Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications | NVIDIA Technical Blog

This post introduces BEVPoolV3, detailing four algorithmic optimizations for BEV pooling on NVIDIA GPUs: reduced duplicate depth loads, a five-array INT32 scatter map, precomputed indices to eliminate runtime integer division, and interval-owned output writes. It also outlines a practical workflow for optimizing gather/scatter kernels: classifying memory regimes (DRAM-bound vs. L2-resident), removing redundant scatter traffic, mapping kernel implementation to target GPU architecture, and validating bottlenecks with Nsight Compute. Performance benchmarks on RTX A6000 and RTX PRO 6000 Blackwell Max-Q show significant speedups (up to 22x FP16 on DRAM-bound, up to 42x FP8 on L2-resident) over previous versions.