KNN Accelerator Partial Distance Computation for Energy Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing nearest neighbor search methods, such as k-dimensional trees and Voronoi diagrams, are inefficient and power-intensive for large data sets, especially in high-dimensional spaces, due to their exhaustive examination and computational complexity.
Innovation Solution
A k-nearest neighbor (kNN) accelerator that employs partial distance computation circuits and a minimum sort network to iteratively refine distance calculations, using lower accuracy initial computations to eliminate candidates and higher accuracy later on, reducing energy consumption and improving efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If exhaustive examination methods (k-dimensional trees, Voronoi diagrams) are used for nearest neighbor search, then search completeness is improved, but computational complexity and power consumption increase significantly
Solution Approach 1:
The patent applies partial distance computation by calculating only the necessary bits of distance (e.g., top 20-30 bits out of 53 bits for double precision) rather than computing the complete distance. This partial action approach eliminates the need for exhaustive examination of all data points while maintaining sufficient accuracy for nearest neighbor identification, thereby significantly reducing power consumption and computational complexity.
Solution Approach 2:
The patent changes the precision parameter of distance computation from full precision (53 bits) to reduced precision (20-30 bits). This parameter change allows the system to achieve acceptable search results with much lower computational cost, resolving the contradiction between search completeness and power consumption by finding an optimal balance point in precision requirements.
2Reliability
If exhaustive examination methods are used for nearest neighbor search, then search completeness is improved, but computational complexity increases
Solution Approach 1:
The patent implements partial distance computation that processes only the essential bits needed for nearest neighbor identification. By computing partial distances (top 20-30 bits) instead of full distances, the system reduces computational complexity from examining all data points exhaustively to a more efficient partial examination, while maintaining search reliability.
Solution Approach 2:
The patent extracts only the necessary computational components for nearest neighbor search, separating the essential distance computation from unnecessary computational overhead. By taking out and computing only the critical partial distance values rather than all possible distance metrics, the system reduces computational complexity while preserving search completeness.
3Measurement precision
If higher accuracy distance computation is performed for all candidates, then measurement precision is improved, but energy consumption and computation time increase
Solution Approach 1:
The patent applies partial distance computation by calculating only the top 20-30 bits of precision needed for nearest neighbor identification, rather than computing full 53-bit double precision distances for all candidates. This partial action approach provides sufficient measurement precision for accurate nearest neighbor selection while dramatically reducing energy consumption and computation time.
Solution Approach 2:
The patent performs preliminary partial distance computation to eliminate obviously non-nearest candidates before performing more accurate computations on remaining candidates. This preliminary action filters out the majority of data points that cannot be nearest neighbors, allowing higher precision computation to be applied only where necessary, thus optimizing the balance between measurement precision and energy consumption.
Data Source
AI summary
Systems, apparatuses, and methods for k-nearest neighbor (KNN) searches are described. In particular, embodiments of a KNN accelerator and its uses are described. In some embodiments, the KNN accelerator includes a plurality of vector partial distance computation circuits each to calculate a partial sum, a minimum sort network to sort partial sums from the plurality of vector partial distance computation circuits to find k nearest neighbor matches and a global control circuit to control aspects of operations of the plurality of vector partial distance computation circuits.