Abstract: A computing system, preferably including a scratchpad, a plurality of computation units, and a register file, and optionally including a controller. The computing system, or any suitable elements thereof, can optionally be integrated into a processor unit, wherein one or more such processor units can optionally be integrated into a larger-scale computing system. Some or all elements of the larger-scale computing system can be collocated and/or codefined on a single semiconductor chip, can be located and/or defined on separate chips, and/or can be otherwise located and/or defined. A method of operation, preferably including performing computation unit I/O operations and/or performing scratchpad I/O operations. The method of operation is preferably performed using the computing system, but can additionally or alternatively be performed using any other suitable systems.

Abstract: A system can include a processing device operatively coupled with the one or more memory devices, to perform operations that include writing data to the one or more memory devices and performing one or more scan operations on a management unit containing the data to determine a current value of a chosen data state metric. Each scan operation can be performed using a corresponding predetermined read-time parameter value. The operations can include determining whether the current value of the chosen data state metric satisfies a criterion, and can also include, responsive to determining that the current value of the chosen data state metric satisfies the criterion, selecting a remedial operation by determining whether redundancy metadata is included in a fault tolerant data stripe on the one or more memory devices. The operations can also include performing the remedial operation with respect to the management unit.

Abstract: Methods and storage unit allocators for allocating one or more portions of a storage unit to a plurality of tasks for storing at least two types of data. The method includes receiving a request for one or more portions of the storage unit to store a particular type of data of the at least two types of data for a task of the plurality of tasks; associating the request with one of a plurality of virtual partitionings of the storage unit based on one or more characteristics of the request, each virtual partitioning allotting none, one, or more than one portion of the storage unit to each of the at least two types of data; and allocating the requested one or more portions of the storage unit to the task from the none, one, or more than one portion of the storage unit allotted to the particular type of data in the virtual partitioning associated with the request.

Abstract: System and techniques for silent cache line eviction are described herein. A memory device receives a memory operation from a host. The memory operation establishes data and metadata in a cache line of the memory device upon receipt. The metadata is stored in a memory element that corresponds to the cache line. Later, an eviction trigger to evict the cache line is identified. Then, in response to the eviction trigger, current metadata of the cache line is compared with the metadata in the memory element to determine whether the metadata has changed. the cache line can be evicted without writing to backing memory in response to the metadata being unchanged.

Abstract: A memory device supports low power operation by facilitating staggered access to row segments within a row of a memory bank. Upon receiving an activate command to activate a row, the memory device sequentially activates a plurality of local wordlines associated with the row with a stagger interval between activations. Upon receiving an access command associated with the activated row, the memory device sequentially initiates column operations for respective row segments with the same stagger interval between the column operations. The memory device may furthermore facilitate error correction code operations in a staggered manner by sequentially performing computations associated with the different row segments.

Abstract: An example methodology includes, by a computing device, polling a power monitoring device coupled to a storage system for power consumption by the storage system and whether the power consumption exceeds a power consumption threshold. The method also includes, responsive to a determination that the power consumption exceeds the power consumption threshold, managing, by the computing device, the power consumption by the storage system to cause the power consumption to not exceed the power consumption threshold.

Abstract: A method includes error encoding data to produce a plurality of data slices. Metadata is determined for a data slice of the plurality of data slices. The metadata is stored in a metadata storage tree. The data slice is stored in a slice storage location indicated by the metadata. Based on determining to access the data slice, the metadata for the data slice is accessed in the metadata storage tree to determine the slice storage location for the data slice, and the data slice is accessed in the slice storage location based on determining the slice storage location for the data slice via accessing the metadata storage tree.

Abstract: Various example embodiments of a capability for supporting reconfigurable partitioning of high bandwidth memory (HBM) in a multi-processor computing system are presented. The capability for supporting reconfigurable partitioning of HBM in a multi-processor computing system, where the multi-processor computing system includes a multi-processor system-on-chip (SoC) including a set of processors supporting respective sets of processor memory channels and an HBM interconnected via an HBM bus supporting a set of HBM bus channels, may support dynamic mapping of the processor memory channels of the processors of the multi-processor SoC to the HBM bus channels of the HBM, thereby overcoming various limitations of a static mapping of the processor memory channels of the processors of the multi-processor SoC to the HBM bus channels of the HBM.

Abstract: An example device can include at least one network controller configured to receive a data request and to retrieve data based on the data request, and a cache agent configured to receive a data access parameter based on the data request, and reconfigure a cache for at least one memory cache based on the data access parameter. The data request can be received from a computer device and the data can be retrieved from at least one memory device. An example data access parameter can include a latency of at least one network-attached memory device to retrieve data from the at least one memory device based on the data request. An example device can further comprises a flit profiler configured to determine the data access parameter. Various other methods, systems, and computer-readable media are also disclosed.

Abstract: Provided is a device that includes an interface operatively coupled to a locked redundant array of independent disks including M (M>1) memory dice, the M memory dice stores stripes of data, and each stripe spanning over the M memory dice; and control circuitry that performs data compression on data to generate compressed data; stores the compressed data in the stripes; generates parity data of each stripe; determines, for each stripe, memory dice required to store the compressed data; determines, for each stripe, whether the memory dice required to store the compressed data in each stripe is N memory dice or less; where N is an integer less than M; and determines, for each stripe, which stripes will store the parity data of a respective stripe based on the determination of whether the memory dice required to store the compressed data in the respective stripe is N memory dice or less.

