Deep Learning Method for Soft Read Thresholds and LLR Generation

1 . A method for estimating parameters for reading data from a memory having a plurality of NAND program-voltage (PV) states, comprising:

determining respective counts of memory cell read patterns obtained by reading pages of the data from the memory with corresponding pre-determined read thresholds, wherein for a NAND PV state, three thresholds (RT1, RT2, RT3) are obtained according to used page read thresholds;

using the at least three read thresholds (RT1, RT2, RT3) and the corresponding survival function values (SF1, SF2, SF3) to determine parametric values for a PV distribution of at least one NAND PV state; and

utilizing at least one deep neural network (DNN) to infer a center read threshold between two of the NAND PV states, a soft read interval Δ, and LLR values, where the DNN takes, as an input, the parametric values for the PV distribution of the at least one NAND PV state.

2 . The method of claim 1 , wherein the reading pages of the data from the memory with voltages away from the at least three read thresholds (RT1, RT2, RT3) comprises reading the memory with voltages greater than the at least three read thresholds (RT1, RT2, RT3).

3 . The method of claim 1 , wherein the determining of the respective counts of the memory cell read patterns comprises:

reading a least significant bit (LSB) page, a center significant bit (CSB) page, a most significant bit (MSB) page, and an upper significant bit (USB) page of the memory; and

for each of sixteen patterns of 1s and 0s read from the LSB page, the CSB page, the MSB page, and the USB page, counting numbers of cells showing each of the sixteen patterns.

4 . The method of claim 3 , wherein the determining corresponding survival function values comprises:

for each page of the LSB page, the CSB page, the MSB page, and the USB page, calculating as the survival functions SF1, SF2, SF3 respective ratios of a) a number of cells associated with the PV state that have threshold voltages larger than the read thresholds RT1, RT2, and RT3 according to the sixteen patterns to b) an average cell count for each NAND PV state.

5 . The method of claim 1 , wherein the parametric values used by the at least one DNN comprises skew normal distribution parameters (ξ 1 , ω 1 , α 1 ) and (ξ 2 , ω 2 , α 2 ) from adjacent NAND PV states having a read valley in between, wherein ξ is a location, ω is a scale factor, and α is a shape factor for a skew normal distribution.

6 . The method of claim 5 , further comprising:

with the at least one DNN, producing a skew normal distribution model of each NAND PV state; and

determining, by the at least one DNN using the skew normal distribution model, the center read threshold, the soft read interval Δ, and the LLR values.

7 . The method of claim 1 , wherein the parametric values used by the at least one DNN comprise non-central t distribution parameters df, nc, loc, scale, where df stands for degrees of freedom, nc is a noncentrality parameter, loc is a location parameter that shifts the distribution, and scale is scale parameter that stretches the non-central t distribution.

8 . The method of claim 7 , further comprising:

besides determining the read thresholds RT1, RT2, and RT3 and the survival functions SF1, SF2, and SF3, obtaining a fourth read threshold (RT4) and a corresponding fourth survival function (SF4).

9 . The method of claim 8 , further comprising utilizing for the at least one DNN a first trained DNN which takes, as the input, a) the read thresholds RT1, RT2, RT3, and RT4 and b) the survival functions SF1, SF2, SF3, and SF4, and outputs the non-central t distribution parameters df, nc, loc, scale.

10 . The method of claim 9 , further comprising utilizing for the at least one DNN a second trained DNN which takes, as the input, the non-central t distribution parameters df, nc, loc, scale, and outputs the center read threshold, the soft read interval Δ, and the LLR values.

11 . A memory system, comprising:

a storage having a plurality of NAND program-voltage (PV) states therein; and

a controller in communication with the storage and configured to:

set at least three read thresholds (RT1, RT2, RT3) away from respective peaks of NAND PV states;

determine respective counts of memory cell read patterns obtained by reading pages of the data from the memory with corresponding pre-determined read thresholds, wherein for a NAND PV state, three thresholds (RT1, RT2, RT3) are obtained according to used page read thresholds;

use the at least three read thresholds (RT1, RT2, RT3) and the corresponding survival function values (SF1, SF2, SF3) to determine parametric values for a PV distribution of at least one NAND PV state; and

utilize at least one deep neural network (DNN) to infer a center read threshold between two of the NAND PV states, a soft read interval Δ, and LLR values, where the DNN takes, as an input, the parametric values for the PV distribution of the at least one NAND PV state.

12 . The memory system of claim 11 , wherein the controller is configured to read the memory with voltages greater than the at least three read thresholds (RT1, RT2, RT3).

13 . The memory system of claim 12 , wherein the controller is configured to:

read a least significant bit (LSB) page, a center significant bit (CSB) page, a most significant bit (MSB) page, and an upper significant bit (USB) page of the memory; and

for each of sixteen patterns of 1s and 0s read from the LSB page, the CSB page, the MSB page, and the USB page, count numbers of cells showing each of the sixteen patterns.

14 . The memory system of claim 13 , wherein the controller is configured to:

for each page of the LSB page, the CSB page, the MSB page, and the USB page, calculate as the survival functions SF1, SF2, SF3 respective ratios of a) a number of cells associated with the PV state that have threshold voltages larger than the read thresholds RT1, RT2, and RT3 according to the sixteen patterns to b) an average cell count for each NAND PV state.

15 . The memory system of claim 11 , wherein the parametric values used by the at least one DNN comprises skew normal distribution parameters (ξ 1 , ω 1 , α 1 ) and (ξ 2 , ω 2 , α 2 ) from adjacent NAND PV states having a read valley in between, wherein ξ is a location, ω is a scale factor, and α is a shape factor for a skew normal distribution.

16 . The memory system of claim 15 , wherein the controller is configured to:

with the at least one DNN, produce a skew normal distribution model of each NAND PV state; and

determine, by the at least one DNN using the skew normal distribution model, the center read threshold, the soft read interval Δ, and the LLR values.

17 . The memory system of claim 16 , wherein the parametric values used by the at least one DNN comprise non-central t distribution parameters df, nc, loc, scale, where df stands for degrees of freedom, nc is a noncentrality parameter, loc is a location parameter that shifts the distribution, and scale is scale parameter that stretches the non-central t distribution.

18 . The memory system of claim 14 , wherein the controller is configured to, besides determining the read thresholds RT1, RT2, and RT3 and the survival functions SF1, SF2, and SF3, obtain a fourth read threshold (RT4) and a corresponding fourth survival function (SF4).

19 . The memory system of claim 18 , wherein the controller is configured to utilize for the at least one DNN a first trained DNN which takes, as the input, a) the read thresholds RT1, RT2, RT3, and RT4 and b) the survival functions SF1, SF2, SF3, and SF4, and outputs the non-central t distribution parameters df, nc, loc, scale.

20 . The memory system of claim 19 , wherein the controller is configured to utilize for the at least one DNN a second trained DNN which takes, as the input, the non-central t distribution parameters df, nc, loc, scale, and outputs the center read threshold, the soft read interval Δ, and the LLR values.