XR Neural Network Task Segmentation for Upload Latency and Power Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of properly segmenting computing tasks of neural network models in XR services to reduce data upload and latency is unresolved.
Innovation Solution
A method involving a radio access network device that determines a target segmentation point based on the amount of to-be-transmitted data, computation amount, and channel status between a terminal device and the network, segmenting the computing task to allocate it to different devices for execution, optimizing latency and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the XR terminal performs preprocessing and uploads data to the server, then the amount of data uploaded is reduced, but the transmission latency increases due to network transmission requirements
Solution Approach 1:
The computing task of the neural network model is segmented into multiple parts using segmentation points. The terminal device executes computing tasks for neural network layers before the segmentation point, while the server executes computing tasks for layers after the segmentation point. This segmentation allows the terminal to perform local preprocessing to reduce upload data volume, while the server handles the remaining heavy computation, thereby balancing data reduction benefits with latency constraints.
2Loss of time
If more neural network layers are executed by the terminal device, then the computation amount at the terminal increases, but the transmission latency decreases due to less data needing network transmission
Solution Approach 1:
The segmentation point is dynamically determined based on real-time channel status, amount of to-be-transmitted data, and computation amount. When channel conditions are poor or terminal computing resources are limited, the segmentation point is adjusted to execute fewer layers at the terminal, reducing power consumption. When channel conditions improve or terminal computing capacity increases, more layers can be executed locally, reducing transmission latency. This dynamic adjustment resolves the contradiction between latency and power consumption.
3Use of energy by moving object
If the computing task is fully executed at the server, then the terminal device power consumption is reduced, but the transmission latency increases due to large data upload requirements
Solution Approach 1:
By segmenting the computing task and assigning the initial neural network layers to the terminal device for local execution, the system enables the terminal to perform preprocessing on input data (such as images or video frames) before uploading. This reduces the volume of data that needs to be transmitted to the server while still leveraging the server's powerful computing resources for the remaining layers, thus balancing power consumption and transmission latency.
4Adaptability or versatility
If the segmentation point is changed to optimize for different conditions, then the computing task allocation adapts to channel status and requirements, but the system complexity increases due to dynamic determination requirements
Solution Approach 1:
The radio access network device determines the segmentation point based on feedback regarding channel status, amount of to-be-transmitted data, and computation amount. This feedback mechanism allows the system to adaptively adjust the segmentation point to optimize performance under different conditions. The radio access network device receives information about terminal computing capabilities and channel conditions, then determines the appropriate segmentation point, and provides this information back to both the terminal and server for coordinated execution.
Data Source
AI summary
A computing task segmentation method and a related apparatus are provided, and may be applied to an extended reality (XR) service or another low-latency service. The method includes: obtaining an amount of to-be-transmitted data and a computation amount of a terminal device that correspond to a first computing task, where the first computing task is obtained by segmenting a computing task of a neural network model based on a segmentation point; determining that the segmentation point is a target segmentation point based on the amount of to-be-transmitted data, the computation amount, and a status of a channel between the terminal device and a radio access network device; and sending indication information to the terminal device, where the indication information indicates the target segmentation point.