Programmable Activation Function Execution Unit for Neural Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural processing units (NPUs) face challenges in efficiently processing various activation functions due to high power consumption, increased chip area, and the inability to handle new or modified activation functions without hardware modifications, leading to inference accuracy issues and increased computational complexity.
Innovation Solution
A programmable activation function execution unit (PAFE) is integrated into the NPU, allowing for the programming of non-linear activation functions using segment data that considers the characteristics of the activation function, approximation errors, and hardware options, enabling flexible and efficient processing of different activation functions without hardware changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural processing unit is designed to process complex mathematical activation functions, then inference accuracy is improved, but chip area and power consumption increase significantly
Solution Approach 1:
The patent applies parameter changes by transforming complex activation functions into piecewise linear approximations with adjustable parameters. Different regions of the activation function are represented by different linear equations with specific parameters, allowing the system to maintain high inference accuracy while using simpler hardware operations that reduce chip area requirements
Solution Approach 2:
The patent segments the activation function into multiple regions, where each region is processed by a dedicated linear approximation module. This segmentation allows the complex non-linear function to be broken down into simpler linear segments that can be processed efficiently with reduced hardware complexity and smaller chip area
2Measurement precision
If a neural processing unit is designed to process complex mathematical activation functions, then inference accuracy is improved, but power consumption increases significantly
Solution Approach 1:
By changing the functional form from complex non-linear operations to parameter-based linear approximations, the patent reduces computational complexity and power consumption while maintaining accuracy through carefully selected parameters that best represent the original activation function in each region
Solution Approach 2:
Segmenting the activation function allows the processor to use simpler linear operations in each segment rather than computing the full complex function, thereby reducing the energy required for each activation operation while maintaining overall accuracy through the combination of segments
3Adaptability or versatility
If the NPU hardware is modified to handle new activation functions, then adaptability is improved, but device complexity and manufacturing cost increase
Solution Approach 1:
The patent creates a universal activation function processing unit that can handle multiple types of activation functions through parameter configuration rather than hardware modification. The same hardware structure can process different activation functions by loading different parameter sets, making the system multi-functional without increasing hardware complexity
Solution Approach 2:
The system achieves adaptability to new activation functions through parameter changes rather than hardware changes. By modifying the parameters that define the piecewise linear approximation, the same hardware can adapt to process different activation functions, maintaining low device complexity while improving versatility
Data Source
AI summary
A method of programming an activation function is provided. The method includes generating a segment data for segmenting the activation function; segmenting the activation function into a plurality of segments using the segment data; and approximating at least one segment of the plurality of segments to a programmable segment.