Message-based processing with assignment of neural network layers to processor clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing message-based multi-processor systems for deep neural networks require substantial memory for look-up tables to determine destination processor clusters, which is inefficient and limits scalability.
Innovation Solution
A message-based multi-processor system with a logic module and message generator that uses control data to determine destination processor clusters, reducing memory requirements by storing only modest amounts of control data and enabling efficient message transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a look-up table is used to determine destination processor clusters, then message routing is simplified, but memory requirements increase substantially
Solution Approach 1:
The patent extracts only the essential routing information needed for message transmission and stores it in a compact format. Instead of using a complete look-up table that maps every source-destination pair, the system extracts and stores only the necessary control data (such as destination cluster identifiers and routing parameters) that can be used to dynamically determine message routes, thereby reducing memory requirements while maintaining routing functionality
Solution Approach 2:
The patent changes the representation parameters of routing information from a comprehensive look-up table format to a compact control data format. By parameterizing the routing information in terms of essential attributes (destination identifiers, cluster mappings) rather than exhaustive route tables, the system achieves significant memory reduction while preserving the ability to determine destination processor clusters efficiently
2Adaptability or versatility
If more processor clusters are assigned to neural network layers, then network scalability improves, but memory usage for routing increases
Solution Approach 1:
The patent segments the routing information storage by processor cluster, with each cluster maintaining its own compact control data structure. This segmentation allows the system to scale to more processor clusters without proportionally increasing total memory usage, as each cluster only stores routing information relevant to its local context rather than maintaining a global look-up table
Solution Approach 2:
The patent introduces a hierarchical dimension to routing information storage, organizing control data by processor cluster layers rather than by individual source-destination pairs. This dimensional reorganization allows the system to accommodate more processor clusters in a scalable manner, where memory requirements grow linearly with the number of clusters rather than exponentially
Data Source
AI summary
Systems and methods described herein relate to a multi-processor system for processing neural networks. The multi-processor system includes multiple processor clusters, each comprising processor cluster elements, with neural network layers assigned to one or more processor clusters. In response to an activation signal associated with a processor cluster element of a source processor cluster, the multi-processor system performs computations using control data for a set of destination processor clusters. The control data includes an offset computed using coordinates associated with the source and destination processor clusters. Based on these computations, the multi-processor system can selectively identify target destination processor clusters from the set of destination processor clusters and transmit output messages to them.