Abstract: A storage device includes a controller that can dynamically adjust the zone active limit (ZAL) for a zoned namespace (ZNS). Rather than assuming a worst-case scenario for the ZNS, the ZAL can be dynamically adjusted, even after providing the ZAL to a host device. In so doing, device behavior changes due to factors such as temperature, failed or flipped bit count, and device cycling can be considered as impacting the ZAL. The ZAL can then be adjusted over time, and the new ZAL can be communicated to the host device. As such, rather than a fixed, worst-case ZAL, the host device will receive updated ZAL values over time as the device performs.

Abstract: Methods and apparatus for use with non-volatile memory (NVM) arrays having single-level cell (SLC) layers and multi-level cell (MLC) layers, such as triple-level cell (TLC) layers, provide for a coupled SLC/MLC write operation where SLC write protection is combined into a MLC write flow. In an illustrative example, data is written concurrently to SLC and TLC. The SLC data provides a backup for the TLC data in the event the TLC data is defective. The TLC data is verified using, for example, write verification. If the data is successfully verified, the SLC block can be erased or otherwise overwritten with new data. If not, the SLC block can be used to recover the data for storage in a different TLC block. The coupled SLC/MLC write operation may be performed in conjunction with a quick pass write (QPW).

Abstract: Methods and apparatus are disclosed for managing the storage of static and dynamic neural network data within a non-volatile memory (NVM) die for use with deep neural networks (DNN). Some aspects relate to separate trim sets for separately configuring a static data NVM array for static input data and a dynamic data NVM array for dynamic synaptic weight data. For example, the static data NVM array may be configured via one trim set for data retention, whereas the dynamic data NVM array may be configured via another trim set for write performance. The trim sets may specify different configurations for error correction coding, write verification, and read threshold calibration, as well as different read/write voltage thresholds. In some examples, neural network regularization is provided within a DNN by setting trim parameters to encourage bit flips to avoid overfitting. Some examples relate to managing non-DNN data, such as stochastic gradient data.

Abstract: Exemplary methods and apparatus are provided for implementing a deep learning accelerator (DLA) or other neural network components within the die of a non-volatile memory (NVM) apparatus using, for example, under-the-array circuit components within the die. Some aspects disclosed herein relate to configuring the under-the-array components to implement feedforward DLA operations. Other aspects relate to backpropagation operations. Still other aspects relate to using an NAND-based on-chip copy with update function to facilitate updating synaptic weights of a neural network stored on a die. Other aspects disclosed herein relate to configuring a solid state device (SSD) controller for use with the NVM. In some aspects, the SSD controller includes flash translation layer (FTL) tables configured specifically for use with neural network data stored in the NVM.

Abstract: Systems, apparatus and methods for generation of XOR signature metadata and XOR signature management are presented. In one or more embodiments, a storage device controller includes a host interface, configured to receive one or more string lines (SLs) of data from a host, the one or more SLs to be programmed into a non-volatile memory (NVM), and processing circuitry. The processing circuitry is configured to, for each of the one or more SLs, generate signature metadata and provide the signature metadata in a header of the SL. The processing circuitry is still further configured to XOR two or more of the SLs with their respective signature metadata to generate a snapshot, and write the snapshot to the NVM.

Abstract: Technology for physical defect detection in an integrated memory assembly having a control semiconductor die and a memory semiconductor die is disclosed. The control die compares actual current usage during a memory operation (such as a program operation) with expected current usage. In the event that the actual current usage deviates from the expected current usage by more than a threshold, a region of the memory structure is suspected as having a physical defect. For example, the selected word that is connected to the memory cells that were programmed may be suspected as having a physical defect. If a region is suspected as having a physical defect, a data integrity check may be performed in that region. If the data integrity check fails, the region may be marked as ineligible to store data.

Abstract: A storage system and method for reducing read-retry duration are provided. In one embodiment, a storage system is provided comprising a memory and a controller. The controller is configured to: receive a command from a host; and in response to an interruption in processing of the command: select a time for the host to retry the command, wherein the time is selected based on an expected host response time; and communicate the selected time to the host. Other embodiments are provided.

Abstract: A method and apparatus for cache management and eviction polices using unsupervised reinforcement learning schemes is disclosed.

Abstract: The present disclosure generally relates to relocating data in a storage device and updating a compressed logical to physical (L2P) table in response without invalidating cache entries of the L2P table. After relocating data from a first memory block associated with a first physical address to a second memory block associated with a second physical address, a version indicator of a cache entry corresponding to the first physical address in the L2P table is incremented. One or more cache entries are then added to the L2P table associating the relocated data to the second physical block without invaliding the cache entry corresponding to the first physical address. When a command to read or write the relocated data is received, the storage device searches the L2P table and reads the data from either the first memory block or the second memory block.

