by the decoder.

The calculated LLR is then provided to the decoder 1400, or optionally the deinterleaver, during step 1290. The calculated LLR can optionally be used to make a final decision on the read data, for example based on the sign of the LLR. International patent application PCT / US09, entitled "Methods and Apparatus for Storing Data in a Multi-Level Cell Flash Memory Device with Cross-Page Sectors, Multi-Page Coding and Per-Page Coding", filed March 11, 2009 Note that as described in / 36810, every bit in a cell (or every page in a wordline) can be jointly encoded and decoded. In another embodiment, bits in a cell (or all pages in a wordline) may be separately encoded and decoded, as also described in International Patent Application PCT / US09 / 36810.

Decoder 1400 LDPC  avatar

The following background description of LDPC codes and LDPC decoding is described in AJ Blanksby and CJ Howland, "A 690-mW 1-Gb / s 1024-b, Rate-1 / 2 Low-Density Parity-Check", incorporated herein by reference. Decoder, "IEEE J. Solid-State Circuits, Vol. 37, 404-412 (March 2002). For a more detailed discussion, see Blanksby and Howland's full text.

LDPC  Graph representation of code

The LDPC code can also be represented using a bipartite graph in which one set of nodes represents a parity check limit and the other set represents data bits. 13 is an illustrative two-letter graph representation 1300 of an LDPC code. The parity check matrix is the incidence matrix of the graph, and if the entry h _ji in H is set, i.e. not zero, the bit node i corresponding to column i in H is connected to the check node j corresponding to row j in H.

One algorithm used to decode the LDPC code is known as the sum-product algorithm. For good decoding performance with this algorithm, it is important that the length of the cycle in the graphical representation of the LDPC code is as long as possible. In the exemplary representation of FIG. 13, an exemplary short cycle of length 4 is shown. Short cycles, such as the length-4 cycles shown in FIG. 13, degrade the performance of the sum-product algorithm. Another well known algorithm for decoding LDPC codes is the min-sum algorithm.

Sum-product algorithm

The sum-product algorithm is a recursive algorithm for decoding LDPC codes. Sum-product algorithms are also known as message passing algorithms or trust propagation. For a more detailed discussion of the sum-product algorithm, see, for example, AJ Blanksby and CJ Howland, "A 690-mW 1-Gb / s 1024-b, Rate 1/2 Low-Density," which is incorporated herein by reference. Parity-Check Decoder, "IEEE J. Solid-State Circuits, Vol. 37, 404-412 (March 2002), D.E. Hocevar, "LDPC Code Construction With Flexible Hardware Implementation," IEEE Int'l Conf. on Comm. (ICC), Anchorage, AK, 2708-2712 (May 2003), and RNS Ratnayake, EF Haratsch and Gu-Yeon Wei, "A Bit-node centric architecture for low-density parity check decoders, '" IEEE Global Telecommunications Conference (Globecom), Washington, DC, 265-270 (November 2007).

The message Q _{i , j} from bit node i to check node j is given by:

Where L is the extrinsic LLR provided by the soft demapper / soft data generator for bit i. The message R _{j , i} from the check node j to the bit node i is given by

이고, here

ego,

이다.

to be.

The epigenetic information value Λ _i , also referred to as the a-poeteriori log-likelihood ratio (LLR) for bit i, is given by:

L _{a , i} for bit i provided to the soft demapper / soft data generator for recursive demapping and decoding

Where B _i is a set of check nodes connected to bit node i, and C _j is a set of bit nodes connected to check node j.

LDPC  Decoder-Hardware Sharing Decoder Architecture

A great challenge in implementing a sum-product algorithm for decoding LDPC codes is managing the delivery of messages. Since the functionality of both the check node and the bit node is relatively simple, only a few gates are involved in their respective implementations. An important issue is the implementation of the bandwidth required for message transfer between functional nodes.

14 is a block diagram of an example hardware shared LDPC decoder architecture 1400. As shown in FIG. 14, the generalized LDPC decoder architecture 1400 stores a number of functional units 1410 and 1420, each implementing check or bit node functionality, and a memory fabric that stores messages and realizes graph connectivity. 1450). The control logic 1430 controls the configuration of the memory fabric 1450. For a detailed discussion of the implementation of the hardware shared LDPC decoder architecture 1400, see, for example, E. Yeo et al., "VLSI Architectures for Iterative Decoders in Magnetic Recording Channels," IEEE Trans. On Magnetics, Vol. 37, No. 2, 748-755 (March 2001).

It has been recognized that this hardware sharing architecture reduces the area of the decoder.

15 illustrates an example flash memory system 1500 with soft data generation in accordance with one embodiment of the present invention. As shown in FIG. 15, an exemplary flash memory system 1500 includes a flash memory block 1510. As will be discussed later, hard or soft data values (or both) are typically assigned by flash memory block 1510 and communicated for further decoding and processing to flash control system 1520 via interface 1515. . Exemplary flash control system 1520 includes an LLR generator 1550, described below in connection with FIG. 16, statistics generator (s) 1570 and decoder 1530, described below in the “Collection of Statistics” section. Statistics generated by the statistics generator (s) 1570 are optionally recorded in one or more statistics table (s) 1560, described below in connection with, for example, FIGS. 17A-17C and 18, or alternatively in real time. Can be generated.

The statistics generated by the statistics generator (s) 1570 are used by the LLR generator 1550 to form soft data, for example in the form of LLR ( L _e ). Initially, LLR L _e is based on soft or hard reads (or both) and corresponding statistics from flash memory 1510. The LLR L _e is processed by the decoder 1530 to generate new soft information L _a which is fed back to the LLR generator 1550 in a recursive manner until the recursive process converges to the final decision.

Decoder 1530 may again be implemented using an LDPC decoding algorithm, such as, for example, trust propagation, message delivery, sum-product, or minimum-sum algorithm. The functionality of the statistics generator 1570 and LLR generator 1550 described herein is implemented in one or more of the flash control system 1520, the decoder 1530, and the read channel 825 (see, eg, FIG. 8). Note that it can be.

Calculation of Soft Data (LLR) Using Reading Statistics

A prior log likelihood ratio (LLR) L _a for bit c may be defined as

Where P (...) is the probability.

Similarly, the LLR of bit c conditioned on flash output r is calculated as follows.

Where L _e (c) is the external LLR or soft information passed to the subsequent decoder, and p (...) is the probability density function (PDF).

