faiss

This work addresses the lack of systematic optimization and end-to-end performance evaluation of graph-based vector search algorithms on modern GPU architectures. It presents the first comprehensive taxonomy framework for GPU-accelerated graph-structured approximate nearest neighbor search (ANNS) algorithms, offering in-depth analysis of the mapping between algorithmic tasks and GPU hardware resources. Through implementations of six representative algorithms across eight large-scale datasets and fine-grained performance profiling, the study identifies distance computation and CPU–GPU data transfer as critical performance bottlenecks and elucidates the trade-offs between scalability and memory usage. The paper further provides practical design guidelines for deploying GPU-based ANNS systems and releases a comprehensive open-source benchmark to support future research in the community.

To address the growing demand for efficient billion-scale embedding vector retrieval driven by AI applications, this paper systematically analyzes Faiss’s trade-offs among accuracy, latency, and memory overhead, and proposes a modular index architecture with hardware-aware optimizations. Methodologically, it introduces a unified indexing framework integrating product quantization (PQ), inverted file (IVF), and multi-level indexing; a standardized API supporting CPU/GPU heterogeneous backends; and low-level optimizations including SIMD acceleration, memory-mapped I/O, and optimized quantization encoding. Evaluated on standard benchmarks, the approach achieves millisecond-latency similarity search over billion-vector datasets, delivering 2–5× higher throughput than prior systems. The resulting infrastructure has been deployed in production recommendation, retrieval, and multimodal search systems at Meta, Netflix, and other industry leaders, establishing a scalable, high-performance paradigm for large-scale vector databases.

This work addresses the challenge of efficiently merging graph indices in distributed systems and real-time vector databases, a problem previously lacking systematic investigation. To this end, the authors propose FGIM, a general and efficient three-stage framework for graph index merging. FGIM first converts input navigable graphs (e.g., HNSW) into k-nearest neighbor graphs (k-NNGs), then enhances neighbor quality and graph connectivity through cross-query candidate extraction and k-NNG refinement, and finally reconstructs a high-quality navigable graph. Extensive experiments on six real-world datasets demonstrate that FGIM achieves up to 3.5× speedup over incremental HNSW construction and averages 7.9× acceleration compared to non-incremental baselines, while maintaining comparable or superior retrieval accuracy.

This study addresses the lack of systematic evaluation of hybrid search mechanisms that combine semantic retrieval with metadata filtering in existing vector databases. We propose a novel relevance metric, Global-Local Selectivity (GLS), construct MoReVec—the first benchmark dataset supporting filtered retrieval—and extend ANN-Benchmarks to enable unified evaluation of hybrid search performance. Through comprehensive experiments integrating diverse filtering strategies into FAISS, Milvus, and pgvector with IVFFlat and HNSW indexes, we demonstrate that engine-level algorithmic integration critically governs performance: Milvus achieves more stable recall via hybrid execution, pgvector’s optimizer often selects suboptimal query plans, and IVFFlat outperforms HNSW under low-selectivity queries. Our findings culminate in practical configuration guidelines that offer both theoretical insights and actionable recommendations for efficient hybrid search deployment.

This work addresses the challenge of achieving both high scalability and efficiency in approximate nearest neighbor search (ANNS) over sparse vectors on conventional CPU architectures. To this end, we propose SpANNS—the first near-memory computing architecture tailored for sparse ANNS—built upon the CXL Type-2 platform. SpANNS integrates a hybrid inverted index, query parsing, clustering-based filtering, and a compute-enabled DIMM co-processing mechanism to perform index traversal and distance computation efficiently near the data. Evaluated against the state-of-the-art CPU baseline, SpANNS achieves a speedup of 15.2× to 21.6×, substantially enhancing both performance and scalability for sparse vector retrieval.

To address the prohibitively long indexing time and poor CPU-architecture compatibility of graph-based indexes in high-dimensional approximate nearest neighbor search (ANNS), this paper proposes Flash—a hardware-aware compact encoding strategy. Flash innovatively integrates vector quantization with key CPU architectural features, including SIMD parallelism and cache locality, enabling efficient distance computation while maintaining bounded quantization error. By jointly optimizing compact encoding, memory access patterns, and cache-friendly graph construction, Flash achieves 10.4×–22.9× speedup in index construction across eight real-world datasets ranging from 10 million to one billion vectors. Crucially, this acceleration comes without sacrificing retrieval accuracy or query latency—indeed, both are preserved or improved. Flash thus bridges the gap between algorithmic efficiency and modern hardware utilization in large-scale ANNS.

