Pith. sign in

REVIEW 4 major objections 5 minor 82 references

Lumina: Real-Time Mobile Neural Rendering by Exploiting Computational Redundancy

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Lumina makes 3D Gaussian Splatting run 4.5x faster and 5.3x more energy-efficient on mobile hardware by reusing sorting results and cached pixel colors, with under 0.2 dB of quality loss.

desk verdict A solid, genuinely novel 3DGS mobile accelerator paper whose main weakness is quality accounting: the abstract's <0.2 dB claim conflicts with the body, and RC's benefits are entangled with retraining; it still deserves serious peer review. read the letter →

arxiv 2506.05682 v1 pith:IJXOV4SG submitted 2025-06-06 cs.AR

Yu Feng , Weikai Lin , Yuge Cheng , Zihan Liu , Jingwen Leng , Minyi Guo , Chen Chen , Shixuan Sun
show 1 more author
Yuhao Zhu
This is my paper
classification cs.AR
keywords 3DGaussianSplattingneuralrenderingmobilearchitectureradiancecachingtemporalcoherencehardware-algorithmco-designsortingsharingwarpdivergence
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

Lumina is a hardware-algorithm co-design that targets the two most expensive steps of 3D Gaussian Splatting (3DGS) rendering on mobile chips: sorting Gaussians by depth and rasterizing them into pixels. The paper claims both steps contain a large amount of redundancy — nearby camera frames produce almost the same sort order, and rays that hit the same sequence of important Gaussians produce almost the same pixel color — and that a modest custom accelerator can harvest that redundancy. Concretely, Lumina reports a 4.5x speedup and a 5.3x energy reduction against a mobile Volta GPU while keeping PSNR within about 0.2 dB of the unmodified 3DGS baseline across synthetic and real-world scenes. If true, this would bring real-time, 90 FPS 3DGS-based rendering within reach of VR/AR-class mobile devices.

What carries the argument

The argument is carried by two algorithms and an accelerator. S2 (Sorting-Shared) exploits temporal coherence: it predicts the next camera pose from the last two poses, speculatively projects and sorts the Gaussian cloud for that predicted pose, and shares that sort across a sharing window of frames, expanding the sorting viewport slightly so tile-edge artifacts do not appear. RC (Radiance Caching) uses a cache whose tag is the concatenation of the first $k$ Gaussian IDs along a ray that pass the significance threshold ($\alpha > 1/255$), and whose value is the finished pixel color; a cache hit terminates the color integration immediately, by skipping the remainder of the equation $C(p)=\sum_{i} \Gamma_i \alpha_i c_i$. LuminCore is the accelerator that makes RC practical: its Neural Rendering Units split transparency computation (frontend) from sparse color integration (backend) to remove GPU warp divergence, and LuminCache is a 4-way set-associative cache using combined tags, low-bit indexing, and a pseudo-LRU replacement policy.

What would settle it

Render a scene dominated by large or elongated Gaussians — or a fast camera move that exposes such Gaussians — through Lumina with and without the cache-aware fine-tuning; if the radiance cache hit rate falls well below the reported 55% and PSNR drops by more than about 0.2 dB relative to the baseline 3DGS, the early-termination premise has failed and the 4.5x and 5.3x headline numbers would not survive.

Watch

Extended reading notes

Core claim

The central discovery is that the expensive per-frame work in 3DGS can be largely reused across time and across rays without visible quality loss. Sorting results stay valid for several consecutive camera poses, so the system predicts a future pose, pre-sorts once for that predicted pose, and shares the result across a fixed window of frames; the paper measures that only about 0.2% of Gaussian orders change between nearby poses. For rasterization, the paper observes that fewer than 10% of the Gaussians a pixel visits are significant contributors, and that pixels whose first few significant Gaussian IDs match have nearly identical final colors. Caching a finished pixel under a tag made of those IDs lets later pixels stop color integration early, and a scale-constrained fine-tuning step keeps large Gaussians from violating the caching assumption. Together the radiance cache skips about 55% of color-integration computation, and the combination of sorting sharing, caching, and a small dedicated accelerator yields the reported 4.5x and 5.3x gains against a mobile Volta GPU.

