Pith. sign in

REVIEW 3 major objections 4 minor 67 references

PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate Nearest Neighbor Search

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read PathWeaver claims pipelined shard handoff makes multi-GPU graph ANNS 3.24x faster at 95% recall.

desk verdict A solid multi-GPU ANNS systems paper with a genuinely new pipelining idea; the headline speedup is plausible but the warm-start assumption that carries it is never directly measured. read the letter →

arxiv 2507.17094 v1 pith:QNKUFNQK submitted 2025-07-23 cs.DC

classification cs.DC
keywords approximatenearestneighborsearchgraph-basedANNSmulti-GPUsystemspipelinedgraphshardingGPUaccelerationproximitydirection-guidedfiltering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PathWeaver sets out to show that graph-based approximate nearest neighbor search (ANNS) can scale across multiple GPUs if the shards stop searching independently. Its central move is pipelining: after a shard's search converges, PathWeaver hands the best match to the adjacent shard, which starts from the nearest node to that match instead of from random nodes, so later stages need fewer iterations. Two supporting techniques, ghost staging and direction-guided selection, attack the cold start of the first stage and the wasted distance calculations inside each iteration. If the claims hold, multi-GPU ANNS stops treating extra GPUs as mere memory expansion and instead gets throughput that grows with GPU count without sacrificing 95% recall.

What carries the argument

The load-bearing object is the inter-shard edge mapping $I(u)=\arg\min_{w\in G_{(i+1)\bmod N}}\mathrm{dist}(u,w)$, a one-way connection from every node in shard $i$ to its nearest node in the adjacent shard. The mapping turns a collection of independent shard graphs into a ring and lets the pipeline warm-start each shard. The other two mechanisms are the ghost shard, a small sampled graph that acts as a hub for locating entry points, and the compressed direction table, a bit-sign encoding of edge directions used to filter neighbors before distance computation. These mechanisms do the work of reducing search iterations, removing the first-stage cold start, and cutting L2 distance calculations.

What would settle it

Construct a test dataset whose shards are disjoint, well-separated clusters and whose queries come from a single cluster: then the top-1 result found in shard i maps under I to a node in shard i+1 that is no closer to the query than a random node, so the pipelined warm start should not reduce per-stage iterations. Measuring per-stage iteration counts and QPS on such a dataset would distinguish whether PathWeaver's speedup comes from cross-shard locality or from its other two techniques.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a sharded proximity-graph search has a path structure worth reusing: the top-1 result in one shard, mapped through the precomputed inter-shard edge $I(u)=\arg\min_{w\in G_{(i+1)\bmod N}}\mathrm{dist}(u,w)$ to the nearest node in the next shard, is a better starting point for that shard than a random sample. PathWeaver builds a ring of such unidirectional connections among shards and pipelines search results around it, so each stage continues the previous stage's path rather than beginning a new one. Ghost staging adds a small auxiliary graph of sampled nodes to locate a good entry point for the first stage, and direction-guided selection uses a compressed sign-bit direction table to skip neighbors that point away from the query, removing most of the distance computations the paper measures as wasted. Together the three techniques yield the paper's reported 3.24x geomean speedup, up to 5.30x, over state-of-the-art multi-GPU ANNS baselines at 95% recall@10.

Load-bearing premise

The load-bearing premise is that the nearest node in the next shard to the current shard's best result is usually close to the query, so a warm start converges in fewer iterations than a random start; the paper supports this empirically on Sift, Gist, Deep, and Wiki datasets, but offers no formal or statistical guarantee.

