Pith. sign in

REVIEW 5 major objections 6 minor 56 references

RVANNS: Mixed-Precision Indexing and Locality-Aware Graph Traversal on RISC-V

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A RISC-V ANNS engine claims 3.4–4.9x speedups by co-designing mixed-precision decoding and graph layout.

desk verdict Serious RVV ANNS engineering with a reproducible ROrder problem: Algorithm 3 does not implement the stated locality score. read the letter →

arxiv 2608.09077 v1 pith:6V7OAQKM submitted 2026-08-10 cs.IR

Chengying Huan , Yudong Liu , Jianguo Wang , Lizheng Chen , Renling Yin , Weijia Chen , Ji Qi , Jiageng Yu
show 4 more authors
Junjie Xu Jie Zhang Chen Tian Yanjun Wu
This is my paper · ORCID
classification cs.IR
keywords approximatenearestneighborsearchRISC-VVectorExtensionmixed-precisionindexinggraphreorderingmemory-levelparallelismvector-length-agnosticexecutionHNSWSIMDdecoding
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

Approximate nearest neighbor search on CPUs is limited less by arithmetic speed than by moving and decoding candidate vectors, and the paper argues this is especially true on RISC-V vector processors. It claims that two co-designed mechanisms—MPMI, a mixed-precision index that stores each vector as a dense 8-bit affine base plus sparse FP16/FP32 residuals, and ROrder, a traversal-aware graph reordering—turn the RISC-V Vector Extension's register grouping and length-agnostic loops into large end-to-end speedups. On real 128-bit and 256-bit RISC-V vector hardware, the engine reports 3.39x and 4.94x speedups over scalar execution, and 2.27–2.76x over a SIMD+FP32 HNSW baseline at Recall@100=0.99. A sympathetic reader would care because it identifies the memory-side bottlenecks that remain after vectorization and offers a portable recipe for addressing them.

What carries the argument

The load-bearing machinery is the pair (MPMI, ROrder) executed on an RVV kernel layer. MPMI is a mixed-precision multi-layer index: each vector is a dense 8-bit affine base $b$ plus sparse FP16 and FP32 residual pools selected by per-dimension error, and the decoder streams the base, gathers selected residuals via bitmap/prefix lookup, and fuses reconstruction with $vfmacc$ accumulation in LMUL-grouped tiles. ROrder is a locality transformation: it scores node pairs by $S(u,v)=|N^-(u)\cap N^-(v)|+1[(u,v)\in E]+1[(v,u)\in E]$, builds a permutation with a windowed greedy algorithm, co-locates node payloads in 64-byte-aligned rows, and sorts adjacency lists by permuted IDs so traversals become address-monotone. The RVV kernel layer supplies vsetvl-based variable-length execution and LMUL grouping that translate these two mechanisms into throughput.

What would settle it

Measure ROrder's QPS gain on a graph whose traversal order is decoupled from topology, e.g., random edge rewiring that preserves degrees, or compare traversal-log co-visitation against the S(u,v) score; if the proxy does not track actual co-visits, the gain should shrink or vanish, and if permuted sorting changes Recall@100 at fixed efSearch, the claim that reordering is outcome-neutral fails.

Watch

Extended reading notes

Core claim

The central discovery is that ANNS on RVV is memory-bound in two specific ways—the decode path for low-precision vectors dilutes RVV's register-group widening, and irregular graph traversal scatters payload accesses enough to neutralize LMUL scaling—and that both can be fixed by co-designing representation and layout. MPMI encodes each vector as a dense 8-bit affine base with sparse FP16/FP32 corrections, so reconstruction and distance accumulation fuse in one LMUL-aligned loop without materializing an FP32 vector. ROrder permutes graph nodes so that likely co-visited nodes get adjacent IDs and sorts each adjacency list by the new IDs, making payload probes predominantly forward-moving address streams. Together they deliver the speedups above, with the largest gains on RVV because vsetvl-based variable-length execution and LMUL register groups convert reduced memory traffic and better locality into sustained throughput.

