The architectural implications of autonomous driving: constraints and acceleration

There are several defined levels of automation, with level 2 being ‘partial automation’ in which the automated system controls steering and acceleration/deceleration under limited driving conditions. At level 3 the automated system handles all driving tasks under limited conditions (with a human driver taking over outside of that). By level 5 the system is fully automated.

An autonomous driving system should be able to process current traffic conditions within a latency of 100ms at a frequency of at least once every 100ms. The system also needs extremely predictable performance, which means that long latencies in the tail are unacceptable.  Thus 99th, or 99.99th, percentile latency should be used to evaluate performance.

We show that GPU- FPGA- and ASCI-accelerated systems can reduce the tail latency of [localization, object detection, and object tracking] algorithms by 169x, 10x, and 93x respectively… while power-hungry accelerators like GPUs can predictably deliver the computation at low latency, their high power consumption, further magnified by the cooling load to meet the thermal constraints, can significantly degrade the driving range and fuel efficiency of the vehicle.