16 is a trellis 1600 representing error probabilities p and q for an exemplary binary channel. Note that in the context of binary channels, p represents an error probability and p (...) represents a probability density function. If p ≠ q, this binary channel is asymmetric. If p = q, this binary channel is symmetric. As shown in Fig. 16, p is the error probability for binary 0 (i.e., the probability of reading 1 when 0 is written). Similarly, q is the error probability for binary 1 (ie, the probability of reading 0 when 1 is written). The probability of properly reading binary zeros (ie, the probability of reading zero when zero is written) can be expressed as 1-p. Similarly, the probability of properly reading binary 1 (ie, the probability of reading 1 when 1 is written) can be expressed as 1-q.

Cross product for binary asymmetric channels ( Extrinsic ) LLR

L _e (c), which is the external LLR of bit c for the binary asymmetric channel defined by grating 1600, can be expressed as follows.

=0 에 대한 외적 LLR인 L _e (c)는 다음과 같이 계산된다.Read bit

L _e (c), the cross product LLR for = 0, is calculated as follows.

=1 에 대한 외적 LLR인 L _e (c)는 다음과 같이 계산된다.Read bit

L _e (c), the cross product LLR for = 1, is calculated as follows.

Cross product LLR for binary symmetric channel (if p = 1 = P ₀ )

=0 에 대한 외적 LLR인 L _e (c)는 다음과 같이 계산된다.Read bit

L _e (c), the cross product LLR for = 0, is calculated as follows.

=1 에 대한 외적 LLR인 L _e (c), 이 LLR은 다음과 같이 계산된다. Read bit

L _e (c), which is the cross product LLR for = 1, is calculated as follows.

Cross product for soft output from flash memory LLR

For a 2-bit / cell flash memory, the external LLR for the soft value r received from the flash memories 810, 910 can be calculated as follows.

In general, for any number of bits per cell, the outer LLR for bit C _i can be expressed as follows.

here,

r: received signal

s: The original storage state or level given by the stored bits (c ₀ , c ₁ , ..., c _m )

c: coded bits

M bits per cell

: 선험적 LLR

A priori LLR

: 외적 LLR

External LLR

: 비트 라벨이 위치 i에서 값 C_i=c_i 를 갖는 상태 또는 수준의 부부분집합

: Bit label value in the position i C _i = c _i with the state or level of the sub-subset

L _a (C _i ) is provided by a decoder such as, for example, LDPC decoder 1090 or 1400. In the first iteration, L _a (C _i ) may be initialized to zero.

Next equation

Using, the expression for the external LLR can also be written as

This equation can be made simpler as

This equation is mathematically equivalent to the above equation if the probability of all states or levels is equally likely.

Pattern dependent cross product for soft output from flash memory LLR

에 대해, 다음을 보일 수 있다. One or more soft values r for the target cell received from the flash memory 810, 910 and one or more values for the attacker cell (s).

For, we can show:

는 주위 셀(들) 또는 타겟 셀에 외란을 유발하는 다른 셀에 저장된 데이터 패턴이다. 예를 들어, here,

Is a data pattern stored in the surrounding cell (s) or other cells causing disturbance to the target cell. E.g,

는 LLR이 계산되고 있는 위치 (k,l)에서의 타겟 셀에 인접한 모든 공격자 셀을 나타낸다.

Denotes all attacker cells adjacent to the target cell at position (k, l) where the LLR is being calculated.

는 예를 들어 공격자 셀로부터 하드 데이터를 독출하여 획득될 수 있다. pattern

For example, may be obtained by reading the hard data from the attacker cell.

The expression for the external LLR can also be written as

This equation can be simplified as:

This equation is mathematically equivalent to the above equation if the probability of all states or levels is equally likely.

From flash memory hard  Cross product for the output ( extrinsic ) LLR

)만을 제공하는 경우, 외적 LLR은 아래와 같이 산출될 수 있다: Soft data is not available from the flash memory, and the hard data is a state or level in which the flash memory is allocated to the data stored by the flash memory.

If only) is provided, the cross product LLR can be calculated as follows:

}은 하드 값(

)을 가정할 때 (전압과 같은) 소프트 값(r)에 대한 기대값 또는 소프트 값(r)에 대한 다른 어떤 추산치이다. P(

|s)는 상태 또는 수준(s)이 본래 기록되거나 저장되었다고 가정할 때(상태 또는 수준과 같은) 하드 값(

)이 판독될 확률이다. Wherein E {r |

} Is the hard value (

) Is an expected value for soft value r (such as voltage) or some other estimate of soft value r. P (

| s) is a hard value (such as a state or level) assuming a state or level (s) was originally recorded or stored.

) Is the probability of reading.

Alternatively, the cross product LLR can be calculated as follows:

)는 (상태 또는 수준과 같은) 하드 값(

)이 판독된다고 가정할 때 상태 또는 수준(s)이 기록되거나 저장되었을 확률이다. Wherein P (s |

) Is the hard value (such as status or level)

Assuming that is read, it is the probability that the state or level (s) has been recorded or stored.

From flash memory hard  Pattern-dependent on the output ( Pattern - Dependent ) Cross product LLR

)만을 제공하는 경우, 외적 LLR은 공격자 셀 내에 저장된 패턴(

)에 기초하여 아래와 같이 산출될 수 있다:Soft data is not available from the flash memory, and the hard data is a state or level in which the flash memory is allocated to the data stored by the flash memory.

), Only the external LLR is used to store the pattern (

Can be calculated as follows:

,

|s)는 상태 또는 수준(s)이 본래 기록되거나 저장되었을 것을 가정할 때 (상태 또는 수준과 같은) 하드 값(

)이 판독되고 공격자 셀 내 패턴이

일 확률이다.

는, 타겟 셀 상 외란을 야기하는 ㅈ주위 셀(들) 또는 다른 셀 내에 저장된 데이터 패턴이다. 예를 들어:Where P (

,

| s) is a hard value (such as state or level), assuming a state or level (s) was originally recorded or stored.

) Is read and the pattern in the attacker cell

Is the probability.

Is the data pattern stored in the surrounding cell (s) or other cells causing disturbance on the target cell. E.g:

은 LLR이 산출되고 있는 위치(k, l)에서의 타겟 셀에 인접한 모든 공격자 셀을 나타낸다.

Denotes all attacker cells adjacent to the target cell at the position (k, l) where the LLR is being calculated.

)은 예를 들어 공격자 셀로부터의 하드 데이터를 판독함으로써 획득할 수 있다. pattern(

) Can be obtained, for example, by reading hard data from an attacker cell.

Alternatively, the pattern-dependent LLR can be calculated as follows:

,

)는 (상태 또는 수준과 같은) 하드 값(

)이 판독되고 공격자 셀 내 패턴이

인 것을 가정할 때 상태 또는 수준(s)이 본래 기록되거나 저장되었을 확률이다. Wherein P (s |

,

) Is the hard value (such as status or level)

) Is read and the pattern in the attacker cell

Assuming that is the probability that the state or level (s) was originally recorded or stored.

