Abstract: Techniques related to a QoS-aware decoder architecture for data storage are described. In an example, QoS specifications include a QOS latency specification indicative of an acceptable latency for completing the processing of a data read command. The decoder may store this QOS latency specification. In operation, the decoder generates a latency measurement indicative of the actual latency for the processing. If a comparison of the latency measurement and QOS latency specification indicates a violation of the QOS latency specification, the decoder can terminate the decoding and generate a decoding failure.

Abstract: Techniques related to improving a performance related to at least data reads from a memory are described. In an example, a computer system hosts a regression model that includes a neural network. The neural network is trained based on training data that is measured under different combinations of operational conditions and storage conditions. In operation, actual operational and storage conditions associated with the memory are input to the regression model. The neural network outputs a voltage read threshold based on these actual conditions. The computer system uses the voltage read threshold to read data stored in the memory.

Abstract: Techniques related to improving a performance related to at least data reads from a memory are described. In an example, a computer system hosts a regression model that includes a neural network. The neural network is trained based on training data that is measured under different combinations of operational conditions and storage conditions. In operation, actual operational and storage conditions associated with the memory are input to the regression model. The neural network outputs a voltage read threshold based on these actual conditions. The computer system uses the voltage read threshold to read data stored in the memory.

Abstract: Adaptive read-threshold schemes for a memory system determine read-threshold with the lowest BER/UECC failure-rates while continuing to serve the host-reads with the required QoS. When it is determined that the QoS or other quality metric is not met for a particular read-threshold, which may be an initial, default, read-threshold, the performance of other read-thresholds are estimated. These estimates may then be used to determine an optimal read-threshold. During the iterative process, selection variables, e.g., how many times, and for which read commands, to use each of the non-default read-thresholds in future read-attempts may be determined on-the-fly.

Abstract: Techniques are described for memory writes and reads according to a chip-kill scheme that allows recovery of multiple failed wordlines. In an example, when reading data from a superblock of the memory, where the decoding of multiple wordlines failed, a computer system schedules the decoding of failed wordlines based on quantity of bit errors and updates soft information based on convergence or divergence of the scheduled decoding. Such a computer system significantly reduces decoding failures associated with data reads from the memory and allows improved data retention in the memory.

Abstract: A system includes memory cells arranged in blocks and a memory controller. The memory controller receives a read command to read a first block. The first block can be associated with a first read count and a first read threshold. The first read count is incremented when the first block is read, and when the first read count reaches the read threshold, a read reclaim test is performed. The first read count is set to zero after a power off or a read reclaim operation. When the first read count is zero, an adaptive read threshold is selected based on the number of bit errors. Further, in a read reclaim test, the number of bit errors is tested against an adaptive error threshold to determine whether a garbage collection operation is performed.

Abstract: Techniques are described for performing a bit-flipping decoding scheme on a G-LDPC codeword. In an example, a decoding system uses two syndrome tables. The first syndrome table identifies a predefined syndrome for a component codeword that protects a bit of the G-LDPC codeword. This predefined syndrome is identified based on a location of the bit and is used to update a current syndrome of the component codeword. The second syndrome table identifies one or more bit error locations for the component codeword. The bit error locations are identified from the second syndrome table based on the current syndrome of the component codeword, as updated. In an example, the error locations are used to update a reliability of the bit if its location corresponds to one of the error locations. A bit flipping decision is made for the bit based on its reliability.

Abstract: Disclosed are devices, systems and methods for improving performance of a block of a memory device. In an example, performance is improved by implementing soft chipkill recovery to mitigate bitline failures in data storage devices. An exemplary method includes encoding each horizontal row of cells of a plurality of memory cells of a memory block to generate each of a plurality of codewords, and generating a plurality of parity symbols, each of the plurality of parity symbols based on diagonally positioned symbols spanning the plurality of codewords.

Abstract: Disclosed is a computer-implemented method for optimizing read thresholds of a memory device using a deep neural network engine, comprising reading, using a set of read threshold voltages applied to the memory device, data from the memory device under a first set of operating conditions that contribute to read errors in the memory device, producing a labeled training data set using the set of read threshold voltages under the first set of the operating conditions, determining, based on characteristics of the memory device, a number of layers, a size of each layer, and a number of input and output nodes of the deep neural network engine, training the deep neural network engine using the labeled training data set, and using the trained deep neural network engine to compute read thresholds voltage values under a second set of operating conditions.

Abstract: A memory system, a controller including a bit-flipping (BF) decoder and a min-sum (MS) decoder that may be included in the memory system and operating methods thereof in which the controller determines a quality metric as a function of initial syndrome weight and information of the BF decoder after a set number of decoding iterations by the BF decoder in a test period. After the test period, the controller applies the quality metric to each codeword to determine whether to send that codeword first to the BF decoder for decoding or directly to the MS decoder for decoding.

