Abstract: The present disclosure generally relates to reducing boot latency of memory devices in a dual boot system. The boot code is loaded to the data storage device controller in a flexible manner by being able to receive chunks of the boot code from two separate locations, the host memory buffer (HMB) and the memory device, which may be a NAND device. Part of the boot code may be received from the HMB and another part of the boot code may be received from the memory device. If either the HMB or the memory device can deliver the chunks faster than the other, then the controller can receive the chunks from the faster location and periodically confirm the speed of delivery to ensure the boot code latency is optimized.

Abstract: A data storage device includes a memory and a controller coupled to the memory device. The controller is configured to be coupled to a host device. The controller is further configured to receive a plurality of commands, generate logical block address (LBA) to physical block address (PBA) (L2P) mappings for each of the plurality of commands, and store data of the plurality of commands to a respective PBA according to the generated L2P mappings. Each of the L2P mappings are generated based on a result of a deep learning (DL) training model using a neural network (NN) structure. The controller includes a NN command interpretation unit and a L2P mapping generator coupled to the NN command interpretation unit. The controller is configured to fetch training data and NN parameters from the memory device.

Abstract: The present disclosure generally relates to utilizing improved DL training models stored in non-volatile memory to optimize data transfer and storage. The proposed system would identify workloads of DNN training and occasionally check the difference rate between successive data transfers (representing successive training iterations of the model). Comparing the difference rate to given thresholds could indicate “recommendation-system” typical use case. In such a case the NAND operating system would apply systematic compression of the data by saving only the changed parameters between successive iteration cycles (“batches”). The host may indicate the checkpoint storage configuration of the training model (every iteration, every several iterations etc. . . . ) and other elements. The system may be efficiently utilized combining the NAND based DNN training interface, adding the checkpoint configuration information to the dedicated interface.

Abstract: A method and system for maintaining coherency between DMA and NVMe data paths are disclosed. As DMA requests are received in the PMR region, a device controller will translate these into NVMe commands with a dedicated queue that is hidden from a host that has higher priority than the corresponding host (NVMe) commands. The payload returned from an internally executed NVMe command is stored in a buffer used to complete the DMA request. As memory reads are submitted, the controller will mark corresponding LBA ranges for overlap, ensuring coherency between these reads and writes from other queues. Since the internal PMR queue has a higher priority than host-facing queues (e.g., NVMe), and the PMR is read-only, read coherency against host writes to the same region may be achieved.

Abstract: A data storage device and method for providing an adaptive data path are disclosed. In one embodiment, a data storage device is in communication with a host comprising a first processor (e.g., a graphics processing unit (GPU)), a second processor (e.g., a central processing unit (CPU)), and a queue. The data storage device chooses a data path to use to communicate with the queue based on whether the queue is associated with the first processor or with the second processor. Other embodiments are possible, and each of the embodiments can be used alone or together in combination.

Abstract: The present disclosure generally relates to improving data transfer in a data storage device. In double data rate (DDR) systems that include a data bus inversion (DBI) functionality, bit flip events can be more prevalent. To mitigate the effect of enhanced erroneous bit flip rate related to DBI bit flip events, the DBI bit can stay static for a predetermined number of consecutive clock cycles, the error correction module can be informed of reduced reliability due to active DBI bit events, the DBI bit can be set to 0, or combinations thereof. Setting the DBI bit to 0 effectively cancels DBI functionality. Informing the error correction module permits a more robust error correction to occur. Forcing the DBI bit to remain static reduces the probability of an unrecognized bit flip event of a full byte. In so doing, data transfer reliability is improved when using DBI functionality.

Abstract: Apparatuses and methods of directly accessing a memory space of a storage device by a host are provided. In one embodiment, a method of driverless access of a non-volatile memory of a non-volatile memory device by a host includes initializing a PCIe memory space mapping a portion of the non-volatile memory of the non-volatile memory device to a host memory space. The non-volatile memory is mapped through a PCIe link between the host and the non-volatile memory device. Load/store commands are sent to the PCIe memory space for driverless access. The method further includes negotiating an alignment size of the minimum transaction packet size to complete the load/store commands.

Abstract: A method and apparatus for video processing on a data storage device. A chip bound architecture includes a CMOS coupled to one or more NAND die, the CMOS including one or more processors, memories, and error correction code (ECC) engines capable of processing video data. According to certain embodiments, macroblocks are correlated between two I-frames, including motion vectors to define different locations of correlated macroblocks. A P-frame may be determined from a previous I-frame and its correlated macroblocks and motion vectors, while a B-frame may be determined from two or more adjacent I-frames with concomitant macroblocks and motion vectors, as well as P-frames associated with an adjacent I-frame.

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

Abstract: An apparatus, system, and method for detecting compromised firmware in a non-volatile storage device. A control bus of a non-volatile storage device is monitored. The non-volatile storage device includes a processor and electronic components coupled to the control bus. Signal traffic on the control bus is analyzed for events and/or triggers related to storage operations initiated on the control bus by the processor. Storage operations include one or more commands directed to at least one of the electronic components. If the latency for the storage operation satisfies an alert threshold a host is notified of compromised firmware.