External product with no soft decoder feedback on soft output from flash LLR

If there is no soft output from the decoder used in the soft demapper / soft data generator (ie, L _a (C _i ) = 0), the soft demapper / soft data generator external LLR can be calculated as follows:

This external LLR is then sent to the decoders shown in FIGS. 10 and 11. The LDPC may apply a message-passage decoding algorithm, for example for local iteration in the decoder, until the data bits are decoded. In this case, there is no global detection / decoding iteration between the soft demapper / soft data generator that is performed to reduce the overall computational complexity.

The pattern-dependent LLR in this case can be calculated as follows:

From flash hard  Cross product with no soft decoder feedback on output LLR

If no soft data is available from the flash memory and no soft output from the decoder is used to reduce the computational complexity, the external LLR can be calculated as follows:

}는 하드 값(

)을 가정할 때 (전압과 같은) 소프트 값(r)에 대한 기대값 또는 소프트 값(r)에 대한 다른 어떤 추산치이다. P(

|s)는 상태 또는 수준(s)이 본래 기록되거나 저장된 것을 가정할 때 (상태 또는 수준과 같은) 하드 값(

)이 판독되는 확률이다. In which E {r |

} Is the hard value (

) Is an expected value for soft value r (such as voltage) or some other estimate of soft value r. P (

| s) is a hard value (such as state or level), assuming that state or level (s) is originally recorded or stored.

) Is the probability of reading.

In another embodiment, the LLR can be calculated as follows:

)는 (상태 또는 수준과 같은) 하드 값(

)이 판독되는 것을 가정할 때 상태 또는 수준(s)이 본래 기록되거나 저장되었을 확률이다. Where P (s |

) Is the hard value (such as status or level)

Assuming that is read, it is the probability that the state or level (s) was originally recorded or stored.

The pattern-dependent LLR in this case can be calculated as follows:

Cross product for soft output from flash LLR Gaussian approximation of approximation )

If the soft output from the flash memory (such as the read threshold voltage) is modeled with Gaussian variance, then the conditional to the soft output p (r) assuming the level (s) originally stored or written. The PDF of p (r | s) can be expressed as:

Where σ (s) is the standard deviation and E {r | s} is the average or expected value of the soft output (such as the threshold voltage) for state s.

The cross product LLR can then be calculated as follows:

If the voltage variances for all states have the same standard deviation (σ (s) = σ), this equation can be simplified to the following expression:

This formula can be further simplified:

From flash hard  External to output LLR Gaussian approximation

If the soft output from the flash memory is not available, assuming that the soft output is Gaussian distributed, the LLR can be calculated as follows:

}은, 플래시 메모리에 의해 할당되고 제공되는 상태 또는 수준인, 하드 출력(

)에 대한 (문턱 전압과 같은) 소프트 출력(r)의 평균 또는 기대값이다. E {r | s} in the above equation is the average or expected value of the soft output r (such as the threshold voltage) for state s, and E {r |

} Is the hard output (state or level allocated and provided by flash memory).

Is the average or expected value of the soft output r (such as the threshold voltage).

If the voltage variances for all states have the same standard deviation (σ (s) = σ), this equation can be simplified to the following expression:

This formula can be further simplified:

From flash memory hard  Pattern-dependent cross product for output LLR Gaussian approximation

If the variance for the soft output is modeled as a Gaussian, then the distal pattern-dependent LLR at the hard output can be calculated as follows:

는 위에서 정의된 바와 같이 공격자 셀에 저장된 패턴이고, σ(s,

)는 상태(s)와 패턴(

)에 대한 소프트 출력의 분산의 표준 편차이다. In the above formula

Is the pattern stored in the attacker cell as defined above, and σ (s,

) Is the state (s) and the pattern (

Is the standard deviation of the variance of the soft output.

), 이 수식은 아래의 표현과 같이 단순화될 수 있다:If the voltage variances for all states and patterns have the same standard deviation, then σ (s,

This formula can be simplified to the following expression:

This formula can be further simplified:

Gaussian approximation for cross product LLRs without soft decoder feedback on soft output from flash memory

If soft decoder feedback is not used, the external LLR can calculate as follows using a Gaussian approximation for soft output variance when soft output from flash memory is available:

If the voltage variances for all states have the same standard deviation (σ (s) = σ), this equation can be simplified to the following expression:

This representation can be further simplified:

From flash memory hard  Gaussian approximation to the cross product LLR without soft decoder feedback on the output

If no soft decoder feedback is used, then the external LLR can be calculated using the Gaussian approximation for soft output variance when only hard output from flash memory is available:

If the voltage variances for all states have the same standard deviation (σ (s) = σ), this equation can be simplified to the following expression:

This formula can be further simplified:

The corresponding pattern dependent LLR can be calculated as follows:

If the voltage variances for all states and patterns have the same standard deviation (σ (s) = σ), this equation can be simplified to the following expression:

This can be further simplified:

Reading Statistics Table read statistics tables )

)) 수준에 대해, 기록 수준(s)이 기록된 경우, 판독 수준(

)이 판독된 횟수를 나타내는 셀 기반 통계 카운트 표(1700)의 예이다. 예를 들어, 판독 수준(

=00)은, 기록 수준(s) 또한 00인 경우, 10617번 판독되었다. 또한, 판독 수준(

)은, 기록 수준(s)이 01인 경우, 148번 에러로 판독되었다. 카운트 표(1700)는 또한 선택적으로 각 행(row)과 열(column)의 합을 나타낸다. 카운트 표(1700) 내의 값은 도 21, 23, 25 및 28을 참조하여 아래에서 설명되는 여러 셀 기반 통계 프로세스에 의해 사용된다. 17A-17C are examples of cell-based statistical tables that record statistics for reading data from flash memory. Figure 17A shows a given pair of (written (s) and read (

For levels), if the recording level (s) is recorded, then the reading level (

) Is an example of a cell-based statistical count table 1700 that represents the number of times it has been read. For example, read level (

= 00) was read out 10601 times when the recording level s was also 00. Also, the reading level (

) Was read with error 148 when the recording level s was 01. Count table 1700 also optionally represents the sum of each row and column. The values in the count table 1700 are used by various cell based statistical processes described below with reference to FIGS. 21, 23, 25, and 28.

