The present disclosure provides a processing device including: a coarse-grained pruning unit configured to perform coarse-grained pruning on a weight of a neural network to obtain a pruned weight, an operation unit configured to train the neural network according to the pruned weight. The coarse-grained pruning unit is specifically configured to select M weights from the weights of the neural network through a sliding window, and when the M weights meet a preset condition, all or part of the M weights may be set to 0. The processing device can reduce the memory access while reducing the amount of computation, thereby obtaining an acceleration ratio and reducing energy consumption.

A memory module includes a device memory configured to store data and including a first memory area and a second memory area, and a controller including an accelerator circuit. The controller is configured to control the device memory, transmit a command to exclude the first memory area from the system memory map to a host processor in response to a mode change request, and modify a memory configuration register to exclude the first memory area from the memory configuration register. The accelerator circuit is configured to use the first memory area to perform an acceleration operation.

Embodiments are directed to a deterministic streaming system with one or more deterministic streaming processors each having an array of processing elements and a first deterministic memory coupled to the processing elements. The deterministic streaming system further includes a second deterministic memory with multiple data banks having a global memory address space, and a controller. The controller initiates retrieval of first data from the data banks of the second deterministic memory as a first plurality of streams, each stream of the first plurality of streams streaming toward a respective group of processing elements of the array of processing elements. The controller further initiates writing of second data to the data banks of the second deterministic memory as a second plurality of streams, each stream of the second plurality of streams streaming from the respective group of processing elements toward a respective data bank of the second deterministic memory.

Disclosed herein includes a system, a method, and a device for reading and writing sparse data in a neural network accelerator. A mask identifying byte positions within a data word having non-zero values in memory can be accessed. Each bit of the mask can have a first value or a second value, the first value indicating that a byte of the data word corresponds to a non-zero byte value, the second value indicating that the byte of the data word corresponds to a zero byte value. The data word can be modified to have non-zero byte values stored at an end of a first side of the data word in the memory, and any zero byte values stored in a remainder of the data word. The modified data word can be written to the memory via at least a first slice of a plurality of slices that is configured to access the first side of the data word in the memory.

An apparatus and method are described for distributed and cooperative computation in artificial neural networks. For example, one embodiment of an apparatus comprises: an input/output (I/O) interface; a plurality of processing units communicatively coupled to the I/O interface to receive data for input neurons and synaptic weights associated with each of the input neurons, each of the plurality of processing units to process at least a portion of the data for the input neurons and synaptic weights to generate partial results; and an interconnect communicatively coupling the plurality of processing units, each of the processing units to share the partial results with one or more other processing units over the interconnect, the other processing units using the partial results to generate additional partial results or final results. The processing units may share data including input neurons and weights over the shared input bus.

The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

An energy-efficient sequencer comprising inline multipliers and adders causes a read source that contains matching values to output an enable signal to enable a data item prior to using a multiplier to multiply the data item with a weight to obtain a product for use in a matrix-multiplication in hardware. A second enable signal causes the output to be written to the data item.

A caching system including a first sub-cache and a second sub-cache in parallel with the first sub-cache, wherein the second sub-cache includes a set of cache lines, line type bits configured to store an indication that a corresponding cache line of the set of cache lines is configured to store write-miss data, and an eviction controller configured to flush stored write-miss data based on the line type bits.

Embodiments of the disclosure provide methods and systems for memory management. The method can include: receiving a request for allocating target node data to a memory space, wherein the memory space includes a buffer and an external memory and the target node data comprises property data and structural data and represents a target node of a graph having a plurality of nodes and edges; determining a node degree associated with the target node data; allocating the target node data to the memory space based on the determined node degree.

Techniques for providing an overlap data buffer to store portions of tiles between passes of chained layers of a neural network are described. One accelerator circuit includes one or more processing units to execute instructions corresponding to the chained layers in multiple passes. In a first pass, the processing unit(s) receives a first input tile of an input feature map from a primary buffer and performs a first operation on the first input tile to obtain a first output tile. The processing unit stores the first output tile in the primary buffer and identifies a portion of the first output tile as corresponding to overlap data between tiles of the input feature map. The processing unit stores the portion in a secondary buffer. In a second pass, the processing unit retrieves the portion to avoid fetching the portion that overlaps and computing the overlap data again.