Editorial extensions

If this is right

  • At 95% recall@10, PathWeaver reports a 3.24x geomean speedup and up to 5.30x over the strongest prior multi-GPU baseline, with 62% scaling efficiency from one to four GPUs.
  • Because each shard starts from the previous shard's top-1 result, total search iterations grow far more slowly than shard count.
  • Ghost staging narrows the cold-start bottleneck of the first stage, and in single-GPU settings it and direction-guided selection alone yield a 3.43x speedup over the strongest prior GPU method.
  • Direction-guided selection discards most neighbor distance calculations while keeping recall loss below about 0.003 in the reported comparisons.
  • The extra graph-build cost of the three mechanisms stays below 15% in the reported multi-GPU cases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This suggests the pipelining benefit may extend beyond four GPUs; because the communication volume is only one index per query per hop, larger rings should remain cheap as long as cross-shard locality holds.
  • A natural testable extension is to replace the fixed multi-GPU ring with a variable-length pipeline that stops after the stage where the warm-start gain saturates, which could be useful on datasets with weak cross-shard correlation.
  • Direction-guided selection could be combined with vector compression or learned quantizers, since it already reduces the distance-computation term that dominates the remaining cost.
  • The ghost-staging result that smaller sampling ratios work better suggests the auxiliary shard can be made tiny enough for streaming or incremental updates, though PathWeaver itself targets static graphs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. PathWeaver is a multi-GPU graph-based approximate nearest neighbor search system that shards the dataset across GPUs and adds three optimizations: pipelining-based path extension (handing each shard's top-1 local result to the adjacent shard via inter-shard edges, Eq. 3), ghost staging (a small auxiliary graph of sampled nodes to find better entry points), and direction-guided selection (filtering each visited node's neighbors by the sign-bit alignment of their direction with the query). The paper reports a 3.24x geomean and up to 5.30x speedup over CAGRA w/Sharding and GGNN at 95% recall@10 on Deep-10M, Deep-50M, and Wiki-10M, plus single-GPU speedups over CAGRA on Sift-1M, Gist-1M, and Deep-10M. The evaluation includes QPS-recall curves, ablations, execution-time breakdowns, and sensitivity studies for ghost-node ratio and neighbor-discarding rate.

Significance. If the results hold, PathWeaver is a useful step toward making sharded multi-GPU ANNS scale beyond memory-capacity extension: the pipeline idea of carrying a converged local search result across shards is simple and plausible, and the paper backs it with a broad benchmark suite and a public implementation. The work is also commendable for reporting scale efficiency (62% at 4 GPUs), build-time overhead, and communication-cost analysis rather than only QPS-recall curves. However, the central speedup claim rests on an empirical warm-start assumption that is not directly isolated, and several core parameters (top-n for direction-guided selection, cool-down ratio, priority-queue size l and expansion count r, number of handoff results) are chosen empirically without fully reporting their effect on the headline result. These gaps are fixable with targeted experiments, so the paper is promising but not yet ready in its current form.

major comments (3)
  1. [Section 3.1.2 / Eq. (3), Fig. 9b, Fig. 13] The headline 3.24x/5.30x speedup is attributed to the claim that I(z) in Eq. (3) lands close enough to the query that later shard stages converge in fewer iterations, but no experiment isolates this warm-start quality. Fig. 9b compares the full pipelined PathWeaver against independent-shard PathWeaver, and Fig. 13 compares whole PathWeaver against the whole sharded baseline; in both cases the handoff structure and the quality of the starting node vary simultaneously. Please add a control where I(z) is replaced by a uniformly random node in the adjacent shard while keeping the same pipeline, and report QPS, recall, and iterations for both variants. Also report a direct statistic such as the distance ratio d(q, I(z))/d(q, I_rand(z)) or the rank of I(z) among adjacent-shard nodes sorted by distance to q. Without this control, the speedup cannot be attributed to the warm-start mechanism rather than to the pipeline structure alone.
  2. [Section 4 / Eq. (3)] The implementation described in Section 4 builds the lookup table by performing an approximate top-1 search in the adjacent shard for every node, which is not the exact argmin defined in Eq. (3). Because the entire warm-start argument is about the quality of the handed-off node, this approximation error is not a minor implementation detail. No experiment measures how often the approximate I(z) differs from the exact nearest cross-shard node, nor how much the distance to the query increases as a result. Please add an offline validation on a sample of source nodes comparing the approximate table against the exact nearest neighbor in the adjacent shard, and report the effect of using the exact table on the final QPS-recall curves.
  3. [Sections 3.3.2, 5.6.3, and 5.6.2] Several load-bearing parameters appear to be hand-tuned without a stated criterion or a full sensitivity analysis. In particular, the top-n candidates retained per visited node in direction-guided selection and the cool-down ratio (30% by default) directly control the accuracy/speed trade-off, and the priority-queue size l and expansion count r are central to the search kernel. The paper only reports sensitivity for the ghost-node sampling ratio (Fig. 14) and for the neighbor discard/cool-down ratios in isolation (Figs. 15 and 16), not end-to-end QPS-recall curves as n, l, r, or the number of handoff candidates vary. Please report the chosen values and an end-to-end sensitivity study over these parameters, or justify that the results are insensitive to them.
minor comments (4)
  1. [Section 7.2] The last paragraph claims pipelining-based path extension 'achieves a proportional speedup with the number of machines used,' but Fig. 9a reports 2.47x speedup on 4 GPUs (62% scale efficiency), which is not proportional. Please soften this to 'improved scaling efficiency' or report the measured scaling law explicitly.
  2. [Fig. 9b] The figure legend and caption make it difficult to tell which bar is Naive PathWeaver and which is PathWeaver across datasets and recall targets; please add a clearer legend or a table with the QPS values.
  3. [Section 5.7] The sentence 'overall overhead is less than 10% for datasets targeting a single GPU' is clear, but the multi-GPU overhead of 15% for Deep-50M is later called 'small'; please add a sentence explaining why 15% is acceptable relative to the search-time reduction, or re-state the threshold.
  4. [Section 5.1] The CAGRA baseline is extended by the authors for multi-GPU sharding; please state explicitly whether the out-degree of the GGNN graphs was also fixed or what graph-degree setting was used for GGNN, so that the comparison is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported speedups are measured against external baselines and isolated by ablation, with only an empirical warm-start assumption that is not circular.