)) 수준에 대해, 판독 수준(

)이 판독되었을 것을 조건으로 하여 기록 수준(s)이 기록된 가능성(p(s|

))을 나타내는 셀 기반 통계표의 예이다. 도 17C는, 주어진 한 쌍의 (기록(s)과 판독(

)) 수준에 대해, 기록 수준(s)이 기록되었을 것을 조건으로 하여 판독 수준(

)이 판독된 가능성(p(

|s))을 나타내는 셀 기반 통계표(1740)의 예이다. Figure 17B shows a given pair of (write s and read (

)) For level, read level (

The probability that the recording level (s) has been written, provided that

Is an example of a cell-based statistical table. Figure 17C shows a given pair of (write s and read (

For the level), provided that the recording level (s) has been recorded

) Has been read (p (

(s)) is an example of a cell-based statistical table 1740.

)이 있을 때, 주어진 한 쌍의 (기록(s)과 판독(

)) 수준에 대해서, 기록 수준(s)이 기록되었을 것을 조건으로 하여, 패턴(

)이 존재할 때 판독 수준(

)이 판독된 확률(p(

,

|s))을 나타낸다. 18 is an example of a pattern dependent cell based statistics table 1800 that records pattern dependent statistics for reading data from flash memory when a given pattern is present. Exemplary table 1800 shows a given pattern (

), Given a pair of (write (s) and read (

With respect to the level), the pattern (

) When present, read level (

) Reads probability (p (

,

| s)).

Collection of statistics

standard( reference ) Collect statistics using cells

19 illustrates the example flash cell array of FIG. 3 in more detail. As shown in FIG. 19, in order to provide reliable channel estimation or statistics under all operating conditions, the exemplary flash cell array 1900 may include a plurality of reference cells 1900-ref ₁ to 1900-ref _N (hereinafter referred to as reference). Collectively referred to as cell 1920).

Exemplary reference cell 1920 is shown in FIG. 19 with an underlined background. The reference cell 1920 can be programmed periodically or intermittently using known patterns, such as known bit patterns or known symbol patterns. Note that the reference cell 1920 can be spread between the flash cell arrays 1900 using any preferred method, for example, an unchanging or varying number of cells in each wordline. For example, the position of the reference cell 1920 can change or be fixed over time to avoid a cell that has been warped out or damaged. In one implementation, the position of the reference cell 1920 is fixed and the performance of the same cell 1920 can be observed over time. In this fixed position implementation, the reference cell 1920 may optionally be written only once or read and read as many times as comparable to other cells in the flash memory array.

In another variation, the position of the reference cell 1920 changes over time so that the performance of the reference cell 1920 reflects the performance of all the arrays 1900. In another variation, statistics may be obtained from the reference cell 1920 in the plurality of different arrays 1900, and then the results may be averaged.

As described further below, the reference cell 1920 can be read and compared to a known pattern. For example, an estimate of the probability of detecting an error (p ₀ ) can be obtained as follows:

Programming and reading of the reference cell can be optionally combined with a wear-level algorithm that spreads wear and tear over memory, in known manner.

In various embodiments, reference cell 1920 may store all possible levels, have a periodic pattern (alternate), and may also be written or read periodically over time.

As discussed below, various embodiments of the present invention collect and use bit based statistics, cell based or pattern dependent statistics. In an embodiment using bit-based statistics, bit error performance is measured. In an embodiment using cell based statistics, read statistics are measured on a cell basis. For pattern dependent statistics, the read statistics also take into account the data patterns stored in the attacker cell.

1. Bit-based Statistics Using Reference Cells

20 is a flowchart illustrating an example implementation of a bit-based statistics generation process 2000 for a reference cell embodiment of the present invention. In general, the bit-based statistics generation process 2000 calculates a probability p ₀ for detecting bit errors. The probability of detecting an error p ₀ is then used by the LLR generator 1550 (FIG. 15) to produce the desired soft data. Initially, during step 2010 the statistics generation process 2000 writes the known pattern into one or more reference cells 1920. As mentioned above, the known pattern may be, for example, a known bit pattern or a known symbol pattern.

Thereafter, the reference cell is read during step 2020. Then, during step 2030, the statistics generation process 2000 determines an error metric, such as the number of error bits in the reference cell 1920. As discussed above, the reference cell 1920 that was read during step 2020 may be compared to a known pattern.

The statistics generation process 2000 calculates the error probability during step 2040 as follows:

2. Cell Based Statistics Using Reference Cells

21 is a flowchart illustrating an example implementation of a cell-based statistics generation process 2100 for a reference cell embodiment of the present invention. As shown in FIG. 21, cell-based statistical generation process 2100 initially writes one or more known voltage levels to reference cell 1920 during step 2110.

또는 LVL_read)이 판독된 횟수를 단계(2130) 동안 카운트한다. The cell based statistics generation process 2100 then reads the voltage level during step 2120 from the reference cell 1920. For each possible recording level (s or LVL _writ ), the cell-based statistics generation process 2100 may determine that each recording level (s or LVL _writ ) has been recorded.

Or LVL _read ) is counted during step 2130.

Error probability statistics are calculated during step 2140 as follows:

In the above formula, # means number.

Alternatively, the error probability statistic can be calculated as follows during step 2140 (inversely):

Other normalization terms may be used in the denominator for the equations produced during step 2140.

Statistics Collection Using Decoded Codewords

In a decoded codeword embodiment of the present invention, soft data is generated for a memory device, such as flash memory devices 810 and 910, using data obtained from the decoded codeword as a reference cell. In general, hard data from a memory device such as flash memory is decoded and error metrics such as the number of decoded bits in error are obtained. For example, the number of erroneous decoded bits can be obtained by comparing the decoded bits to hard data obtained from the memory device. In this way, the decoded codeword can be assumed to be correct and can be used as the aforementioned reference cell.

1. Bit-based Statistics Using Decoded Codewords

22 is a flowchart illustrating an example implementation of a bit-based statistics generation process 2200 for a decoded codeword embodiment of the present invention. In general, the bit-based statistics generation process 2200 uses the decoded codeword to calculate the probability of detection of an error p ₀ . The error detection probability p ₀ can then be used by the LLR generator 1500 (FIG. 15) to yield the desired soft data. Initially, the statistics generation process 2000 obtains hard data from flash memory during step 2210.

The bit-based statistics generation process (decoded codeword) 2200 then decodes the hard data during step 2220. Error metrics, such as the number of error bits from flash memory, are determined during step 2230. The number of error bits can be determined, for example, by comparing the decoded bits (assuming correct) to hard data from flash memory.

The statistical generation process 2200 calculates error probability statistics during step 2240 as follows:

2. Cell-Based Statistics Using Decoded Codewords

23 is a flow diagram illustrating an example implementation of a cell-based statistics generation process (decoded codeword) 2300 incorporating features of the present invention. In general, the statistics generation process 2300 uses the decoded codewords to calculate cell based error probabilities. Initially, cell based statistics generation process 2300 obtains hard data from flash memory during step 2310.

