Pith. sign in

REVIEW 2 major objections 5 minor 45 references

This paper shows that replacing global training barriers with primitive-level synchronization speeds distributed point-based neural rendering by 1.22× on average while preserving reconstruction quality.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:23 UTC pith:KWYLS5LZ

load-bearing objection Odin's primitive-level synchronization idea is new and the speedup looks real, but the throughput-first quality guarantee rests on a gate that doesn't bound the actual skipped update, so the paper needs a reframed claim plus code and variance before I'd trust it. the 2 major comments →

arxiv 2607.19893 v1 pith:KWYLS5LZ submitted 2026-07-22 cs.DC

Odin: Primitive-Level Synchronization for Distributed Point-Based Neural Rendering

Zhenxiang Ma , Zeyu He , Yuanzhen Zhou , Zhenyu Yang , Yuchang Zhang , Miao Tao , Rong Fu , Jidong Zhai
show 1 more author
Hengjie Li
This is my paper
classification cs.DC
keywords distributed trainingpoint-based neural renderingprimitive-level synchronizationcommunication-computation overlapbarrier eliminationruntime validationGaussian splattingscheduling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Point-based neural renderers train 3D scenes by having each camera view read and update only the visible primitives, yet distributed systems currently synchronize all GPUs after every iteration. This paper argues that these global barriers are far coarser than the actual dependencies, which live at the level of individual primitive indices. It presents a scheduler that predicts which views will touch which primitives, overlaps work that does not conflict, and validates that published updates are visible before a later view reads state. On four rendering pipelines and 13 non-city scenes, it reports a 1.22× average throughput gain, 82% of critical-path wait hidden, and reconstruction quality within a ±1% band. The work matters because as scenes scale and renderers get faster, synchronization—not rendering—is becoming the bottleneck in distributed 3D training.

Core claim

The central claim is that replacing task- or iteration-level barriers with primitive-level publication checks removes the synchronization bottleneck in distributed point-based neural rendering without changing the underlying model. The system builds a relative locality graph from stable view co-visibility to pre-order training views and place overlap windows, then at runtime validates, before any state read, that all updates a later view might observe are either disjoint or already published. A throughput-first mode additionally admits read-after-write overlaps only when the delayed scope is small and the producer primitives have low gradient magnitude, relying on compositing and occlusion t

What carries the argument

The central objects are the relative locality graph (RLG), a weighted undirected graph on training views whose edge weights are the Jaccard overlap of stable 3D reconstruction tracks, and the shadow-graph execution model, which gives each in-flight unit private logical views of data and gradient staging over shared physical tensors. The RLG drives a three-step ahead-of-time schedule: logical partitioning, static data ordering that rotates through weakly coupled groups, and static asynchronous scheduling that marks candidate overlap windows. At runtime, dynamic asynchronous scheduling revalidates each planned window using conservative read/update primitive scopes—checking RAW, WAW, and WAR ha

Load-bearing premise

The throughput-first path's safety rests on the assumption that a read-after-write overlap with small delayed scope and small producer gradients will not harm training quality, a claim supported only by aggregate quality deltas within the ±1% band on the evaluated scenes, not by a convergence argument.