Load-bearing premise

The load-bearing premise is that pixels whose rays pass through the same first few significant Gaussians really do end up with nearly identical final colors, so the cached value can replace the rest of the integration; the paper itself shows large Gaussians violate this, and the scale-constrained retraining that restores it may not transfer to scenes outside the training data.

Editorial extensions

If this is right

  • If the central claim holds, 3DGS reaches 218.5 FPS on the synthetic traces and 97.9 FPS on the real-world traces, crossing the 90 FPS VR/AR target.
  • The S2 algorithm alone skips projection and sorting for a whole sharing window, which by itself reduces energy by about 20% even on an unmodified GPU.
  • RC reduces color-integration computation by 55% on average, and on LuminCore that translates into a further 2.5x speedup of the rasterization stage.
  • The full Lumina system reports an average 81% energy saving, rising to 93% on synthetic scenes and 80% on real-world scenes if the target is exactly real-time frame rate.
  • When configured with the GSCore accelerator units for fair comparison, the Lumina pipeline reports a 29.6x speedup over the mobile GPU baseline on the evaluated datasets.

Reading between the lines

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

  • If RC generalizes as the paper anticipates, the same tag-and-value idea could be applied to newer rendering primitives such as 2D Gaussians, voxels, or future learned representations, by replacing 'significant Gaussian IDs' with whatever primitives a ray hits first.
  • The NRU's frontend-backend design is essentially a parallel sparse-accumulation engine, and the paper notes that pattern also appears in sparse linear algebra and graph neural networks, so a programmable version of LuminCore could plausibly accelerate those workloads on mobile SoCs, though the paper does not demonstrate this.
  • The cache hit rate could itself serve as a live measurement of temporal scene redundancy; a testable extension would be to adapt the sharing window size and cache capacity in real time as hit rate rises or falls with camera speed and scene geometry.
  • Because cache tags are a hash of ray similarity, the mechanism could combine naturally with foveated rendering, using a lower confidence threshold in the visual periphery to trade quality for energy exactly where the eye is not looking.
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

4 major / 5 minor

Summary. The paper presents Lumina, a hardware-algorithm co-designed system for accelerating 3D Gaussian Splatting (3DGS) rendering on mobile SoCs. It proposes two algorithmic techniques: S2, which shares sorting results across nearby camera poses and speculatively pre-sorts for future frames, and Radiance Caching (RC), which caches pixel colors keyed by the first few significant Gaussian IDs so that cache hits can skip most color integration. The paper also designs LuminCore hardware with Neural Rendering Units and a specialized LuminCache. The system is evaluated on a mobile Volta GPU baseline with measured GPU execution, RTL synthesis and cycle-accurate simulation, and reports 4.5x speedup, 5.3x energy reduction, and <0.2 dB average PSNR loss across synthetic and real-world datasets, with a user study supporting similar visual quality.

Significance. The paper addresses an important and timely problem: making 3DGS rendering real-time on mobile platforms. Its evaluation is more thorough than many architecture papers in this area: GPU baseline measurements on actual hardware, RTL synthesis with area/power estimates, cycle-accurate simulation, quality metrics on multiple datasets, and a user study. The proposed S2 technique for hiding sorting latency is a plausible and useful contribution. The RC mechanism is a creative use of the sparsity structure of 3DGS color integration, and the LuminCore architecture is described in enough detail to be reproducible in simulation. If the quality and performance claims hold after addressing the confounds noted below, this would be a solid contribution to neural rendering acceleration.

