Neural Network Fixed-Point Conversion for Short Bit-Length Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional general-purpose processors are limited in computing floating-point numbers of a single format, leading to unnecessary power consumption and complex computations, particularly in applications like neural network algorithms that require varying levels of accuracy.
Innovation Solution
An apparatus and method utilizing a fixed-point number converter to convert floating-point numbers to fixed-point numbers of short bit length, processed by a neural network processor to generate results, optimizing computations and reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point numbers with high accuracy are used in neural network operations, then computational precision is improved, but power consumption increases unnecessarily
Solution Approach 1:
The patent changes the numerical representation parameter from floating-point to fixed-point format, and adjusts the bit length parameter to match the actual precision requirements of neural network operations. This parameter transformation maintains sufficient computational precision while significantly reducing power consumption by eliminating the overhead of floating-point representation and computation.
Solution Approach 2:
The patent applies partial action by using only the necessary precision level for neural network operations rather than full floating-point precision. By using fixed-point numbers with appropriate bit length, the system provides just enough precision needed for activation functions and neural network computations without the excessive precision of floating-point formats, thereby reducing power consumption.
2Measurement precision
If floating-point number computations are performed, then computational accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent transforms the computational parameter from floating-point arithmetic to fixed-point arithmetic, fundamentally simplifying the computational complexity. Fixed-point operations can be performed using simple integer arithmetic units without requiring complex floating-point instruction sets, thereby reducing device complexity while maintaining adequate accuracy for neural network applications.
3Device complexity
If a single floating-point format is used in general-purpose processors, then device simplicity is maintained, but adaptability to different accuracy requirements deteriorates
Solution Approach 1:
The patent introduces dynamic adaptability by allowing the fixed-point bit length to be adjusted according to different neural network operation requirements. The system can dynamically select appropriate bit lengths for different layers or operations within the neural network, providing versatility in accuracy levels while maintaining the simplicity of fixed-point processing architecture.
Data Source
AI summary
Aspects for neural network operations with fixed-point number of short bit length are described herein. The aspects may include a fixed-point number converter configured to convert one or more first floating-point numbers to one or more first fixed-point numbers in accordance with at least one format. Further, the aspects may include a neural network processor configured to process the first fixed-point numbers to generate one or more process results.