What would settle it

Run the throughput-first path at τ=0.2 on a scene where many primitives each have small per-view gradients but collectively dominate a later view—for example, a large occluded region that becomes visible after many delayed reads—and check whether PSNR, SSIM, or LPIPS deltas exceed the ±1% reporting band; if they do, the admission gate is not safe as stated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, distributed training of point-based scenes can scale further without waiting for all GPUs to finish each iteration; synchronization no longer needs to be a scaling ceiling.
  • The same primitive-level publication rules apply to both data-parallel and partitioned-state execution, so the technique composes with existing placement and communication-sparsification systems.
  • Because quality is preserved within a reporting band without changing kernels or optimizers, the approach can be adopted incrementally as a scheduling layer over existing renderers.
  • The 82% critical-path wait hiding suggests that as renderers continue to get faster, the bottleneck shifts decisively to synchronization, making this class of techniques more important over time.
  • The quality-first path offers a conservative mode that preserves synchronized-training visibility, giving practitioners a drop-in option when quality guarantees are strict.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The admission gate—delayed-scope ratio plus gradient threshold—is heuristic; a formal cumulative-error bound on delayed reads would turn the empirical ±1% safety into a guarantee and could justify larger τ on quality-critical workloads.
  • The same predict-and-validate pattern could extend to other sparse explicit-state workloads beyond rendering, such as online reconstruction, neural mapping, and object- or voxel-level world models, where mutable state is indexed rather than layer-structured.
  • Because the scheduler relies on stable scene metadata from 3D reconstruction tracks, it may be less effective in fully online settings where reconstruction and training happen simultaneously; a streaming locality prior would be a natural extension.
  • The reported gains average over 13 non-city scenes; the city-scale case hints at larger gains at 64 GPUs, but the dense-scene fallback (98.6% overlap rejection in one scene) indicates that scenes with broad view coupling will see little benefit.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. Odin replaces global task/iteration barriers in distributed point-based neural rendering (PBNR) with primitive-index-scoped synchronization. An ahead-of-time scheduler builds a Relative Locality Graph from SfM track co-visibility, partitions views into logical groups, orders task units, and places candidate overlap windows; a runtime layer, using Shadow Graph staging, validates primitive publication before later state reads. Two execution modes are offered: a quality-first path that admits only disjoint or already-published primitive scopes, and a throughput-first path that additionally admits RAW delayed reads when a delayed-scope ratio and per-primitive producer-gradient magnitudes are below a parameter τ. The evaluation across 3DGS, 2DGS, TamingGS, and DashGS on 13 non-city 8-GPU scenes reports 1.22× average throughput and 82% critical-path wait hiding, with aggregate PSNR/SSIM/LPIPS deltas within a ±1% band; a MatrixCity mixed-parallel case study reports up to 1.89× over Grendel at 64 GPUs. The paper claims these gains come without changing renderer kernels, optimizers, training budgets, or model capacity.

Significance. The paper identifies a real and increasingly important bottleneck: as PBNR renderers become faster, global synchronization takes over the critical path, while view dependencies are sparse at primitive-index granularity. The predict-and-validate design is principled, and the system contribution is well scoped: it addresses when primitive updates become visible rather than where state is placed or how much state moves. The evaluation is internally controlled, keeping kernels, optimizers, budgets, and model capacity fixed, and the ablations attempt to attribute gains to primitive-scoped synchronization rather than fewer bytes or unscoped asynchrony. The two-graph scheduler plus Shadow Graph staging is a plausible and potentially reusable systems mechanism. However, the manuscript does not release code or artifacts, reports no variance or multi-seed results, and the throughput-first admission gate does not, as written, bound the actual skipped update in data-parallel training. These issues are load-bearing for the central quality-preservation claim, so the paper needs a major revision rather than acceptance.

major comments (2)
  1. [Section 4.3, throughput-first admission; Section 5 (post-backward hook)] The producer-gradient gate uses ∥g_A(p)∥, the per-primitive gradient magnitude reported by the post-backward hook before communication/reduction. In data-parallel execution, the update that a delayed read skips is the all-reduced global gradient and the resulting optimizer (Adam) parameter delta, not A's local gradient. A primitive can have small g_A(p) while other replicas contribute large gradients, so the actual skipped update can be large. Even a correct global-gradient bound would not by itself bound B's rendering sensitivity under opacity/transmittance changes. Section 4.3 asserts that small producer gradients imply small impact, but provides no formal or empirical link; the ±1% aggregate quality results cannot establish the claimed 'only small, low-impact delayed reads' guarantee.
  2. [Section 6.1/6.3, Figure 15] Quality-preservation evidence consists of normalized aggregate deltas from a single seed. No variance or multi-seed results are reported, and no per-scene worst-case table is given. Given the admission gate in Section 4.3 does not formally bound the skipped update, these aggregates are the only support for the central claim of 'preserving reconstruction quality' under the throughput-first path. Please report per-scene PSNR/SSIM/LPIPS deltas over at least three seeds and state the worst-case scene; a scatter plot or table would allow readers to verify the ±1% band.
minor comments (5)
  1. [Section 6.1/6.2, headline throughput] The headline 1.22× is the arithmetic mean of scene-level Odin/base ratios, but Figure 8 shows grouped labels rather than per-scene values. Please include the underlying per-scene ratios and clarify whether the mean is over scenes or over runs.
  2. [Section 6.2, MatrixCity] Clarify that the 1.89× speedup over Grendel is for the single-region MatrixCity MP case at 64 GPUs, and not necessarily for a full city-scale training run, to avoid overgeneralization.
  3. [Section 6.4, Figure 18 ablation] Specify the 'partitionless HOGWILD-style control' in more detail: does it retain the RLG/SDS order, Shadow Graph staging, and any validation? Without this, the conclusion that gains come from the coordinated design rather than from a single component is underdetermined.
  4. [Section 4.1.1/Table 3, SfM dependence] State the fallback behavior when SfM tracks are unavailable (e.g., synthetic or non-SfM inputs), and report whether Table 3's AOT overhead includes RLG construction from tracks or only graph construction and schedule compilation after the track mapping is computed.
  5. [Figure 16, Shadow Graph overhead] Define the 'replication-based alternative' precisely, including whether it replicates full optimizer state, and describe the operation-cost measurement. The 'up to 257×' figure needs a clear baseline to be interpretable.