Load-bearing premise

The load-bearing premise is that shared in-neighbors plus direct adjacency reliably predicts which HNSW nodes are actually visited together during beam search, and that permuting node IDs and sorting adjacency lists leaves the search outcome unchanged; the 18.77% QPS gain from ROrder rests on this proxy.

Editorial extensions

If this is right

  • If the central claim is right, a single RVV binary can get portable acceleration across 128-bit and 256-bit vector lengths, with MPMI+ROrder adding 2.76x–3.63x over each platform's SIMD+FP32 baseline.
  • The same memory-side mechanisms transfer partially to SVE and AVX-512, yielding 1.18x–1.59x over their SIMD+FP32 HNSW baselines, so the techniques are not RISC-V-exclusive.
  • Index memory footprint drops by roughly 58–72% versus NEON HNSW baselines because MPMI reduces per-candidate bytes, shifting the bottleneck away from DRAM traffic.
  • On a 10M-scale dataset, the engine reports higher QPS per watt than GPU baselines, suggesting CPU-resident vector search can be the more power-efficient operating point at high recall.

Reading between the lines

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

  • The locality proxy could be checked directly: if one recorded actual co-visitation during beam search and compared it with the shared-in-neighbor score, ROrder's 18.77% gain would stand or fall on that match; the paper does not provide such a traversal-log validation.
  • ROrder's gains should shrink on workloads where beam search drifts from topology, e.g., query distributions that visit nodes adversarially; testing with adversarial or clustered query sets would reveal how general the reordering is.
  • The same fused-decode idea could be extended to learned compression or product quantization, but that would require gather-heavy codebook lookups that MPMI deliberately avoids, so the extension is not immediate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. RVANNS targets approximate nearest neighbor search on RISC-V vector processors. It contributes two co-designed mechanisms: MPMI, a mixed-precision representation with a dense 8-bit affine base and sparse FP16/FP32 residuals aligned to RVV LMUL widening, and ROrder, a traversal-aware graph permutation plus adjacency-list sorting intended to create forward-moving payload streams. The paper reports within-platform speedups up to 3.39x and 4.94x over scalar on SG2044 and BPI-F3, 2.27-2.76x over SIMD+FP32 HNSW on RVV, cross-ISA transferability results on SVE/AVX-512, and a GPU energy-efficiency comparison on Cohere10M.

Significance. At a high level the paper makes a useful contribution: it is a serious, hardware-driven co-design of an ANNS engine for RVV, with unusually detailed PMU-based motivation and within-platform ablations that isolate VLA execution, LMUL sensitivity, MPMI versus conventional SQ8, and ROrder versus generic graph orderings. The same binary running on 128-bit and 256-bit RVV platforms is a nice demonstration of RVV VLA portability. If the mechanisms are as described, the results are of interest to both the ANNS and RISC-V systems communities.