Cell-based statistics generation process (decoded codeword) 2300 then decodes the hard data during step 2320. The decoded bits are mapped to corresponding voltage levels during step 2325.

또는 LVL_read)이 판독된 횟수를 단계(2330) 동안 카운트한다. Then, for each possible voltage level (s or LVL _decod ), the cell based statistical process (decoded codeword) 2300 may determine if each decoded level s LVL _decod was decoded.

Or LVL _read ) is counted during step 2330.

Error probability statistics are calculated during step 2340 as follows:

Alternatively, the error probability statistics can be calculated as follows during step 2340 (inversely):

Condition-specific ( condition - specific Error probability

As discussed above, error probability statistics may be selectively obtained for different conditions such as different locations of the memory array, different patterns of attacker cells, different numbers of program / erase or read cycles, and the like. Then, if the same condition is observed, soft data can be obtained using appropriate condition dependent statistics or probabilities.

As discussed below with reference to FIGS. 24 and 25, example location specific statistics generation processes 2400 and 2500 use bit-based statistics and cell-based statistics, respectively, for error probabilities for different locations in the memory array. Acquire statistics.

Bit-based position-specific statistics

24 is a flow diagram illustrating an example bit-based location specific statistics generation process 2400 for estimating the detection probability of bit error for a number of different locations in a memory array. For example, the probability of detecting an error (p _{0. LOC)} is different from the page location of one or more word line position, (such as even and odd bit lines), the bit line position and (MSB (most significant bit) and LSB (least can be obtained for different bits in a multi-level cell (such as significant bit). As shown in FIG. 24, an exemplary bit-based location specific statistics generation process 2400 may determine the number of error bits in a desired location of a decoded codeword or reference cell based on desired location specific statistics during step 2430. Decide early. For example, if the location specific statistics are for the MSB, the number of error MSB bits is evaluated during step 2430. Note that if MSB statistics are being obtained, for example, any other bits in each cell can be selectively ignored.

The location specific statistics generation process 2400 then calculates the location specific error probability statistics during step 2440 as follows:

Cell-based location specific stats

For cell-based location specific implementations, different interest locations in memory may include, for example, one or more wordline locations or bitline locations (such as even and odd bitlines).

LVL_read)이 판독된 횟수를 카운트한다. 25 illustrates example cell based location specific statistics for obtaining error probability statistics for a number of different locations in a memory array 1900, such as one or more different wordline locations or bitline locations (such as even and odd bitlines). A flow diagram illustrating a creation process 2500. As shown in FIG. 25, for each possible reference voltage level s or LVL _ref , an exemplary cell-based location specific statistics generation process 2500 initially begins with reference level s or LVL _ref during step 2530. Is decoded or recorded, the voltage level within each desired position (

Count the number of times LVL _read ) has been read.

The cell-based location specific statistics generation process 2500 then calculates the location specific error probability statistics during step 2540 as follows:

)는 전술한 바와 같이 산출될 수 있다. In contrast, p (s |

) May be calculated as described above.

In various implementations of the invention, separate bit-based, cell-based, or pattern-based page statistics can be generated for each wordline or memory array, or for a page, wordline, or memory array group (eg, within a wordline). For different page levels, or for lower, middle, and upper word lines in a memory array. It is also possible to average the statistics across a plurality of pages, word lines or memory arrays, and then the average statistics can be used for this page, leader or memory array.

Pattern-dependent statistics

,를 산출하고, 여기에서

는 (서라운딩 셀(들)과 같은) 공격자 셀 내에 저장된 데이터 패턴이다. As noted above, various embodiments of the present invention provide for cross product LLRs, for one or more soft values, L _e , for target cells and one or more values, r, for attacker cell (s),

, And here

Is the data pattern stored within the attacker cell (such as the surrounding cell (s)).

FIG. 26 illustrates an exemplary set 2600 of probability density functions 2610 for a given target cell 710 of the exemplary multilevel cell flash memory 600, based on all possible values of each attacker cell 720. Illustrated. An exemplary multilevel cell flash memory has four levels (2 bits) per cell, and one attacker cell 720 is considered for data dependent pdfs. The number of probability density functions applicable to each possible level of a given target cell 710 is the number of possible levels for each attacker cell 720 which is the number of attacker cells 720 affecting a given target cell 210. . As noted above, in an exemplary embodiment, each cell can have one of four possible values, one attacker cell 720 per target cell 710, and each attacker cell 720 has four It can have one of the possible levels. Thus, for example, the set of probability density functions 2600 includes four probability density functions 2610-1-2610-4 for data or voltage level 0 due to a pattern of attacker cells. There are also four probability density functions for each of the other data levels 1, 2 and 3. It will be apparent to those skilled in the art that the present invention can be extended to multi-level cell flash memory 600 having any number of levels per cell and any number of attacker cells 720.

In general, each of the probability density functions in FIG. 26 indicates, among other noise and disturbance effects, the ICI effect on a given target cell 710 for a given value of the corresponding attacker cell 720. In another embodiment of the invention, the data dependent probability density function may represent other data dependent distortions instead of or in conjunction with the ICI. As discussed below, in various embodiments, the probability density function is predefined and fixed, adapted based on real-time observations, or measured with respect to the attacker cell 720, such as a Gaussian function. Or as a function of the sensed value h.

According to one aspect of the invention, the disturbance in the flash memory device may be characterized by obtaining one or more probability density functions indicative of pattern dependent disturbance of one or more attacker cells on the at least one target cell. Other disturbances may include, for example, back pattern dependencies, intercell interference, program disturb, read disturb, and / or additional noise. The probability density function may be updated based on one or more data determination information. The probability density function may be represented as a stored table and / or expression.

)에 기초하여 선택된다. 그 후, 선택된 확률 밀도 함수는, 공지의 기술을 이용하여, 수신된 타겟 셀 값(r)에 기초하여, 최근의 발생(occurrence)으로 (예를 들어, 대응하는 카운터를 증가시킴으로써) 업데이트된다. The table entry or function parameter of the probability density function may be selectively and adaptively updated, for example based on the received data determination information. For example, the probability density function can be used to determine the

Is selected based on The selected probability density function is then updated using a known technique, based on the received target cell value r, with a recent occurrence (eg, by incrementing the corresponding counter).

