Hardware Utilization and Inference Performance of Edge Object Detection Under Fault Injection
As deep learning models are increasingly deployed on resource-constrained edge platforms, particularly in the domain of autonomous driving systems, understanding the behavior of hardware under resource degradation becomes essential. A recent study, detailed in arXiv:2604.09631v1, presents a comprehensive characterization of various hardware metrics while utilizing TensorRT-optimized deep learning models for object detection tasks.
This research focuses on the performance of three neural network architectures: YOLOv10s, YOLOv11s, and YOLO2026n, which were executed on the NVIDIA Jetson Nano platform. The study employed a large-scale fault injection campaign aimed at assessing both lane-following and object detection tasks. The faults were synthesized using a decoupled framework that integrates large language models (LLMs) and latent diffusion models (LDMs), utilizing original data collected from the JetBot platform.
Key Findings
The results of this investigation reveal several crucial insights regarding hardware performance under adverse conditions:
- Stable GPU Occupancy: Across both tasks and model architectures, the inference engines maintained stable GPU utilization, which is critical for ensuring efficient processing of real-time data.
- Controlled Temperature Rise: The thermal behavior of the system remained within safe limits, indicating that the hardware could handle extended periods of intense computational load without overheating.
- Consistent Power Consumption: The study found that power draw remained stable, falling within acceptable thresholds, thereby ensuring the longevity and reliability of the hardware.
- Memory Usage Patterns: After an initial warm-up phase, memory usage exhibited a consistent release pattern, suggesting efficient management of resources even under varying conditions.
- Variability in Object Detection: While object detection tasks displayed more variability in terms of memory and thermal behavior compared to lane-following tasks, both categories confirmed the robustness of the TensorRT pipelines in handling degraded input data.
Conclusion
The findings from this study provide a hardware-level perspective on the reliability of deep learning models deployed at the edge, complementing existing research focused on inference performance. As the demand for efficient and reliable AI systems in autonomous driving continues to grow, understanding the interaction between hardware constraints and model performance will be vital for future developments.
This research underscores the importance of considering hardware behavior in the design and deployment of AI models, especially in scenarios where resource constraints are prevalent. By ensuring that systems can perform reliably under fault conditions, developers can enhance the safety and effectiveness of autonomous driving technologies.