major comments (5)
  1. [Sec. 5.1 / Algorithm 3] Section 5.1 defines the objective S(u,v)=|N^-(u)∩N^-(v)| + 1[(u,v)∈E] + 1[(v,u)∈E] and states that the greedy selects each next node by maximizing the window sum of S. Algorithm 3, however, initializes s[v]=|N^-(v)| for every node (line 5), so every ExtractMax is biased by raw in-degree in addition to the windowed S terms. Moreover, IncrUpdate (lines 17-19) updates only out-neighbors of the entering node and out-neighbors of its in-neighbors; it adds the shared-in-neighbor term and the forward-adjacency term, but it never adds the reverse-adjacency term 1[(·,v)∈E] for unplaced in-neighbors of the node just placed. The pseudocode therefore does not implement the greedy described in the text, and the results in Table 13 (including the 18.77% gain over the unordered layout and the 9.97% gain over GOrder) were obtained with an undocumented scoring rule. Please correct the algorithm description or describe and analyze the actual score, and state which variant was used for the reported measurements.
  2. [Sec. 4.3 / Algorithm 2] Section 4.3 describes residual fusion as "vluxei* loads the selected residuals, vrgather packs the corresponding base lanes, and vsuxei* writes the corrected values into a temporary buffer; the downstream distance kernel then reloads this buffer," which implies a materialized reconstructed tile. In contrast, Algorithm 2 (lines 24 and 26) applies a masked vfadd and then accumulates with vfmacc "no store," and the text claims decoding "without materializing a persistent full FP32 vector." This is an internal inconsistency in the description of the central MPMI fused-decode mechanism; the authors should clarify the actual data path, since the fused versus materializing distinction is exactly what MPMI's performance claims rely on.
  3. [Sec. 7.5 / Table 12] The text states that MPMI delivers 1.35x higher QPS than HNSW and 3.43x higher than IVF on SIFT1M, and 1.20x and 9.12x on GIST1M, but Table 12 reports +91.38% (≈1.91x) and +230.14% (≈3.30x) for SIFT1M, and +124.14% (≈2.24x) and +751.02% (≈8.51x) for GIST1M. Please reconcile the table and the prose; as published, the text numbers are inconsistent with the data.
  4. [Sec. 7.10 / Table 16] The GPU comparison measures SG2044 platform power over PSU1+PSU2 with ipmitool, while the GPU baselines report NVIDIA device power. Comparing these numbers to compute QPS/W makes the 1.82-2.27x energy-efficiency claim not a like-for-like comparison, because platform power includes DRAM and other system components that device power excludes. Please either measure SG2044 device/package power, report both scopes transparently, or restrict the claim to the power scope actually measured.
  5. [Sec. 5.2] The claim that sorting each adjacency list by remapped IDs "changes layout within each neighbor list, not the underlying graph, so recall remains unchanged under fixed HNSW parameters" is not guaranteed: HNSW beam search visits neighbors in stored order, and reordering a neighbor list can change the order in which candidates are evaluated, which can alter the candidate set at a fixed ef_search. The paper provides no ablation showing recall invariance; please provide such a verification or qualify the claim.
minor comments (6)
  1. [Fig. 2 / Algorithm 2] The figure and Algorithm 2 label the base as "FP8" while the text and Equation (1) describe a dense 8-bit affine base with integer byte values; please use a consistent name to avoid confusion with IEEE FP8 formats.
  2. [Table 4] The sequential LMUL=1 run (344.3M cycles) is slower than the irregular LMUL=1 run (289.7M cycles); please explain why the two microbenchmarks are not directly comparable or what accounts for the reversal.
  3. [Algorithm 3] Line 11 applies IncrUpdate(P[i-w],-1) for the node leaving the window; please state whether leaving-node effects on scores of already-placed nodes are intentionally ignored, since the described windowed objective only sums over the current window.
  4. [References] Reference [4] is a survey on FPGAs but is cited to support graph processing on vector processors; this citation appears mismatched.
  5. [sec. 6.1 / Artifacts] The paper mentions a released artifact with 58,082 SLOC but gives no repository or DOI; please include the artifact link.
  6. [Fig. 8] The axis label in Figure 8(a) renders the thresholds as "Precision Threshold ( 32, 16)" with missing Greek letters and units; please fix the formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RVANNS speedups are measured against independent baselines, and the design motivations are not reused as evidence of the results.

full rationale

RVANNS's headline speedups are measured on real hardware against independent baselines (Milvus/FAISS, cuVS, GOrder, RabbitOrder, RCM, AVX-512/SVE/NEON); no headline number is obtained by substituting a fitted constant into the paper's equations. MPMI's rate-distortion thresholds (τ32, τ16) are tuned, but the reported gains are end-to-end QPS measurements, not predictions computed from the thresholds. ROrder's locality score S(u,v) is a topology-defined heuristic, not fitted to traversal logs, and the 18.77% QPS gain is presented as a measurement of the resulting layout rather than as a consequence of maximizing S. The PMU profiling on SG2044 motivates the design but is not used as evidence that the final speedups are achieved—the speedups are separately measured. The only self-citation (Milvus [43]) identifies the integration target and is not load-bearing. The noted discrepancy between Sec. 5.1's score definition and Algorithm 3's update rule is an internal consistency/reproducibility concern, not a circular derivation, and it does not change the circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central performance claims rest on measured engineering, so the ledger is dominated by tuning parameters, not invented physics. The main soft spot is the empirical proxy linking graph topology to traversal co-visitation.

