Approximate nearest neighbor search (ANNS) on GPUs has become increasingly important for workloads that handle massive high‑dimensional vectors. Graph‑based indexes deliver high recall and throughput but suffer from long build times and large storage footprints. Cluster‑based methods build quickly and scale well, yet often require many probes to achieve high recall, stressing memory bandwidth and compute resources.
To achieve fast index construction, high‑throughput search, high recall, and low storage on GPUs, we introduce IVF‑RaBitQ (GPU), which tightly integrates the IVF clustering index with RaBitQ quantization into a fully GPU‑native build and search pipeline. For index building, we devise a scalable GPU‑native RaBitQ quantizer that provides rapid, accurate low‑bit encoding at large scale. For search, we develop GPU‑native distance computation schemes for RaBitQ codes and a fused search kernel that attains high throughput while preserving recall.
Implemented within NVIDIA's cuVS library and evaluated on the cuVS Bench across multiple datasets, IVF‑RaBitQ demonstrates a strong performance frontier. At a recall of approximately $0.95$, its queries‑per‑second (QPS) are $3.0\times$ higher than the state‑of‑the‑art graph method CAGRA, and index build time is $14.7\times$ faster on average. Compared with the cluster‑based IVF‑PQ, IVF‑RaBitQ achieves over $4.5\times$ higher throughput and avoids accessing raw vectors for re‑ranking.
Blogger's Review: IVF‑RaBitQ combines the lightweight nature of clustering indexes with low‑bit quantization and GPU‑optimized computation, effectively breaking the traditional trade‑off between build speed and query throughput. It represents a noteworthy milestone for GPU‑accelerated ANNS.