Abstract: Techniques are described for memory writes and reads according to a chip-kill scheme that allows recovery of multiple failed wordlines. In an example, when reading data from a superblock of the memory, where the decoding of multiple wordlines failed, a computer system schedules the decoding of failed wordlines based on quantity of bit errors and updates soft information based on convergence or divergence of the scheduled decoding. Such a computer system significantly reduces decoding failures associated with data reads from the memory and allows improved data retention in the memory.

Abstract: A system includes a plurality of memory cells. Each memory cell is programmed to a data state corresponding to one of multiple cell programmed voltages. The memory cells are read to determine a programmed data state of each memory cell. Error correction decoding is performed to determine a corrected data state of each memory cell. The corresponding cell levels, or programmed voltages, are determined based on the programmed data state and the corrected data state. A first error count represents a total number of error cells that have a higher cell level for the corrected data state than the programmed data state. A second error count represents a total number of error cells that have a lower cell level for the corrected data state than the programmed data state. The system is configured to perform a memory operation based on the first error count and the second error count.

Abstract: A memory system, a bit-flipping (BF) low-density parity check (LDPC) decoder included in the memory system and operating methods thereof in which such decoder or decoding has a reduced error floor. Such a BF LDPC decoder is configured using a deep learning framework of trained and training neural networks and data separation that exploits the degree distribution information of the constructed LDPC codes.

Abstract: A memory system includes a memory device and a controller. The controller determines a target word line group to which a target word line corresponding to a read command belongs. The controller identifies a reference voltage corresponding to the target word line group. The controller controls the memory device to perform a read operation on a target page coupled to the target word line, using the reference voltage.

Abstract: Techniques are described for performing a check node update (CNU) as part of iterative decoding of a low density-parity check (LDPC) code. The CNU uses a min-sum decoding approach that monitors whether two values received in messages from two variable nodes connected to a check nodes are equal and are the minimum value among the values received by the check nodes from other variable nodes connected thereto. Upon detecting such an event, the minimum value is adjusted by reducing it by an adjustment value to generate an adjusted minimum value. This adjusted minimum value approximates the minimum value that a sum-product algorithm (SPA) decoding approach would have generated. The adjusted minimum value is included in a response message sent from the check node to a variable node. The bit corresponding to that variable node is decoded based on this adjusted minimum value.

Abstract: A method is provided for controlling a storage system, which can include a plurality of memory cells arranged in blocks and a memory controller coupled to the plurality of memory cells for controlling data write and read in the plurality of memory cells. The method includes identifying a block as a good block, if a count of bad pages in the block is zero, identifying the block as a degraded block if the count of bad pages is below a threshold number, and identifying the block as a bad block if the count of bad pages is above or equal to the threshold number. The method includes using good blocks and degraded blocks for read and program operations, and not using the bad blocks.

Abstract: A dynamic interleaver performs a read operation to identify bit lines with high failures, and form groups of data bits for parity bits computation, such that each group includes at most one data bit from the bit lines with high failures. Thus, the interleave selects the bit lines with high failures based on a most recent read test, and can be adjusted according to the conditions of the storage device.

Abstract: An apparatus of a semiconductor memory system and an operating method thereof include: a plurality of memory devices; and a controller coupled with the memory devices, the controller including a training data storage, a classifier trainer, and a decoder, is configured to perform decoding iterations, wherein the training data storage configured to collect and store at least training data, the classifier trainer configured to train classifiers at least with the training data, and the decoder configured to decode code-bits in accordance with rules of the classifier.

Abstract: A memory system includes a memory device and a controller. The controller performs a test read operation on a read data set of the memory device, using multiple read threshold entries and determines which are good read threshold entries based on results of the read operation. The controller selects a best read threshold entry among the multiple read threshold entries based on a result of the test read operation, partitions the read data set into a good data set decodable by the best read threshold entry and a bad data set undecodable by the best read threshold entry, and sets the bad data set as a new read data set.

Abstract: Adaptive read-threshold schemes for a memory system determine read-threshold with the lowest BER/UECC failure-rates while continuing to serve the host-reads with the required QoS. When it is determined that the QoS or other quality metric is not met for a particular read-threshold, which may be an initial, default, read-threshold, the performance of other read-thresholds are estimated. These estimates may then be used to determine an optimal read-threshold. During the iterative process, selection variables, e.g., how many times, and for which read commands, to use each of the non-default read-thresholds in future read-attempts may be determined on-the-fly.