Distributed Neural Network Convolution for Low-Latency Matrix Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing matrix processing approaches are inefficient for complex operations like matrix multiplication and convolution, particularly in artificial intelligence and machine learning, due to processor and memory latency, and inability to utilize all available resources effectively.

Innovation Solution

A distributed matrix processing architecture that optimizes matrix operations across multiple processing resources with cyclical or hierarchical configurations, enabling 100% processing efficiency and scalable parallel execution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If matrix operations are performed using traditional processor architectures, then processing capability is provided, but processor latency and memory latency limit performance

Engineering Contradiction:
Improvematrix operation throughputVSAvoidprocessor and memory latency
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent divides the matrix processing system into multiple independent processing resources (tiles) arranged in a grid configuration. Each processing resource handles a specific portion of the matrix operation, allowing parallel execution and eliminating the bottleneck of single-processor sequential processing. This segmentation enables the system to achieve higher throughput by distributing the computational workload across multiple units.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from traditional single-processor or linear multi-processor architectures to a two-dimensional grid arrangement of processing resources. This dimensional change enables more efficient data flow patterns and allows simultaneous access to multiple processing units, thereby reducing latency and improving overall matrix operation performance.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If more processing resources are added to handle complex matrix operations, then processing power increases, but resource utilization becomes inefficient with wasted or idle processing time

Engineering Contradiction:
Improveprocessing powerVSAvoididle processing time
Core Design Contradiction:
ProductivityVSLoss of energy

Solution Approach 1:

The patent implements a continuous data flow architecture where input data is continuously fed to processing resources and results are continuously generated. The grid arrangement with intermediate buffers ensures that processing resources remain continuously occupied without idle periods, maintaining 100% utilization. Each processing resource operates without interruption, eliminating wasted processing time.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The patent pre-arranges the grid configuration of processing resources and pre-allocates input data to appropriate processing units before operations begin. This preliminary organization ensures that when processing starts, all resources are immediately engaged with their designated tasks, avoiding any initial idle time and ensuring continuous productive operation throughout the matrix computation.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If matrix operations are distributed across multiple processing resources, then processing efficiency improves, but coordination and communication overhead increases

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidsystem coordination complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent combines multiple processing resources into a unified grid architecture with standardized interfaces and shared memory structures. This merging approach allows processing resources to work together as an integrated system rather than independent units, simplifying coordination through uniform data flow patterns and reducing communication overhead through shared intermediate buffers and systematic data exchange protocols.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3346391B1Distributed convolution for neural networks
Publication Date: 2025.07.02 INTEL CORP
  • EP3346391B1 patent drawingFigure 1
  • EP3346391B1 patent drawingFigure 2A
  • EP3346391B1 patent drawingFigure 2B~2C

AI summary

In one embodiment, a matrix operation may be performed using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations. The plurality of input matrices may be partitioned into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements. The plurality of input partitions may be distributed among a plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements. A plurality of partial matrix operations may be performed using the plurality of processing elements, and partial matrix data may be transmitted between the plurality of processing elements while performing the plurality of partial matrix operations. A result of the matrix operation may be determined based on the plurality of partial matrix operations.