Compute-in-memory using neural networks stored in a non-volatile memory for predictive block health assessment of the non-volatile memory
Predictive block health assessment of a non-volatile memory can be performed by storing a pre-trained neural network for this purpose on the memory device itself. As operational errors occur during operation of the memory device, this defect data is saved by the non-volatile memory and used as input data for the neural network to identify and retire potential bad blocks before data loss occurs.
1 . A non-volatile memory device, comprising:
a control circuit configured to connect to one or more arrays of non-volatile memory cells storing weights of a first neural network, the control circuit is configured to:
perform memory operations, including reading and writing of data, on the one or more arrays;
determine error indication data while performing a plurality of the memory operations;
write the determined error indication data to the memory cells of the one or more arrays;
perform a compute-in-memory inference operation with the first neural network using the written determined error indication data as input values; and
determine defective portions of the one or more arrays based on output of the inference operation.
2 . The non-volatile memory device of claim 1 , wherein the control circuit is formed on a control die, the non-volatile memory device further comprising:
a memory die including the one or more arrays of non-volatile memory cells, the memory die separate from and bonded to the control die.
3 . The non-volatile memory device of claim 1 , the one or more arrays of non-volatile memory cells further storing error indication data determined prior to performing the memory operations on the one or more arrays, and, wherein to perform the compute-in-memory inference operation the control circuit is further configured to:
perform the compute-in-memory inference operation with the first neural network using the written determine error indication data and the error indication data determined prior to performing the memory operations as input values.
4 . The non-volatile memory device of claim 1 , wherein the control circuit is further configured to:
perform error correction code (ECC) operations on data read the one or more arrays, wherein the error indication data determined while performing the memory operations includes ECC results.
5 . The non-volatile memory device of claim 1 , wherein the error indication data determined while performing the memory operations includes data writing errors.
6 . The non-volatile memory device of claim 1 , wherein the memory operations further include erasing of data and the error indication data determined while performing the memory operations includes data erasing errors.
7 . The non-volatile memory device of claim 1 , the one or more arrays of non-volatile memory cells further storing weights of a second neural network, the control circuit is further configured to:
receive input vectors for the second neural network; and
perform compute-in-memory vector-matrix multiplications between the input vectors for the second neural network and the weights of the second neural network.
8 . The non-volatile memory device of claim 1 , wherein the weights of the first neural network are stored in a first of one or more arrays of non-volatile memory cells and wherein to write the determined error indication data to the memory cells of the one or more arrays the control circuit is further configured to:
write the determined error indication data to the memory cells of the first array not storing weights of the first neural network.
9 . The non-volatile memory device of claim 1 , further comprising:
one or more memory dies each comprising one or more of the arrays, each of the arrays having a plurality of blocks of memory cell have a NAND architecture, wherein, to determine defective portions of the one or more memory, the memory control circuit is further configured to:
determine defective portions of the one or more arrays at an individual block level.
10 . The non-volatile memory device of claim 9 , wherein the memory control circuit is further configured to:
assign blocks determined to be bad to a bad block pool.
11 . The non-volatile memory device of claim 9 , wherein the weights of the first neural network are stored in a first of the one or more arrays and wherein to perform the compute-in-memory inference operation the control circuit is further configured to:
convert the written determined error indication data as input values into a first input vector; and
apply the input vector to the first array, wherein values of the first input vector are analog values.
12 . The non-volatile memory device of claim 11 , wherein the analog values of the input vector are encoded as voltage level amplitudes.
13 . The non-volatile memory device of claim 11 , wherein non-zero values of the analog values of input vector have a non-zero voltage amplitude encoded as a time duration.
14 . The non-volatile memory device of claim 9 , wherein the weights of the first neural network are stored in a first of the one or more arrays, wherein each of the blocks of non-volatile memory cells has a NAND architecture in which the memory cells are connected along word lines, and wherein to perform the compute-in-memory inference operation the control circuit is further configured to:
convert the written determined error indication data as input values into an input vector encoded as a set of bias levels; and
apply the input vector to the first array as the set of bias levels to word lines of a block of the first array.
15 . The non-volatile memory device of claim 9 , wherein the weights of the first neural network are stored in a first of the one or more arrays, wherein each of the blocks of non-volatile memory cells has a NAND architecture in which NAND strings of memory cells are connected along bit lines, and wherein to perform the compute-in-memory inference operation the control circuit is further configured to:
convert the written determined error indication data as input values into an input vector encoded as a set of bias levels; and
apply the input vector to the first array as the set of bias levels to bit lines of a block of the first array.
16 . The non-volatile memory device of claim 9 , wherein the weights of the first neural network are stored in a first of the one or more arrays, wherein each of the blocks of non-volatile memory cells has a three dimensional NAND architecture in which NAND strings extend vertically above a substrate through a plurality of horizontal word line layers, along which memory cells of the NAND strings are connected, and through a select gate layer, along which a select gate of each of the NAND strings is connected, the select gate layer having multiple individually biasable sections corresponding to sub-sets of the NAND strings of the array, and wherein to perform the compute-in-memory inference operation the control circuit is further configured to:
convert the written determined error indication data as input values into an input vector encoded as a set of bias levels; and
apply the input vector to the first array as the set of bias levels to a first plurality of the individually biasable sections of the select gate layer.
17 . The non-volatile memory device of claim 1 , further comprising:
a plurality of memory dies each comprising one or more of the arrays, wherein the weights of a first neural network are stored in an array of a first of the memory dies and the determined error indication data for all of the memory dies is written to the first of the memory dies.
18 . A method, comprising:
receiving test data on defects from a plurality of examples of a non-volatile memory die;
training a neural network using the test data on defects;
receiving a plurality of the non-volatile memory die;
programming a set of weights of the trained neural network into the plurality of the non-volatile memory die;
individually testing each of the received plurality of the non-volatile memory die to determine corresponding defect data; and
programming the corresponding determined defect data into each of the non-volatile memory dies; and
subsequent to programming a set of weights of the trained neural network into the plurality of the non-volatile memory die and programming the corresponding determined defect data into each of the non-volatile memory dies, providing the plurality of the non-volatile memory die for incorporation into a memory system.
19 . The method of claim 18 , further comprising:
subsequent providing the plurality of the non-volatile memory die for incorporation into the memory system, for one or more of the non-volatile memory die, performing memory operations, including reading and writing of data, on the memory die;
determining error indication data while performing a plurality of the memory operations;
writing the determined error indication data to the memory die;
performing a compute-in-memory inference operation with the trained neural network using the written determined error indication data and the corresponding determined defect data as input values; and
determining defective portions of the memory die based on output of the inference operation.
20 . A non-volatile memory system, comprising:
a plurality of non-volatile memory dies each comprising one or more arrays of non-volatile memory cells, a first array of a first of the of non-volatile memory dies storing a set of weights of a neural network pre-trained for determining defective for portions of the memory dies; and
a non-volatile memory controller connected to the plurality of memory dies and configured to:
perform memory operations on the plurality of memory dies, including reading and writing of data on the one or more arrays of each of the memory dies;
collect error indication data for the plurality of memory dies while performing the memory operations;
write the determined error indication data to the first array;
perform a compute-in-memory inference operation with the neural network using the written determined error indication data as input values; and
determine defective portions of the plurality of the memory dies based on the inference operation.