Neural Network Partitioning Across Edge Nodes Under Power-Latency Limits

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Edge devices with limited resources and power constraints struggle to execute complex neural networks within predefined power and latency thresholds, especially under varying ambient conditions, leading to incomplete execution or failure to meet service level agreements.

Innovation Solution

A neural network partitioning system that divides the network into portions, with one portion executed on a first Edge node and another on a second Edge node, based on energy consumption, latency, and resource availability, using processor circuitry to determine optimal partitioning and communication strategies.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If the entire neural network is executed on a single Edge node, then computation latency is reduced, but power consumption exceeds available power thresholds

Engineering Contradiction:

Improvecomputation latencyVSAvoidpower consumption

Core Design Contradiction:

Loss of timeVSUse of energy by moving object

Solution Approach 1:

The neural network is divided into multiple portions that can be executed on different Edge nodes. The system identifies split points in the network architecture and distributes computation across multiple nodes, allowing each node to handle only the portion it can execute within power constraints while maintaining overall network functionality.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The solution transitions from a single-node execution model to a multi-node distributed execution model, adding the dimension of spatial distribution across the Edge cloud network. This allows the system to trade off local computation latency for distributed computation that respects power constraints.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Use of energy by moving object

If the neural network is partitioned across multiple Edge nodes, then power consumption is reduced, but communication overhead and transmission time increase

Engineering Contradiction:

Improvepower consumptionVSAvoidtransmission time

Core Design Contradiction:

Use of energy by moving objectVSLoss of time

Solution Approach 1:

The system performs preliminary analysis to identify optimal split points in the neural network that minimize communication overhead. By pre-determining which layers should be executed on which nodes based on network architecture and resource constraints, the system reduces runtime communication overhead and transmission time.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If the neural network is executed on resource-constrained Edge devices, then service level agreements on latency are met, but the system fails under varying ambient conditions

Engineering Contradiction:

Improveservice level agreement complianceVSAvoidadaptability to ambient conditions

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system dynamically adjusts the neural network partitioning and execution strategy based on real-time ambient conditions, available power thresholds, and resource availability. This allows the system to adapt to varying environmental conditions while maintaining service level agreements through flexible redistribution of computation across the Edge cloud.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS20260074969A1Apparatus, articles of manufacture, and methods to partition neural networks for execution at distributed edge nodes

Publication Date: 2026.03.12 INTEL CORP

AI summary

Methods, apparatus, systems, and articles of manufacture are disclosed to partition neural network models for executing at distributed Edge nodes. An example apparatus includes interface circuitry, machine-readable instructions, and at least one processor circuit to be programmed by the machine-readable instructions. One or more of the at least one processor circuit is to partition a neural network model into a first portion to be executed at an edge of a network and a second portion to be executed at a cloud based on a transmission metric.