full rationale

PathWeaver's central claims are empirical benchmark results, not derivations from fitted parameters. The headline 3.24x/5.30x speedups are measured against CAGRA (official implementation, extended to sharding as described in Section 5.1) and GGNN (out-of-box multi-GPU), and the ablation in Fig. 11 attributes speedup component-by-component to pipelining-based path extension, ghost staging, and direction-guided selection. The one analytically weak link, the warm-start assumption around Eq. (3) that I(z) lies close to the query, is an externally testable empirical hypothesis rather than a definitional or self-citational reduction; the paper supports it indirectly through iteration counts (Fig. 13) and stage breakdowns (Fig. 5). There are no self-citations used as load-bearing evidence, no uniqueness theorem imported from the authors' prior work, and no fitted constant renamed as a prediction. The absence of a random-handoff control is a completeness/correctness gap, not circularity. Score 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 3 invented entities

PathWeaver adds no purely speculative entities: all three mechanisms are concrete data structures or communication patterns with measurable effects. The main honest cost is in free parameters: several knobs (top-n, cool-down ratio, ghost sampling ratio, number of results passed per query) are tuned empirically, and their default values are not given in numeric form. The claims rest on domain assumptions about graph quality and dataset representativeness, which are standard for systems papers.

free parameters (5)
  • Direction-guided selection top-n (neighbors kept per visited node) = Not stated as a numeric constant; paper gives a figure example with n=2 and reports discarding ratios up to 0.7.
    The number of neighbors selected after direction sorting controls the accuracy/speed trade-off and is an empirical knob.
  • Cool-down ratio (fraction of iterations with full neighbor expansion) = 30% of max iterations as default.
    Selected to balance accuracy and speed; the paper itself reports ablating it (Figures 15, 16).
  • Ghost node sampling ratio = Default not stated explicitly; sensitivity study sweeps 1e-4 to 0.1.
    Controls the size of the auxiliary ghost graph and is fit empirically to maximize QPS.
  • Number of candidate results sent per query between shards = 1
    The paper says 'we empirically choose to send only one to minimize communication overhead' without showing the supporting experiment.
  • Priority queue size l and top-r expansion count r = Not stated numerically.
    These control search quality and cost and are implicitly tuned; the paper only states top-r nodes are expanded each iteration.
assumptions (3)
  • domain assumption CAGRA's graph construction produces a graph with reachability and convexity properties that make beam search converge reliably.
    Invoked in Section 2.2 and Section 3.3.1 to argue that direction-guided selection losing a few candidates only increases iterations instead of hurting accuracy.
  • domain assumption The datasets used (Sift, Gist, Deep, Wiki) are representative of real ANNS workloads.
    All claims about speedup and recall generalization rest on this standard systems-benchmarking assumption (Section 5.1).
  • domain assumption A top-1 result from a shard, when mapped to the nearest node in the adjacent shard, gives a better starting point than a random node for the next shard's beam search.
    This is the core premise of pipelining-based path extension (Section 3.1.2, Eq. 3). Not proven, only empirically observed.
invented entities (3)
  • Inter-shard edges (mapping I from each node to its nearest node in the adjacent shard) independent evidence
    purpose: Enables warm-start handoff of search results between shards for pipelining-based path extension.
    The mapping is defined by the paper (Eq. 3) as a nearest-neighbor lookup. It is a stored edge table, and its effect is measurable: the paper reports faster convergence with PPE than without (Fig. 13). This is a real, testable artifact, not a metaphysical construct.
  • Ghost graph (auxiliary shard of sampled nodes with its own edges and links to the main graph) independent evidence
    purpose: Provides a small search space for finding good entry points into the main graph.
    Its effect is measured in the sensitivity study (Fig. 14) and ablation (Fig. 11). It is a concrete data structure.
  • Compressed direction table (sign bits of neighbor direction vectors) independent evidence
    purpose: Enables direction-guided selection of neighbors without cosine-like distance computations.
    The table is precomputed from the graph and its usefulness is directly compared against exact and random neighbor selection (Fig. 15).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate Nearest Neighbor Search." pith.science (2026). https://pith.science/paper/QNKUFNQK

