Calculation device, calculation method, and chip

Abstract

The invention provides an arithmetic device. The arithmetic device comprises a screening unit, which is used for screening out a characteristic map for calculating and obtaining the characteristic mapof the output neuron and corresponding weight values according to a connection state array of the characteristic map composed of the output neuron and the input neuron, and outputting the result to the arithmetic unit. And/or for selecting a characteristic map row/column for calculating and obtaining the output neuron and a corresponding weight value row/column according to the connection state array of each row/column in the characteristic map composed of the output neuron and the input neuron, and outputting the obtained result to an arithmetic unit; And an arithmetic unit which performs corresponding artificial neural network operation supporting structure clipping on the data output from the screening unit according to the instruction to obtain an output neuron. The invention avoids redundancy of operation amount and redundancy of accessing memory caused by all input neurons and weights participating in network operation, solves the problems of insufficient CPU and GPU operation performance and large front-end decoding overhead.

Description

technical field [0001] The invention belongs to the technical field of data processing, and relates to a computing device, a computing method and a chip. Background technique [0002] Artificial neural network (ANN), referred to as neural network (NN), is a mathematical model or operational model that imitates the structure and function of a biological neural network. Artificial neural network has a wide range of applications in many fields, such as image recognition, computer vision, speech recognition, natural language processing and other fields, and has achieved excellent results in these fields. [0003] With the development of the neural network, its network framework gradually becomes larger, the network parameters gradually increase, and its computing-intensive and storage-intensive features become more and more prominent. For example, for a caffe-based network architecture, the size of GoogLeNet Caffemodel is about 50MB, AlexNet Caffemodel and ResNet -152Caffemodel...

AI Technical Summary

Problems solved by technology

[0004] At present, the development of neural network is mainly limited to the following two aspects. In terms of hardware, the calculation-intensive and storage-intensive nature of artificial neural network has led to high requirements for hardware equipment to run artificial neural network. Currently, general-purpose processors ( CPU) or graphics processing unit (GPU) to perform artificial neural network operations is still very difficult, resulting in high front-end decoding overhead

Images