As discussed above, the number of attacker cells 720 affecting a given target cell 710 may be reduced or ignored based on a plurality of factors. In this way, the number of probability density functions that need to be considered can be reduced. For example, in an exemplary embodiment of migrating ICI, if the diagonal coupling coefficient (k _xy ) is much smaller than the other coupling coefficients (as is often the case), the cells are located diagonally. ICI from can be ignored. In addition, the programming sequence can affect the number of attacker cells 720 to be considered. For example, if a wordline is always written in a fixed order, e.g., from a bottom up approach, there may be no disturbance ICI contribution from the cells in the lower wordline. In addition, when the disturbance ICI is symmetric about the left and right adjacent to the target cell 710, the number of probability density functions that need to be characterized is reduced by half.

As mentioned above, in one exemplary implementation, the probability density function may be approximated using a Gaussian probability density function. In another variant, if the probability density function is based, for example, on a histogram, improved performance may be obtained instead of adding complexity. If the probability density function is implemented using a histogram, the probability density function can be adaptively updated using a successfully decoded wordline to train the histogram.

In another embodiment, the probability density function and its approximation can be used by trellis based sensing algorithms such as Viterbi, Soft Output Viterbi (SOVA) and BCJR to sense read data.

1. Bit-based pattern dependent statistics

또는 PATT)에 대한 비트 에러를 감지하는 확률(p_{0 , PATT})을 추산하는 예시적인 비트 기반 패턴 의존형 통계 생성 프로세스(2700)를 도시하는 흐름도이다. 초기에, 단계(2720) 동안 비트 기반 패턴 의존형 통계 생성 프로세스(2700)는 기준 타겟 셀(710)과 잠재적으로(potentially) 연관 공격자 셀(들)(720)을 판독한다. 또한, 각 판독 타겟 비트에 대해, 연관 공격자 셀(720)의 패턴(PATT)은 단계(2725) 동안 식별(identify)된다. 패턴은, 예를 들어, 기록된 알려진 패턴을 평가함으로써 또는 디코딩된 코드워드 또는 기준 셀의 실제 판독 동작에 기초하여, 단계(2725) 동안 식별된다. 27 illustrates a given pattern of one or more attacker cells 720 associated with at least one target cell 710 (FIG. 7).

Or an example bit-based pattern dependent statistics generation process 2700 that estimates the probability (p _{0 , PATT} ) of detecting a bit error for _PATT ). Initially, during step 2720 the bit-based pattern dependent statistics generation process 2700 reads potentially associated attacker cell (s) 720 with the reference target cell 710. In addition, for each read target bit, the pattern PATT of the associated attacker cell 720 is identified during step 2725. The pattern is identified during step 2725, for example, by evaluating a known pattern recorded or based on the decoded codeword or actual read operation of the reference cell.

For one or more identified patterns, the number of error target bits having a corresponding pattern is determined during step 2730. The error probability statistics are then calculated during step 2740 as follows:

It will be apparent to those skilled in the art that the above described techniques can be selectively integrated to obtain location specific, pattern dependent statistics. Further, in other variations, read statistics may additionally or alternatively be obtained as a function of endurance, read cycles, retention, temperature, or other parameters of the memory device.

2. Cell-Based Pattern-Dependent Statistics

또는 PATT)은 단계(2825) 동안 식별된다. 28 is a flow diagram illustrating an example cell-based pattern dependent statistics generation process 2800 that estimates the probability of detecting an error for a given pattern of one or more attacker cells associated with at least one target cell. As shown in FIG. 28, the cell based pattern dependent statistics generation process 2800 initially reads one or more target cells during step 2820. Then, the pattern of associated attacker cell (s) (

Or PATT) is identified during step 2825.

또는 LVL_ref)에 대해서, 셀 기반 패턴 의존형 통계 생성 프로세스(2800)는 기준 수준(s 또는 LVL_ref)이 디코딩되었거나 기록되었던 경우 각 전압 수준(

또는 LVL_read)이 판독된 횟수를 단계(2830) 동안 카운트한다. Then, for one or more identified patterns, and for each possible reference voltage level (

Or LVL _ref ), the cell-based pattern-dependent statistical generation process 2800 may determine each voltage level (if the reference level s or LVL _ref ) has been decoded or written.

Or LVL _read ) is counted during step 2830.

The pattern dependent error probability statistics are calculated during step 2840 as follows:

Asymmetric Error Probability Statistics

As mentioned above, in certain channels, such as NAND flash memory channels, the probability of detecting an error for possible different binary values, such as binary 0 and binary 1, can vary significantly. Thus, the present invention optionally provides the probability of detecting an error for an asymmetric channel. 29 and 30 provide exemplary statistical generation processes 2900 and 30000 for estimating error probabilities p and q for two possible binary values, such as binary 1 and binary 0. As discussed further below, FIG. 29 estimates asymmetric statistics using a reference cell and FIG. 30 estimates asymmetric statistics using decoded codewords. The present invention thus provides an asymmetric LLR for each possible binary value based on hard data from flash memory.

Asymmetric Error Probability-Reference Cell

As mentioned above, in certain channels, such as NAND flash memory channels, the probability of detecting errors for possible different binary values, such as binary 0 and binary 1, can vary significantly. Thus, the present invention optionally provides the probability of detecting an error for an asymmetric channel. FIG. 29 is a flow diagram illustrating an exemplary asymmetric statistical generation process 2900 for estimating error probabilities for two possible binary values for a reference cell embodiment of the present invention.

As shown in FIG. 29, the asymmetric statistics generation process 2900 initially writes the known pattern to the reference cell 1920 during step 2910, and then reads the reference cell 1920 during step 2020. do. The asymmetric statistic generation process 2900 determines the number of error bits with binary zeros in the reference data during step 2930, and then calculates the error probability statistics for binary zeros during step 2940 as follows:

The asymmetric statistics generation process 2900 then determines the number of error bits with binary 1 in the reference data during step 2950, and then calculates the error probability statistics for binary 1 during step 2960 as follows. do:

Asymmetric Error Probability-Decoded Codeword

30 is a flow diagram illustrating an exemplary asymmetric statistical generation process 3000 for estimating error probabilities for two possible binary values in a decoded codeword embodiment of the present invention. As shown in FIG. 30, asymmetric statistical generation process 3000 initially acquires hard data from flash memory during step 3010 and decodes hard data during step 3020.

The asymmetric statistics generation process 3000 then determines, during step 3030, the number of error bits from the flash memory with binary zeros in the decoded data. The error probability statistics for binary 0 are then calculated during step 3040 as follows:

Similarly, the number of error bits from the flash memory having binary 1 in the decoded data is then determined during step 3050. Error probability statistics for binary 1 are then calculated during step 3060 as follows:

In one embodiment, statistics may be collected, calculated, and stored while the NAND flash memory is idle (ie, not actively recording or reading user data).

