Dynamic relay-based infinite routing method and architecture for on-chip neural network

Abstract

The invention belongs to the technical field of routing implementation of brain-like computing chips, and relates to an on-chip neural network-oriented dynamic relay-based infinite routing method and architecture, and the method comprises the following steps of: applying a single-stage relay structure mode, a multi-stage relay structure mode, a multi-path relay structure mode and a dynamic relay structure mode or a combination of the modes; a path between a source node and a target node is opened, and long-distance communication and even infinite-distance communication between routes are achieved. According to the method, the utilization rate of resources in the on-chip neural network is greatly improved in a dynamic relay mode, reachability between neurons under a huge neural network topological structure can be supported, and the problem that data transmission is limited due to the fact that paths between the neurons under the huge neural network topological structure are long is solved.

Description

technical field [0001] The invention belongs to the technical field of routing implementation of brain-inspired computing chips, and relates to an infinite routing method and architecture based on dynamic relays for on-chip neural networks. Background technique [0002] In recent years, the effects of "memory wall" and "power consumption wall" have become increasingly serious, and the von Neumann architecture followed by traditional computers is facing huge challenges. In the post-Moore era, the semiconductor industry urgently needs to find new architectures and methods to meet the needs of the electronics industry for ever-increasing computing performance and extremely low power consumption. With the development of brain science, people gradually understand that the human brain is a computer with extremely high energy efficiency, and brain-like computing came into being. The combination of memory and computing unit fundamentally eliminates the "memory wall" problem of the ...

AI Technical Summary

Problems solved by technology

However, due to the limited function of a single neuron, only millions of neurons can work together to show unique advantages in specific intelligent information processing. For the network topology of millions of neurons, due to the huge number of connections, neurons Paths between are often long, a challenge for chips with limited routing

Images