Western Digital Technologies, Inc.

Sandisk Technologies Inc.

Alexander Bazarsky

Ariel Navon

Eran Sharon

Ran Zamir

Yehuda Hahn

Shay Benisty

Idan Alrod

David Avraham

Omer Fainzilber

Stella Achtenberg

Tomer Tzvi Eliash

Idan Goldenberg

Karin Inbar

Rami Rom

Michael Micha Ionin

Dudy David Avraham

Ofir Pele

Alon Marcu

Daniel Joseph Linnen

Yuval Grossman

Michael Ionin

Ishai Ilani

Yoav Markus

Nika Yanuka

William Bernard Boyle

Avichay Haim Hodes

Evgeny Mekhanik

Opher Lieber

David Gur

Sergey Anatolievich Gorobets

Nian Niles Yang

Grishma Shah

Jack Frayer

Tien-Chien Kuo

Tz-Yi Liu

Gadi Vishne

Philip David Reusswig

Tomer Eliash

David Rozman

Alon Eyal

Ofer Shapira

Joshua Lehmann

Roi Kirshenbaum

Alexander Kalmanovich

Micha Yonin

Itay Busnach

Yan Li

Yu-Chung Lien

Tianhong Yan

Amit Kumar Sharma

Julian Vlaiko

Rohit Sehgal

Abhinandan Venugopal

Xinmiao Zhang

Vinayak Bhat

Yan Dumchin

Eyal Sobol

Noam Presman

Yuri Ryabinin

Danny Berler

Mahim Raj Gupta

A Harihara Sravan

Noga Deshe

Arthur Shulkin

Lola Grin

Aki Bleyer

Lior Avital

Lior Bublil

Yuval Bahar

Gilad Koren

Tsiko Shohat Rozenfeld

Ofer Privman

Martin Booth

Jonathan Sokolowski

Sanel Alterman

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.

Storage controller having data augmentation components for use with non-volatile memory die

Splitting an address translation cache (ATC) into two portions can reduce costs and maintain efficient retrieval of data. One portion can be disposed in a first location while a second portion can be disposed in a second location distinct from the first location. The first location can be in the controller. The second location can be in a host memory buffer (HMB) or in a memory device separate from the controller. To obtain translated addresses, untranslated addresses can be searched in the first portion and the corresponding translated addresses can be retrieved from the second portion. When invalidating untranslated addresses, the untranslated addresses of the first portion can be deleted without a need to delete corresponding translated addresses in the second portion. To improve ATC storage capacity, grouping of untranslated addresses is possible using most significant bytes (MSBs).

Hybrid address translation cache using DRAM

A storage system has an inference engine that can infer a read threshold based on a plurality of parameters of the memory. The read threshold can be used in reading a wordline in the memory during a regular read operation or as part of an error handling process. Using a machine-learning-based approach to infer a read threshold can provide significant improvement in read threshold accuracy, which can reduce bit error rate and improve latency, throughput, power consumption, and quality of service. In another embodiment, a circuit-bounded array is used to manage updates to time and temperature tag information and to infer read thresholds.

Storage system and method for circuit-bounded-array-based time and temperature tag management and inference of read thresholds

Instead of ignoring workflow or priority of tenant, modify the write path according to the tenant behavior and system priority using the multi-tenant storage device. When writing information, the write path usually includes RAM memory where the host or device copies the data. The data is copied to the NVM storage element. The RAM memory may be either part of the storage controller (either SRAM or DRAM) or part of the host (like DRAM in HMB) that may be controlled by the storage controller. The RAM memories are used for different goals by the storage controller and optimizing the overall system performance (write/read) is a priority of the storage device. In multi-tenant architectures, the different tenants may have different priorities to have their requirements fulfilled. The tenant priority is passed by the host to the storage device through an interface and either be static or change dynamically during the device operation. Optimizing the write paths according to tenant workload and priority may be done by a module in the storage controller.

Optimizing usage of multiple write paths on multi-tenant storage devices

Recurrent Neural Networks (RNNs) wherein a non-volatile memory (NVM) array provides a memory bank for the RNN. The RNN may be a Neural Turing Machine (NTM) and the memory bank may be an NTM matrix stored in the NVM array. In some examples, an NTM controller sets the size of the NTM matrix based on a storage access granularity of the NVM array. For instance, if the NVM reads and writes data in flash memory unit (FMUs), the NTM controller sets the size of the NTM matrix to correspond to the size of an integer number of FMUs. In some examples, the NVM array includes on-chip NTM circuitry configured to perform at least some NTM read head and write head operations. Threshold-based processing is described that can reduce an amount of NTM data read from the NVM array. In other examples, volatile memory is employed rather than an NVM array.

