How Centralized Radar Processing on NVIDIA DRIVE Enables Safer, Smarter Level 4 Autonomy

NVIDIA's centralized radar processing on DRIVE AGX Thor moves raw ADC data from edge radar sensors into a central compute platform, enabling ~100x more signal information than traditional point-cloud outputs. A dedicated Programmable Vision Accelerator (PVA) handles all radar DSP stages (range-FFT, Doppler-FFT, angle-FFT) without consuming GPU resources, freeing the GPU for AI perception and planning. Validated with ChengTech production hardware at GTC 2026, the pipeline supports five radar units at 30 FPS with 540 MB/s aggregate throughput. This approach enables richer ML models trained on raw radar representations, early sensor fusion with cameras and lidar, and coherent multi-radar fusion — capabilities critical for Level 4 autonomy stacks. Hardware benefits include 30%+ unit cost reduction and ~20% lower system power consumption.