free parameters (3)
  • MPMI residual thresholds (tau32, tau16) = tau32=1%, tau16=10% of dimensions
    Chosen from sensitivity sweep in Fig 8(a) as the highest compression with recall loss under 0.002; they control per-vector code size (D + 2|S16| + 4|S32|) and are a tunable knob.
  • ROrder window w = 5
    Selected after sensitivity sweep (Fig 8(b)); larger windows increase offline cost linearly with diminishing QPS gains; w=5 reaches 99.2% of best observed QPS.
  • LMUL for decode/distance kernel = 4
    Chosen from Table 11 as the best-performing register-group multiplier on SG2044; LMUL=8 is slower due to register pressure.
assumptions (4)
  • domain assumption HNSW beam search visits nodes with shared in-neighbors or direct adjacency within a short time window
    Used to justify ROrder's additive locality score in Sec 5.1. No traversal-log data is shown to verify the correlation.
  • domain assumption Sorting adjacency lists by permuted IDs does not change the set of visited candidates
    Stated in Sec 5.2 as leaving recall unchanged; in priority-queue traversal, insertion order can affect tie-breaking, so this is not a formal guarantee.
  • domain assumption The PMU profile on SG2044 is representative of the RISC-V vector processors RVANNS targets
    Used to derive the two design principles in Sec 2.3; only two RVV implementations are tested (SG2044 and BPI-F3).
  • domain assumption Base-8 affine quantization of each dimension with top-tau residual correction preserves recall
    Validated empirically on four datasets; no error bound is derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RVANNS: Mixed-Precision Indexing and Locality-Aware Graph Traversal on RISC-V." pith.science (2026). https://pith.science/paper/6V7OAQKM

@misc{pith2026260809077,
  author       = {Pith},
  title        = {Pith review of: RVANNS: Mixed-Precision Indexing and Locality-Aware Graph Traversal on RISC-V},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6V7OAQKM}},
  note         = {Machine review of arXiv:2608.09077}
}
read the original abstract

Approximate nearest neighbor search (ANNS) on CPUs is increasingly constrained by candidate-vector movement and decoding rather than peak arithmetic throughput. Although the RISC-V Vector Extension (RVV) provides vector-length-agnostic execution and LMUL-based register grouping, generic low-precision decoding still incurs conversion overhead, while irregular graph traversal generates scattered accesses that degrade cache locality and memory-level parallelism. We present RVANNS, an RVV-oriented ANNS engine that jointly optimizes vector representation and graph locality. Its Mixed-Precision Multi-Layer Index (MPMI) represents each vector with a dense 8-bit affine base and sparse FP16/FP32 residuals, fusing reconstruction with distance accumulation and aligning widening with LMUL-sized register groups. ROrder co-locates likely co-visited graph nodes and sorts remapped adjacency lists, transforming scattered payload probes into denser, predominantly forward-moving address streams. Integrated into Milvus, RVANNS achieves 3.39x and 4.94x speedups over scalar execution on real 128-bit and 256-bit RVV processors, respectively. Under controlled HNSW configurations, it improves throughput by 2.27--2.76x over RVV SIMD+FP32 and by 1.18--1.59x over the corresponding AVX-512 and SVE baselines. On Cohere10M, it further delivers 1.82--2.27x higher QPS/W than the evaluated GPU baselines.

Figures

Figures reproduced from arXiv: 2608.09077 by the authors.

