Autonomous vehicles process immense amounts of data and they have to do it in real-time. There are dimensional, energy and thermal limitations on the amount of computation which reasonably can be placed in a typical sedan or minivan. Out of this contradiction various design compromises emerge. This article reveals some specific challenges and idiosyncrasies; looks into some important design choices and how they define the overall system architecture; argues the importance of a camera interface; talks about the benefits and perils of using video compression; attempts to define some useful video-related software abstractions; points out a qualitative-changing role of hardware timestamps; describes advantages HDR-cameras bring and shows associated computational burden.