When would you use CUDA programming or custom hardware-level optimizations for an autonomy perception pipeline at Anduril Industries?
