NVIDIA cuVS

NVIDIA RAPIDS cuVS is a GPU-accelerated library for vector search and clustering with CUDA-optimized HNSW, IVF, CAGRA, and PQ implementations. Supports batch queries for high QPS, suited for large-scale similarity search in GenAI apps. Achieves up to 12x faster indexing and lower latency vs CPU-only alternatives like FAISS CPU.

Memory-bounded GPU-accelerated framework for graph-based ANN vector search using CUDA and LibTorch, optimized for large-scale workloads beyond GPU memory. Features batch processing for high efficiency; outperforms CPU-only ANN in speed for similarity search in vector databases.

GPU-accelerated vector query processing system using CUDA to handle datasets larger than GPU memory via reordered pipelining and cluster-based retrofitting. Supports batch queries with up to 135x speedup over traditional GPU methods and 23x vs CPU-only for large-scale similarity search and MIPS.

An efficient CPU/GPU cooperative processing architecture for billion-scale approximate nearest neighbor search. FusionANNS achieves up to 13.1× higher QPS compared to SPANN and can handle billion-vector datasets with over 12,000 QPS while maintaining 15ms latency using only one entry-level GPU.

RAFT is a suite of GPU-accelerated libraries for data science, including support for vector search and similarity operations, often used in vector database scenarios.

ASPLOS 2024 paper introducing Juno, a system that accelerates high-dimensional approximate nearest neighbor search using sparsity-aware algorithms and GPU ray-tracing (RT) core mapping for hardware-level computation acceleration.