Researchers at the University of California, Irvine, have made significant strides in the realm of artificial intelligence within the automotive industry with the publication of a new technical paper titled “Performance Implications of Multi-Chiplet Neural Processing Units on Autonomous Driving Perception.” This research focuses on the integration of chiplet-based Neural Processing Units (NPUs) to enhance vehicular AI perception capabilities, especially within the constraints of automotive environments.
The study highlights the growing importance of chiplet technology in modern vehicular architectures, as it presents a cost-effective solution that balances performance, modularity, and customisation. The necessity for such advancements is propelled by the fact that perception models used in autonomous driving are among the most computationally intensive tasks encountered in these systems.
The researchers used Tesla's Autopilot perception pipeline as a real-world case study, meticulously breaking down its constituent models and analysing their performance when run on various chiplet accelerators. Their investigation led to the proposal of an innovative scheduling strategy, designed to optimise the deployment of perception workloads across multi-chip AI accelerators.
Through extensive experiments conducted using the DNN performance simulator known as MAESTRO, the research team reported significant enhancements in processing efficiency. According to their findings, the proposed scheduling strategy achieved an impressive 82% increase in throughput, along with a 2.8 times greater utilisation of processing engines in comparison to traditional monolithic accelerator designs.
The paper, authored by Mohanad Odema, Luke Chen, Hyoukjun Kwon, and Mohammad Abdullah Al Faruque, underscores the potential of chiplet technology to revolutionise AI applications in the automotive sector. As the industry looks to further embrace AI for improvements in safety and efficiency, the insights provided in this research may serve as a catalyst for future developments in autonomous driving technologies.
The technical paper is currently available as a preprint, with a formal publication anticipated in November 2024.
Source: Noah Wire Services