Autonomous Driving Compute Distribution for Low-Latency Node Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional dynamic compute distribution in autonomous driving vehicles is unsustainable due to the high volume and diversity of data, with existing solutions failing to efficiently distribute computational resources across vehicles, network edges, and clouds, leading to inefficiencies and high latency in critical use cases.
Innovation Solution
A dynamic model that combines resources from in-vehicle platforms, network edge platforms, and cloud platforms as a single pool, using AI CPUs and multi-access edge computing (MEC) to distribute machine learning workloads based on latency and real-time requirements, with 5G connectivity enabling end-to-end solutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If full reliance on cloud model is used for computational requirements, then centralized processing capability is improved, but latency and response time worsen
Solution Approach 1:
The patent segments the computational workload across multiple locations: in-vehicle platforms handle immediate processing, network edge platforms handle regional processing, and cloud platforms handle aggregate processing. This segmentation allows critical functions to be processed closer to the vehicle, reducing latency while maintaining centralized capabilities for non-time-critical tasks.
Solution Approach 2:
The patent introduces a spatial dimension to the computational architecture by distributing processing across three distinct layers (vehicle, edge, cloud) rather than relying solely on centralized cloud processing. This multi-dimensional distribution enables simultaneous optimization of both processing power and response time by routing tasks to appropriate layers based on their latency requirements.
2Loss of time
If computational resources are distributed across multiple platforms, then latency is reduced, but system complexity increases
Solution Approach 1:
The patent creates a universal computational framework where the same three-platform architecture (vehicle, edge, cloud) serves multiple autonomous driving use cases with different latency requirements. Each platform type can handle various computational tasks, and the system provides a unified interface for task distribution, reducing the effective complexity despite the distributed nature of the system.
3Speed
If cloud resources are stretched to network edge and vehicles, then processing speed is improved, but resource coordination difficulty increases
Solution Approach 1:
The patent implements dynamic resource allocation where computational tasks are routed to different platforms based on real-time conditions such as latency requirements, available capacity, and task priority. This dynamic approach allows the system to adapt to changing conditions and optimize processing speed while managing coordination complexity through flexible, condition-based task distribution rather than static resource assignment.
Data Source
AI summary
The disclosed embodiments generally relate to methods, systems and apparatuses for directing Autonomous Driving (AD) vehicles. In one embodiment, an upcoming condition is assessed to determine the computational needs for addressing the condition. A performance value and latency requirement value are assigned to the upcoming condition. A database of available nodes in the network is then accessed to select an optimal node to conduct the required computation. The database may be configured to maintain real time information concerning performance and latency values for all available network nodes. In certain embodiments, all nodes are synchronized to maintain substantially the same database.