@misc{pith2026250717094,
  author       = {Pith},
  title        = {Pith review of: PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate Nearest Neighbor Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNKUFNQK}},
  note         = {Machine review of arXiv:2507.17094}
}
abstract

Graph-based Approximate Nearest Neighbor Search (ANNS) is widely adopted in numerous applications, such as recommendation systems, natural language processing, and computer vision. While recent works on GPU-based acceleration have significantly advanced ANNS performance, the ever-growing scale of datasets now demands efficient multi-GPU solutions. However, the design of existing works overlooks multi-GPU scalability, resulting in naive approaches that treat additional GPUs as a means to extend memory capacity for large datasets. This inefficiency arises from partitioning the dataset and independently searching for data points similar to the queries in each GPU. We therefore propose PathWeaver, a novel multi-GPU framework designed to scale and accelerate ANNS for large datasets. First, we propose pipelining-based path extension, a GPU-aware pipelining mechanism that reduces prior work's redundant search iterations by leveraging GPU-to-GPU communication. Second, we design ghost staging that leverages a representative dataset to identify optimal query starting points, reducing the search space for challenging queries. Finally, we introduce direction-guided selection, a data selection technique that filters irrelevant points early in the search process, minimizing unnecessary memory accesses and distance computations. Comprehensive evaluations across diverse datasets demonstrate that PathWeaver achieves 3.24$\times$ geomean speedup and up to 5.30$\times$ speedup on 95% recall rate over state-of-the-art multi-GPU-based ANNS frameworks.

Figures

Figures reproduced from arXiv: 2507.17094 by the authors.

Figure 2
Figure 2. Execution time breakdown analysis of baseline [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance scalability of prior work on Sift-1M [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Illustration of the Pipelining-based path extension [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Execution time breakdown analysis with 4 GPUs [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Illustration of the direction-guided selection design. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Performance comparison on multi-GPU environment. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Performance comparison of Recall@10 between [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Performance comparison on a single GPU. Deep-50M has five times more vertices than that of Deep￾10M, the number of iterations until convergence is similar, which supports the motivation for pipelining-based path exten￾sion. However, having a wider vector per vertex in…
Figure 11
Figure 11. Figure 11: Ablation study of PathWeaver on (a) multi-GPU [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 13
Figure 13. Figure 13: Accuracy comparison on different number of itera [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]
Figure 14
Figure 14. Figure 14: Sensitivity study on the relationship between the [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 17
Figure 17. Figure 17: Graph build time overhead analysis. in [PITH_FULL_IMAGE:figures/full_fig_p011_17.png]
Figure 16
Figure 16. Figure 16: Comparison of neighbor selection strategies by [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]
Figure 18
Figure 18. Figure 18: Comparing ghost staging and GPU-based HNSW. [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 58 canonical work pages

  1. [1]

    https:// big-ann-benchmarks.com/neurips21.html

    Billion-Scale Approximate Nearest Neighbor Search Challenge: NeurIPS’21 competition track. https:// big-ann-benchmarks.com/neurips21.html

  2. [2]

    Chroma.https://trychroma.com/

  3. [3]

    MongoDB.https://www.mongodb.com/

  4. [4]

    https://www

    NVIDIA NVLink Bridge. https://www. nvidia.com/en-us/products/workstations/ nvlink-bridges/

  5. [5]

    Pinecone.https://www.pinecone.io/

  6. [6]

    Weaviate.https://weaviate.io/

  7. [7]

    https://docs.rapids.ai/api/ cuvs/stable/cuvs_bench/wiki_all_dataset/

    Wiki-all Dataset. https://docs.rapids.ai/api/ cuvs/stable/cuvs_bench/wiki_all_dataset/

  8. [8]

    Amsaleg and H

    L. Amsaleg and H. Jégou. Datasets for approximate nearest neighbor search. http://corpus-texmex. irisa.fr/

Show all 67 references
  1. [9]

    Accelerated Nearest Neighbor Search with Quick ADC

    Fabien André, Anne-Marie Kermarrec, and Nicolas Le Scouarnec. Accelerated Nearest Neighbor Search with Quick ADC. InInternational Conference on Multi- media Retrieval (ICMR), 2017

  2. [10]

    Efficient index- ing of billion-scale datasets of deep descriptors

    Artem Babenko and Victor Lempitsky. Efficient index- ing of billion-scale datasets of deep descriptors. InIEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2016

  3. [11]