Circularity Check

0 steps flagged

No circular derivation found; the paper's claims are empirical systems measurements with an explicit tunable gate, not fitted re-statements of the target results.

full rationale

Odin's derivation chain is not a mathematical derivation that reduces to its inputs. The quality-first path is explicitly defined to preserve synchronized-training visibility under conservative primitive scopes, and the equivalence claim is a design property of the admission rule rather than a fitted result. The throughput-first path uses a tunable admission parameter τ to gate delayed-scope ratio and producer-gradient magnitude; the headline throughput and quality results are then measured under that fixed setting, not derived from the gate. No equation in the paper is fitted to the target speedup or quality band. The paper makes no load-bearing self-citation: references [12] and [13], cited for the observation that compositing/occlusion give primitive interactions numerical weights, are external works, and no uniqueness theorem or prior-work ansatz is invoked to force the design. The 82% wait-hiding and 1.22× measurements are standard systems evaluations under Odin's own scheduling policy, which is not circular. The skeptic concern about pre-all-reduce gradient magnitudes is a correctness/robustness gap, not a circularity, because the gate is not claimed to be equivalent to the quality measurement. Overall, no specific reduction of a claimed result to its own inputs could be identified.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

Odin's gains come from scheduling parameters K and τ and from domain assumptions about sparse primitive scopes and bounded-delay safety. No target metric is fitted by equations; the main quantitative claims are empirical. The RLG and Shadow Graph are constructed components rather than externally evidenced entities.

free parameters (4)
  • K (logical region count) = 4 (default)
    Controls logical partitioning granularity in the scheduler; swept in Figure 14 and fixed for main results.
  • τ (admission parameter) = 0.2 (default)
    Bounds delayed-scope ratio and producer-gradient gate in the throughput-first path; chosen via sensitivity sweep to stay within the quality band.
  • top-k neighbor retention in RLG = unspecified
    RLG construction retains only the top-k neighbors per image; affects scheduling quality though not publication safety.
  • track-length cap = unspecified
    Used to limit outliers in SfM co-visibility counts; ranking-only effect.
axioms (5)
  • domain assumption Each PBNR view reads and updates only a sparse, view-dependent subset of primitive indices; conflicts can be represented by primitive scopes.
    Core premise for primitive-level synchronization; introduced in Section 2.1 and Figure 2.
  • domain assumption Compositing, occlusion, and transmittance make primitive interactions numerically weighted, so small delayed reads with small producer gradients have low training impact.
    Throughput-first admission rule in Section 4.3 relies on refs [12,13] and empirical quality, not a proof.
  • domain assumption SfM co-visibility is a stable predictor of view-pair coupling during training despite densification and pruning.
    RLG construction in Section 4.1.1; used for ahead-of-time scheduling, with safety delegated to runtime validation.
  • ad hoc to paper Bounded delayed reads do not break convergence of the PBNR optimizer beyond the reported quality band.
    Not proven; supported only by aggregate empirical deltas on evaluated scenes (Section 6.3, Figure 15).
  • domain assumption Structural versions (densification, pruning, opacity reset) can be detected and synchronized safely.
    Runtime relies on structural version metadata to force barriers; described in Section 5.
invented entities (2)
  • Relative Locality Graph (RLG) no independent evidence
    purpose: Ahead-of-time ranking of view-pair coupling from SfM tracks to place low-conflict overlap windows.
    Internal scheduling structure; correctness does not depend on it because runtime validates planned windows.
  • Shadow Graph no independent evidence
    purpose: Versioned logical views and staged gradient publication over shared physical tensors so overlapped task units do not alias.
    Runtime mechanism; overhead is measured but it provides no external falsifiable prediction beyond the paper's own benchmarks.

