# US 2023/0418738 A1 · Memory device with latch-based neural network weight parity detection and trimming

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0418738 A1
- Document ID: US20230418738A1
- Status: Published application
- Kind: A1
- Application number: 17/847,089
- Filed: 2022-06-22
- Published: 2023-12-28
- Earliest claimed: 2022-06-22
- Assignees: Western Digital Technologies, Inc.
- Applicants: Western Digital Technologies, Inc.
- Inventors: Daniel Joseph Linnen; Ramanathan Muthiah; Kirubakaran Periyannan
- Web page: https://patentgrep.com/patent/US20230418738A1

## CPC

- G06F 12/0246: {in block erasable memory, e.g. flash memory}
- G06F 3/0616: {in relation to life time, e.g. increasing Mean Time Between Failures [MTBF]}
- G06F 3/0658: {Controller construction arrangements}
- G06F 3/0688: {Non-volatile semiconductor memory arrays}
- G06F 11/1004: {to protect a block of data words, e.g. CRC or checksum (G06F11/1076 takes precedence; security arrangements for protecting computers or computer systems against unauthorized activity G06F21/00)}
- G06F 12/0207: {with multidimensional access, e.g. row/column, matrix}
- G06N 3/02: Neural networks
- G06N 3/063: using electronic means
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/10: Interfaces, programming languages or software development kits, e.g. for simulating neural networks
- G06F 3/0619: {in relation to data integrity, e.g. data losses, bit errors}
- G06F 3/065: {Replication mechanisms}
- G06F 3/0653: {Monitoring storage devices or systems}
- G06F 3/0679: {Non-volatile semiconductor memory device, e.g. flash memory, one time programmable memory [OTP]}
- G06F 12/0868: Data transfer between cache memory and other subsystems, e.g. storage devices or host systems
- G06F 2212/1008: Correctness of operation, e.g. memory ordering
- G06F 2212/1024: Latency reduction
- G06F 2212/1036: Life time enhancement
- G06F 2212/214: Solid state disk
- G06F 2212/313: In storage device
- G06F 2212/452: Instruction code
- G06F 2212/7208: Multiple device management, e.g. distributing data over multiple flash devices
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Latch-based methods and apparatus for performing neural network weight parity detection on the die of a non-volatile memory (NVM) array to detect bit flip errors within neural network weight data are described, particularly for use with floating point number values. Upon detection of a parity error in a neural network weight, the erroneous weight is set to zero to trim the corresponding neuron from the network, thus preventing the erroneous value from significantly affecting the network, particularly in situations where the bit flip would otherwise affect the magnitude of a floating-point weight value. The exemplary latch-based procedures described herein are linear procedures that do not require logic decisions. Procedures are also described that assess an amount of degradation in the NVM array based on parity bit data collected in the latches. Multiple plane and multiple die NVM array implementations are also described for massive parallel processing.

## Claims

### Claim 1

A device, comprising:
a non-volatile memory (NVM) array formed on a die;
a plurality of latches formed on the die; and
processing circuitry formed on the die and configured to:
read a neural network weight from the NVM array into at least one latch of the plurality of latches, and
perform a linear sequence of latch operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise, wherein the linear sequence of latch operations corresponds to a single processing flow without decision logic and without separate processing branches.

### Claim 2 (depends on 1)

The device of claim 1, wherein the processing circuitry is further configured to re-store the neural network weight in the NVM array.

### Claim 3 (depends on 1)

The device of claim 1, wherein the processing circuitry is further configured to perform a machine learning operation using the neural network weight as a floating-point number.

### Claim 4 (depends on 1)

The device of claim 1, wherein the neural network weight is represented as a parity-encoded binary sequence.

### Claim 5 (depends on 4)

The device of claim 4, wherein the processing circuitry is configured to perform the linear sequence of operations by being further configured to:
perform a first series of linear operations using one or more latches of the plurality of latches to reduce the neural network weight down to a single bit within one of the latches, where the single bit is representative of the parity of the neural network weight;
perform a second series of linear operations to propagate the parity bit into the one or more of the other latches to yield a binary sequence within one of the latches, which, when combined with the neural network weight, results in (a) the neural network weight if parity was not violated and results in (b) binary zeroes if parity was violated; and
combine the binary sequence with the neural network weight.

### Claim 6 (depends on 1)

The device of claim 1, wherein the processing circuitry is further configured to determine an amount of degradation in the NVM array.

### Claim 7 (depends on 6)

The device of claim 6, wherein the processing circuitry is further configured to perform at least one operation to mitigate the degradation.

### Claim 8 (depends on 7)

The device of claim 7, wherein the processing circuitry is configured to mitigate the degradation by performing one of more of: (a) re-generating the plurality of neural network weights; (b) re-storing the plurality of neural network weights in the NVM array; and (c) performing one or more NVM maintenance operations.

### Claim 9 (depends on 1)

The device of claim 1, wherein the NVM array comprises a plurality of planes, and wherein the processing circuitry is further configured to concurrently perform the linear sequence of operations for a plurality of neural network weights in parallel, wherein the plurality of neural network weights are obtained from the plurality of planes.

### Claim 10 (depends on 1)

A multiple die apparatus comprising a plurality of the devices of claim 1, wherein the multiple die apparatus is configured to concurrently perform the linear sequence of operations for a plurality of neural network weights in parallel on a plurality of dies.

### Claim 11

A method for use with a device comprising a non-volatile memory (NVM) array comprising a plurality of planes formed on a die, the method comprising:
reading a neural network weight from the NVM array into at least one latch of a plurality of latches on the die; and
performing a linear sequence of latch operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise, wherein the linear sequence of latch operations corresponds to a single processing flow without decision logic and without separate processing branches.

