Winograd Convolution Compute Circuit for Lower-Power AI Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing operation apparatuses for artificial intelligence tasks consume high power and take long operation times due to the large number of matrix multiplications required in convolution operations.
Innovation Solution
The operation apparatus performs winograd convolution operations by disassembling transformation operations into summation operations, replacing numerous multiplication operations with addition operations to accelerate processing and reduce resource consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional convolution operations are performed using matrix multiplications, then the convolution can be completed, but power consumption increases and operation time extends
Solution Approach 1:
The patent changes the mathematical parameters of the convolution operation by transforming it into a winograd convolution, which replaces standard matrix multiplications with a different set of operations (disassembly into summation operations). This parameter change in the operation methodology reduces computational complexity and resource consumption while maintaining the convolution functionality.
Solution Approach 2:
The patent substitutes the traditional mechanical system of matrix multiplications with a different operational mechanism - the winograd transformation that breaks down the convolution into summation operations. This substitution of the computational mechanism reduces the number of operations required and thereby lowers power consumption and execution time.
2Productivity
If traditional convolution operations are performed using matrix multiplications, then the convolution can be completed, but operation time increases
Solution Approach 1:
The patent changes the operational parameters by implementing winograd convolution instead of standard convolution. This parameter change transforms the computation into a more efficient form that requires fewer operations, directly reducing the time needed to complete the convolution while maintaining accuracy.
Solution Approach 2:
The patent replaces the time-consuming matrix multiplication mechanism with a more efficient summation-based mechanism through the winograd transformation. This substitution of the computational mechanism reduces the total number of operations and thereby decreases operation time.
3Reliability
If more computation resources are allocated to convolution operations, then processing accuracy can be maintained, but power consumption and resource usage increase
Solution Approach 1:
The patent changes the computational parameters by using winograd convolution, which maintains processing accuracy while reducing the quantity of operations required. The transformation preserves the mathematical equivalence of the convolution but with fewer and more efficient operations, thereby maintaining reliability while reducing resource consumption.
Solution Approach 2:
The patent substitutes the resource-intensive matrix multiplication system with a more efficient summation-based system through winograd transformation. This substitution maintains the computational accuracy while significantly reducing the quantity of operation resources needed, as the transformed operations require fewer computational steps.
Data Source
AI summary
The present disclosure includes an operation apparatus configured to perform a winograd convolution operation. A control circuit of the operation apparatus is configured to send a control instruction to instruct a compute circuit to perform the winograd convolution operation. The computer circuit is configured to extract data from the storage circuit for the winograd convolution operation in response to the control instruction and disassembles a transformation operation into multiple summation operations.