Multi-Accelerator Arithmetic Processing With Adaptive Data Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing arithmetic processing devices with multiple accelerators face delays in data transfer due to uneven busy rates and cache inefficiencies, leading to overall execution time delays.
Innovation Solution
Implement a system that monitors execution statuses across accelerators and compresses data with defined accuracy to mitigate delays, allowing for timely data transfer and execution, with optional up-conversion to original accuracy when necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If data is transferred uncompressed from main storage to accelerators, then data transfer speed is improved, but data transfer volume increases causing memory bandwidth consumption and transfer time to worsen
Solution Approach 1:
The patent changes the parameter of data representation by applying compression algorithms to reduce the volume of data transferred between main storage and accelerators. This transforms the data from uncompressed format to compressed format, achieving faster transfer speeds while reducing the quantity of data moved across the memory interface.
Solution Approach 2:
The patent introduces a compression intermediary layer between the main storage device and accelerators. This intermediary component compresses data before transfer and can decompress it at the destination, acting as a mediator that reduces the burden on the memory interface while maintaining data accessibility for computation.
2Productivity
If data is compressed before transfer, then data transfer volume is reduced improving efficiency, but data accuracy may deteriorate
Solution Approach 1:
The patent dynamically changes the compression parameter (accuracy level) based on the specific computation requirements. Different compression levels are applied to different data types or computational tasks, allowing the system to optimize between transfer efficiency and accuracy preservation by adjusting the compression parameter according to needs.
Solution Approach 2:
The patent applies different compression accuracy levels to different parts of the data or different computational workloads. Critical data that requires high precision uses lower compression (higher accuracy), while less critical data uses higher compression (lower accuracy), thereby achieving local optimization of the balance between transfer efficiency and data accuracy.
3Loss of time
If compression processing is applied to all data, then overall data transfer time is reduced, but processing overhead increases
Solution Approach 1:
The patent applies compression processing selectively rather than universally - only to data that benefits from compression or where compression is feasible. This partial action approach avoids the excessive processing overhead that would result from compressing all data, while still achieving time savings on the subset of data where compression is effective.
Solution Approach 2:
The patent implements self-service mechanisms where the system automatically determines which data requires compression based on metadata, data characteristics, or computational requirements, eliminating the need for manual intervention or complex external control systems to manage the compression process.
Data Source
AI summary
An arithmetic processing device includes: a memory; and a processor coupled to the memory and configured to: monitor execution statuses for operations in a plurality of accelerators that execute the operations on data stored in a main storage device; and in a case where a delay in data transfer is detected in at least one accelerator among the plurality of accelerators in a monitoring result for the execution statuses, transfer the data compressed with predetermined accuracy to the at least one accelerator from the main storage device.