### Claim 12 (depends on 11)

The method of claim 11, further comprising re-storing the neural network weight in the NVM array.

### Claim 13 (depends on 11)

The method of claim 11, wherein the neural network weight is a parity-encoded binary sequence.

### Claim 14 (depends on 13)

The method of claim 13, wherein performing the linear sequence of operations comprises:
performing a first series of linear operations using one or more latches of the plurality of latches to reduce the neural network weight down to a single bit within one of the latches, where the single bit is representative of the parity of the neural network weight;
performing a second series of linear operations to propagate the parity bit into the one or more of the other latches so as to yield a binary sequence within one of the latches, which, when combined with the neural network weight, results in (a) the neural network weight if parity was not violated and results in (b) binary zeroes if parity was violated; and
combining the binary sequence with the neural network weight.

### Claim 15 (depends on 11)

The method of claim 11, further comprising determining an amount of degradation in the NVM array.

### Claim 16 (depends on 15)

The method of claim 15, further comprising performing at least one operation to mitigate the degradation.

### Claim 17 (depends on 16)

The method of claim 16, further comprising mitigating the degradation by performing one or more of: (a) re-generating the plurality of neural network weights; (b) re-storing the plurality of neural network weights in the NVM array; and (c) performing one or more NVM maintenance operations.

### Claim 18

A device, comprising:
a memory formed on a die;
a plurality of latches formed on the die; and
processing circuitry formed on the die and configured to:
read parity-encoded data from the memory into at least one latch of the plurality of latches, and
perform a linear sequence of latch operations using the plurality of latches to set the data to zero if there is a parity error in the data while leaving the data unchanged otherwise, wherein the linear sequence of latch operations corresponds to a single processing flow without decision logic and without separate processing branches.

### Claim 19 (depends on 18)

The device of claim 18, wherein the processing circuitry is configured to perform the linear sequence of operations by being further configured to:
perform a first series of linear operations using one or more latches of the plurality of latches to reduce the data down to a single bit within one of the latches, where the single bit is representative of the parity of the data;
perform a second series of linear operations to propagate the parity bit into the one or more of the other latches so as to yield a binary sequence within one of the latches, which, when combined with the data, results in (a) the data if parity was not violated and results in (b) binary zeroes if parity was violated; and
combine the binary sequence with the data.

### Claim 20 (depends on 18)

The device of claim 18, wherein the processing circuitry is further configured to determine an amount of degradation in the memory.

### Claim 21

An apparatus for use with a device comprising a non-volatile memory (NVM) array formed on a die, the apparatus comprising:
means for reading a neural network weight from the NVM array into at least one latch of a plurality of latches on the die; and
means for performing a linear sequence of latch operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise, wherein the linear sequence of latch operations corresponds to a single processing flow without decision logic and without separate processing branches.

### Claim 22 (depends on 1)

The device of claim 1, wherein the processing circuitry is further configured to perform the latch operations by being further configured to perform latch operations comprising one or more of: a LOAD operation, an OR operation, a XOR operation, an AND operation, a LEFT SHIFT operation, a RIGHT SHIFT operation, and a COPY operation.

### Claim 23 (depends on 11)

The method of claim 11, wherein the performing the latch operations comprises performing one or more of: a LOAD operation, an OR operation, a XOR operation, an AND operation, a LEFT SHIFT operation, a RIGHT SHIFT operation, and a COPY operation.

### Claim 24 (depends on 18)

The device of claim 18, wherein the processing circuitry is further configured to perform the latch operations by being further configured to perform latch operations comprising one or more of: a LOAD operation, an OR operation, a XOR operation, an AND operation, a LEFT SHIFT operation, a RIGHT SHIFT operation, and a COPY operation.

### Claim 25

A device, comprising:
a non-volatile memory (NVM) array formed on a die, wherein the NVM array comprises a plurality of planes;
a plurality of latches formed on the die; and
processing circuitry formed on the die and configured to:
read a neural network weight from the NVM array into at least one latch of the plurality of latches, and
perform a linear sequence of operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise; and
wherein the processing circuitry is further configured to concurrently perform the linear sequence of operations for a plurality of the neural network weights in parallel, wherein the plurality of neural network weights are obtained from the plurality of planes.

### Claim 26

A method for use with a device comprising a non-volatile memory (NVM) array including a plurality of planes formed on a die, the method comprising:
reading a neural network weight from the NVM array into at least one latch of a plurality of latches on the die; and
performing a linear sequence of operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise; and
wherein performing the linear sequence of operations comprises concurrently performing the linear sequence of operations for a plurality of the neural network weights in parallel, wherein the plurality of neural network weights are obtained from the plurality of planes.

### Claim 27

A multiple die apparatus comprising a plurality of devices, each device comprising:
a non-volatile memory (NVM) array formed on a die;
a plurality of latches formed on the die; and
processing circuitry formed on the die and configured to:
read a neural network weight from the NVM array into at least one latch of the plurality of latches, and
perform a linear sequence of operations using the plurality of latches to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise; and
wherein the multiple die apparatus is configured to concurrently perform the linear sequence of operations for a plurality of the neural network weights in parallel on a plurality of the dies.

### Claim 28

A method for use with a multiple die apparatus comprising a plurality of devices, each including a non-volatile memory (NVM) array formed on a die, the method comprising:
for each die of the plurality of devices,
reading a neural network weight from the NVM array of the die into at least one latch of a plurality of latches on the die; and
performing a linear sequence of operations using the plurality of latches of the die to set the neural network weight to zero if there is a parity error in the neural network weight while leaving the neural network weight unchanged otherwise; and
wherein the multiple die apparatus concurrently performs the linear sequence of operations for a plurality of the neural network weights in parallel on the plurality of dies.