major comments (4)
  1. [Abstract and Sec. 6.1] The abstract's claim of '<0.2 dB peak signal-to-noise ratio reduction' is internally inconsistent with the paper's own numbers in Fig. 20a: the average synthetic PSNR for Lumina is 33.2 dB versus 33.5 dB for the baseline, a 0.3 dB loss, and Sec. 6.1 states 'minimal quality losses of 0.2dB and 0.3dB' for RC-only and Lumina respectively. The abstract should either use the actual measured loss or report an aggregate across datasets that is consistent with the figure.
  2. [Sec. 3.3, Eq. (4), and Fig. 21] The claim that RC achieves 'minimal impact on quality' is not isolated from the cache-aware fine-tuning introduced in Eq. (4). The quality numbers in Fig. 20 are obtained on models retrained with the scale-constrained loss, and Fig. 21 shows that this loss changes cache hit rate and improves PSNR by 0.6 dB. Therefore the reported quality loss is the combined effect of caching plus a changed scene model, not the error of caching alone. The paper should provide an ablation that applies RC without the scale-constrained retraining, or otherwise quantify the caching-only error, and should discuss whether the headline speedup transfers to pretrained 3DGS scenes that have not been fine-tuned.
  3. [Sec. 6.4 and Fig. 25] The GSCore comparison in Fig. 25 is not on the same footing as the main evaluation. The 'baseline hardware' in that comparison incorporates GSCore's CCU and GSU units, so the 9.6x speedup attributed to this baseline is not the same as the NRU+GPU baseline in Sec. 6.2, which reports 1.9x speedup. The text should state clearly that Fig. 25 uses a different baseline configuration and explain why this is a fair comparison, or the 29.6x headline speedup for Lumina against GSCore should be qualified accordingly.
  4. [Sec. 6.3 and Fig. 23] The sensitivity study shows that S2's quality and speedup depend strongly on the sharing window and expanded margin, but the default settings (expanded margin 4, skipped window 6) are only evaluated on a single synthetic scene (Drums). The paper should report the default-configuration S2 quality/speed trade-off on real-world scenes as well, since real scenes have lower frame rates and larger inter-frame motion (as acknowledged in Sec. 6.1), which may make the chosen default less representative.
minor comments (5)
  1. [Sec. 5 and Sec. 6] No code, models, or artifacts are released; the 'website' link in Sec. 6.1 is shown only as 'link'. Given the user study and reproducibility claims, providing at least the retrained models and a rendering script would strengthen the paper.
  2. [Sec. 6] The opening of Sec. 6 contains a duplicated sentence: 'Finally, we show that Lumina performs better against GSCore' appears twice. One occurrence should be removed.
  3. [Sec. 3.2, Fig. 12] The color-difference analysis in Fig. 12 reports values on a 0-255 scale, but the y-axis label says 'RGB difference' without units; stating the scale explicitly would prevent confusion.
  4. [Sec. 5, Experimental Setup] The LuminCache is described as caching '64x64 pixels to share across 4x4 image tiles with a size of 16x16'; this phrasing is ambiguous about whether the cache holds 64x64 pixels in total or per tile, and should be clarified.
  5. [Sec. 6.1, Fig. 19] The user-study methodology says each participant reviews 4 traces, but the number of scenes and the comparison conditions are not fully specified; adding a short description of the traces shown and the baseline condition would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central speedup and quality claims are measured outcomes against an unchanged baseline, and the cache-aware fine-tuning is disclosed co-design rather than a fitted value renamed as a prediction.

