Time-division-multiplexing general neural network processor

Abstract

Provided in the invention is a time-division-multiplexing general neural network processor comprising at least one storage unit (100), at least one storage unit controller (101), at least one arithmetic logic unit (103), and a control unit (102). To be specific, the at least one storage unit (100) is used for storing an instruction and data. The at least one storage unit controller (101) corresponds to the at least one storage unit (100) and accesses the corresponding storage unit (100). The at least one arithmetic logic unit (103) is used for executing neural network computing. The control unit (102) connected with the at least one storage unit controller (101) and the at least one arithmetic logic unit (103) can obtain the instruction stored by the at least one storage unit (100) by the at least one storage unit controller (101) and parse the instruction to control the at least one arithmetic logic unit (103) to execute computation. The provided general neural network processor with high universality is suitable for computation of a large-scale neural network.

Description

technical field [0001] The present invention relates to an accelerator and a processor architecture and a design method, in particular to a hardware acceleration technology for an artificial neural network (Artificial Neural Network, referred to as ANN). Background technique [0002] Artificial neural network, referred to as neural network, is a computing model consisting of a large number of nodes (or neurons) connected to each other. Each node represents a specific output function, also known as an activation function (activation function), which can be a linear function, a ramp function, a threshold function, a sigmoid function, a bipolar sigmoid function, and the like. The connection between each two nodes represents a weight for the signal passing through the connection, called weight, which is equivalent to the memory of the neural network. The output of the neural network is different according to the connection method, weight and activation function of the network. ...

AI Technical Summary

Problems solved by technology

This method has high computational efficiency, but its shortcomings are also prominent: on the one hand, with the popularity of deep learning, the scale of neural networks used in practical applications is getting larger and larger, and the method of topological mapping is difficult to apply; on the other hand, Accelerators designed using topological mapping methods are only suitable for neural networks with a specific structure, but not for neural networks with other structures, that is, different neural network algorithms cannot be run on the same hardware accelerator.

[0005] In summary, the existing neural network accelerators are not suitable for computing large-scale neural networks and have poor versatility

Images