Abstract: A Data Storage Device (DSD) receives weights for a plurality of layers of a neural network with layer information associating the weights with one or more layers. The received weights are stored in at least one Non-Volatile Memory (NVM) of the DSD using different storage characteristics based at least in part on the received layer information. The different storage characteristics include at least one of different storage locations, different storage techniques, and different storage maintenance settings. In another aspect, a first group of weights is requested by a host device for processing one or more first layers. The first group of weights is received by the host device and loaded into at least one memory of the host device. A second group of weights is requested from the DSD for processing one or more additional layers of the neural network before computations complete for the one or more first layers.

Abstract: An apparatus is provided that includes a control circuit coupled to a plurality of non-volatile memory cells disposed in a plurality of planes. The control circuit is configured to concurrently erase a block of memory cells in each of the plurality of planes, determine that the concurrent erase failed, disconnect a first one of the planes from the plurality of planes to form first remaining planes, and concurrently erase a block of memory cells in each of the first remaining planes.

Abstract: Devices and techniques are disclosed wherein a data storage device (DSD) generates ranking information corresponding to user data stored at a non-volatile memory of the DSD. The ranking information can be used by the DSD to form a frequently used files list, which can be read by a host system upon initialization with the host system and displayed to a user at the host system.

Abstract: A data storage device is enabled to independently self-format, without requiring a connected host device during the active formatting process. The storage device includes a data interface configured to receive power from the host device or a wall charger, non-volatile storage media, and control circuitry. The control circuitry is configured to receive first power from the host device, receive instructions from the host device to perform a format operation, save the instructions to perform the format operation, and cease receiving the first power from the host device. The control circuitry is further configured to receive second power from the wall charger and, in response to retrieving the saved instructions, initiate the format operation on the non-volatile storage media.

Abstract: Various embodiments of this disclosure are directed to a mixed digital and analog domain approach to computational storage or memory applications. The mixed approach enables certain compute operations to be advantageously performed in the analog domain, achieving power saving. In some embodiments, an analog compute core is implemented based on a first set of memory elements that are made available with a second set of memory elements for digital data storage. A controller coupled to both sets of memory elements is able to selectively direct computational tasks to either the analog compute core or a digital processor coupled with the controller, based on one or more parameters including power, precision, and workload. In certain embodiments involving neural network tasks, the controller is configured to route certain tasks to the analog compute core based on neural network based factors such as network layer positioning and input signal type.

Abstract: Instead of using programmable block size aggregation, a lower multiple of page, and down to a page size aggregation is used. A bad block prediction unit in a controller is able to predict when a programmable block has a bad page. The bad block prediction unit can lower the aggregation size of a programmable block by monitoring the life cycle of the programmable block through bad block statistic collection. When the accumulation size passes a threshold, the bad block prediction unit lowers the aggregation size. The bad block prediction unit can also predict when to lower aggregation size based on the number of reconstructions. An aggregate size level is set at a page boundary, and once the number of reconstructions reaches that page boundary, the bad block prediction unit lowers the aggregation size to page aggregation. The bad block prediction unit is able to predict both life cycle threshold changes and reconstructions changes.

Abstract: A data storage device and method for race-based data access in a multiple host memory buffer system are provided. In one embodiment, the data storage device stores data in a plurality of host memory buffers in the host instead of in just the host memory buffer usually associated with the data. To read the data, the data storage device sends read commands to all of the host memory buffers. That way, even if some of the host memory buffers are busy, the data can be returned from another one of the host memory buffers. In future reads in similar workloads, a read command can be sent to the host memory buffer that returned the data. Other embodiments are possible, and each of the embodiments can be used alone or together in combination.

Abstract: The present disclosure generally relates to more effective utilization of write and read bandwidth in submission queues (SQs). The data storage device treats a SQ as two separate SQs: one write SQ and one read SQ. Rather than a single fetch pointer for the entire SQ, the write SQ has a write fetch pointer (WFP) while the read SQ has a separate read fetch pointer (RFP). So long as the individual pointers are less than a queue pointer (QP), the data storage device can still process commands for either read or write SQ even if the other SQ has run out of credits. In so doing, read and write bandwidths can be effectively utilized.

Abstract: A data storage device receives a speculative read command from a host identifying logical block addresses. The speculative read command is not required be to executed within a certain amount of time or even at all. The data storage device at least partially executes the speculative read command in response to determining that such execution will not reduce performance of the data storage device. At least partially executing the speculative read command causes data associated with at least some of the logical block addresses to be read from the non-volatile memory and stored in at least one buffer. Other embodiments are possible, and each of the embodiments can be used alone or together in combination.

Abstract: Instead of a system with no awareness to the specific properties of the described system files, such as atomicity of different types of system files, utilize the special characteristics of the corresponding system files to optimize storage handling. A host marks a certain logical block address (LBA) range as belonging to an atomic file. That entire range will be treated as a single atomic unit. Conversely, an LBA range being used to append to a log file may have very small atomic units, allowing for incremental updates without changing the atomicity of the rest of the media. When a write command is passed, the write command will have a certain length. Depending on the length of the write command, the device can disassemble the write command into smaller write sectors of the smallest possible write portion. The device will then write the small write portions to a storage location, while keeping an atomic principle of each of the small write portions.

