Abstract: A method of exploiting activation sparsity in deep neural networks is described. The method includes retrieving an activation tensor and a weight tensor where the activation tensor is a sparse activation tensor. The method also includes generating a compressed activation tensor comprising non-zero activations of the activation tensor, where the compressed activation tensor has fewer columns than the activation tensor. The method further includes processing the compressed activation tensor and the weight tensor to generate an output tensor.

Abstract: Various implementations described herein are related to a device having multiplier circuitry with an array of summation result cells that holds summation bit values for shifted arrays added together. The device may include latch circuitry having one or more gated elements disposed between the summation result cells, and the gated elements may be adapted to provide a portion of the summation bit values based on a gating signal.

Abstract: Provided are an integrated circuit chip apparatus and a related product, the integrated circuit chip apparatus being used for executing a multiplication operation, a convolution operation or a training operation of a neural network. The present technical solution has the advantages of a small amount of calculation and low power consumption.

Abstract: Method and system relating generally to convolution is disclosed. In such a method, an image patch is selected from input data for a first channel of a plurality of input channels of an input layer. The selected image patch is transformed to obtain a transformed image patch. The transformed image patch is stored. Stored is a plurality of predetermined transformed filter kernels. A stored transformed filter kernel of the plurality of stored predetermined transformed filter kernels is element-wise multiplied by multipliers with the stored transformed image patch for a second channel of the plurality of input channels different from the first channel to obtain a product. The product is inverse transformed to obtain a filtered patch for the image patch.

Abstract: A device for computing an inner product includes an index unit, a storage operation unit, a redundant to 2's complement (RTC) converter, a mapping table, and a multiplier-accumulate (MAC) module. The index unit, storing index values, is coupled to word lines. The storage operation unit includes the word lines and bit lines and stores data values. The mapping table stores coefficients corresponding to the index values. The index unit enables the word line according to a count value and the index value, such that the storage operation unit accumulates the data values corresponding to the bit lines and the enabled word line, thereby generating accumulation results. The RTC converter converts the accumulation results into a total data value in 2's complement format. The MAC module operates based on the total data value and the coefficient to generate an inner product value.

Abstract: Some embodiments provide a neural network inference circuit for executing a neural network with multiple layers. The neural network inference circuit includes a set of processing circuits for executing the layers of the neural network, a set of memories for storing data used by the set of processing circuits to execute the neural network layers, and a read controller for retrieving the data from the set of memories and storing the data in a cache for use by the set of processing circuits. The read controller retrieves the data in one of (i) a first mode for retrieving the data from sequential memory locations within the set of memories to store in the cache and (ii) a second mode for retrieving the data from non-sequential memory locations within the set of memories to store in the cache.

Abstract: A computing device, comprising: a computing module, comprising one or more computing units; and a control module, comprising a computing control unit, and used for controlling shutdown of the computing unit of the computing module according to a determining condition. Also provided is a computing method. The computing device and method have the advantages of low power consumption and high flexibility, and can be combined with the upgrading mode of software, thereby further increasing the computing speed, reducing the computing amount, and reducing the computing power consumption of an accelerator.

Abstract: A processing device with dynamically configurable operation bit width, characterized by comprising: a memory for storing data, the data comprising data to be operated, intermediate operation result, final operation result, and data to be buffered in a neural network; a data width adjustment circuit for adjusting the width of the data to be operated, the intermediate operation result, the final operation result, and/or the data to be buffered; an operation circuit for operating the data to be operated, including performing operation on data to be operated of different bit widths by using an adder circuit and a multiplier; and a control circuit for controlling the memory, the data width adjustment circuit and the operation circuit. The device of the present disclosure can have the advantages of strong flexibility, high configurability, fast operation speed, low power consumption or the like.

Abstract: The present disclosure relates to a processing device including a memory configured to store data to be computed; a computational circuit configured to compute the data to be computed, which includes performing acceleration computations on the data to be computed by using an adder circuit and a multiplier circuit; and a control circuit configured to control the memory and the computational circuit, which includes performing acceleration computations according to the data to be computed. The present disclosure may have high flexibility, good configurability, fast computational speed, low power consumption, and other features.

Abstract: Some embodiments provide a method for a circuit that executes a neural network including multiple nodes. The method loads a set of weight values for a node into a set of weight value buffers, a first set of bits of each input value of a set of input values for the node into a first set of input value buffers, and a second set of bits of each of the input values into a second set of input value buffers. The method computes a first dot product of the weight values and the first set of bits of each input value and a second dot product of the weight values and the second set of bits of each input value. The method shifts the second dot product by a particular number of bits and adds the first dot product with the bit-shifted second dot product to compute a dot product for the node.