Systems and methods for use with recurrent neural networks

Detecting the removal of a data storage device from a storage system involves first determining that a shorter pin of an electrical connector of a storage device is disconnected from a mating electrical connector, such as by sensing a voltage drop on that pin, then determining at a later time that a longer pin of the connector is disconnected from the mating connector. Responsive to determining that the longer pin was disconnected after a predetermined period of time after the shorter pin, a conclusion may be made that the storage device has been removed from the system as opposed to being subject to a simple device power aberration. Thus, responsive data destruction action(s) may be taken to render the data stored on the device inaccessible to the attacker thereby protecting the device even after the device is removed from the storage system.

Detection of data storage device removal

Data storage devices configured to exploit generative-adversarial-networks (GANs). The GANs include super-resolution GANs (SRGANs). In some examples, a GAN-based reconstruction procedure is implemented within a data storage controller to replace or supplement an error correction coding (ECC) decoding procedure. In other examples, soft bit information is exploited using GANs during decoding. A dissimilarity matrix may be generated to represent differences between an initial image and a GAN-reconstructed image, with matrix values mapped into low-density parity check (LDPC) codewords to facilitate LDPC decoding of data. In still other examples, confidence information obtained from a GAN is incorporated into image pixels. In some examples, GAN reconstruction of data is limited to modifying valley bits. Multiple GANs may be used in parallel with their outcome aggregated.

Data storage device configured for use with a generative-adversarial-network (GAN)

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.

Non-volatile memory die with deep learning neural network

Data storage devices configured to exploit generative-adversarial-networks (GANs). The GANs include super-resolution GANs (SRGANs). In some examples, a GAN-based decoding (reconstruction) procedure is implemented within a data storage controller to replace or supplement an error correction coding (ECC) decoding procedure to permit a reduction in the number of parity bits used while storing the data. In other examples, soft bit information is exploited using GANs during decoding. A dissimilarity matrix may be generated to represent differences between an initial image and a GAN-reconstructed image, with matrix values mapped into low-density parity check (LDPC) codewords to facilitate LDPC decoding of data. In still other examples, confidence information obtained from a GAN is incorporated into image pixels. In some examples, GAN reconstruction of data is limited to modifying valley bits. Multiple GANs may be used in parallel with their outcome aggregated.

Instead of incorporating a single interface towards the host for transferring data, utilizing a designated write-only storage logging device. The write-only storage logging device can accept sequential streams and automatically overwrite. The controller will read the log material in a secure manner using a different and separate physical connection than the one used for write. The storage device may have LBA ranges that work as write-only as well as other LBA ranges, which are normal (both reads and writes are enabled). Both options will allow for a traditional file system as well as sharing the storage, but will still protect the log areas that would be used for events that should not be read out.

Read-protected storage device with sequential logging

More efficient memory device usage is possible by altering the memory device management. For example, when the full storage capacity of the memory device will not be used, certain portions of the memory device can be shut off and then turned on when the storage capacity is needed. When less capacity is needed, data can be consolidated and certain portions of the memory device can be shut off. Additionally, rather than operating in multilevel cell (MLC) memory, the memory device can start in single level cell (SLC) memory and transition to MLC memory over time. If there is a determination that less memory is needed, the memory device can transition from MLC memory to SLC memory. In so doing, the storage capacity of the memory device is more appropriately utilized.

Adaptive use of multiple channels in a storage device

A DNA-based storage system includes an error correction system operable to: (a) identify a DNA codeword from a DNA sequencing operation; (b) calculate an initial syndrome weight; (c) determine that the initial syndrome weight is greater than a predetermined threshold; (d) perform an alignment alteration in the information segment by: (i) selecting a skew point within the information segment; (ii) performing an indel operation on the information segment at the skew point; (iii) calculating a modified syndrome weight; (iv) comparing the initial syndrome weight with the modified syndrome weight; and (v) incorporating the indel operation into the information segment when the comparing indicates an improvement in the modified syndrome weight; (e) decode the modified codeword; and (f) transmit the contents of the output file to a computing device, the output file representing user data stored within the DNA molecule.

Error correction systems and methods for DNA storage

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.

Memory (NVM) devices for use with recurrent neural networks

Rather than using low power mode since there is no indication of a sleep mode type to the data storage device, the sleep mode type is inferred by the data storage device, or supplied by the host. In so doing, the sleep type communication is improved. After the system returns power to the data storage device, there may be a read workload, depending on the sleep type. The workload is characterized by a low-queue depth (QD) sequential read from a specific area of the storage medium that was written to just prior to shut down. In response to inference or host cue, the data storage device will modify an operation so that data storage device is optimized to the sleep mode type, resulting in improved performance and power consumption.

Optimized power up in storage memory devices

The present disclosure generally relates to improved handling of write commands. The host memory buffer (HMB) or other storage space can be utilized to delay execution of host write commands which will improve write performance in different use cases and will also allow having more concurrent streams than open blocks without impacting write or read performance. Generally, once a write command is received, the write command is revised as a new write command that is logically equivalent to the original write command. The revised write command is moved to the HMB along with the data. In so doing, the write command is coalesced and write command handling is improved.

Write coalescing via HMB to optimize write performance

Growing community of inventors

Holon, Israel

Alexander Bazarsky