Real-Number Full-Connection Units for Neural Network Accelerator Reuse
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing number of units in neural network accelerators leads to hardware resource overconsumption, hindering efficient data processing, particularly for voice data.
Innovation Solution
Implementing real-number full-connection operations on a neural network accelerator to reuse hardware units efficiently, allowing multiple operations with minimal hardware logic by converting complex operations into real-number operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If more units are configured in the neural network accelerator, then the processing capability is improved, but the hardware resource consumption increases and chip area is occupied
Solution Approach 1:
The patent makes the full-connection unit capable of performing multiple operations including complex convolution, complex full-connection, and both real-number operations by adding a conversion unit. This allows one hardware unit to replace multiple dedicated units, reducing overall hardware resource consumption while maintaining processing capability for various neural network operations.
2Productivity
If more units are configured in the neural network accelerator, then the processing capability is improved, but the chip area is occupied
Solution Approach 1:
By enabling the full-connection unit to perform convolution operations through conversion, the patent reduces the need for separate dedicated convolution units, thereby reducing chip area occupation while maintaining comprehensive processing capability for different neural network operations.
3Reliability
If complex operations are processed using dedicated units, then the operation accuracy is maintained, but the hardware resource consumption increases
Solution Approach 1:
The patent changes the numerical parameter type by converting complex numbers to real numbers for processing in the full-connection unit. The conversion unit transforms complex convolution operations and complex full-connection operations into real-number operations, allowing the full-connection unit to handle diverse operations accurately while reducing hardware resource requirements.
Data Source
AI summary
A data processing method for a neural network accelerator, an electronic device and a storage medium are provided. The technical solution includes: obtaining data to be processed and an operation to be executed; obtaining a real-number full-connection operation corresponding to the operation to be executed; and performing the real-number full-connection operation on the data based on a real-number full-connection unit of the neural network accelerator to obtain a result of the operation to be executed for the data.