Neural network accelerator based on multichannel systolic array

A neural network accelerator based on a multi-channel systolic array comprises a novel systolic array composed of basic operation units with the configurable number, and the systolic array comprises weight data flow channels and feature data flow channels with the configurable number and direction. The multi-channel systolic array can carry out arbitration distribution on data flow channels and basic arithmetic units according to the scale of neural networks, and parallel four-thread acceleration of four neural networks can be achieved at most. The multi-channel systolic array sends a row of feature data in each calculation period, and a data stream formed by discrete data is improved by adopting a data stream formed by the whole row of data, so that the acceleration processing time is greatly reduced. Neural network acceleration operation is realized based on the multi-channel systolic array. Compared with a traditional systolic array, the multi-channel systolic array has the advantages that the acceleration