Abstract: A processor and an operation method thereof may access a memory in multiple segments. The processor includes a vector register file (VRF) and a load-store device. When the load-store device performs a multi-segment load on the memory, the load-store device reads a plurality of data elements from a source segment of the memory, and then writes the data elements in the VRF within a single write cycle, so that the data elements are written in a same location in different vector registers of the VRF. When the load-store device performs a multi-segment store on the memory, the load-store device reads a plurality of data elements from the VRF within a single read cycle (the data elements are data elements of a same location in different vector registers), and then writes the data elements in a target segment of the memory.

Abstract: Disclosed herein are system, apparatus, article of manufacture, method and/or computer program product embodiments, and/or combinations and sub-combinations thereof, for utilizing a single storage buffer for a dynamic number of players, each using a dynamically sized virtual buffer. For example, a system includes a buffer management controller that receives a request to initiate at least one player instance for displaying a content item. The buffer management controller creates a virtual buffer having a maximum capacity for the player instance. Finally, the buffer management controller identifies one or more available regions of the single storage buffer and maps the virtual buffer to the one or more available regions in response to determining that address space in the one or more available regions matches or exceeds the maximum capacity of the virtual buffer.

Abstract: A storage system includes a NAND storage media and a nonvolatile storage media as a write buffer for the NAND storage media. The write buffer is partitioned, where the partitions are to buffer write data based on a classification of a received write request. Write requests are placed in the write buffer partition with other write requests of the same classification. The partitions have a size at least equal to the size of an erase unit of the NAND storage media. The write buffer flushes a partition once it has an amount of write data equal to the size of the erase unit.

Abstract: A data processing system that implements an improved read-modify-write operation for lossless tiling in convolution networks is presented. The data processing system includes runtime logic that is configured to execute a graph to generate at the output of the graph, a plurality of tiles of a tensor, initialize a memory comprising all zeros for storing the plurality of tiles, determine a current memory region in the memory for storing a current tile of the plurality of tiles, wherein the current memory region comprises an overlapping memory region that stores data from write operations of previously stored neighboring tiles of the plurality of tiles and a remaining memory region, and perform a read-modify-write operation on the data from the overlapping memory region using the data from the overlapping memory region and first tile data of the current tile for storing in the overlapping memory region.

Abstract: A memory device and in-memory searching method are provided. The memory device is, for example, a three dimensional NAND flash memory circuit, and provides a storage media with high-performance and high-capacity. The memory device includes an address scanner, a searching data transmitter, a readout data sensor and comparator, an error bit detector, and a matching status processor. The address scanner provides a scanned address information within a search setting range. The readout data sensor and comparator compares a readout data with a search data bit by bit to generate a comparison result. The error bit detector determines a matching information of the readout data and the search data according to the comparison result. The matching status processor generates a matching address information.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed to reduce read-modify-write cycles for non-aligned writes. An example apparatus includes a memory that includes a plurality of memory banks, an interface configured to be coupled to a central processing unit, the interface to obtain a write operation from the central processing unit, wherein the write operation is to write a subset of the plurality of memory banks, and bank processing logic coupled to the interface and to the memory, the bank processing logic to determine the subset of the plurality of memory banks to write based on the write operation, and determine whether to cause a read operation to be performed in response to the write operation based on whether a number of addresses in the subset of the plurality of memory banks to write satisfies a threshold.

Abstract: A method and device for accessing data in a host memory are provided. The method includes: in response to writing first data corresponding to a first Remote Dictionary Server (Redis) instance of a plurality of Redis instances into the host memory, determining whether a remaining host memory capacity for the first Redis instance is sufficient to store the first data, wherein the remaining host memory capacity is based on a difference between maximum capacity of the host memory that is capable of being used by the first Redis and capacity of the host memory that has been used by the first Redis instance; and in response to determining that that the remaining host memory capacity is not sufficient to store the first data, evicting second data of the first Redis instance from the host memory to a Computing Express Link (CXL) device and storing the first data in the host memory.

Abstract: A memory controller capable of improving the performance of a read operation may include: a read command generator for generating read commands instructing a memory device to read, in advance, data to be read-requested from a host; a cache memory for storing data read from the memory device according to the read commands; a read operation controller for receiving, from the cache memory, data corresponding to a read request of the host in response to the read request, and outputting the received data to the host; and a cache memory size controller for controlling a size of the cache memory, based on a size of an area assigned for prefetching data in the data stored in the cache memory and a size of the data corresponding to the read request.

Abstract: A geometry-aware method of writing parity, performed by a storage system is provided. The method includes writing data, using error correction coding, to a RAID stripe comprising a plurality of allocation units of storage memory of the storage system, and writing one or more parity pages comprising parity data for a block having a portion of the data of the RAID stripe in one of the plurality of allocation units of the RAID stripe to a targeted differing word line of the block or a differing block, the parity data for the block distinct from the error correction coding of the RAID stripe.

Abstract: Systems and methods are provided for durable asynchronous replication in a distributed storage system. A storage system may commit a transaction comprising saving a data object to a persistent object storage location and saving a replication marker to a first persistent storage directory for pending replication tasks, wherein the replication marker comprises replication data for replicating the data object to a remote storage system, and wherein the first persistent storage directory comprises a replication queue; determine, based on a presence of the replication marker in the first persistent storage directory, to dequeue the replication marker from the replication queue; move the replication marker to a second persistent storage directory for in-process replication tasks; attempt replication of the data object to the remote storage system; and move the replication marker to a third persistent storage directory.