    Connolly, Dimitri Lisin, Jimmy Briggs, and Mario E

    Nandan Banerjee, Ryan C. Connolly, Dimitri Lisin, Jimmy Briggs, and Mario E. Munich. View management for lifelong visual maps. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019

  4. [12]

    Finger: Fast inference for graph-based approximate nearest neighbor search

    Patrick Chen, Wei-Cheng Chang, Jyun-Yu Jiang, Hsiang- Fu Yu, Inderjit Dhillon, and Cho-Jui Hsieh. Finger: Fast inference for graph-based approximate nearest neighbor search. InACM Web Conference (WWW), 2023

  5. [13]

    Approximate Nearest Neighbor Search under Neural Similarity Metric for Large-Scale Recommendation

    Rihan Chen, Bin Liu, Han Zhu, Yaoxuan Wang, Qi Li, Buting Ma, Qingbo Hua, Jun Jiang, Yunlong Xu, Hongbo Deng, and Bo Zheng. Approximate Nearest Neighbor Search under Neural Similarity Metric for Large-Scale Recommendation. InACM International Conference on Information & Knowle...

  6. [14]

    Vector and line quanti- zation for billion-scale similarity search on GPUs.Fu- ture Generation Computer Systems, 2019

    Wei Chen, Jincai Chen, Fuhao Zou, Yuan-Fang Li, Ping Lu, Qiang Wang, and Wei Zhao. Vector and line quanti- zation for billion-scale similarity search on GPUs.Fu- ture Generation Computer Systems, 2019

  7. [15]

    Clarkson

    Kenneth L. Clarkson. An algorithm for approximate closest-point queries. InSymposium on Computational Geometry (SCG), 1994

  8. [16]

    Graph reordering for cache- efficient near neighbor search

    Benjamin Coleman, Santiago Segarra, Alex Smola, and Anshumali Shrivastava. Graph reordering for cache- efficient near neighbor search. InInternational Confer- ence on Neural Information Processing Systems (NIPS), 2022

  9. [17]

    Indexing by latent semantic analysis.Journal of the American society for information science, 1990

    Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. Indexing by latent semantic analysis.Journal of the American society for information science, 1990

  10. [18]

    The Faiss library.arXiv:2401.08281, 2024

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The Faiss library.arXiv:2401.08281, 2024

  11. [19]

    P. G. Emma. Understanding Some Simple Processor- performance Limits.IBM Journal of Research and De- velopment, 1997

  12. [20]

    Efanna : An extremely fast approximate nearest neighbor search algorithm based on knn graph.arXiv:1609.07228, 2016

    Cong Fu and Deng Cai. Efanna : An extremely fast approximate nearest neighbor search algorithm based on knn graph.arXiv:1609.07228, 2016

  13. [21]

    Fast approximate nearest neighbor search with the navi- gating spreading-out graph.Proceedings of the VLDB Endowment (VLDB), 2019

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. Fast approximate nearest neighbor search with the navi- gating spreading-out graph.Proceedings of the VLDB Endowment (VLDB), 2019

  14. [22]

    High-Dimensional Ap- proximate Nearest Neighbor Search: with Reliable and Efficient Distance Comparison Operations.Proceed- ings of the ACM on Management of Data (PACMMOD), 2023

    Jianyang Gao and Cheng Long. High-Dimensional Ap- proximate Nearest Neighbor Search: with Reliable and Efficient Distance Comparison Operations.Proceed- ings of the ACM on Management of Data (PACMMOD), 2023

  15. [23]

    Opti- mized Product Quantization for Approximate Nearest Neighbor Search

    Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. Opti- mized Product Quantization for Approximate Nearest Neighbor Search. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013

  16. [24]

    Filtered-DiskANN: Graph Algorithms for Approximate Nearest Neighbor Search with Filters

    Siddharth Gollapudi, Neel Karia, Varun Sivashankar, Ravishankar Krishnaswamy, Nikit Begwani, Swapnil Raz, Yiyong Lin, Yin Zhang, Neelam Mahapatro, Premkumar Srinivasan, Amit Singh, and Harsha Vard- han Simhadri. Filtered-DiskANN: Graph Algorithms for Approximate Nearest Neighb...

