Neural Data Transfer Compression for Tile-Based Feature Map Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data processing systems for neural networks face challenges in efficiently transferring large amounts of feature map data between local and external storage, leading to high power consumption due to numerous memory accesses.
Innovation Solution
A method is introduced that compresses data by reading it in units along a first dimension, filling partial units, and compressing on a tile-by-tile basis before transferring it to external storage, and subsequently decompressing and writing it back to local storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If data is transferred between local storage and external storage without compression, then data transfer is simple and fast, but power consumption increases due to numerous memory accesses
Solution Approach 1:
The patent divides the data transfer process into segments: compression phase (where data is compressed and stored in external storage) and decompression phase (where data is decompressed and transferred to local storage). This segmentation allows the system to optimize each phase independently, reducing overall power consumption despite added complexity.
Solution Approach 2:
The patent introduces compression and decompression operations as intermediary processes between local storage and external storage. These intermediaries transform the data format to reduce the volume of data transferred, thereby reducing power consumption during memory accesses.
2Use of energy by moving object
If data is compressed and decompressed during transfer, then power consumption decreases, but processing time increases
Solution Approach 1:
The patent performs compression as a preliminary action before data transfer to external storage. By compressing the data beforehand, the system reduces the amount of data that needs to be transferred and processed later, offsetting the initial compression time with reduced transfer and decompression time.
3Ease of operation
If data is transferred in uncompressed form, then processing is simpler, but the number of memory accesses increases leading to higher power consumption
Solution Approach 1:
The patent changes the parameter of data representation by applying compression algorithms. This parameter change transforms the data from an uncompressed state to a compressed state, reducing the volume of data transferred between storage units and thereby reducing power consumption during memory accesses.
Data Source
AI summary
A method of compressing data for transfer between a local storage of a processor and an external storage. The data is formed in an array of three or more dimensions and the method comprises sequentially reading data stored in the local storage to a compressor in units of data. Each unit has a predetermined unit size corresponding to an integer number of a tile size. At an extremity of the array, a partial unit of data is read in a case that the array size is not an integer multiple of the unit size. The partial unit of data is filled at the compressor and the filled data is compressed on a tile-by-tile basis to form compressed data. The compressed data associated with the unit of data is transferred to the external storage.