BEVPoolV3 is a new CUDA kernel optimization for bird's-eye-view (BEV) pooling used in autonomous vehicles and robotics. The post walks through a practical GPU optimization workflow: classify whether the working set fits in L2 cache, remove redundant scatter traffic via a five-array INT32 scatter map, implement interval-owned scatter-reduce to avoid atomics, and validate with NVIDIA Nsight Compute. On RTX PRO 6000 Blackwell Max-Q (large L2), BEVPoolV3 FP8 achieves up to 42x speedup over the V2 baseline. On RTX A6000 (small L2, DRAM-bound), the adapted FP16 path reaches 19x speedup. The post also explains why FP8 outperforms NVFP4 for L2-resident scatter-reduce workloads, and how the same methodology applies to sparse embeddings, voxelization, and other irregular memory-bound kernels.