  17. [25]

    Fabian Groh, Lukas Ruppert, Patrick Wieschollek, and Hendrik P. A. Lensch. GGNN: Graph-Based GPU Near- est Neighbor Search.IEEE Transactions on Big Data, 2023

  18. [26]

    Ac- celerating large-scale inference with anisotropic vector quantization

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. Ac- celerating large-scale inference with anisotropic vector quantization. InInternational Conference on Machine Learning (ICML), 2020

  19. [27]

    Query-aware locality-sensitive hashing for approximate nearest neighbor search.Proceedings of the VLDB Endowment (VLDB), 2015

    Qiang Huang, Jianlin Feng, Yikai Zhang, Qiong Fang, and Wilfred Ng. Query-aware locality-sensitive hashing for approximate nearest neighbor search.Proceedings of the VLDB Endowment (VLDB), 2015

  20. [28]

    Approximate nearest neighbors: towards removing the curse of dimension- ality

    Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of dimension- ality. InACM Symposium on Theory of Computing (STOC), 1998

  21. [29]

    OOD-DiskANN: Efficient and Scal- able Graph ANNS for Out-of-Distribution Queries

    Shikhar Jaiswal, Ravishankar Krishnaswamy, Ankit Garg, Harsha Vardhan Simhadri, and Sheshansh Agrawal. OOD-DiskANN: Efficient and Scal- able Graph ANNS for Out-of-Distribution Queries. arXiv:2211.12850, 2022

  22. [30]

    CXL-ANNS: Software-Hardware Collaborative Mem- ory Disaggregation and Computation for Billion-Scale Approximate Nearest Neighbor Search

    Junhyeok Jang, Hanjin Choi, Hanyeoreum Bae, Se- ungjun Lee, Miryeong Kwon, and Myoungsoo Jung. CXL-ANNS: Software-Hardware Collaborative Mem- ory Disaggregation and Computation for Billion-Scale Approximate Nearest Neighbor Search. InUSENIX Annual Technical Conference (ATC), 2023

  23. [31]

    Billion- Scale Similarity Search with GPUs.IEEE Transactions on Big Data, 2021

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion- Scale Similarity Search with GPUs.IEEE Transactions on Big Data, 2021

  24. [32]

    Product quantization for nearest neighbor search.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2011

    Hervé Jégou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2011

  25. [33]

    Locally Opti- mized Product Quantization for Approximate Nearest Neighbor Search

    Yannis Kalantidis and Yannis Avrithis. Locally Opti- mized Product Quantization for Approximate Nearest Neighbor Search. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014

  26. [34]

    Mining visual phrases for long-term visual SLAM

    Tanaka Kanji, Chokushi Yuuto, and Ando Masatoshi. Mining visual phrases for long-term visual SLAM. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2014

  27. [35]

    Generalization through Memorization: Nearest Neighbor Language Models

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through Memorization: Nearest Neighbor Language Models. In International Conference on Learning Representations (ICLR), 2020

  28. [36]

    Accelerating large-scale graph-based nearest neighbor search on a computational storage platform.IEEE Transactions on Computers, 2023

    Ji-Hoon Kim, Yeo-Reum Park, Jaeyoung Do, Soo- Young Ji, and Joo-Young Kim. Accelerating large-scale graph-based nearest neighbor search on a computational storage platform.IEEE Transactions on Computers, 2023

  29. [37]

    Andersen, and Yuxiong He

    Conglong Li, Minjia Zhang, David G. Andersen, and Yuxiong He. Improving Approximate Nearest Neighbor Search through Learned Adaptive Early Termination. InInternational Conference on Management of Data (SIGMOD), 2020

  30. [38]

    Approximate Nearest Neighbor Search on High Dimensional Data — Experi- ments, Analyses, and Improvement .IEEE Transactions on Knowledge & Data Engineering, 2020

    Wen Li, Ying Zhang, Yifang Sun, Wei Wang, Mingjie Li, Wenjie Zhang, and Xuemin Lin. Approximate Nearest Neighbor Search on High Dimensional Data — Experi- ments, Analyses, and Improvement .IEEE Transactions on Knowledge & Data Engineering, 2020

  31. [39]

    Ting Liu, Charles Rosenberg, and Henry A. Rowley. Clustering Billions of Images with Large Scale Nearest Neighbor Search. InIEEE Workshop on Applications of Computer Vision (WACV), 2007

  32. [40]

    Malkov and D

    Yu A. Malkov and D. A. Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020

  33. [41]

    Approximate nearest neigh- bor algorithm based on navigable small world graphs

    Yury Malkov, Alexander Ponomarenko, Andrey Logvi- nov, and Vladimir Krylov. Approximate nearest neigh- bor algorithm based on navigable small world graphs. Information Systems, 2014

  34. [42]

    ParlayANN: Scalable and Deterministic Parallel Graph-Based Approximate Nearest Neighbor Search Algorithms

    Magdalen Dobson Manohar, Zheqi Shen, Guy Blelloch, Laxman Dhulipala, Yan Gu, Harsha Vardhan Simhadri, and Yihan Sun. ParlayANN: Scalable and Deterministic Parallel Graph-Based Approximate Nearest Neighbor Search Algorithms. InACM SIGPLAN Annual Sympo- sium on Principles and Pr...

