Device and method for neural network operation supporting floating point number with few digits

Abstract

The invention provides a device and a method for executing artificial neural network forward operation. The device comprises a floating point data statistical module, a floating point data conversion unit and a floating point data operation module, wherein the floating point data statistical module is used for performing statistical analysis on all types of needed data to obtain index digit offset and length EL of index digits; the floating point data conversion unit is used for realizing conversion from a long-digit floating point data type to a short-digit floating point data type according to the index digit offset and the length EL of the index digits; and the floating point data operation module is used for performing artificial neural network forward operation on short-digit floating point data. According to the device, the data in multi-level artificial neural network forward operation is expressed with short-digit floating points, and the corresponding floating point operation module is used, so that forward operation of the short-digit floating points of an artificial neural network is realized, and accordingly the performance-to-consumption ratio of hardware is greatly improved.

Description

technical field [0001] The invention relates to the technical field of artificial neural networks, and more particularly to a device and method for supporting neural network operations of floating-point numbers with fewer digits. Background technique [0002] Multi-layer artificial neural networks are widely used in the fields of pattern recognition, image processing, function approximation, and optimization calculations. In recent years, multi-layer artificial neural networks have been favored by academic circles and industry is getting more and more attention. [0003] Traditional artificial neural network forward operations usually use 32-bit floating-point data types for data storage, but in most forward operations of artificial neural networks, the same type of data will be concentrated within a certain data range. However, the range of data that can be represented by 32-bit floating point is much larger than the data range of the same type of data in the forward opera...

AI Technical Summary

Problems solved by technology

[0003] Traditional artificial neural network forward operations usually use 32-bit floating-point data types for data storage, but in most forward operations of artificial neural networks, the same type of data will be concentrated within a certain data range. However, the data range that can be represented by 32-bit floating point is much larger than the data range of the same type of data in the forward operation. There is a lot of redundancy in the use of 32-bit floating point data type, which increases the area overhead of hardware.

Images