Accelerator Control Apparatus for Reducing Data Transfer Costs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing accelerator control systems experience decreased processing speed when data is divided and processed across multiple accelerators due to the need for frequent data transfer back to a control unit, resulting in high data transfer costs.
Innovation Solution
The system determines whether the width of boundary data items is equal to or larger than a reference width and synchronizes them with adjacent segment data items only when necessary, allowing accelerators to continuously process data without frequent synchronization, thereby reducing data transfer costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data is divided and processed by multiple accelerators with frequent synchronization, then processing accuracy is maintained, but processing speed decreases due to continuous data transfer back to control unit
Solution Approach 1:
The patent segments the data processing task by dividing data into multiple segments that can be processed independently by different accelerators. Each accelerator processes its assigned segment continuously without needing to return intermediate results to the control unit, thereby maintaining processing accuracy while improving overall processing speed.
Solution Approach 2:
The patent applies preliminary action by pre-dividing the data into segments and pre-assigning them to different accelerators before processing begins. This allows each accelerator to start processing immediately without waiting for synchronization points, eliminating the need for continuous data transfer back to the control unit during the processing phase.
2Productivity
If data is divided into multiple segments for parallel processing, then processing speed increases, but data transfer cost increases due to frequent synchronization with control unit
Solution Approach 1:
The patent extracts the synchronization requirement from the data processing flow by allowing accelerators to process data segments independently. The control unit only performs data division at the beginning and collects final results at the end, removing the need for repeated data transfer and synchronization operations during processing, thereby significantly reducing data transfer cost while maintaining high processing speed.
Solution Approach 2:
The patent ensures continuity of useful action by enabling each accelerator to process its assigned data segment continuously without interruption for synchronization. This continuous processing eliminates the stop-start pattern caused by frequent data transfers, maintaining high processing speed while minimizing energy loss associated with data transfer operations.
3Stability of the object's composition
If boundary data items are synchronized frequently, then data consistency is maintained, but processing efficiency decreases due to unnecessary synchronization operations
Solution Approach 1:
The patent applies partial action by synchronizing boundary data items only when absolutely necessary - specifically at the beginning of processing and at the end when results are collected. During the main processing phase, boundary data items are not synchronized, eliminating unnecessary synchronization operations while still maintaining data consistency through the initial and final synchronization points.
Data Source
AI summary
Provided is an accelerator control apparatus including: an accelerator that is configured to store at least one segment data item of a plurality of segment data items obtained by dividing data, and a boundary data item that is data item being included in a segment data item adjacent to the at least one segment data item; and data management unit to determine whether a width of data which is included in the boundary data item and is consistent with the segment data item adjacent to the at least one segment data item, is equal to or larger than a reference width representing a width of data referred to in processing executed by the accelerator.


