Methods, systems, and devices for improved data management for memory are described. An apparatus may include a memory array including one or more blocks of memory cells. Data read from a block of memory cells may be written to a buffer, to support providing the data to a host system or modification of the data by the host system. If a quantity of read commands performed at the block of memory cells satisfies a threshold, the data may be written from the buffer to a different block of memory cells, rather than the block from which the data was previously read.

Implementations disclosed describe methods and systems to perform the methods of optimizing a size of memory used for accumulation of neural node outputs and for supporting multiple computational paths in neural networks. In one example, a size of memory used to perform neural layer computations is reduced by performing nodal computations in multiple batches, followed by rescaling and accumulation of nodal outputs. In another example, execution of parallel branches of neural node computations include evaluating, prior to the actual execution, the amount of memory resources needed to execute a particular order of branches sequentially and select the order that minimizes this amount or keeps this amount below a target threshold.

A computer-implemented method, according to one embodiment, includes: storing records in an input data buffer, where each of the records include a key which is appended to payload data in the respective record. Moreover, for each of the records: shearing the key associated with the record from the payload data, normalizing the sheared key, and storing the normalized sheared key in a first target area of memory. A determination is made as to whether a size of the payload data in the record is outside a predetermine range, and in response to determining that the size of the payload data in the record is outside the predetermine range, the payload data is stored in a second target area of memory. A data locator is also appended to the normalized sheared key in the first target area of memory to form a sheared record.

A memory device includes a memory die bonded to a logic die. A logic die that is bonded to a memory die via a wafer-on-wafer bonding process can receive signals indicative of input data from a global data bus of the memory die and through a bond of the logic die and memory die. The logic die can also receive signals indicative of kernel data from local input/output (LIO) lines of the memory die and through the bond. The logic die can perform a plurality of operations at a plurality of vector-vector (VV) units utilizing the signals indicative of input data and the signals indicative of kernel data.

Aspects of the disclosure are directed to hardware interconnects and corresponding devices and systems for non-coherently accessing data in shared memory devices. Data produced and consumed by devices implementing the hardware interconnect can read and write directly to a memory device shared by multiple devices, and limit coherent memory transactions to relatively smaller flags and descriptors used to facilitate data transmission as described herein. Devices can communicate less data on input/output channels, and more data on memory and cache channels that are more efficient for data transmission. Aspects of the disclosure are directed to devices configured to process data that is read from the shared memory device. Devices, such as hardware accelerators, can receive data indicating addresses for different data buffers with data for processing, and non-coherently read or write the contents of the data buffers on a memory device shared between the accelerators and a host device.

An device is disclosed. A first buffer to store a query data point, and a second buffer to store a matrix of candidate data points. A processing element may process the query data point and the matrix of candidate data points to identify candidate data points in the matrix of candidate data points that are nearest to the query data point.

A reconfigurable compute fabric of a system can include multiple nodes, and each node can include multiple, communicatively coupled tiles with respective processing and storage elements. In an example, a tile-based processor can be configured to perform operations comprising receiving a first stencil that defines input data for a first operation. The stencil can have a height corresponding to N rows in a main memory and a stencil width corresponding to M columns in the main memory. The processor can perform operations comprising establishing N buffers in a tile memory, each buffer having M buffer elements, and populating the M buffer elements of the N buffers using respective information, defined by the first stencil, from the main memory. Tile-based stencil operations can use information from the N buffers and provide compute results in an output array.

A memory device includes an array of memory cells configured on a die or chip and coupled to sense lines and access lines of the die or chip and a respective sense amplifier configured on the die or chip coupled to each of the sense lines. Each of a plurality of subsets of the sense lines is coupled to a respective local input/output (I/O) line on the die or chip for communication of data on the die or chip and a respective transceiver associated with the respective local I/O line, the respective transceiver configured to enable communication of the data to one or more device off the die or chip.

A storage device according to the present technology may include a memory device for storing data, a buffer memory configured to temporarily store data to be stored in the memory device, and a memory controller configured to determine a delay time based on a plurality of parameters upon receipt of a write request from a host, and transmit a data request to the host after the delay time has elapsed.

A system includes a memory device, and a processing device, operatively coupled to the memory device, to perform operations including receiving a request to sequentially write data to a block of a memory device, in response to receiving the request, writing the data to the block to obtain sequentially written data, initiating accumulation of logical-to-physical (L2P) mapping data corresponding to the sequentially written data, determining that a criterion for terminating the accumulation of the L2P mapping data is satisfied, in response to determining that the criterion is satisfied, terminating the accumulation of the L2P mapping data to obtain accumulated L2P mapping data, and updating an L2P mapping data structure based on the accumulated L2P mapping data.