Compressed LUT Mapping for AI Bandwidth and Storage Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited transmission bandwidth between a host and a device in AI applications necessitates reducing data transmission amounts, particularly in scenarios where large look-up tables are used, requiring efficient compression methods to minimize storage space requirements.
Innovation Solution
A look-up table compression method that involves retrieving original data, applying a reconstruction condition to obtain a check result, converting the data into reconstructed data, and writing it to a compressed table with a corresponding mapping table, allowing for reduced storage space and efficient data retrieval.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If a large look-up table is stored in the device to reduce data transmission amount, then the data transmission bandwidth requirement is reduced, but the storage space requirement increases
Solution Approach 1:
The original look-up table is segmented into multiple compressed tables based on different compression conditions. Each compressed table stores data in a compressed format, and the original data can be reconstructed by combining data from multiple compressed tables according to the mapping relationship stored in the mapping table. This segmentation allows the device to store compressed tables with smaller total storage space while still being able to retrieve the original data when needed.
Solution Approach 2:
The patent applies parameter changes by transforming the storage format of the look-up table from uncompressed to compressed format. Different compression algorithms and compression ratios are used based on the characteristics of the data, allowing the system to adjust the compression level and achieve optimal balance between storage space and data transmission reduction. The mapping table stores the parameters needed to reconstruct the original data from the compressed format.
2Volume of stationary object
If data is compressed into a smaller format to reduce storage space, then the storage space requirement is reduced, but the complexity of data retrieval increases
Solution Approach 1:
The mapping table is pre-generated and stored, containing all the information needed to reconstruct the original data from the compressed tables. This preliminary action includes pre-calculating and storing the mapping relationships, compression conditions, and reconstruction parameters. When data retrieval is needed, the system simply needs to query the mapping table and follow the pre-established reconstruction process, significantly reducing the complexity of real-time data retrieval operations.
3Ease of operation
If the original look-up table is used without compression, then the data retrieval process is simple, but the storage space and data transmission amount increase
Solution Approach 1:
The mapping table serves as an intermediary between the compressed tables and the original look-up table. It contains the mapping relationships and reconstruction information that bridge the compressed storage format and the original data format. This intermediary structure allows the system to maintain simple data retrieval operations by providing a clear lookup path through the mapping table, while simultaneously achieving significant reduction in storage space and data transmission amounts through the compressed tables.
Data Source
AI summary
The disclosure provides a look-up table (LUT) compression method and a LUT reading method for computation equipment and its host and device. In a LUT compression phase, the host retrieves an original data from an original LUT by using an original table address, checks the original data according to a reconstruction condition to obtain a check result (bitmap), converts the original data into a reconstructed data according to the check result, writes the reconstructed data to a compressed LUT by using a compressed table address, writes a relationship among the original table address, the compressed table address, and the check result (bitmap) to a mapping table, and stores the compressed LUT to the device. In a LUT reading phase, the host retrieves a compressed table address and the bitmap from the mapping table by using an original table address, and the device retrieves a reconstructed data from the compressed LUT by using the compressed table address, and converts the reconstructed data into an original data according to the bitmap and a reconstruction condition.