cuSBF: GPU Bloom filter speeds genomic indexing up to 234x vs CPU

Genomic sequence analysis relies on efficient indexing of k-mers (short DNA substrings), but traditional bloom filters struggle with high false-positive rates and poor GPU performance due to irregular memory access and register pressure. The new cuSBF library, detailed in a preprint on arXiv (2606.24417), introduces a minimizer-aware Super Bloom Filter (SBF) that reimagines GPU parallelism for this task. Key innovations include sectorized shards, cooperative shared-memory tiling, and warp-level shard sharing that transform the locality of overlapping k-mers into scalable parallelism.

Benchmarked on NVIDIA RTX PRO 6000 Blackwell and GH200 systems, cuSBF achieves throughput records among all evaluated sequence-capable AMQ structures. It outperforms cuCollections' blocked Bloom filter by up to 9.1× for insertion and 7.7× for query, and surpasses multi-threaded CPU SBF by up to 234×. Against GPU-based Cuckoo, Two-Choice, and Quotient filters, speedups range from 1.5× to 3400× depending on workload. A parameter sweep identifies (s=28, m=16, H=4) as Pareto-optimal for k=31 nucleotides, delivering significantly lower false-positive rates at matched memory budgets. The library is open-source, header-only, and designed to sustain 85% SM utilization even when the filter exceeds cache capacity, proving that sequence locality—not raw bandwidth—is the key bottleneck for GPU-accelerated genomic indexing.