Multi-Compression Data Storage for Memory Bandwidth Bottlenecks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Memory bandwidth limitations in computer systems hinder the acceleration of iterative linear solvers for solving linear systems to a high degree of accuracy, particularly in large sparse matrix calculations, where existing compression methods may introduce precision errors and increase computation time.
Innovation Solution
Storing multiple copies of data with different compression ratios and schemes, allowing selection based on estimated error and precision requirements, and using a stream processor to perform computations with decompressed data for accurate results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data compression is applied to reduce memory bandwidth usage, then memory bandwidth consumption is reduced, but precision errors increase and computation time increases
Solution Approach 1:
The patent segments the data storage into multiple copies with different compression ratios (first copy with higher compression, second copy with lower compression). This allows the system to selectively access appropriate precision levels based on computational needs, resolving the contradiction between memory bandwidth efficiency and calculation precision.
Solution Approach 2:
The patent changes the compression ratio parameter of data copies to create a spectrum of precision options. By storing data at multiple compression levels, the system can adjust the effective precision parameter dynamically, balancing memory bandwidth consumption against calculation accuracy requirements.
2Quantity of substance
If high compression ratio is used to reduce memory bandwidth usage, then memory bandwidth consumption is reduced, but computation time increases due to decompression overhead
Solution Approach 1:
The patent performs preliminary compression of data at multiple different compression ratios before computation. This preliminary action creates ready-to-use data copies at various precision levels, eliminating the need for real-time decompression during computation and reducing overall computation time while maintaining memory bandwidth efficiency.
Solution Approach 2:
The patent creates multiple copies of the data set, each with different compression ratios. Instead of decompressing highly compressed data during computation, the system can directly use appropriately compressed copies, eliminating decompression overhead and reducing computation time while still achieving memory bandwidth reduction.
3Productivity
If multiple copies of data are stored with different compression ratios, then memory bandwidth limitations are mitigated, but device complexity increases
Solution Approach 1:
The patent creates a universal data storage structure where multiple compressed copies serve multiple functions: highly compressed copies for bandwidth-critical operations and less compressed copies for precision-critical operations. This multi-functional approach improves computational efficiency without requiring entirely separate systems for different computational needs.
Data Source
AI summary
There is provided a method of, and apparatus for, processing a computation on a computing device comprising at least one processor and a memory, the method comprising: storing, in said memory, plural copies of a set of data, each copy of said set of data having a different compression ratio and/or compression scheme; selecting a copy of said set of data; and performing, on a processor, a computation using said selected copy of said set of data. By providing such a method, different compression ratios and/or compression schemes can be selected as appropriate. For example, if high precision is required in a computation, a copy of the set of data can be chosen which has a low compression ratio at the expense of processing time and memory transfer time. In the alternative, if low precision is acceptable, then the speed benefits of a high compression ratio and/or lossy compression scheme may be utilised.