Figure 1
Figure 1. ANNS query latency and stall profile on SG2044. (A) Query-phase breakdown of graph-based ANNS. (B) PMU￾based stalled-cycle breakdown of distance evaluation. (C) PMU counters with HNSW on GIST1M. IPC denotes instruc￾tions per cycle; FE/BE stall and MPKI report frontend/back￾end stalls and cache/branch misses. Compression Choice for SIMD ANNS: SQ over PQ. As noted in FLASH [44], SQ is generally more suitable than prod… view at source ↗
Figure 2
Figure 2. System architecture of RVANNS. The offline phase constructs MPMI for register-group-aligned widening and applies ROrder for address-monotone memory access. During online search, the shared RVV kernel consumes the reordered graph and mixed-precision payloads through fused decode-and-distance execution, translating both into higher query throughput. 3 RVANNS Engine Framework 3.1 Overview Guided by the design principle… view at source ↗
Figure 3
Figure 3. Mixed-Precision Multi-Layer Index (MPMI). (A) Sensitivity Analysis: High-error (MSE) dimensions are as￾signed to the FP32 or FP16 residual tier. (B) Logical Model: A cache-aware pyramid with a dense 8-bit affine base and sparse FP16/FP32 residuals. (C) Physical Layout and LMUL￾Aligned RVV Decode: the contiguous 8-bit affine base is widened from an LMUL=1 base group to an LMUL=4 FP32 decode group, with sparse FP16/FP… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: QPS against Arm SVE and x86 AVX-512 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Index memory footprint comparison. RISC-V x86 AVX-512 Arm SVE 1 2 4 8 16 32 64 Thread 0.2 0.4 0.6 0.8 1.0 Normalized QPS/core 0.814 0.235 0.316 (a) Thread Scalability 1 2 4 8 16 32 64 Thread 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 L2 Cache MPKI +3.1% +2.6% +30.1% (b) Cache…
Figure 7
Figure 7. Figure 7: Cross-ISA thread scalability on GIST1M. (a) Nor￾malized per-core throughput, 𝑄 (𝑇 ) 𝑇 ·𝑄 (1) ; (b) Cache miss, L2D MPKI. (c) Microarchitectural efficiency, IPC. decode and FP32 accumulation map to LMUL-grouped widen￾ing, while ROrder supplies a more locality-friendly s…
Figure 8
Figure 8. Figure 8: shows robust, near-optimal performance across broad ranges of ROrder window 𝑤 and MPMI thresholds (𝜏32, 𝜏16). ROrder Window Size Sensitivity. ROrder uses a win￾dowed greedy procedure; larger 𝑤 expands the search space and increases offline cost roughly linearly [PITH_…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 38 canonical work pages

  1. [1]

    Fabien André, Anne-Marie Kermarrec, and Nicolas Le Scouarnec. 2021. Quicker ADC: Unlocking the Hidden Potential of Product Quantiza- tion with SIMD.IEEE Transactions on Pattern Analysis and Machine Intelligence43, 5 (2021), 1666–1677.https://doi.org/10.1109/TPAMI. 2019.2952606

    arXiv 2021

  2. [2]

    Junya Arai, Hiroyuki Shiokawa, Takuya Yamamuro, Masashi Onizuka, and Sotetsu Iwamura. 2016. Rabbit Order: Just-in-Time Parallel Re- ordering for Fast Graph Analysis. In2016 IEEE International Parallel and Distributed Processing Symposium (IPDPS). IEEE, 22–31.https: //doi.org/10.1109/IPDPS.2016.15

    work page doi:10.1109/ipdps.2016.15 2016

  3. [3]

    Artem Babenko and Victor Lempitsky. 2016. Efficient indexing of billion-scale datasets of deep descriptors. InProceedings of the IEEE 14 Conference on Computer Vision and Pattern Recognition. 2055–2063

    work page 2016

  4. [4]

    Maciej Besta, Michał Podstawski, Linus Groner, Edgar Solomonik, and Torsten Hoefler. 2019. Graph processing on FPGAs: Taxonomy, survey, challenges.CoRRabs/1903.06697 (2019).https://arxiv.org/abs/1903. 06697

    work page Pith review arXiv 2019

  5. [5]

    Alina Beygelzimer, Sham Kakade, and John Langford. 2006. Cover trees for nearest neighbor. InProceedings of the 23rd international conference on Machine learning. 97–104

    2006

  6. [6]

    Sung-Hyuk Cha. 2002. Comprehensive survey on distance/similarity measures between probability density functions.International Journal of Mathematical Models and Methods in Applied Sciences1, 4 (2002), 300–307

    work page 2002

  7. [7]

    Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Q Yan, Meghan Cowan, Haichen Shen, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. 2018. TVM: An automated end-to-end optimizing compiler for deep learning. In13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18). 578–594

    work page 2018

  8. [8]

    Cohere AI. 2023. Cohere Embed v3: Large-Scale Text Embedding Model.https://cohere.com/embedCohere Embed v3 generates 768- dimensional embeddings for text retrieval

    work page 2023

Show all 56 references
  1. [9]

    Benjamin Coleman, Santiago Segarra, Anshumali Shrivastava, and Alex Smola. 2021. Graph Reordering for Cache-Efficient Near Neighbor Search. arXiv:2104.03221 [cs.LG]https://arxiv.org/abs/2104.03221

    2021 arXiv

  2. [10]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  3. [11]

    Elizabeth Cuthill and James McKee. 1969. Reducing the bandwidth of sparse symmetric matrices. InProceedings of the 1969 24th national conference. ACM, 157–172

    1969

  4. [12]

    Leonardo Dagum and Ramesh Menon. 1998. OpenMP: an industry standard API for shared-memory programming.IEEE Computational Science and Engineering5, 1 (1998), 46–55

    1998

  5. [13]

    Arnaldo Carvalho De Melo. 2010. The new Linux ’perf’tools. InSlides from Linux Kongress, Vol. 18. 1–42

    2010

  6. [14]

    Agner Fog. 2016. Instruction tables: Lists of instruction latencies, throughputs and micro-operation breakdowns for Intel, AMD and VIA CPUs.Copenhagen University College of Engineering93 (2016)

    2016

  7. [15]

    Free Software Foundation. 2025. GCC, the GNU compiler collection. Free Software Foundation(2025). Version 15.1

    2025

  8. [16]

    Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized product quantization for approximate nearest neighbor search. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2946–2953

    2013

  9. [17]

    Yunchao Gong, Svetlana Lazebnik, Albert Gordo, and Florent Per- ronnin. 2012. Iterative quantization: A procrustean approach to learn- ing binary codes for large-scale image retrieval.IEEE Transactions on Pattern Analysis and Machine Intelligence35, 12 (2012), 2916–2929

    2012

  10. [18]

    Jorge González-Vázquez, Salvador Bermúdez, and José María García

  11. [19]

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. 2020. Accelerating large-scale inference with anisotropic vector quantization. InInternational Conference on Machine Learning. PMLR, 3887–3896

    2020

  12. [20]

    Guseul Heo, Sangyeop Lee, Jaehong Choi, Minsu Cho, Hyunmin Kim, Sanghyeon Park, Hyungkyu Yoo, Gwangsun Kim, Darrel Kwon, and Jongse Kim. 2024. NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing. InProceedings of the 29th ACM International Conference on Arc...

    2024

  13. [21]

    Piotr Indyk and Rajeev Motwani. 1998. Approximate nearest neigh- bors: towards removing the curse of dimensionality. InProceedings of the thirtieth annual ACM symposium on Theory of computing. 604–613

    1998

  14. [22]

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnawamy, and Rohan Kadekodi. 2019. DiskANN: Fast accurate billion-point nearest neighbor search on a single node. InAdvances in Neural Information Processing Systems, Vol. 32

    2019

  15. [23]

    Hervé Jégou, Matthijs Douze, and Cordelia Schmid. 2011. Product Quantization for Nearest Neighbor Search.IEEE Transactions on Pat- tern Analysis and Machine Intelligence33, 1 (2011), 117–128.https: //doi.org/10.1109/TPAMI.2010.57

    2011 doi

  16. [24]

    Hervé Jégou, Romain Tavenard, Matthijs Douze, and Laurent Amsaleg

  17. [25]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2017. Faiss: A library for efficient similarity search. In2017 IEEE International Conference on Big Data (Big Data). IEEE, 3488–3491

    2017

  18. [26]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data7, 3 (2019), 535–547.https://doi.org/10.1109/TBDATA.2019.2921572

    2019

  19. [27]

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)....

