AI Convolution Acceleration via Zero-Value Skipping and Identical Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI chips are not energy-efficient and struggle to accelerate deep learning algorithms, particularly in the convolution part, due to high computational and storage complexity, which leads to significant power consumption and performance bottlenecks.
Innovation Solution
A data processing method and device that optimizes convolution calculations by reading M*N feature map data and weights, skipping calculations where values are zero, and selecting one value for multiple identical feature map data, thereby reducing power consumption and improving performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithms use multi-layer large-scale neural networks with matrix product and convolution operations, then calculation accuracy is improved, but computational complexity and power consumption increase significantly
Solution Approach 1:
The patent extracts and processes only the essential data elements for convolution operations. By reading only M*N feature map data of all input channels and weights of a preset number of output channels, the system extracts necessary computational elements while discarding redundant data, thereby reducing power consumption while maintaining calculation accuracy.
Solution Approach 2:
The patent implements partial action by selectively performing convolution calculations only when necessary. The system skips calculations when feature map data or weights are zero, and selects one from multiple identical values for calculation, performing partial computations rather than exhaustive calculations, thus reducing energy consumption while preserving essential computational accuracy.
2Reliability
If convolution calculations are performed for all feature map data and weights, then calculation completeness is improved, but calculation time and productivity are reduced
Solution Approach 1:
The patent applies skipping by omitting unnecessary convolution calculations. When feature map data or weights are zero, or when multiple feature map data have the same values, the system skips redundant calculations and directly produces results, thereby maintaining calculation completeness for essential data while significantly improving calculation speed through selective omission of redundant operations.
Solution Approach 2:
The patent performs preliminary analysis of feature map data and weights before executing convolution calculations. By pre-identifying zero values and identical values, the system prepares optimization strategies in advance, enabling it to skip redundant calculations and improve overall processing efficiency while ensuring complete processing of necessary data.
3Adaptability or versatility
If M*N feature map data of all input channels and weights are read into memory, then data availability for convolution is improved, but storage complexity and power consumption increase
Solution Approach 1:
The patent extracts only the necessary data elements required for convolution operations. Instead of reading all possible feature map data and weights into memory, the system reads only M*N feature map data of all input channels and weights of a preset number of output channels, extracting essential computational elements while reducing memory storage requirements and associated power consumption.
Data Source
AI summary
A data processing method and device, a storage medium and an electronic device are disclosed. The method includes: reading M*N feature map data of all input channels and weights of a preset number of output channels, here a value of M*N and a value of the preset number are respectively determined by preset Y*Y weights; inputting the read feature map data and the weights of the preset number of output channels into a multiply-add array of the preset number of output channels for a convolution calculation; here a mode of the convolution calculation includes: not performing the convolution calculation in a case that the feature map data or the weights of the output channels are zero, and selecting one from same values for the convolution calculation in a case that there are a plurality of feature map data with the same values; and outputting a result of the convolution calculation.