While the exemplary embodiment uses statistics collection using reference cell or decoder feedback, the adaptive method is also used to detect statistics using sensed or decoded data, eg, using a minimum mean square error criterion. Estimate

In another embodiment, statistics or corresponding LLRs may be used for worst-case operating conditions (eg, in terms of number of program / erase cycles, retention time and temperature), eg For example, it can precompute based on experimental characterization of the flash memory and then use it for bad channel conditions. In this way, more accurate statistics or LLRs are available where the error probability is highest. In other words, pre-defined statistics or corresponding LLRs can be calculated in advance for predefined operating conditions.

In another variation, soft data may be generated repeatedly based on different statistics (eg, error probability) until decoding is successful. Statistics may change over a range until successful detection or decoding of the data. This variant of the invention provides a virtual reread of the data. Even if the data is not actually read back from the flash memory, the data is successfully decoded into other soft information.

Dissatisfaction parity ( parity Error performance based on performance )

One aspect of the present invention recognizes that an unsatisfactory parity check can be used as a performance metric to obtain soft data. (N, K, J, L) LDPC code-N is the codeword length, K is the uncoded codeword length (user data length in the codeword), and J and L are the columns of the parity check matrix, respectively. ) And row weights. If the (N, K, J, L) LDPC codeword is stored or passed with an error probability p ₀ , the probability that the check sum fails in the first iteration can be expressed as follows:

This probability can be estimated as follows:

So, the error probability p ₀ can be estimated as:

In this process, the channel and initial LLR values can be estimated before performing the iterative decoding. Channel estimation complexity and latency are less than that of one iteration of soft-decision decoding with a significant gain for hard decision decoding of the LDPC code. Hardware added in comparison to the standard implementation of soft-decision decoding is a block that performs the following calculations:

31 is a flowchart illustrating an example implementation of a statistics generation process 3100 using an unsatisfactory parity check in accordance with an aspect of the present invention. In one embodiment, an unsatisfactory parity check after the first iteration is used. In general, the statistics generation process 3100 uses an unsatisfactory parity check to calculate the probability of detecting an error p ₀ . The error detection probability p ₀ can then be used by the LLR generator 1500 (FIG. 15) to calculate the desired soft data.

Initially, the statistics generation process 3100 obtains the number of unsatisfactory parity checks during step 3110. The statistics generation process 3100 then calculates the error probability statistics during step 3120 as follows:

Location specific stats-dissatisfaction parity check

32 is a flow diagram illustrating an example location specific statistics generation process 3200 for obtaining error probability statistics for a plurality of different locations within a memory array 1900 using unsatisfactory parity check. For example, error probability statistics can include one or more different page positions, wordline positions, bitline positions (such as even and odd bitlines), and multilevel cells (such as Most Significant Bits (LBS) and Least Significant Bits (LSB)). Can be obtained for different bits. In general, location specific statistics may be obtained using an unsatisfactory parity check by using a codeword that locates the bits in the desired location (step 3210).

As shown in FIG. 32, the example location specific statistics generation process 3200 then obtains the number of unsatisfactory parity checks for the codeword during step 3220. The location specific error probability statistics are then calculated during step 3230 as follows:

Asymmetric Statistics Based on Unsatisfactory Parity Check

)이 불만족 패리티 체크에 기초하여 산출될 수 있다는 것을 인식한다(여기에서

). p와 q에 대한 값은 평균 확률 확률(

)과, 에러 확률(p와 q)의 비율(k)에 기초하여 산출될 수 있다. 33 is a flow diagram illustrating an example asymmetric statistics generation process 3300 that estimates the probability of detecting an error for two possible binary values using an unsatisfactory parity check. This aspect of the invention relates to an average error probability (

It can be appreciated that can be calculated based on the dissatisfaction parity check (where

). The values for p and q are the average probability probabilities (

) And the ratio k of the error probabilities p and q.

The ratio k of the error probabilities p and q can be obtained using data analysis, eg, the decoded codeword technique described above. Alternatively, the ratio k of error probabilities p and q is, for example, the international patent application "Methods and Apparatus for Soft Data Generation for memory devices Using Reference Cells, which is filed concurrently with, and incorporated herein by reference. Can be obtained using the reference cell technique described in " The ratio k of the error probabilities p and q is typically calculated offline and will be stored, for example, in a table. As shown in FIG. 33, an exemplary asymmetric statistics generation process (dissatisfaction parity check) 3300 initially obtains the ratio k of error probabilities p and q during step 3310.

)은 도 16을 참조하여 전술한 기술을 사용하여 단계(3320) 동안 획득할 수 있다. 특히, 평균 에러 확률 확률(

)은 아래와 같이 추산될 수 있다:Mean error probability probability (

) May be obtained during step 3320 using the techniques described above with reference to FIG. 16. Specifically, the mean error probability probability (

) Can be estimated as:

The error probability statistics p for binary 0 are then calculated during step 3330 as follows:

Then, the error probability statistics q for binary 1 are calculated during step 3340 as follows:

Note that the error probability statistics p and q calculated by the asymmetric statistics generation process (dissatisfaction parity check) 3300 are optionally location specific and / or pattern dependent.

Processes, systems and products ( article of manufacture ) Details

While a number of flowcharts herein illustrate an exemplary order of steps, it is also an embodiment of the present invention that the order may vary. Various permutations of the algorithm are contemplated as another embodiment of the present invention. As will be apparent to one skilled in the art, although exemplary embodiments of the present invention have been described in connection with processing steps in a software program, various functions may be performed in the digital domain as processing steps in the software program, in hardware by circuit components, or It may be implemented in a state machine or a combination of both hardware and software. Such software may be used, for example, in digital signal processors, application specific integrated circuits, microcontrollers or general-purpose computers. Such hardware and software may be implemented within circuitry implemented within integrated circuits.

Thus, the functionality of the present invention may be implemented in the form of methods and apparatus for carrying out such a method. One or more aspects of the invention are embodied in the form of program code, for example, stored on a storage medium, loaded on a machine, and / or executed by a machine, or transmitted over any transmission medium, the program code being a computer When loaded into and executed by a machine, such a machine becomes an apparatus for implementing the invention. If implemented in a general-purpose processor, the program code segment may be combined with the processor to provide an apparatus that operates similar to a particular logic circuit. The invention may also be implemented in the form of one or more integrated circuits, digital signal processors, microprocessors and microcontrollers.

