NOC-based reconfigurable neural network operation method and circuit

Abstract

The invention provides an NOC-based reconfigurable neural network operation method and an NOC-based reconfigurable neural network operation circuit. The NOC-based reconfigurable neural network operation circuit comprises a resource search unit, a resource list storage unit, a network configuration dispatching unit, a configuration information storage unit, a network mapping storage unit, a main storage unit, neural network processing circuits, NOC interface processing units and a network interconnection unit. Through arrangement of the NOC interface processing units, different neural network processing circuits are interconnected, and the NOC interface processing units can convert data calculated by the different neural network processing circuits into the same standard format for interaction, so that the neural network circuits can expand any new processing circuit, and data exchange between the various neural network processing circuits is more unified and efficient.

Description

technical field [0001] The invention relates to the field of neural network chip circuits, in particular to a NOC-based reconfigurable neural network operation method and device. Background technique [0002] With the rise of the artificial intelligence industry, deep learning chips are also developing rapidly. With the widespread use of deep learning applications, more and more devices begin to support deep learning functions. As the core of deep learning equipment, deep learning chips are becoming more and more important. At present, a large number of deep learning chips are aimed at certain characteristics or functions of the neural network structure, and the structure and connection relationship cannot be changed during use, resulting in the inability to adapt to the rapid development of deep learning and the rapid update of new network structures. In addition, in the existing technology, the data interaction between the network level of the neural network structure an...

AI Technical Summary

Problems solved by technology

[0003] To this end, it is necessary to provide a technical solution for reconfigurable neural network operations based on NOC, to solve the problem that the current neural network structure cannot change the structure and connection mode, resulting in the inability to adapt to the rapid update and use of the network structure, and because the network layer The data between them are in proprietary format, which leads to the problem of low data transmission efficiency when the network structure is complex

Images