Neural network accelerator for facing multi-variant LSTM and data processing method thereof

Abstract

The invention relates to a neural network accelerator for facing a multi-variant LSTM. The neural network accelerator comprises the components of a storage unit which is used for storing neuron data and weight data of the LSTM and the variant network and outputting; a matrix multiplication unit which is used for receiving data from the storage unit, executing a vector multiplying and accumulatingoperation on the received data and outputting an operation result; a multifunctional operating unit which is used for receiving data from a vector multiplication operation unit, performs a specific operation which corresponds with the LSTM or the variant network for aiming at the received data and outputs an operation result; an activating unit which is used for receiving data from the multifunctional operating unit and the storage unit, performs an activation operation for aiming at the received data and outputs an activation result; and a vector parallel multiplying and adding unit which isused for receiving data from the activating unit and the storage unit and performing a multiplying operation and an adding operation for aiming at the received data.

Description

technical field [0001] The invention relates to the computing field, in particular to a multi-variant LSTM-oriented neural network accelerator and a data processing method. Background technique [0002] Neural network is one of the perception models with a high level of development in the field of artificial intelligence. Once it appeared, it became a research hotspot in academia and industry. With the deepening of research, different types of neural networks have been proposed one after another. For example, long-term and short-term Memory network (LSTM, Long Short-Term Memory). [0003] The LSTM network is a time-recursive cyclic neural network that can learn long-term dependent information. It is often used to learn language translation, robot control, image analysis, etc. The main part of the calculation process is the multiplication and accumulation operation of each gate value vector and The iterative operation process of each layer; in order to adapt to different app...

AI Technical Summary

Problems solved by technology

[0003] The LSTM network is a time-recursive cyclic neural network that can learn long-term dependent information. It is often used to learn language translation, robot control, image analysis, etc. The main part of the calculation process is the multiplication and accumulation operation of each gate value vector and The iterative operation process of each layer; in order to adapt to different application requirements, LSTM has many variants, such as threshold repeat unit (GRU) variants, etc., while the neural network accelerator in the prior art is usually only for one LSTM or LSTM variant monolithic design, which severely reduces computing resource utilization

Images