pith-pipeline@v1.3.0-alltime-deepseek · 17148 in / 13978 out tokens · 145532 ms · 2026-08-01T11:23:05.447485+00:00 · methodology

0 comments
read the original abstract

Point-based neural rendering (PBNR) represents 3D scenes as explicit, trainable primitives and underpins high-quality reconstruction and emerging embodied AI and world-model pipelines. Unlike layer-structured neural networks, PBNR has primitive-indexed dependencies: each view reads and updates only a sparse, view-dependent subset of mutable scene state. As large scenes require distributed training and optimized renderers reduce per-view computation, global task- or iteration-level barriers increasingly place synchronization, rather than rendering, on the critical path. We present Odin, a distributed PBNR training system that replaces global barriers with primitive-level synchronization. Its ahead-of-time scheduler uses stable locality and phase order to identify low-conflict overlap windows, while the runtime validates primitive publication before later work observes mutable state. Odin provides a quality-first path that preserves synchronized-training visibility and a throughput-first path that uses overlap and gradient evidence to admit only small, low-impact delayed reads; structural changes and high-impact cases remain synchronized. Across four existing PBNR pipelines and 13 non-city scenes on 8 GPUs, Odin improves throughput by 1.22 times on average and hides 82% of critical-path wait while preserving reconstruction quality. In a MatrixCity mixed-parallel case study scaling to 64 GPUs, Odin improves throughput over Grendel by up to 1.89 times without changing renderer kernels, optimizers, training budgets, or model capacity.

Figures

Figures reproduced from arXiv: 2607.19893 by Hengjie Li, Jidong Zhai, Miao Tao, Rong Fu, Yuanzhen Zhou, Yuchang Zhang, Zeyu He, Zhenxiang Ma, Zhenyu Yang.

Figure 1
Figure 1. Figure 1: Motivating synchronization overhead in distributed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Running four-view locality example. Left: four num [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Locality-scope ratios in real scenes after 7,000 itera [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Primitive-index synchronization over aligned ten [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative locality graph (RLG) for the four-view [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Shadow Graph execution for overlapped task units. Phase labels 𝑃/𝑅/𝐵/𝑈 /𝐶 denote preprocessing, rendering, backward computation, optimizer update, and communi￾cation; suffixes 0/1 identify two task units. Each unit uses private logical views for local execution and staged gradient publication while physical model state remains shared. Num￾bered arrows summarize private read, local autograd, staged update, … view at source ↗
Figure 10
Figure 10. Figure 10: sweeps GPU count and global batch size. Odin improves throughput throughout the sweep, while relative gain narrows as 𝐵 increases because larger batches create fewer synchroniza￾tion boundaries per sample. This is the expected signature of a primitive-level synchronization optimization: it helps most when communication remains on the critical path and enough later work exists to overlap with it [PITH_FUL… view at source ↗
Figure 8
Figure 8. Figure 8: End-to-end 8-GPU throughput, excluding Matrix [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: Mixed-parallel scaling across datasets. Dashed line [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 15
Figure 15. Figure 15: evaluates the default throughput-first setting. At 𝐾 = 4, 𝜏 = 0.2, all plotted PSNR, SSIM, and LPIPS aggregate deltas stay within the ±1% reporting band. This matches the admission rule: delayed reads are limited to small nonzero overlap scopes and weak producer-gradient updates, which typically correspond to [PITH_FULL_IMAGE:figures/full_fig_p010_15.png] view at source ↗
Figure 13
Figure 13. Figure 13: Multi-node MP throughput on single-region Ma [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Scheduling-parameter sensitivity. Left sweeps [PITH_FULL_IMAGE:figures/full_fig_p010_14.png] view at source ↗
Figure 19
Figure 19. Figure 19: Runtime breakdown on a representative 8-GPU [PITH_FULL_IMAGE:figures/full_fig_p011_19.png] view at source ↗
Figure 18
Figure 18. Figure 18: Ablation on a representative 8-GPU 3DGS run [PITH_FULL_IMAGE:figures/full_fig_p011_18.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 canonical work pages

