"Amping Up" Autonomous Vehicle Safety Design: Benchmarking on the NVIDIA Ampere Architecture
, Tata Consultancy Services
, Tata Consultancy Services Limited
Unsafe lane-change and lane-merge maneuvers by human drivers cause about 7% of all U.S. crash fatalities. In a mixed-mode environment of autonomous and human-driven vehicles, aggressive cut-in behavior of an autonomous vehicle (AV) might put pressure on nearby human-driven vehicles. Conversely, a seemingly unpredictable move by a human driver could put the AV control system outside its normal operating region. AV safety systems should track the vehicles around the ego vehicle and make real-time predictions to decelerate or apply emergency brakes. Modeling and training systems to predict such events pose unique challenges to designers. We have benchmarked our training and inference process against Volta-based GPU and the NVIDIA A100 Tensor Core GPU platform. The training process using A100 was significantly faster, which enabled us to tune the system with even bigger grid sizes. A100's multi-instance GPU technology allowed us to scale up the inference throughput significantly.