full rationale

Lumina's central claims (4.5x speedup, 5.3x energy reduction, 55% of color integration avoided) are presented as measured or simulated results over real GPU backends and a cycle-accurate simulator, not as quantities derived from fitted parameters. The RC mechanism's key premise, that rays sharing the first k significant Gaussian IDs have similar colors, is empirically justified in Figs. 11-12 and its known failure mode (large Gaussians, Fig. 13) is explicitly admitted. The mitigation is the scale-constrained loss in Eq. (4), which retrains the scene model so that the caching assumption holds. This is disclosed as end-to-end fine-tuning, and the reported PSNR numbers are measured on the resulting (retrained) Lumina/RC pipeline compared against the original 3DGS baseline. Retraining to make a cache effective is co-design, not circular reasoning: the quality number is not forced by construction, it is an experimental result. S2's trajectory prediction is explicitly attributed to prior work (Cicero) and disclaimed as a contribution, so the self-citation is not load-bearing. No uniqueness theorem is invoked, and no fitted constant is renamed as a prediction. The abstract's '<0.2 dB' claim does appear internally inconsistent with Fig. 20a's 0.3 dB average synthetic loss, but that is a correctness/consistency concern, not evidence of circularity. The empirical and architectural results stand independently of this inconsistency.

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

The central claims rest on several tuned hyperparameters (sharing window, expanded margin, alpha-record length, scale-loss weights) and on two empirical regularities: temporal sorting stability and the sufficiency of significant Gaussian IDs as a ray fingerprint. The quality and speedup numbers also depend on a simulator and on retraining with a scale constraint. No invented physical or mathematical entities are introduced.

free parameters (4)
  • S2 sharing window N = 6 (default)
    Number of consecutive frames reusing one sorting result; set by sensitivity study in Sec. 6.3, not derived. Larger N raises speedup but lowers PSNR.
  • S2 expanded margin = 4 pixels per dimension (default)
    Size of viewport expansion at the predicted sorting pose; chosen from the trade-off in Sec. 6.3 and needed to avoid artifacts (Fig. 8).
  • RC alpha-record length k = 5 (default)
    Number of significant Gaussian IDs used as the cache tag; tuned in Fig. 24 and controls the trade-off between quality and cache hit rate.
  • Scale loss threshold theta and weight alpha = not reported
    Hyperparameters in Eq. 4 that constrain Gaussian scale; not specified, yet the quality results depend on them.
assumptions (4)
  • domain assumption Two rays that intersect the same sequence of Gaussians, or at least share the first k significant Gaussians, will have approximately equal pixel colors.
    Core of RC in Sec. 3.2. It is validated empirically on four scenes, but not proven; the paper's Fig. 13 shows it can fail and requires fine-tuning.
  • domain assumption The depth order of Gaussians stays stable across nearby camera poses, so sorting can be shared over a window of frames.
    Core of S2 in Sec. 3.1. The paper measures 0.2% order changes for significant Gaussians but provides no worst-case bound; Sec. 8 admits rapid head rotations are a pathological case.
  • domain assumption The cycle-accurate simulator plus RTL synthesis, scaled from 16nm to 12nm, faithfully models the performance and energy of a real mobile SoC.
    All headline numbers in Sec. 6.2 come from simulation, with no chip measurement. Latency excludes parallel execution between sorting and rasterization, which may overstate the speedup.
  • ad hoc to paper Retraining the 3DGS model with the scale-constrained loss does not materially change scene representational quality, so comparing against an unretrained baseline is fair.
    Eq. 4 modifies the model to make the cache reliable; the paper reports 0.6 dB average improvement for RC-only with the loss, but does not compare against a baseline retrained with the same loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lumina: Real-Time Mobile Neural Rendering by Exploiting Computational Redundancy." pith.science (2026). https://pith.science/paper/IJXOV4SG

@misc{pith2026250605682,
  author       = {Pith},
  title        = {Pith review of: Lumina: Real-Time Mobile Neural Rendering by Exploiting Computational Redundancy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJXOV4SG}},
  note         = {Machine review of arXiv:2506.05682}
}
read the original abstract