  1. [1]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis

  2. [2]

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao

  3. [3]

    Guangchi Fang and Bing Wang. 2024. Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerland, Cham, 165–181

    2024

  4. [4]

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. 2025. EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerland, Cham, 54–71

    2025

  5. [5]

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park

  6. [6]

    Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. 2024. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering.ACM Trans. Graph.43, 4, Article 64 (July 2024), 17 pages. doi:10.1145/3658187

    work page doi:10.1145/3658187 2024

  7. [7]

    Jad Abou-Chakra, Krishan Rana, Feras Dayoub, and Niko Sünderhauf. 2024. Physically Embodied Gaussian Splatting: A Realtime Correctable World Model for Robotics.arXiv preprint arXiv:2406.10788(2024)

    Pith/arXiv arXiv 2024

  8. [9]

    Hexu Zhao, Haoyang Weng, Daohan Lu, Ang Li, Jinyang Li, Aurojit Panda, and Saining Xie. 2025. On Scaling Up 3D Gaussian Splatting Training. InComputer Vision – ECCV 2024 Workshops, Alessio Del Bue, Cristian Canton, Jordi Pont-Tuset, and Tatiana Tommasi (Eds.). Springer Nature Switzerland, Cham, 14–36

    2025

  9. [10]

    Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. 2025. CityGaussian: Real-Time High-Quality Large-Scale Scene Render- ing with Gaussians. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerland, Cham, 265–282

    2025

  10. [11]

    Guanxing Lu, Baoxiong Jia, Puhao Li, Yixin Chen, Ziwei Wang, Yansong Tang, and Siyuan Huang. 2025. GWM: Towards Scalable Gaussian World Models for Robotic Manipulation. arXiv:2508.17600 [cs.RO] https://arxiv.org/abs/2508.17600

    arXiv 2025

  11. [12]

    Shiyong Liu, Xiao Tang, Zhihao Li, Yingfan He, Chongjie Ye, Jianzhuang Liu, Binxiao Huang, Shunbo Zhou, and Xiaofei Wu. 2025. OccluGaussian: Occlusion- Aware Gaussian Splatting for Large Scene Reconstruction and Rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 26643–26652

    2025

  12. [13]

    Chai-Rong Lee, Ting-Yu Yen, Kai-Wen Hsiao, Shih-Hsuan Hung, Sheng-Chi Hsu, Min-Chun Hu, Chih-Yuan Yao, and Hung-Kuo Chu. 2024. ODA-GS: Occlusion- and Distortion-aware Gaussian Splatting for Indoor Scene Reconstruction. InSIG- GRAPH Asia 2024 Technical Communications (SA ’24). Association for Computing Machinery, New York, NY, USA, Article 29, 4 pages. doi...

    arXiv 2024

  13. [14]

    Hexu Zhao, Xiaoteng Liu, Xiwen Min, Jianhao Huang, Youming Deng, Yanfei Li, Ang Li, Jinyang Li, and Aurojit Panda. 2025. Scaling Point-based Differentiable Rendering for Large-scale Reconstruction. arXiv:2512.20017 [cs.DC] https: //arxiv.org/abs/2512.20017

    arXiv 2025

  14. [16]

    James B McQueen. 1967. Some methods of classification and analysis of multi- variate observations. InProc. of 5th Berkeley Symposium on Math. Stat. and Prob. 281–297

    1967

  15. [17]

    Schönberger and Jan-Michael Frahm

    Johannes L. Schönberger and Jan-Michael Frahm. 2016. Structure-from-Motion Revisited. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 4104–4113. doi:10.1109/CVPR.2016.445

    doi:10.1109/cvpr.2016.445 2016

  16. [18]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. 2022. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5460–5469. doi:10.1109/CVPR52688.2022.00539

    arXiv 2022

  17. [19]