Abstract: A non-volatile storage system that is implementing a storage region (e.g., a persistent memory region) which is accessible to a host (e.g., via a PCIe connection) and a cache for the storage region shares details of the structure of the storage region and/or the cache (e.g., cache segment size). With awareness of the shared details of the structure of the storage region and/or the cache, the host arranges and sends out requests to read data from the persistent memory region in a manner that takes advantage of parallelism within the non-volatile storage system. For example, the host may initially send out one read request per cache segment to cause the non-volatile storage system to load the cache. Subsequently, additional read requests are made to the non-volatile storage system, with the data already loaded (or starting to load) in the cache, thereby increasing performance.

Abstract: The present disclosure generally relates to host memory buffer (HMB) cache management. HMB is transient memory and may not always be available. For example, when the link between the data storage device and the host device is not active, the data storage device can't access the HMB. Placing an HMB log in the HMB controller that is disposed in the data storage device provides access to data that would otherwise be inaccessible in the HMB. The HMB log contains any deltas that have occurred since either the last copying to an HMB cache in the memory device or any delta that have occurred since the link became inactive. The HMB cache mirrors the HMB. In so doing, the data of the HMB is available to the data storage device not only when the link is active, but also when the link is not active.

Abstract: The present disclosure generally relates to host memory buffer (HMB) cache management in DRAM-less SSDs. HMB is transient memory and may not always be available. For example, when the link between the data storage device and the host device is not active, the data storage device can't access the HMB. Placing an HMB log in the HMB controller that is disposed in the data storage device provides access to data that would otherwise be inaccessible in the HMB. The HMB log contains any deltas that have occurred since either the last copying to an HMB cache in the memory device or any delta that have occurred since the link became inactive. The HMB cache mirrors the HMB. In so doing, the data of the HMB is available to the data storage device not only when the link is active, but also when the link is not active.

Abstract: The present disclosure generally relates to a storage snapshot management system. When updated data is written to the memory device, rather than rewriting all of the data, only the updated data is written to a new namespace. A snapshot of the new namespace indicates which LBAs in the new namespace contain data. New namespaces are added each time data is updated. When the updated data is to be read, the data storage device reads the updated LBA from the new namespace, and also gathers the non-updated data from the previous namespace. Eventually, the number of namespaces for the data reaches a threshold, and thus some namespaces need to be evicted. To evict a namespace, the updated data in the namespace is moved to a different namespace, or the non-updated data is moved to a namespace that contains updated data. In either case, the now unused namespaces are evicted.

Abstract: The present disclosure generally relates to data storage devices, such as solid state drives, and effective power management of the data storage device. The data storage device includes a controller, where the controller is configured to predict when a host device will send a command to enter a low power state, prepare the data storage device to enter the low power state, and receive a command to enter the low power state after the predicting and preparing. If the data storage device is idled for greater than a threshold value, then the data storage device prepares to transition to a low power state but will wait to enter the lower power state until receiving a request from a host device.

Abstract: Aspects of the present disclosure generally relate to data storage devices, systems, and related methods that group commands of doorbell transactions from host devices into a plurality of groupings. A controller of a data storage device is configured to receive a plurality of submission doorbell transactions comprising a plurality of commands from a host device. The controller is configured to group the plurality of commands of the plurality of submission doorbell transactions into a plurality of groupings having a grouping order. Each grouping of the plurality of groupings corresponds to a single doorbell transaction of the plurality of submission doorbell transactions. The controller is configured to send one or more completion doorbell transactions to the host device. Each completion doorbell transaction of the one or more completion doorbell transactions identifies a completed grouping of the plurality of groupings.

Abstract: Recurrent Neural Networks (RNNs) wherein a non-volatile memory (NVM) array provides a memory bank for the RNN. The RNN may include a Neural Turning Machine (NTM) and the memory bank may be an NTM matrix stored in the NVM array. In some examples, a data storage device (DSD) that controls the NVM array includes both a data storage controller and a separate NTM controller. The separate NTM controller accesses the NTM matrix of the NVM array directly while bypassing flash translation layer (FTL) components of the data storage controller. Additionally, various majority wins error detection and correction procedures are described, as well as various disparity count-based procedures.

Abstract: Recurrent Neural Networks (RNNs) wherein a non-volatile memory (NVM) array provides a memory bank for the RNN. The RNN may include a Neural Turning Machine (NTM) and the memory bank may be an NTM matrix stored in the NVM array. In some examples, a data storage device (DSD) that controls the NVM array includes both a data storage controller and a separate NTM controller. The separate NTM controller accesses the NTM matrix of the NVM array directly while bypassing flash translation layer (FTL) components of the data storage controller. Additionally, various majority wins error detection and correction procedures are described, as well as various disparity count-based procedures.

Abstract: An adaptive-feedback-based read-look-ahead management system and method are provided. In one embodiment, a method for stream management is presented that is performed in a storage system. The method comprises performing a read look ahead operation for each of a plurality of streams; determining a success rate of the read look ahead operation of each of the plurality of streams; and allocating more of the memory for a stream that has a success rate above a threshold than for a stream that has a success rate below the threshold. Other embodiments are provided.

Abstract: A data storage device and method for file-based interrupt coalescing are provided. In one embodiment, a data storage device is provided comprising a memory and a controller. The controller is configured to execute a plurality of read commands read from a submission queue in a host; write a plurality of completion messages to a completion queue in the host; and coalesce interrupts to inform the host that plurality of completion messages were written to the completion queue; wherein the submission queue and the completion queue are dedicated to read commands from a host application and are separate from a submission queue and a completion queue for read and write commands from an operating system of the host. Other embodiments are possible, and each of the embodiments can be used alone or together in combination.