3D Gaussian Splatting (3DGS) has vastly advanced the pace of neural rendering, but it remains computationally demanding on today's mobile SoCs. To address this challenge, we propose Lumina, a hardware-algorithm co-designed system, which integrates two principal optimizations: a novel algorithm, S^2, and a radiance caching mechanism, RC, to improve the efficiency of neural rendering. S2 algorithm exploits temporal coherence in rendering to reduce the computational overhead, while RC leverages the color integration process of 3DGS to decrease the frequency of intensive rasterization computations. Coupled with these techniques, we propose an accelerator architecture, LuminCore, to further accelerate cache lookup and address the fundamental inefficiencies in Rasterization. We show that Lumina achieves 4.5x speedup and 5.3x energy reduction against a mobile Volta GPU, with a marginal quality loss (< 0.2 dB peak signal-to-noise ratio reduction) across synthetic and real-world datasets.

Figures

Figures reproduced from arXiv: 2506.05682 by the authors.

Figure 1
Figure 1. The computation flow of today’s 3DGS rendering pipeline is highlighted in green. Gaussian points are first projected [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Normalized execution breakdown across scenes. Sorting and Rasterization dominate the execution [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Example of a warp execution during Rasterization, as [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figures from the paper (15 more)
Figure 6
Figure 6. Figure 6: The intuition of S 2 algorithm. The rendering orders, a.k.a, the depth orders, of two spatially closed camera poses, 𝑀 and 𝑁, are the same and can be reused. The example in [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 8
Figure 8. Figure 8: Comparison of the rendering results with and with [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: The intuition behind radiance caching. The first six [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 11
Figure 11. Figure 11: The significance of Gaussian points towards the final radiance. Points are sorted by their contributions. 2 3 4 5 6 7 8 9 10 k Value 0.0 0.5 1.0 1.5 2.0 RGB difference [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 10
Figure 10. Figure 10: The overall procedure of cached Rasterization step. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 14
Figure 14. Figure 14: The overall LuminSys system. Our S 2 algorithm predicts the future pose and proactively computes sorting which is later shared across multiple frames. Meanwhile, our radiance caching accelerates the Rasterization step by exploiting similarities across rays. The framew…
Figure 15
Figure 15. Figure 15: An example of the sparse distribution of cache-hit [PITH_FULL_IMAGE:figures/full_fig_p007_15.png]
Figure 16
Figure 16. Figure 16: Hardware support for radiance caching. The lower [PITH_FULL_IMAGE:figures/full_fig_p008_16.png]
Figure 17
Figure 17. Figure 17: The overall SoC architecture design. We integrate our [PITH_FULL_IMAGE:figures/full_fig_p009_17.png]
Figure 18
Figure 18. Figure 18: The detailed architecture of a three-stage PE. Each PE [PITH_FULL_IMAGE:figures/full_fig_p009_18.png]
Figure 19
Figure 19. Figure 19: User study of Lumina compared against original 3DGS. Among 27% of participants who notice differences, we achieve a 50%-50% tie. use sharing window to denote the number of frames that share a single sorting result. We use expanded margin to denote the number of pixels…
Figure 21
Figure 21. Figure 21: Rendering quality and cache hit rate of RC-only with and without scale-constrained loss, 𝐿𝑠𝑐𝑎𝑙𝑒 , in Sec. 3.3. (a) Normalized speedup. Higher is better. 1.1 1.2 1.2 (b) Normalized energy. Lower is better [PITH_FULL_IMAGE:figures/full_fig_p011_21.png]
Figure 22
Figure 22. Figure 22: Speedup and normalized energy consumptions of [PITH_FULL_IMAGE:figures/full_fig_p011_22.png]
Figure 20
Figure 20. Figure 20: Image quality comparison. Both S 2 -only and Lumina configure the expanded margin (i.e., the number of pixels that the sorting viewport expands) of 4 and the skipping window (i.e., the number of frames share a single sorting result) of 6. DS-2 first renders a 2× downs…
Figure 24
Figure 24. Figure 24: The sensitivity of ren￾dering quality and normal￾ized speedup to the number of significant Gaussians [PITH_FULL_IMAGE:figures/full_fig_p012_24.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 58 canonical work pages