Abstract: The present disclosure generally relate to dynamically changing predictive latency related attributes to increase the deterministic window (DTWIN) of operation. The host device workload characteristics as well as the memory device's current condition provide valuable information for the duration of the DTWIN. If the memory device is near the end of life, then the DTWIN duration will be smaller. Additionally, if the workload from the host device is heavy, then the DTWIN duration will also be smaller. Rather than utilizing a fixed DTWIN duration based upon worst case scenarios for host device workload and memory device condition, dynamically adjusting the DTWIN duration based upon the workload and condition will provide a DTWIN duration that can gradually decrease over time from a much longer DTWIN duration than is currently available.

Abstract: A storage system and method for reducing read-retry duration are provided. In one embodiment, a storage system is provided comprising a memory and a controller. The controller is configured to: receive a command from a host; and in response to an interruption in processing of the command: select a time for the host to retry the command, wherein the time is selected based on an expected host response time; and communicate the selected time to the host. Other embodiments are provided.

Abstract: Systems, apparatus and methods for generation of XOR signature metadata and XOR signature management are presented. In one or more embodiments, a storage device controller includes a host interface, configured to receive one or more string lines (SLs) of data from a host, the one or more SLs to be programmed into a non-volatile memory (NVM), and processing circuitry. The processing circuitry is configured to, for each of the one or more SLs, generate signature metadata and provide the signature metadata in a header of the SL. The processing circuitry is still further configured to XOR two or more of the SLs with their respective signature metadata to generate a snapshot, and write the snapshot to the NVM.

Abstract: A non-volatile storage system, configured to use a protocol that supports predictable latency, including: a memory array storing a data in a block of memory; a controller coupled to the memory array, where the controller is configured to: in response to determining that predictable latency is enabled, operate the storage system using a first mode for a duration of time, where during the first mode, the storage system operates such that a read latency is below a read latency threshold; and after the duration of time, operate, the storage system using a second mode for a second duration of time, where during the second mode: the storage system performs a management operation based on a second set of thresholds that are different from a first set of threshold used during the first mode.

Abstract: An arrangement for providing a combined data and control signal for a multi die flash, comprising, a memory arrangement, the memory arrangement comprising at least two dies, a controller configured to send and receive signals to the memory arrangement and a common line connected to the memory arrangement and the controller and configured to transmit the signals from the controller to the at least two dies, wherein the arrangement is configured to provide a combined data and combined control signals to the multi-die flash.

Abstract: Calibrating read reference voltages is disclosed. In an aspect, a control die calibrates read reference voltages for reading the non-volatile memory cells. The control die is bonded to a memory die that contains memory cells. In one aspect, a tiered approach to calibrating read reference voltages is taken. For example, first the control die may attempt to determine new values for read reference voltages. If the new read reference voltages are satisfactory, then the control die may use the new read reference voltages. The control die could use one or more different techniques to determine new read reference voltages. If the new read reference voltages determined by the control die are unsatisfactory, then a memory controller in communication with the control die may calibrate the read reference voltages. By the control die determining the new read reference voltages, the memory controller is substantially less burdened with such tasks.

Abstract: A method and apparatus for cache management and eviction polices using unsupervised reinforcement learning schemes is disclosed.

Abstract: A memory system configured to decode a data set may pause a convergence process to update reliability metric values. The memory system may utilize a positive feedback system that updates the reliability metric values by analyzing current a posteriori reliability metric values to calculate average estimated reliability characteristic values associated with a memory error model. The updates to the reliability metric values may provide increased error correction capability and faster decoding.

Abstract: Systems, apparatus and methods for generation of XOR signature metadata and XOR signature management are presented. In one or more embodiments, a storage device controller includes a host interface, configured to receive one or more string lines (SLs) of data from a host, the one or more SLs to be programmed into a non-volatile memory (NVM), and processing circuitry. The processing circuitry is configured to, for each of the one or more SLs, generate signature metadata and provide the signature metadata in a header of the SL. The processing circuitry is still further configured to XOR two or more of the SLs with their respective signature metadata to generate a snapshot, and write the snapshot to the NVM.

Abstract: A non-volatile storage system, configured to use a protocol that supports predictable latency, including: a memory array storing a data in a block of memory; a controller coupled to the memory array, where the controller is configured to: in response to determining that predictable latency is enabled, operate the storage system using a first mode for a duration of time, where during the first mode, the storage system operates such that a read latency is below a read latency threshold; and after the duration of time, operate, the storage system using a second mode for a second duration of time, where during the second mode: the storage system performs a management operation based on a second set of thresholds that are different from a first set of threshold used during the first mode.

Abstract: Methods and apparatus are disclosed for implementing data augmentation within a storage controller of a data storage device based on machine learning data read from a non-volatile memory (NVM) array of a memory die. Some particular aspects relate to configuring the storage controller to generate augmented versions of training images for use in training a Deep Learning Accelerator of an image recognition system by rotating, translating, skewing, cropping, etc., a set of initial training images obtained from a host device and stored in the NVM array. Other aspects relate to controlling components of the memory die to generate noise-augmented images by, for example, storing and then reading training images from worn regions of the NVM array to inject noise into the images. Data augmentation based on data read from multiple memory dies is also described, such as image data spread across multiple NVM arrays or multiple memory dies.