  35. [43]

    Nene and Shree K

    Sameer A. Nene and Shree K. Nayar. A Simple Algo- rithm for Nearest Neighbor Search in High Dimensions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1997

  36. [44]

    CAGRA: Highly Parallel Graph Construction and Approximate Nearest Neighbor Search for GPUs

    Hiroyuki Ootomo, Akira Naruse, Corey Nolet, Ray Wang, Tamas Feher, and Yong Wang. CAGRA: Highly Parallel Graph Construction and Approximate Nearest Neighbor Search for GPUs. InIEEE International Con- ference on Data Engineering (ICDE), 2024

  37. [45]

    LM-DiskANN: Low Memory Footprint in Disk-Native Dynamic Graph- Based ANN Indexing

    Yu Pan, Jianxin Sun, and Hongfeng Yu. LM-DiskANN: Low Memory Footprint in Disk-Native Dynamic Graph- Based ANN Indexing. InIEEE International Confer- ence on Big Data (BigData), 2023

  38. [46]

    Least squares binary quantization of neural networks

    Hadi Pouransari, Zhucheng Tu, and Oncel Tuzel. Least squares binary quantization of neural networks. In IEEE/CVF conference on computer vision and pattern recognition workshops (CVPRW), 2020

  39. [47]

    HM-ANN: effi- cient billion-point nearest neighbor search on heteroge- neous memory

    Jie Ren, Minjia Zhang, and Dong Li. HM-ANN: effi- cient billion-point nearest neighbor search on heteroge- neous memory. InInternational Conference on Neural Information Processing Systems (NeurIPS), 2020

  40. [48]

    Optimised KD-trees for fast image descriptor matching

    Chanop Silpa-Anan and Richard Hartley. Optimised KD-trees for fast image descriptor matching. InIEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2008

  41. [49]

    FreshDiskANN: A Fast and Accurate Graph- Based ANN Index for Streaming Similarity Search

    Aditi Singh, Suhas Jayaram Subramanya, Ravis- hankar Krishnaswamy, and Harsha Vardhan Simhadri. FreshDiskANN: A Fast and Accurate Graph- Based ANN Index for Streaming Similarity Search. arXiv:2105.09613, 2021

  42. [50]

    DiskANN: fast accurate billion-point nearest neighbor search on a single node

    Suhas Jayaram Subramanya, Devvrit, Rohan Kadekodi, Ravishankar Krishaswamy, and Harsha Vardhan Simhadri. DiskANN: fast accurate billion-point nearest neighbor search on a single node. InInternational Conference on Neural Information Processing Systems (NeurIPS), 2019

  43. [51]

    Scalable billion-point approximate nearest neighbor search using SmartSSDs

    Bing Tian, Haikun Liu, Zhuohui Duan, Xiaofei Liao, Hai Jin, and Yu Zhang. Scalable billion-point approximate nearest neighbor search using SmartSSDs. InUSENIX Annual Technical Conference (ATC), 2024

  44. [52]

    FusionANNS: An Efficient CPU/GPU Cooperative Processing Architecture for Billion-scale Approximate Nearest Neighbor Search

    Bing Tian, Haikun Liu, Yuhang Tang, Shihai Xiao, Zhuohui Duan, Xiaofei Liao, Xuecang Zhang, Jun- hua Zhu, and Yu Zhang. FusionANNS: An Efficient CPU/GPU Cooperative Processing Architecture for Billion-scale Approximate Nearest Neighbor Search. arXiv:2409.16576, 2024

  45. [53]

    BANG: Billion-Scale Approximate Nearest Neighbor Search using a Single GPU.arXiv:2401.11324, 2024

    Karthik V ., Saim Khan, Somesh Singh, Harsha Vardhan Simhadri, and Jyothi Vedurada. BANG: Billion-Scale Approximate Nearest Neighbor Search using a Single GPU.arXiv:2401.11324, 2024

  46. [54]

    Gonçalves, Zanoni Dias, and Ricardo da S

    Javier Vargas Muñoz, Marcos A. Gonçalves, Zanoni Dias, and Ricardo da S. Torres. Hierarchical clustering- based graphs for large scale approximate nearest neigh- bor search.Pattern Recognition, 2019

  47. [55]

    Milvus: A Purpose- Built Vector Data Management System

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, et al. Milvus: A Purpose- Built Vector Data Management System. InInternational Conference on Management of Data (SIGMOD), 2021