Abstract: The present disclosure generally relates to utilizing a transparent host memory buffer (HMB) where the host device is granted access to the HMB to obtain data from a mapping table. The data storage device stores the mapping table in HMB and then allows the host device to view the mapping table and retrieve information. The host device sends a command to the data storage device that includes not only a read command, but also mapping table info specific to the read command. Additionally, an indication of the mapping table version from where the information is also provided. The data storage device, upon receiving the command, confirms the version of the information is the most recent version and then, if confirmed, utilizes the mapping information provided with the command. In so doing, accessing the HMB after receiving the command will not be necessary.

Abstract: A data storage device includes a memory device and a controller coupled to the memory device. The controller is configured to receive a shutdown notification, fetch one or more command identifiers from a submission queue of a host device, generate error indications for the one or more command identifiers, and send a completion message, including the generated error indication, for each of the one or more command identifiers to the host device. The controller is further configured to push non-processed pending commands to a completion finite state machine, where the controller generates an error indication for each of the non-processed pending commands and sends a completion message, including the generated error indication, for each of the non-processed pending commands to the host device. While the controller is fetching command identifiers and pushing non-process commands, the controller is configured to continue processing processed commands in parallel.

Abstract: A data storage device includes a memory device, a random access memory (RAM) device, and a controller coupled to the memory device and the RAM device. The controller is configured to determine a workload type of the data storage device, determine to store at least a portion of a compressed logical to physical translation table (ZCAT) in the RAM device based on the workload type, and utilize a remaining portion of the RAM device to perform background operations. The controller is further configured to determine a persistence or temperature of the ZCAT, where the portion of the ZCAT stored to the RAM device is based on the persistence and/or the temperature of the ZCAT. The remaining portion of the ZCAT is stored in the HMB, where the remaining portion may update any out-of-date ZCAT entries.

Abstract: A semiconductor storage device such as an SSD includes a pliable printed circuit board (PCB) having semiconductor memory devices mounted by solder balls on first and second opposed major surfaces. The memory devices are mounted so as to be staggered and/or partially overlapping with each other on the first and second surfaces of the PCB in at least one direction. The staggered arrangement allows the PCB to flex upon warping of the memory devices mounted on the PCB.

Abstract: A three-dimensional memory device includes an alternating stack of insulating layers and electrically conductive layers located over a substrate, memory opening fill structures including a respective vertical semiconductor channel and a respective memory film, and support pillar structures including a respective dummy vertical semiconductor channel, a respective dummy memory film, and a vertical stack of dielectric spacer fins located at levels of the electrically conductive layers and interposed between the electrically conductive layers and the respective dummy memory film.

Abstract: A sacrificial memory opening fill structure for a multi-tier memory device may include a semiconductor fill material portion a metallic fill material portion to enhance control of a vertical cross-sectional profile of an inter-tier memory opening. Multiple inter-tier dielectric layers may be employed to reduce sharp corners in a memory opening fill structure. Alternatively or additionally, a combination of an isotropic etch process followed by an anisotropic etch process may be used to form a first-tier memory opening.

Abstract: In order to ensure that a bandwidth allocated to each tenant of a plurality of tenants of a data storage device is maintained, a controller of the data storage device may split a large read command, received from a host device, into a plurality of chunks, where each chunk corresponds to a distinct portion of the split large read command. Because the allocated bandwidth for a tenant is static, one or more chunks of the plurality of chunks, up to the allocated bandwidth, are executed, such that the bandwidth required to perform the one or more chunks does not exceed the allocated bandwidth for the particular tenant. Split information is added to the plurality of chunks in order to maintain coherency when executing the one or more chunks. Therefore, the agreed-upon allocated bandwidth for each tenant is maintained while performing large read commands requiring more bandwidth than allocated.

Abstract: The present disclosure generally relates to achieving an acceptable uncorrectable bit error rate (UBER) using a dual temporary data protecting approach and a small SLC cache by adding a temporary XOR protection to zone-groups rather than storing another copy of the zone within the drive. The parity data can be stored with the user data (e.g., as part of the zone-group, effectively increasing zone-group size by 1) or in a separate location, e.g., in an SLC block or another separate MLC block.

Abstract: In order to guarantee data validity of data read from a memory device of the data storage device to a host device, a controller of the data storage device may calculate a cyclic redundancy code (CRC) signature of the decoded data and compare the CRC signature of the decoded data with a CRC signature of the data. The CRC signature of the data is generated during a write operation of the data to the memory device. Rather than returning an uncorrectable error correction code error (UECC) error to the host device when the CRC signature of the decoded data does not match the CRC signature of the data, the controller executes the read command again. By using a different buffer to store the decoded data, the controller may confirm whether the error stemmed from the read path or the error was not from the read path.

Abstract: A data storage device can store data and parity information for the data in its memory. In some storage methodologies, data and parity information are striped across a plurality of memory dies (e.g., in a redundant array of independent drives (RAID) configuration). That way, if one of the memory dies fails, the data or the parity information can be reconstructed from the other memory dies. These embodiments recognize that because parity information is used relatively infrequently, the parity information can be stored in locations in the memory that have a relatively-worse performance than other areas of the memory. This can increase performance of the memory in situations where the parity information does not need to be read.