Narrow Bit Width Linear Algebra Computation via Binary Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current edge computing devices face challenges in efficiently performing linear algebra operations with narrow integer data representations due to high computational intensity and resource constraints.
Innovation Solution
The method employs binary segmentation to reduce the computation overhead by extending narrow bit width elements to a clustering bit width, allowing for efficient computation of linear algebra operations such as linear convolution and inner product using a digital circuit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If narrow bit width data representations are used to reduce memory footprint and energy consumption, then data size and energy demands are reduced, but computational intensity increases
Solution Approach 1:
The patent applies segmentation by dividing the narrow bit width operands into multiple segments and processing them through multiple processing elements in parallel. Each processing element handles a portion of the computation, breaking down the complex narrow-bit operation into simpler sub-operations that can be executed efficiently in parallel, thereby reducing overall computational intensity while maintaining energy efficiency
Solution Approach 2:
The patent transitions from processing narrow bit width data in a single dimension (sequential processing) to multi-dimensional parallel processing by distributing computations across multiple processing elements. This dimensional expansion allows simultaneous execution of multiple operations, reducing computational intensity without increasing energy consumption
2Quantity of substance
If narrow bit width elements are used for linear algebra operations, then memory footprint is reduced, but processing overhead increases
Solution Approach 1:
The patent merges multiple narrow bit width operations into a single unified processing framework. By combining several processing elements and their operations into an integrated system, the patent reduces redundant overhead operations and streamlines the processing pipeline, thereby decreasing processing overhead while maintaining reduced memory footprint
Solution Approach 2:
The patent performs preliminary actions by pre-processing and organizing narrow bit width data into optimal formats before main computation. This includes preliminary segmentation, alignment, and preparation of operands, which reduces the complexity and overhead of subsequent processing operations, thereby decreasing processing overhead while maintaining efficient memory usage
Data Source
AI summary
The present invention relates to a method for computing a linear algebra operation of two operands or arrays comprising one or more narrow bit width elements with a digital circuit. The method uses the principle of binary segmentation to reduce the computation overhead of linear algebra operations like linear convolution and inner product of operands such as vectors with narrow bit width components. The invention is also directed to a digital circuit configured to execute the method.