Deterministic Workload Pipelines for ASIL-Certifiable Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving systems face challenges in achieving functional safety integrity levels (ASIL) due to non-deterministic inference models, making certification and ASIL grading difficult.
Innovation Solution
Implementing a computing system with workload processing chiplets that execute tasks in independent deterministic pipelines using a shared memory with a reservation table for dynamic scheduling and out-of-order execution, facilitated by a central chiplet and sensor data input chiplet, ensuring deterministic workload processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If non-deterministic inference models are used for autonomous driving tasks, then computational flexibility and adaptability are improved, but functional safety integrity certification and ASIL grading become difficult
Solution Approach 1:
The system segments autonomous driving workloads into multiple independent pipelines, each handling specific tasks (e.g., perception, planning, control). This segmentation allows deterministic execution within each pipeline while maintaining overall system flexibility, enabling both adaptability and functional safety certification.
Solution Approach 2:
A reservation table mechanism is introduced as an intermediary between workload submission and execution. The reservation table deterministically schedules workloads across pipelines, ensuring predictable execution order and timing while allowing flexible workload types to be processed, thus bridging the gap between adaptability and certification requirements.
2Reliability
If deterministic pipelines are implemented for workload execution, then functional safety integrity and certification are improved, but computational complexity and scheduling overhead increase
Solution Approach 1:
The system performs preliminary actions by pre-reserving pipeline slots and resources in the reservation table before workload execution. This advance scheduling simplifies real-time execution decisions while ensuring deterministic behavior, reducing the complexity of on-the-fly scheduling operations.
Solution Approach 2:
Workloads self-register in the reservation table and automatically proceed to execution when their designated pipeline slots become available. This self-service mechanism reduces the need for complex centralized scheduling logic, as workloads manage their own execution timing based on pre-established reservations.
3Productivity
If out-of-order execution is used for workloads, then processing efficiency and throughput are improved, but execution determinism and predictability are reduced
Solution Approach 1:
The system introduces a new dimension of control by separating workload submission order from execution order through the reservation table. Workloads can be submitted out of order and executed out of order, but the reservation table ensures that execution remains deterministic by tracking dependencies and maintaining predictable timing characteristics.
Solution Approach 2:
The reservation table dynamically adapts to workload dependencies and pipeline availability, allowing flexible out-of-order execution when possible while maintaining determinism when dependencies require ordered execution. This dynamic approach optimizes efficiency without sacrificing predictability.
Data Source
AI summary
A computing system includes a sensor data input chiplet to obtain sensor data from a sensor system of a vehicle, and one or more workload processing chiplets executing workloads in a set of independent pipelines based on the sensor data.