Neural network accelerator and data processing method

Abstract

The invention relates to a neural network accelerator. The neural network accelerator comprises a storage unit for storing neuron data and weight data of an LSTM network and outputting the data, a vector multiplying and accumulating matrix unit used for receiving the data from the storage unit, performing vector multiplying and accumulating operations for the received data and outputting operationresults, an addition unit used for receiving the data from the vector multiplying and accumulating matrix unit and performing an offset addition operation for the received data, an activation unit used for receiving the data from the multifunctional operation unit and / or the storage unit, performing an activation operation and outputting an activation result for the received data, and a vector parallel multiplying and accumulating unit used for receiving the data from the activation unit and / or the storage unit and performing multiplying and accumulating operations on the received data. All the modules are connected end to end to form an assembly line working mechanism to carry out data processing together with input vectors by taking weighted row vectors as a unit.

Description

technical field [0001] The invention relates to the computing field, in particular to an 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 In the iterative operation process of each layer, in the prior art, when performing calculatio...

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 In the iterative operation process of each layer, in the prior art, when performing calculations on LSTM networks, resource usage is often unbalanced. The idle state; and when calculating the amount at the scale of the vector dimension, its computing resource consumption is too large; this leads to a reduction in the utilization of hardware resources

Images