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. 2017. Tanks and temples: benchmarking large-scale scene reconstruction.ACM Trans. Graph. 36, 4, Article 78 (July 2017), 13 pages. doi:10.1145/3072959.3073599

    arXiv 2017

  18. [20]

    S. Lloyd. 1982. Least squares quantization in PCM.IEEE Transactions on Infor- mation Theory28, 2 (1982), 129–137. doi:10.1109/TIT.1982.1056489

    arXiv 1982

  19. [21]

    Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. 2023. MatrixCity: A Large-scale City Dataset for City-scale Neural Rendering and Beyond. In2023 IEEE/CVF International Conference on Computer Vision (ICCV). 3182–3192. doi:10.1109/ICCV51070.2023.00297

    arXiv 2023

  20. [22]

    Feng Niu, Benjamin Recht, Christopher Re, and Stephen J. Wright. 2011. HOG- WILD! a lock-free approach to parallelizing stochastic gradient descent. InPro- ceedings of the 25th International Conference on Neural Information Processing Systems(Granada, Spain)(NIPS’11). Curran Associates Inc., Red Hook, NY, USA, 693–701

    2011

  21. [23]

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. 2018. Deep blending for free-viewpoint image-based rendering.ACM Trans. Graph.37, 6, Article 257 (Dec. 2018), 15 pages. doi:10. 1145/3272127.3275084

    arXiv 2018

  22. [25]

    Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. 2024. Taming 3DGS: High-Quality Radiance Fields with Limited Resources. InSIGGRAPH Asia 2024 Conference Papers(Tokyo, Japan)(SA ’24). Association for Computing Machinery, New York, NY, USA, Article 2, 11 pages. doi:10.1145/36805...

    arXiv 2024

  23. [26]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

  24. [27]

    Junseo Lee, Seokwon Lee, Jungi Lee, Junyong Park, and Jaewoong Sim. 2024. GSCore: Efficient Radiance Field Rendering via Architectural Support for 3D Gaussian Splatting. InProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3(La Jolla, CA, USA)(ASPLOS ’24). Association for C...

    arXiv 2024

  25. [28]

    Houshu He, Gang Li, Fangxin Liu, Li Jiang, Xiaoyao Liang, and Zhuoran Song

  26. [29]

    Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. 2025. Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 21537–21546. doi:10.1109/CVPR52734.2025.02006

    arXiv 2025

  27. [32]

    Lee, and Hongil Yoon

    Donghyun Lee, Dawoon Jeong, Jae W. Lee, and Hongil Yoon. 2026. GS-Scale: Unlocking Large-Scale 3D Gaussian Splatting Training via Host Offloading. In Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2(USA)(ASPLOS ’26). Association for Computing Machinery, New York, NY, US...

    arXiv 2026

  28. [34]

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. 2024. Vast- Gaussian: Vast 3D Gaussians for Large Scene Reconstruction. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5166–5175. doi:10.1109/CVPR52733.2024.00494

    arXiv 2024

  29. [35]

    Lukas Höllein, Aljaž Božič, Michael Zollhöfer, and Matthias Nießner. 2025. 3DGS- LM: Faster Gaussian-Splatting Optimization with Levenberg-Marquardt. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2025

  30. [36]

    Yu Chen and Gim Hee Lee. 2024. DOGS: distributed-oriented Gaussian splatting for large-scale 3D reconstruction via Gaussian consensus. InProceedings of the 38th International Conference on Neural Information Processing Systems(Vancou- ver, BC, Canada)(NIPS ’24). Curran Associates Inc., Red Hook, NY, USA, Article 1087, 26 pages

    2024

  31. [37]

    Abhinav Jangda, Jun Huang, Guodong Liu, Amir Hossein Nodehi Sabet, Saeed Maleki, Youshan Miao, Madanlal Musuvathi, Todd Mytkowicz, and Olli Saarikivi

  32. [38]

    Lin Zhang, Shaohuai Shi, Xiaowen Chu, Wei Wang, Bo Li, and Chengjian Liu

  33. [39]

    Shenggan Cheng, Shengjie Lin, Lansong Diao, Hao Wu, Siyu Wang, Chang Si, Ziming Liu, Xuanlei Zhao, Jiangsu Du, Wei Lin, and Yang You. 2025. Concerto: Automatic Communication Optimization and Scheduling for Large-Scale Deep Learning. InProceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Syst...

