Aspiring Sky Co. Limited

Yujie Wen

Zhijiong Luo

A digital integrated circuit with embedded memory for neural network inferring may include a controller and a matrix of processing blocks and cyclic bidirectional interconnections, where each processing block is coupled to 4 neighboring processing blocks regardless of its position in the matrix. A cyclic bidirectional interconnection may transmit every processing block's output to its upper, lower, left, right neighboring blocks or to its cyclic neighbors of the same row or column in replacement of any missing upper, lower, left or right neighbors. Each processing block may include invariant word buffers, variant word buffers, a multiplexer, and a processing unit. The multiplexer may select one of the 4 neighbor processing blocks' outputs. The processing unit may accept as inputs the multiplexer's selected value, a selected value from the variant word buffers and a selected value from the invariant word buffer and produce output which acts as the processing block's output.

Digital integrated circuit with embedded memory for neural network inferring

Technologies are described for multiple accelerators for a neural network, and methods thereof. In an example implementation, a neural network can be mapped to a system comprising a control unit and multiple accelerators, where the controller unit controls each accelerator's behavior, sends data to and receives data from each accelerator through the interconnections. Sub-networks may be created by grouping several network layers or dividing a network layer into multiple sub-layers depending on data to be processed and memory capacity of each accelerator. Accelerators have internal storage, thus, do not require external memory.

Multiple accelerators for neural network

A circuit for performing parallel convolutional computation for features and kernels of variable sizes may receive inputs of an m&#xd7;n matrix of feature data, an m&#xd7;n matrix of convolution data, and a (2m&#x2212;1)&#xd7;(2n&#x2212;1) matrix of kernel data. A feature manager of the circuit may hold m rows of n data buffers storing the input feature data and rotating values between rows during one restricted convolution calculation. A kernel manager of the circuit may hold a (2m&#x2212;1)&#xd7;(2n&#x2212;1) matrix of data buffers storing the input kernel data in the buffers and cyclically rotating values in upwards, downwards, leftwards and rightwards directions for different restricted convolution calculations. A row convolution engine of the circuit may hold m row convolution processors, each storing and updating input convolution data by multiplication-and-accumulation (MAC) operations on its input feature and kernel data rows. The circuit produces accumulated convolutional data.

Growing community of inventors

Shanghai, China

Yujie Wen