    2020

  20. [28]

    Conglong Li, Minjia Xu, Xin Zhang, and Bolin He. 2020. GRAF: A graph-based approach to fine-grained approximate nearest neighbor search.arXiv preprint arXiv:2003.00863(2020)

    2020 arXiv

  21. [29]

    Yiwei Li, Yuxin Jin, Boyu Tian, Huanchen Zhang, and Mingyu Gao

  22. [30]

    Zihan Liu, Wentao Ni, Jingwen Leng, Yu Feng, Cong Guo, Quan Chen, Chao Li, Minyi Guo, and Yuhao Zhu. 2024. JUNO: Optimizing High- Dimensional Approximate Nearest Neighbour Search with Sparsity- Aware Algorithm and Ray-Tracing Core Mapping. InProceedings of the 29th ACM Interna...

    2024

  23. [31]

    Yury A Malkov and Dmitry A Yashunin. 2020. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs.IEEE transactions on pattern analysis and machine intelligence42, 4 (2020), 824–836.https://doi.org/10.1109/TPAMI.2018. 2889473

    2020 doi

  24. [32]

    NVIDIA RAPIDS Team. 2026. cuVS: Vector Search and Clustering on the GPU.https://docs.rapids.ai/api/cuvs/stable/. Accessed: 2026-06-16

    2026

  25. [33]

    Aude Oliva and Antonio Torralba. 2001. Modeling the shape of the scene: A holistic representation of the spatial envelope.International journal of computer vision42, 3 (2001), 145–175

    2001

  26. [34]

    Christoph Platzer and Markus Pusch-el. 2021. Ten lessons from three generations of the RISC-V vector extension. In2021 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). IEEE, 1–12

    2021

  27. [35]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on ma- chine learn...

    2021

  28. [36]

    2021.RISC-V Vector Extension Specification

    RISC-V International. 2021.RISC-V Vector Extension Specification. Tech- nical Report. RISC-V International.https://docs.riscv.org/reference/ isa/extensions/vector/_attachments/riscv-v-spec.pdfVersion 1.0

    2021

  29. [37]

    RISC-V International. 2025. The RISC-V Instruction Set Manual, 15 Volume I: Unprivileged ISA.https://docs.riscv.org/reference/isa/ _attachments/riscv-unprivileged.pdfVersion 2025-05-08

    2025

  30. [38]

    Anshumali Shrivastava and Ping Li. 2014. Asymmetric LSH (ALSH) for sublinear time maximum inner product search (MIPS).Advances in neural information processing systems27 (2014)

    2014

  31. [39]

    Nigel Stephens, Stuart Biles, Matthias Boettcher, Jacob Eapen, Mbou Eyole, Giacomo Gabrielli, Matt Horsnell, Grigorios Magklis, Alejandro Martinez, Nathanael Premillieu, et al. 2017. The ARM scalable vector extension.IEEE Micro37, 2 (2017), 26–39

    2017

  32. [40]

    Bing Tian, Haikun Liu, Yuhang Tang, Shihai Xiao, Zhuohui Duan, Xi- aofei Liao, Xuecang Zhang, Junhua Zhu, and Yu Zhang. 2024. Fusion- ANNS: An Efficient CPU/GPU Cooperative Processing Architecture for Billion-scale Approximate Nearest Neighbor Search.arXiv preprint arXiv:2409....

    2024 doi arXiv

  33. [41]

    George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioan- nis Partalas, Matthias Zschunke, Michael R Alvers, Dirk Weissenborn, Anastasia Krithara, Sergios Petridis, Dimitris Polychronopoulos, et al

  34. [42]

    Emanuele Venieri, Simone Manoni, Giacomo Madella, Federico Prover- bio, Federico Ficarelli, Luca Benini, and Andrea Bartolini. 2026. Monte Cimone v3: Where RISC-V Stands in High-Performance Computing. https://doi.org/10.48550/arXiv.2605.22831arXiv:2605.22831 [cs.DC]

    2026 doi arXiv

  35. [43]

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, et al

  36. [44]