  48. [56]

    A comprehensive survey and experimental comparison of graph-based approximate nearest neigh- bor search.Proceedings of the VLDB Endowment (VLDB), 2021

    Mengzhao Wang, Xiaoliang Xu, Qiang Yue, and Yuxi- ang Wang. A comprehensive survey and experimental comparison of graph-based approximate nearest neigh- bor search.Proceedings of the VLDB Endowment (VLDB), 2021

  49. [57]

    NDSEARCH: Accelerating Graph-Traversal- Based Approximate Nearest Neighbor Search through Near Data Processing

    Yitu Wang, Shiyu Li, Qilin Zheng, Linghao Song, Zong- wang Li, Andrew Chang, Hai “Helen” Li, and Yiran Chen. NDSEARCH: Accelerating Graph-Traversal- Based Approximate Nearest Neighbor Search through Near Data Processing. InACM/IEEE Annual Interna- tional Symposium on Computer ...

  50. [58]

    Xu, Uri Alon, and Graham Neubig

    Frank F. Xu, Uri Alon, and Graham Neubig. Why do nearest neighbor language models work? InInterna- tional Conference on Machine Learning (ICML), 2023

  51. [59]

    Proxima: Near-storage Acceleration for Graph-based Approximate Nearest Neighbor Search in 3D NAND.arXiv:2312.04257, 2023

    Weihong Xu, Junwei Chen, Po-Kai Hsu, Jaeyoung Kang, Minxuan Zhou, Sumukh Pinge, Shimeng Yu, and Ta- jana Rosing. Proxima: Near-storage Acceleration for Graph-based Approximate Nearest Neighbor Search in 3D NAND.arXiv:2312.04257, 2023

  52. [60]

    Proximity graph maintenance for fast on- line nearest neighbor search.arXiv:2206.10839, 2022

    Zhaozhuo Xu, Weijie Zhao, Shulong Tan, Zhixin Zhou, and Ping Li. Proximity graph maintenance for fast on- line nearest neighbor search.arXiv:2206.10839, 2022

  53. [61]

    PECANN: Parallel Efficient Clustering with Graph-Based Approximate Nearest Neighbor Search

    Shangdi Yu, Joshua Engels, Yihao Huang, and Julian Shun. PECANN: Parallel Efficient Clustering with Graph-Based Approximate Nearest Neighbor Search. arXiv:2312.03940, 2023

  54. [62]

    Gpu-accelerated proximity graph approximate nearest neighbor search and construc- tion

    Yuanhang Yu, Dong Wen, Ying Zhang, Lu Qin, Wen- jie Zhang, and Xuemin Lin. Gpu-accelerated proximity graph approximate nearest neighbor search and construc- tion. InIEEE International Conference on Data Engi- neering (ICDE), 2022

  55. [63]

    DF-GAS: a Distributed FPGA-as-a-Service Architecture towards Billion-Scale Graph-based Approximate Nearest Neigh- bor Search

    Shulin Zeng, Zhenhua Zhu, Jun Liu, Haoyu Zhang, Guo- hao Dai, Zixuan Zhou, Shuangchen Li, Xuefei Ning, Yuan Xie, Huazhong Yang, and Yu Wang. DF-GAS: a Distributed FPGA-as-a-Service Architecture towards Billion-Scale Graph-based Approximate Nearest Neigh- bor Search. InIEEE/ACM...

  56. [64]

    Grasp: Optimizing graph-based nearest neighbor search with subgraph sampling and pruning

    Minjia Zhang, Wenhan Wang, and Yuxiong He. Grasp: Optimizing graph-based nearest neighbor search with subgraph sampling and pruning. InACM International Conference on Web Search and Data Mining (WSDM), 2022

  57. [65]

    SONG: Ap- proximate Nearest Neighbor Search on GPU

    Weijie Zhao, Shulong Tan, and Ping Li. SONG: Ap- proximate Nearest Neighbor Search on GPU. InIEEE International Conference on Data Engineering (ICDE), 2020

  58. [66]

    Guitar: Gradient pruning toward fast neural ranking

    Weijie Zhao, Shulong Tan, and Ping Li. Guitar: Gradient pruning toward fast neural ranking. InInternational ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2024

  59. [67]

    Processing-in- hierarchical-memory architecture for billion-scale ap- proximate nearest neighbor search

    Zhenhua Zhu, Jun Liu, Guohao Dai, Shulin Zeng, Bing Li, Huazhong Yang, and Yu Wang. Processing-in- hierarchical-memory architecture for billion-scale ap- proximate nearest neighbor search. InACM/IEEE De- sign Automation Conference (DAC), 2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.