Streaming AI Convolution Processing via Data Matrix Padding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial intelligence algorithms face bottlenecks due to insufficient pipeline implementation and parallelism, hindering the widespread adoption of AI technologies, particularly in processing large datasets that require extensive computational resources.
Innovation Solution
A streaming-based AI convolution processing method and apparatus that modifies data matrices by adding invalid data to create a multiple of the degree of parallelism, allowing for efficient data transmission and convolution operations through multiple streaming lakes and modules, enhancing parallel processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional processors (CPU/GPU) are used for AI convolution processing, then the system can handle basic computational tasks, but the processing speed and efficiency are insufficient due to limited pipeline implementation and parallelism
Solution Approach 1:
The patent segments the data matrix into multiple sub-matrices and processes them through parallel convolution units. The input data is divided into chunks that can be simultaneously processed by multiple processing elements, increasing the degree of parallelism and overall processing throughput.
Solution Approach 2:
The patent introduces a streaming lake architecture that adds a temporal dimension to data processing. Data flows continuously through the system in a stream, allowing for pipeline parallelism where different stages of convolution processing occur simultaneously at different time steps, thereby improving productivity without proportionally increasing device complexity.
2Quantity of substance
If the data matrix size is increased to process more information, then the computational accuracy and completeness improve, but the processing time and resource requirements increase significantly
Solution Approach 1:
The patent performs preliminary actions by pre-processing and organizing data into optimized matrix formats before convolution operations. Data is pre-arranged in the streaming lake to facilitate efficient parallel access and processing, reducing the time required for actual convolution computations.
Solution Approach 2:
The streaming lake architecture enables continuous data flow and processing without idle cycles. Data is continuously fed into parallel convolution units, and results are continuously generated and output, maximizing the utilization of processing resources and reducing overall processing time for large data volumes.
3Productivity
If the degree of parallelism is increased to improve processing efficiency, then the calculation speed improves, but the system complexity and data transmission requirements increase
Solution Approach 1:
The patent designs a universal streaming lake architecture that can accommodate multiple data types and convolution operations through a single unified interface. The same data transmission and processing infrastructure handles various parallel operations, reducing the need for separate specialized components and thereby limiting the increase in system complexity despite high parallelism.
4Quantity of substance
If more processors are deployed to handle large datasets, then the processing capability increases, but the cost and resource consumption increase proportionally
Solution Approach 1:
The patent merges multiple processing operations into a unified streaming processing pipeline. Data flows through a single optimized path that combines multiple convolution operations, reducing redundant data movement and computation. This consolidation achieves high computational capacity while minimizing energy consumption compared to deploying separate processor systems.
Data Source
AI summary
Provided is a streaming-based artificial intelligence convolution processing method, applied to a processing module. The method includes: adding invalid data to a starting point of a first to-be-processed data matrix stored in a first streaming lake to form a second to-be-processed data matrix, where a number of columns of the second to-be-processed data matrix is an integral multiple of a degree of parallelism of data transmission; using a data transmission module to take out the second to-be-processed data matrix from the first streaming lake in a preset manner for a convolution operation. Also provided are a streaming-based artificial intelligence convolution processing apparatus, a readable storage medium and a terminal.