A data processing circuit for neural network

Abstract

A data processing circuit relates to the field of digital circuits and comprises a first computing circuit and an input control circuit, wherein the first computing circuit comprises one or more sub-computing circuits, each sub-computing circuit comprises a first addition operation circuit, a multiplication circuit, a first comparison operation circuit and a first nonlinear operation circuit, andthe first nonlinear operation circuit comprises at least one of an exponential operation circuit and a logarithmic operation circuit. The input control circuit is used for controlling the first computing circuit to read input data and input parameters, controlling the operation circuit in the sub computing circuits included in the first computing circuit according to the received first instruction, and operating the input data and the input parameters. The data processing circuit not only can realize basic multiplication and accumulation operation, but also can realize more complex operationsuch as logarithmic operation, exponential operation or power operation through the first nonlinear operation circuit, thereby adapting to calculation of more neural network layers.

Description

technical field [0001] The invention relates to the field of digital circuits, in particular to the data processing of neural networks. Background technique [0002] The neural network processor mainly includes three modules: computing engine, control and storage. In order to adapt to the high-density calculation of the neural network, the calculation engine is usually adapted to the convolutional layer and the fully connected layer that carry more than 90% of the calculation of the neural network, that is, the calculation engine is suitable for the multiplication and accumulation of large amounts of data. For other types of calculations, the neural network processor reports the data to be processed to the CPU (Central Processing Unit, central processing unit), and the CPU sends the obtained data results to the neural network processor after the calculation is completed. For every calculation of the neural network layer other than the convolutional layer and the fully conne...

AI Technical Summary

Problems solved by technology

For every calculation of the neural network layer other than the convolutional layer and the fully connected layer, the data needs to be transmitted back and forth between the neural network processor and the CPU, and the resulting transmission delay greatly reduces the performance of the neural network processor. Efficiency in performing neural network-related calculations

Images