Prediction method applied to neural network chip and prediction apparatus thereof

Abstract

The invention provides a prediction method applied to a neural network chip and an apparatus thereof, a server and a readable storage medium. The method comprises the following steps of dividing M output data of a current layer and acquiring M output data distribution intervals of the current layer; carrying out statistics on the output data in each output data distribution interval of the M output data distribution intervals and acquiring a ratio of a number of the output data in each output data distribution interval and a total number of the M output data; based on the ratio corresponding to the M output data distribution intervals and a preset bit width, carrying out a bit width constraint on the M output data and acquiring a first initial bit and a first termination bit of N distribution intervals of the M output data distribution intervals; and based on the first initial bit and the first termination bit, carrying out bit width constraint on the M output data of the current layer so as to realize prediction of the neural network chip. In the invention, a data bandwidth is reduced and calculating efficiency is further increased.

Description

technical field [0001] The invention relates to the technical field of artificial neural network computing, in particular to a prediction method and device, a server and a readable storage medium applied to a neural network chip. Background technique [0002] Artificial Neural Network (ANN) is composed of a large number of nodes (or neurons) connected to each other. Each node represents a specific output function, called the activation function (ActivationFunction). Each connection between every two nodes has a weight that determines its connection strength. The value of this weight determines the state of the neuron and the performance of the entire neural network system. For the same network structure with different weights, the performance behavioral characteristics are often different. [0003] Weights and biases are important parameters that affect the performance of artificial neural network models. In the prior art, in the training phase of the artificial neural ne...

AI Technical Summary

Problems solved by technology

Due to the existence of a large number of convolution operations in the operation process, a large amount of temporary data is generated during the operation process and a large amount of data bandwidth is occupied. Therefore, the calculation time is long and the calculation efficiency is low, and the rapid prediction of the artificial neural network model cannot be realized.

Images