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: Methods and apparatus for managing and optimizing data storage devices that include non-volatile memory (NVM) are described. One such method involves deriving a hint for one or more logical block addresses (LBAs) of a storage device based on information received from a host device and/or physical characteristics of the storage device, such as LBAs that are invalidated together; grouping the LBAs into one or more clusters of LBAs based on the derived hint and a statistical analysis of the physical characteristics of the storage devices; allocating available physical block addresses (PBAs) in the storage device to one of the LBAs based on the one or more clusters of LBAs to achieve optimization of a data storage device.

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: 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: 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: 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: Methods and apparatus for managing and optimizing data storage devices that include non-volatile memory (NVM) are described. One such method involves deriving a hint for one or more logical block addresses (LBAs) of a storage device based on information received from a host device and/or physical characteristics of the storage device, such as LBAs that are invalidated together; grouping the LBAs into one or more clusters of LBAs based on the derived hint and a statistical analysis of the physical characteristics of the storage devices; allocating available physical block addresses (PBAs) in the storage device to one of the LBAs based on the one or more clusters of LBAs to achieve optimization of a data storage device.

Abstract: In an illustrative example, a method includes sensing at least a portion of a representation of a convolutional low-density parity-check (CLDPC) codeword stored at a memory of a data storage device. The method further includes receiving the portion of the representation of the CLDPC codeword at a controller of the data storage device. The method further includes performing one or more management operations associated with the memory based on an estimated number of errors of the portion of the representation of the CLDPC codeword.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.

Abstract: Systems and methods are described for generating location-based read voltage offsets in a data storage device. Optimal read voltage thresholds vary across memory elements of a device. However, data storage devices are often limited in the number of read voltage thresholds that can be maintained in the device. Thus, it may not be possible to maintain optimal read voltage parameters for each memory element within a device. The systems and methods described herein provide for increased accuracy of read voltage thresholds when applied to memory elements within a specific location in a device, by enabling the use of location-based read voltage offsets, depending on a relative location of the memory element being read from. The read voltage offsets can be determined based on application of a neural network to data regarding optimal read voltage thresholds determined from at least a sample of memory elements in a device.

Abstract: Systems and methods are described for compacting operating parameter sets in a data storage device. Data storage device may be configured to maintain multiple operating parameter sets, each of which stores various parameters for interacting with different memory elements within the device. The data storage device may further be limited in the total number of operating parameter sets that can be maintained in the device at any given time. Thus, the data storage device may be required at various times to combine two or more operating parameter sets, to enable creation of a new operating parameter set. Because each operating parameter set can contain a number of parameters, identification of similar sets for combination can be computationally intensive. To identify similar sets in an efficient manner, a device as disclosed herein is enabled to reduce a dimensionality of each set, and locate similar sets under that reduced dimensionality.

Abstract: A data storage device includes a shaping engine and a compression engine. The shaping engine is configured to shape first data to generate second data. The compression engine is configured to compress the second data to generate third data.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.

Abstract: Systems and methods are described for generating location-based read voltage offsets in a data storage device. Optimal read voltage thresholds vary across memory elements of a device. However, data storage devices are often limited in the number of read voltage thresholds that can be maintained in the device. Thus, it may not be possible to maintain optimal read voltage parameters for each memory element within a device. The systems and methods described herein provide for increased accuracy of read voltage thresholds when applied to memory elements within a specific location in a device, by enabling the use of location-based read voltage offsets, depending on a relative location of the memory element being read from. The read voltage offsets can be determined based on application of a neural network to data regarding optimal read voltage thresholds determined from at least a sample of memory elements in a device.

Abstract: Systems and methods are described for compacting operating parameter sets in a data storage device. Data storage device may be configured to maintain multiple operating parameter sets, each of which stores various parameters for interacting with different memory elements within the device. The data storage device may further be limited in the total number of operating parameter sets that can be maintained in the device at any given time. Thus, the data storage device may be required at various times to combine two or more operating parameter sets, to enable creation of a new operating parameter set. Because each operating parameter set can contain a number of parameters, identification of similar sets for combination can be computationally intensive. To identify similar sets in an efficient manner, a device as disclosed herein is enabled to reduce a dimensionality of each set, and locate similar sets under that reduced dimensionality.

Abstract: In an illustrative example, a method includes sensing at least a portion of a representation of a convolutional low-density parity-check (CLDPC) codeword stored at a memory of a data storage device. The method further includes receiving the portion of the representation of the CLDPC codeword at a controller of the data storage device. The method further includes performing one or more management operations associated with the memory based on an estimated number of errors of the portion of the representation of the CLDPC codeword.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.

Abstract: A storage device may program data differently for different memory areas of a memory. In some embodiments, the storage device may use different codebooks for different memory areas. In other embodiments, the storage device may modify bit orders differently for different memory areas. What codebook the storage device uses or what bit order modification the storage device performs for a particular memory area may depend on the bad storage locations specific to that memory area. Where different codebooks are used, optimal codebooks may be selected from a library, or codebooks may be modified based on the bad storage locations of the memory areas.