BEV Pooling Optimization for Physical AI

This feature thread tracks the evolution of optimizing Bird's-Eye-View (BEV) pooling for NVIDIA GPUs, crucial for physical AI applications like autonomous vehicles and robotics. Initial efforts focused on the fundamental BEV pooling operation, which gathers depth-weighted image features and scatters them into BEV grid cells. Subsequent developments, exemplified by BEVPoolV3, introduce algorithmic and implementation optimizations to significantly reduce latency. These include reducing duplicate depth loads, a five-array INT32 scatter map, precomputed indices eliminating runtime integer division, and interval-owned output writes, enabling substantial latency reduction across different GPU memory regimes. The workflow for optimizing BEV pooling operators involves classifying the working set by L2 fit, eliminating redundant scatter traffic, mapping kernel implementation to the target NVIDIA GPU architecture, and validating bottlenecks with NVIDIA Nsight Compute, with optimization strategies selected according to whether the workload is DRAM-bound or L2-resident. Benchmarks on NVIDIA RTX A6000 (Ampere) and NVIDIA RTX PRO 6000 Blackwell Max-Q demonstrate that BEVPoolV3 achieves up to 22x speedup (FP16) on the DRAM-bound path and up to 42x speedup (FP8) on the L2-resident path over V2, with FP8 delivering further latency gains for L2-resident scatter-reduce workloads, while edge-class platforms require architecture-specific evaluation for FP8 adoption.