    Mengzhao Wang, Haotian Wu, Xiangyu Ke, Yunjun Gao, Yifan Zhu, and Wenchao Zhou. 2025. Accelerating Graph Indexing for ANNS on Modern CPUs.Proceedings of the ACM on Management of Data3, 3 (2025), 123:1–123:29.https://doi.org/10.1145/3725260

    2025 doi

  37. [45]

    Yitu Wang, Shiyu Li, Qilin Zheng, Linghao Song, Zongwang Li, An- drew Chang, Hai Li, and Yiran Chen. 2024. NDSEARCH: Accelerat- ing Graph-Traversal-Based Approximate Nearest Neighbor Search through Near Data Processing. In2024 ACM/IEEE 51st Annual Interna- tional Symposium on ...

    2024

  38. [46]

    Hao Wei, Jeffrey Xu Yu, Can Lu, and Xuemin Lin. 2016. Speedup Graph Processing by Graph Ordering. InProceedings of the 2016 International Conference on Management of Data. ACM, 1813–1828.https://doi.org/ 10.1145/2882903.2915220

    2016

  39. [47]

    Xiang Wei, Qiang Zhang, and Yihong Gong. 2020. PQ-Fast-Scan: A fast and accurate product quantization method for approximate nearest neighbor search. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 6342–6349

    2020

  40. [48]

    Yinan Xu, Zihao Yu, Dan Tang, Guokai Chen, Lu Chen, Lingrui Gou, Yue Jin, Qianruo Li, Xin Li, Zuojun Li, Jiawei Lin, Tong Liu, Zhigang Liu, Jiazhan Tan, Huaqiang Wang, Huizhe Wang, Kaifan Wang, Chuanqi Zhang, Fawang Zhang, Linjuan Zhang, Zifei Zhang, Yangyang Zhao, Yaoyang Zho...

    2022

  41. [49]

    Hanchen Ye, Cong Hao, Jianyi Cheng, Hyunmin Jeong, Jack Huang, Stephen Neuendorffer, and Deming Chen. 2022. ScaleHLS: A New Scalable High-Level Synthesis Framework on Multi-Level Intermediate Representation. InIEEE International Symposium on High-Performance Computer Architect...

    2022

  42. [50]

    Florian Zaruba and Luca Benini. 2019. The cost of application-class processing: Energy and performance analysis of a linux-ready 1.7-GHz 64-bit RISC-V core in 22-nm FDSOI technology.IEEE Transactions on Very Large Scale Integration (VLSI) Systems27, 11 (2019), 2629–2640

    2019

  43. [51]

    Shulin Zeng, Zhenhua Zhu, Jun Liu, Haoyu Zhang, Guohao Dai, Zix- uan Zhou, Shuangchen Li, Xuefei Ning, Yuan Xie, Huazhong Yang, and Yu Wang. 2023. DF-GAS: a Distributed FPGA-as-a-Service Ar- chitecture towards Billion-Scale Graph-based Approximate Nearest Neighbor Search. InPr...

    2023

  44. [2011]

    In2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Searching in one billion vectors: re-rank with source coding. In2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 861–864

  45. [2015]

    An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition.BMC Bioinformatics 16, 1 (2015), 138:1–138:28

    2015

  46. [2021]

    In Proceedings of the 2021 International Conference on Management of Data

    Milvus: A purpose-built vector data management system. In Proceedings of the 2021 International Conference on Management of Data. 2614–2627.https://doi.org/10.1145/3448016.3457550

    2021

  47. [2022]

    In2022 IEEE International Symposium on Circuits and Systems (ISCAS)

    OpenVLA: A RISC-V vector accelerator for deep learning infer- ence. In2022 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 1–5

  48. [2025]

    InProceedings of the 52nd Annual International Symposium on Computer Architecture, ISCA 2025, Tokyo, Japan, June 21-25, 2025

    ANSMET: Approximate Nearest Neighbor Search with Near- Memory Processing and Hybrid Early Termination. InProceedings of the 52nd Annual International Symposium on Computer Architecture, ISCA 2025, Tokyo, Japan, June 21-25, 2025. ACM, 1093–1107.https: //doi.org/10.1145/3695053.3731013

    2025

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.