Search International and National Patent Collections

1. (WO2017185412) NEURAL NETWORK OPERATION DEVICE AND METHOD SUPPORTING FEW-BIT FIXED-POINT NUMBER

Pub. No.:    WO/2017/185412    International Application No.:    PCT/CN2016/081839
Publication Date: Fri Nov 03 00:59:59 CET 2017 International Filing Date: Fri May 13 01:59:59 CEST 2016
IPC: G06F 9/44
G06N 3/02
Applicants: CAMBRICON TECHNOLOGIES CORPORATION LIMITED
北京中科寒武纪科技有限公司
Inventors: CHEN, Yunji
陈云霁
LIU, Shaoli
刘少礼
GUO, Qi
郭崎
CHEN, Tianshi
陈天石
Title: NEURAL NETWORK OPERATION DEVICE AND METHOD SUPPORTING FEW-BIT FIXED-POINT NUMBER
Abstract:
A device and method for performing a forward operation of an artificial neural network. The device comprises: a floating-point data statistics module used for performing statistical analysis on various types of required data, to obtain decimal point locations of fixed-point data; a data conversion unit (32) used for implementing conversion from a long-bit floating-point data type to a short-bit fixed-point data type according to the decimal point locations of fixed-point data; a fixed-point data operation module used for performing a forward operation of an artificial neural network on short-bit fixed-point data. By representing data in a forward operation of a multilayer artificial neural network with a short-bit fixed point, and using a corresponding fixed-point operation module, the device implements the forward operation of the short-bit fixed point of the artificial neural network, thereby greatly improving the performance per watt of hardware.