This work addresses the challenges of high memory consumption, substantial computational overhead, and poor adaptability to heterogeneous data distributions in HNSW for billion-scale vector retrieval. To overcome these limitations, the authors propose a density-aware adaptive quantization scheme, a multi-stage re-ranking mechanism, and enhancements to the graph index structure integrated with cross-architecture SIMD-optimized vectorized computation. The resulting approach significantly improves system efficiency while maintaining recall rates above 98%: query throughput (QPS) increases by 2.5–3.3×, index memory usage is reduced by 75%, and indexing speed accelerates by 5×. These advances substantially enhance the scalability and practicality of HNSW in large-scale similarity search scenarios.

This work addresses a critical limitation in existing graph-based disk indexing systems for large-scale high-dimensional vector similarity search: their performance is constrained by overlooking computational overhead, as the true bottleneck in high-dimensional settings lies in computation rather than I/O. The study is the first to reveal the intrinsic nature of this computational bottleneck and proposes a novel computation-optimized disk data layout that fully exploits modern CPU SIMD instructions. The approach integrates degree-based node caching, cluster-driven entry point selection, and an early scheduling strategy. Extensive experiments demonstrate that the proposed method significantly outperforms state-of-the-art disk-based graph index systems across multiple large-scale high-dimensional datasets, achieving performance comparable to—or even surpassing—that of in-memory indexing schemes, thereby transcending the traditional I/O-centric design paradigm.

This work addresses the high latency and service disruption caused by frequent index rebuilds in existing approximate nearest neighbor (ANN) methods under dynamic vector database updates. To overcome these limitations, the authors propose ACRONYM—a co-designed algorithm-hardware platform that leverages a data-distribution-agnostic XOR-and-Accumulate (XAC) systolic array encoder and Hamming-distance-based search, integrated with content-addressable memory (CAM) to enable in-memory parallel computation. A two-stage coarse-to-fine retrieval architecture circumvents CAM dimensionality constraints, allowing continuous, interruption-free updates. Evaluated on million-scale dynamic datasets, ACRONYM achieves over 90% recall, 8 million queries per second throughput, only 32 MB memory footprint, and 2.56 μJ per query energy efficiency—outperforming CPU-based HNSW by 400× and GPU-based FAISS-IVF by 80× in speed.

Existing approximate nearest neighbor (ANN) search methods struggle to simultaneously achieve high query efficiency, fast indexing, low memory consumption, scalability to high dimensions, robustness across varying data scales, and support for online insertions—key requirements in modern AI applications. This work proposes the Projection-Augmented Graph (PAG) framework, which uniquely integrates projection mechanisms into graph-based indexing. By leveraging statistically grounded asymmetric distance comparisons, PAG drastically reduces the need for exact distance computations while jointly optimizing index construction and query processing within a unified architecture. The framework natively supports online insertions and, across six real-world datasets, demonstrates up to 5× higher QPS-recall performance than HNSW, while maintaining efficient indexing, moderate memory overhead, and consistent robustness in both high-dimensional and large-scale settings.

This work addresses the memory bandwidth bottleneck in high-dimensional approximate nearest neighbor search (ANNS) on CPUs and GPUs, where conventional early termination mechanisms struggle to accelerate computation due to slow distance convergence. The authors propose a hardware-software co-design that integrates DIMM-level near-data processing (NDP) with a PCA-statistics-based feature-level early stopping mechanism, employing an estimate-and-correct strategy to accurately approximate full-dimensional distances. Additionally, they introduce bit-level dynamic floating-point compression and data-aware neighbor list mapping to substantially reduce memory access and communication overhead. Evaluated under strict accuracy constraints, the proposed system achieves 8.4× and 1.4× speedups over state-of-the-art CPU and GPU baselines, respectively, and outperforms the latest NDP accelerator, ANSMET, by 1.69×.