As known in the art, the methods and apparatus discussed herein may be distributed as an article of manufacture that itself includes a computer readable medium having computer readable code means implemented thereon. The computer readable program code means may be operable in conjunction with a computer system to perform all or some of the steps for carrying out the method or manufacturing the device discussed herein. Computer-readable media may be recordable media (eg, floppy disks, hard drives, compact discs, memory cards, semiconductor devices, chips, ASICs), and transmission media (eg, optical fibers, worldwide web, cables) Or a network including a radio channel using time division multiple access, code division multiple access, or another radio frequency channel). Any medium known or developed that can store information appropriate for use with a computer system may be used. Computer readable code means is any mechanism that allows a computer to read inline lines and data, such as changes in height on the surface of a compact disc or changes in magnetic properties of magnetic media.

The computer system and server described herein each comprise an associated processor and each includes a memory for implementing the methods, steps, and functions disclosed herein. The memory may be distributed or local, and the process may be distributed or singular. The memory may be implemented as an electrical, magnetic or optical memory or any combination of these or other types of storage devices. In addition, the term “memory” should be construed broadly enough to include any information that can be read or written to or from an address in an addressable space connected by the associated processor. With this definition, information on the network is also in memory because the relevant processor can retrieve information from the network.

It will be appreciated that the embodiments and variations shown and described herein are merely illustrative of the principles of the invention and that various modifications may be made by those skilled in the art without departing from the scope and spirit of the invention.

Claims (37)

A method of generating at least one soft data value in a memory device, the method comprising:
Recording known data into one or more reference cells;
Reading at least one of the reference cells;
Obtaining read statistics based on the read one or more reference cells;
Obtaining the at least one soft data value based on the obtained read statistics.
How to generate soft data values.
The method of claim 1,
The reading step reads one or more of data bits, voltage levels, current levels, and resistance levels.
How to generate soft data values.
The method of claim 1,
The reading step reads one or more of soft data and hard data.
How to generate soft data values. The method of claim 1,
The soft data value includes a soft read value used to generate one or more log likelihood ratios (LLRs).
How to generate soft data values.
The method of claim 1,
The soft data value includes one or more LLRs.
How to generate soft data values.
The method of claim 1,
One or more of the steps may be implemented by one or more of a controller, read channel, signal processing unit and decoder.
How to generate soft data values.
The method of claim 1,
The read statistics are determined by comparing one or more bits read from the one or more reference cells with one or more bits in the known data.
How to generate soft data values.
The method of claim 1,
The known data includes one or more of a bit pattern and a symbol pattern.
How to generate soft data values.
The method of claim 1,
The known data is written to the reference cell using one or more of voltage, current and resistance.
How to generate soft data values.
The method of claim 1,
The read statistics are based on the number of faulty bits relative to the total number of reference bits in the known data.
How to generate soft data values.
The method of claim 1,
The soft data value L _e ,

(

= 1)

(

= 0)
Obtained by
p ₀ is the read statistic and c is the hard bit received from the memory device
How to generate soft data values.
The method of claim 1,
One or more of the read statistics and the soft data values are obtained from a table.
How to generate soft data values.
The method of claim 1,
One or more of the read statistics and the soft data values may include: endurance, number of program / erase cycles, number of read cycles, retention time, temperature, temperature change, process corners, intercell interference effects, the memory Computed as a function of one or more of a location within a device, a location within a wordline, a location within a memory array, or a pattern within an attacker cell
How to generate soft data values.
The method of claim 1,
The read statistics are obtained for one or more desired locations of a memory array, and the writing step includes writing the known data to the one or more desired locations.
How to generate soft data values.
The method of claim 14,
The desired position includes one or more of a given bit, page position, wordline position, and bitline position in a multilevel cell.
How to generate soft data values.
The method of claim 15,
The bit line position includes an even bit line or an odd bit line.
How to generate soft data values.
The method of claim 15,
The given bit in the multilevel cell comprises one or more of a most significant bit (MSB) and a least significant bit (LSB).
How to generate soft data values.
The method of claim 15,
The known data is written to a desired position corresponding to one or more of the bits in the multilevel cell.
How to generate soft data values.
The method of claim 14,
The read statistics are based on the number of faulty bits in the one or more desired positions relative to the total number of bits considered.
How to generate soft data values.
The method of claim 14,
The read statistics for one or more desired locations are employed to decode data associated with the one or more desired locations.
How to generate soft data values.
The method of claim 1,
The read statistics are obtained for a given pattern (PATT) in one or more attacker cells associated with at least one target cell,
The writing step comprises writing the known data to the at least one target cell,
The method
Determining the number of erroneous target bits with the given pattern (PATT) in the one or more attacker cells and obtaining the read statistics for the given pattern based on the number of erroneous target bits. doing
How to generate soft data values.
The method of claim 21,
Reading the one or more attacker cells to determine the pattern PATT.
How to generate soft data values.
The method of claim 21,
Recording a known pattern in the one or more attacker cells;
How to generate soft data values.
The method of claim 21,
The read statistics are based on the number of faulty target bits relative to the total number of target bits.
How to generate soft data values.
The method of claim 21,
The read statistics for the given pattern PATT are employed to decode data when the one or more associated attacker cells include the pattern PATT.
How to generate soft data values.
The method of claim 1,
The read statistics comprise asymmetric read statistics obtained for a plurality of possible values,
The method
Determining the number of erroneous bits having a first value in the known data, determining the number of erroneous bits having a second value in the known data, and the erroneous having the first value. Obtaining read statistics for the first value based on the number of bits, and obtaining read statistics for the second value based on the number of erroneous bits having the second value. doing
How to generate soft data values.
The method of claim 26,
The read statistics are based on the number of faulty bits with each value relative to the total number of reference bits with each value.
How to generate soft data values.
The method of claim 26,
The asymmetric read statistics are employed to decode the data having each value.
How to generate soft data values.
The method of claim 26,
The soft data value L _e is

(if c = 0)

(if c = 1)
Is the read statistic for the first value, q is the read statistic for the second value, and c is the hard bit received from the memory device.
How to generate soft data values.
The method of claim 1,
Providing the soft data value to a decoder.
How to generate soft data values.
31. The method of claim 30,
The soft data value is provided recursively to the decoder.
How to generate soft data values.
The method of claim 1,
The memory device is a flash memory device
How to generate soft data values.
The method of claim 1,
The memory device may store at least two data levels s per cell.
How to generate soft data values.
The method of claim 1,
The reference cell stores virtually all possible levels
How to generate soft data values.
The method of claim 1,
The reference cell has a periodic pattern
How to generate soft data values.
The method of claim 1,
The reference cell is written or read periodically over time
How to generate soft data values.
A system for generating at least one soft data value in a memory device, the system comprising:
A memory,
At least one processor coupled to the memory,
The processor writes known data in one or more reference cells, reads one or more of the reference cells, obtains read statistics based on the read one or more reference cells, and based on the obtained read statistics. Operative to obtain at least one soft data value
Soft data value generation system.

Images