Short-Bit Fixed-Point Neural Network Operations for Lower Power
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 in applications like neural network algorithms, where high accuracy is not always required.
Innovation Solution
Implementing a fixed-point number converter to convert floating-point numbers to fixed-point numbers with specific formats based on data type and layer distribution patterns, using a neural network processor to process these fixed-point numbers for neural network operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point numbers with high accuracy are used in neural network algorithms, then computation 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. This parameter change allows the system to maintain sufficient computation precision for neural network operations while significantly reducing power consumption, as fixed-point arithmetic requires fewer computational resources and simpler hardware implementation.
Solution Approach 2:
The patent applies different precision requirements to different parts of the neural network computation process. By analyzing the actual precision needs of various neural network operations, the system uses fixed-point representation with appropriate bit lengths tailored to specific computational stages, avoiding the uniform high precision of floating-point numbers throughout the entire computation chain.
2Measurement precision
If floating-point number computations are implemented, then computation accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent substitutes the mechanical floating-point computation system with a fixed-point computation system. This replacement eliminates the complex exponentiation and normalization operations required by floating-point arithmetic, replacing them with simpler fixed-point arithmetic operations that are easier to implement in hardware and execute more efficiently.
3Adaptability or versatility
If general-purpose processors are used for neural network computations, then versatility is improved, but power consumption and computational efficiency worsen
Solution Approach 1:
The patent segments the processing function by separating the floating-point to fixed-point conversion operation from the general-purpose processor. This segmentation allows the main processor to handle high-level algorithmic logic while a specialized conversion module handles the numerical representation transformation, enabling the system to achieve both versatility and energy efficiency.
Solution Approach 2:
The patent introduces fixed-point numbers as an intermediary representation between floating-point input data and the neural network computation process. This intermediary format serves as a bridge that allows the system to accept floating-point data while performing computations in a more efficient fixed-point format, reducing the energy burden on the general-purpose processor.
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.