Neural network processing method and system

Abstract

The invention is suitable for the technology of computers, and provides a neural network processing method and system. The system comprises a multicore processing module, an on-chip storage medium, an in-chip address index module and an ALU module, wherein the multicore processing module consists of a plurality of core processing modules. The multicore processing module is used for executing vector multiplication and addition operation in neural network calculation. The ALU module is used for obtaining input data from the multicore processing module or the on-chip storage medium, and executing nonlinear calculation which cannot be completed by the multicore processing module. The plurality of core processing modules share the on-chip storage medium and the ALU module, or the plurality of core processing modules have independent on-chip storage mediums and ALU modules. According to the invention, a multicore design is introduced in the neural network processing system, thereby improving the calculation speed of the neural network processing system, enabling the performance of the neural network processing system to be higher, and enabling the neural network processing system to be higher in efficiency.

Description

technical field [0001] The invention relates to the technical field of computers, and belongs to a neural network processing method and system. Background technique [0002] In the era of big data, more and more devices need to process more and more complex real-time input from the real world, such as industrial robots, self-driving cars and mobile devices. Most of these tasks are biased towards the field of machine learning, and most of the operations are vector operations or matrix operations with a very high degree of parallelism. Compared with traditional general-purpose GPU / CPU acceleration solutions, hardware ASIC accelerators are currently the most popular acceleration solutions. On the one hand, they can provide extremely high parallelism to achieve extremely high performance, and on the other hand, they have extremely high energy efficiency. [0003] Common neural network algorithms include the most popular Multi-Layer Perceptron (MLP), Convolutional Neural Network...

AI Technical Summary

Problems solved by technology

In the prior art, the modules that perform processing from input neurons to output neurons are usually designed with a single core, which is difficult to meet the performance requirements of neural network accelerators

[0005] In summary, there are obviously inconveniences and defects in the actual use of the existing technology, so it is necessary to improve

Images