    arXiv 2025

  34. [40]

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. 2024. A Hierarchical 3D Gaussian Representation for Real-Time Rendering of Very Large Datasets.ACM Trans. Graph.43, 4, Article 62 (July 2024), 15 pages. doi:10.1145/3658160

    doi:10.1145/3658160 2024

  35. [41]

    Yanghua Peng, Yibo Zhu, Yangrui Chen, Yixin Bao, Bairen Yi, Chang Lan, Chuan Wu, and Chuanxiong Guo. 2019. A generic communication sched- uler for distributed DNN training acceleration. InProceedings of the 27th ACM Symposium on Operating Systems Principles(Huntsville, Ontario, Canada) (SOSP ’19). Association for Computing Machinery, New York, NY, USA, 16...

    arXiv 2019

  36. [42]

    Xupeng Miao, Xiaonan Nie, Yingxia Shao, Zhi Yang, Jiawei Jiang, Lingxiao Ma, and Bin Cui. 2021. Heterogeneity-Aware Distributed Machine Learning Training via Partial Reduce. InProceedings of the 2021 International Conference on Management of Data(Virtual Event, China)(SIGMOD ’21). Association for Computing Machinery, New York, NY, USA, 2262–2270. doi:10.1...

    doi:10.1145/3448016 2021

  37. [43]

    Shigang Li, Tal Ben-Nun, Salvatore Di Girolamo, Dan Alistarh, and Torsten Hoefler. 2020. Taming unbalanced training workloads in deep learning with partial collective operations. InProceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming(San Diego, California)(PPoPP ’20). Association for Computing Machinery, New Yor...

    arXiv 2020

  38. [44]

    Weigao Sun, Zhen Qin, Weixuan Sun, Shidi Li, Dong Li, Xuyang Shen, Yu Qiao, and Yiran Zhong. [n. d.]. CO2: Efficient Distributed Training with Full Communication-Computation Overlap. InThe Twelfth International Conference on Learning Representations. 13

  39. [45]

    In2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS)

    DeAR: Accelerating Distributed Deep Learning with Fine-Grained All- Reduce Pipelining. In2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS). 142–153. doi:10.1109/ICDCS57875.2023.00054

    arXiv 2023

  40. [47]

    Le, Yonghui Wu, and Zhifeng Chen

    Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan Firat, Mia Xu Chen, Dehao Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. 2019.GPipe: efficient training of giant neural networks using pipeline parallelism. Curran Associates Inc., Red Hook, NY, USA

    2019

  41. [2018]

    In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 586–595. doi:10.1109/CVPR.2018.00068

    arXiv 2018

  42. [2022]

    InProceedings of the 27th ACM Inter- national Conference on Architectural Support for Programming Languages and Operating Systems(Lausanne, Switzerland)(ASPLOS ’22)

    Breaking the computation and communication abstraction barrier in distributed machine learning workloads. InProceedings of the 27th ACM Inter- national Conference on Architectural Support for Programming Languages and Operating Systems(Lausanne, Switzerland)(ASPLOS ’22). Association for Com- puting Machinery, New York, NY, USA, 402–416. doi:10.1145/350322...

    arXiv

  43. [2023]

    Graph.42, 4, Article 139 (July 2023), 14 pages

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Trans. Graph.42, 4, Article 139 (July 2023), 14 pages. doi:10.1145/3592433

    doi:10.1145/3592433 2023

  44. [2024]

    In ACM SIGGRAPH 2024 Conference Papers(Denver, CO, USA)(SIGGRAPH ’24)

    2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers(Denver, CO, USA)(SIGGRAPH ’24). Association for Computing Machinery, New York, NY, USA, Article 32, 11 pages. doi:10.1145/3641519.3657428

    arXiv 2024

  45. [2025]

    In2025 IEEE International Symposium on High Performance Computer Architecture (HPCA)

    GSArch: Breaking Memory Barriers in 3D Gaussian Splatting Training via Architectural Support. In2025 IEEE International Symposium on High Performance Computer Architecture (HPCA). 366–379. doi:10.1109/HPCA61900.2025.00037

    arXiv 2025