Hybrid precision arithmetic unit for FP16 floating point data and INT8 integer data operation

Abstract

The invention discloses a hybrid precision arithmetic unit for FP16 floating point data and INT8 integer data operation. The mixed precision arithmetic unit comprises a precision conversion module, an arithmetic unit, two input data and one, two or four output data. The input data and the output data are expressed in an EM16 format, and the EM16 format expression is 16 digits and comprises FP16 floating point expression, INT8 integer expression and INT8 double integer expression; the two input data are respectively feature data and parameter data in neural network calculation; the precision conversion module is used for executing precision conversion among expressions of the EM16 format of the feature data according to the external configuration information; and the arithmetic unit is used for executing additive operation or multiplication operation between two pieces of data which are both FP16 floating point data or both INT8 integer data according to external configuration information. According to the hybrid precision arithmetic unit, FP16 floating point data and INT8 integer data can be used in a mixed mode in the calculation process, so that hybrid precision calculation can have the precision of the FP16 and the speed of the INT8 in a neural network calculation task.

Description

technical field [0001] The invention relates to the field of neural network tensor processor architecture, in particular to a mixed-precision arithmetic unit used for FP16 floating-point data and INT8 integer data operations. Background technique [0002] Traditional neural network computing hardware uses fixed computing precision. For example, DSPs usually use 16-bit fixed-point calculations, GPUs usually use 32-bit floating-point or 16-bit floating-point calculations, and AI accelerators usually use 16-bit floating-point or 8-bit integer calculations. [0003] Neural network algorithms usually use 32-bit floating-point precision for model training, and the parameter data obtained from training usually use 32-bit floating-point precision. On the other hand, when using 32-bit floating point or 16-bit floating point precision for inference calculations, you can usually achieve comparable accuracy to training. However, when calculating with 16-bit or 8-bit integer precision, ...

AI Technical Summary

Problems solved by technology

However, when using 16-bit or 8-bit integer precision calculations, it often brings a certain loss of calculation accuracy

Especially under 8-bit integer precision, if no additional techniques such as quantization training are used to compensate, the loss of accuracy caused by 8-bit integer precision calculations will often be huge

Images