Abstract: A universal interferometer for a three-dimensional quantum random number generator (3D QRNG) that includes three input ports. Each input port being configured to receive at least one prepared photon having three dimensions of quantum information in Hilbert space from a preparation stage of the 3D QRNG. The prepared photon is prepared based on a selected probability distribution set selected from probability distribution sets of p1, p2 and p3 that add to 1 and where the p1, the p2 and the p3 are rational numbers less than 1 and greater than zero. The universal interferometer includes three two-dimensional interferometers connected and arranged to measure along an x-axis of a Cartesian coordinate system the prepared photon and preserving the three dimensions of the quantum information of the measured photon. The universal interferometer includes three output ports which produces the measured photon corresponding to a three-dimensional eigenstate.

Abstract: A deep neural network (“DNN”) module compresses and decompresses neuron-generated activation data to reduce the utilization of memory bus bandwidth. The compression unit receives an uncompressed chunk of data generated by a neuron in the DNN module. The compression unit generates a mask portion and a data portion of a compressed output chunk. The mask portion encodes the presence and location of the zero and non-zero bytes in the uncompressed chunk of data. The data portion stores truncated non-zero bytes from the uncompressed chunk of data. A decompression unit receives a compressed chunk of data from memory in the DNN processor or memory of an application host. The decompression unit decompresses the compressed chunk of data using the mask portion and the data portion.

Abstract: A computation method and a computation apparatus exploiting weight sparsity, adapted for a processor to perform multiply-and-accumulate operations on a memory including multiple input and output lines crossing each other. In the method, weights are mapped to the cells of each operation unit (OU) in the memory. The rows of the cells of each OU are compressed by removing at least one row of the cells each mapped with a weight of 0, and an index including values each indicating a distance between every two rows of the cells including at least one cell mapped with a non-zero weight for each OU is encoded. Inputs are inputted to the input lines corresponding to the rows of each OU excluding the rows of the cells with the weight of 0 according to the index and outputs are sensed from the output lines corresponding to the OU to compute a computation result.

Abstract: A system may include a memory array for VMM and includes a matrix of devices. The devices may be configured to receive a programming signal to program a weight to store a matrix of weights. The devices may be configured to receive a digital signal representative of a vector of input bits. The devices may generate an analog output signal by individually multiplying input bits by a corresponding weight. The system may include multiple ADCs electrically coupled to a corresponding device. Each ADC may be configured to convert a corresponding analog output signal to a digital signal based on a current level of the corresponding analog output signal. The system may include registers electrically coupled to a corresponding ADC configured to shift and store an output vector of bits of a corresponding digital output signal based on an order of the vector of input bits received by the corresponding device.

Abstract: Provided are an integrated circuit chip apparatus and a related product, the integrated circuit chip apparatus being used for executing a multiplication operation, a convolution operation or a training operation of a neural network. The present technical solution has the advantages of a small amount of calculation and low power consumption.

Abstract: An architecture is disclosed for an neural processing element having single instruction, multiple data (“SIMD”) compute lanes. The neural processing element includes compute lanes having multipliers configured to multiply a binary operand with another binary operand to generate a binary output. The neural processing element also includes a single adder tree for summing the binary outputs of the hardware binary multipliers. The neural processing element also includes a storage element for storing a binary output of the single hardware binary adder tree.

Abstract: The present disclosure relates to a fused vector multiplier for computing an inner product between vectors, where vectors to be computed are a multiplier number vector {right arrow over (A)}{AN . . . A2A1A0} and a multiplicand number {right arrow over (B)} {BN . . . B2B1B0}, {right arrow over (A)} and {right arrow over (B)} have the same dimension which is N+1.

Abstract: An apparatus to facilitate a computer number format conversion is disclosed. The apparatus comprises a control unit to receive to receive data format information indicating a first precision data format that input data is to be received and converter hardware to receive the input data and convert the first precision data format to a second precision data format based on the data format information.

Abstract: An adaptive matrix multiplier. In some embodiments, the matrix multiplier includes a first multiplying unit a second multiplying unit, a memory load circuit, and an outer buffer circuit. The first multiplying unit includes a first inner buffer circuit and a second inner buffer circuit, and the second multiplying unit includes a first inner buffer circuit and a second inner buffer circuit. The memory load circuit is configured to load data from memory, in a single burst of a burst memory access mode, into the first inner buffer circuit of the first multiplying unit; and into the first inner buffer circuit of the second multiplying unit.

Abstract: A method, system and electronic device for mitigating variance in a two transistor two resistive memory element (2T2R) circuit is provided. The method includes calculating a sum of a number of logical 1's in a column of bitcells in the 2T2R circuit, N, of an input vector, sensing output current values from each current line in the column of bitcells and calculating an inner product, M, of the input vector and the bitcells in the column in the 2T2R circuit based on the sensed output current values.