Answers are grounded in this paper and cite their sources.
Approximate Nearest Neighbor Search (ANNS) systems face significant challenges when processing large-scale vector datasets that exceed GPU memory capacity, typically requiring dataset partitioning and distributed computation. We propose a novel ANNS framework leveraging GPU-Initiated Direct Storage (GIDS) access, enabling GPU threads to directly retrieve vector data from SSDs through user-space I/O mechanisms, bypassing CPU bottlenecks. Our system addresses memory constraints through a fine-grained page-level caching system optimized for high-concurrency GPU workloads. We introduce a two-tier graph search architecture where a lightweight upper-tier graph (1% of dataset) resides in GPU memory and provides query-specific entry points for main graph searches. This approach fundamentally redesigns the storage-compute pipeline by co-locating vectors with neighboring data in page-aligned storage and implementing GPU-controlled cache management. Experimental evaluation on SIFT1B demonstrates 20% improved query throughput compared to state-of-the-art DiskANN while maintaining superior recall rates. On clustered datasets, our hierarchical search achieves up to 48% recall improvement. The framework is modular, dataset-agnostic, and requires no vector quantization, indicating GPU-driven storage control as a promising direction for next-generation large-scale search engines.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: