Autonomous driving systems hardware and software architecture exploration: optimizing latency and cost under safety constraints

With the recent progress of techniques in computer vision and processor design, vehicles are able to perform a greater number of functions, and are reaching higher levels of autonomy. As the list of autonomous tasks that the car is supposed to perform grows, two design questions arise: how to group these tasks into modules and which processors and data buses should instantiate these modules and their links in the physical architecture. Both questions are linked, as the processing capacity of the processors influences how centralized the architecture can be, and the modularization influences the overall system latency as well. Furthermore, our interest lies in designing architectures that perform the tasks rapidly, while minimizing cost. This multiobjective problem is intractable without architecture exploration and an analysis tool. This paper presents a linear optimization formulation to capture these tradeoffs, and to systematically find relevant architectures with optimal latency and cost. The results show that enforcing all safety constraints on the architecture leads to a worst case increase of 17% in latency and 18% component cost per vehicle. The increase in latency is significant at the scale of human driver reaction times.