  1. [1]

    Apple Vision Pro screen refresh rate is up to 100Hz

  2. [2]

    Meta Quest Pro specs

  3. [3]

    Micron 178-Ball, Single-Channel Mobile LPDDR3 SDRAM Features

    work page

  4. [4]

    Micron System Power Calculators

    work page

  5. [5]

    Nvidia reveals xavier soc details

    work page

  6. [6]

    NVIDIA’s Xavier System-on-Chip, HotChips 30

    work page

  7. [7]

    Qualcomm Powers Next-Gen Spatial Computing With XR2 Gen 2 And AR1 Gen 1 Platforms

    work page

  8. [8]

    Qualcomm QCS8550/QCM8550 Processors

    work page

Show all 82 references
  1. [9]

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 58...

    2021

  2. [10]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields.CVPR, 2022

    2022

  3. [11]

    The physics of optimal decision making: a formal analysis of models of perfor- mance in two-alternative forced-choice tasks.Psychological review, 113(4):700, 2006

    Rafal Bogacz, Eric Brown, Jeff Moehlis, Philip Holmes, and Jonathan D Cohen. The physics of optimal decision making: a formal analysis of models of perfor- mance in two-alternative forced-choice tasks.Psychological review, 113(4):700, 2006

    2006

  4. [12]

    Eva2: Exploiting temporal redundancy in live computer vision

    Mark Buckler, Philip Bedoukian, Suren Jayasuriya, and Adrian Sampson. Eva2: Exploiting temporal redundancy in live computer vision. In2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA), pages 533–546. IEEE, 2018

    2018

  5. [13]

    Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1–17, 2020

    Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1–17, 2020

    2020

  6. [14]

    A survey on 3d gaussian splatting.arXiv preprint arXiv:2401.03890, 2024

    Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting.arXiv preprint arXiv:2401.03890, 2024

    2024 arXiv

  7. [15]

    Geometry-guided progressive nerf for generalizable and efficient neural human rendering

    Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yujun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. InEuropean Conference on Computer Vision, pages 222–239. Springer, 2022

    2022

  8. [16]

    Quicktime vr: An image-based approach to virtual envi- ronment navigation

    Shenchang Eric Chen. Quicktime vr: An image-based approach to virtual envi- ronment navigation. InProceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 29–38, 1995

    1995

  9. [17]

    View interpolation for image synthe- sis

    Shenchang Eric Chen and Lance Williams. View interpolation for image synthe- sis. InSeminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 423–432. 2023

    2023

  10. [18]

    Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures

    Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p...

    2023

  11. [19]

    An overview of the sparse basic linear algebra subprograms: The new standard from the blas technical forum

    Iain S Duff, Michael A Heroux, and Roldan Pozo. An overview of the sparse basic linear algebra subprograms: The new standard from the blas technical forum. ACM Transactions on Mathematical Software (TOMS), 28(2):239–267, 2002

    2002

  12. [20]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023

    2023 arXiv

  13. [21]

    Mini-splatting: Representing scenes with a constrained number of gaussians.arXiv preprint arXiv:2403.14166, 2024

    Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians.arXiv preprint arXiv:2403.14166, 2024

    2024 arXiv

  14. [22]

    Real-time gaze tracking with event-driven eye segmentation

    Yu Feng, Nathan Goulding-Hotta, Asif Khan, Hans Reyserhove, and Yuhao Zhu. Real-time gaze tracking with event-driven eye segmentation. In2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 399–408. IEEE, 2022

    2022

  15. [23]

    Fast and accurate: Video enhancement using sparse depth

    Yu Feng, Patrick Hansen, Paul N Whatmough, Guoyu Lu, and Yuhao Zhu. Fast and accurate: Video enhancement using sparse depth. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4492– 4500, 2023

    2023

  16. [24]

    Potamoi: Accelerating neural rendering via a unified streaming architecture.ACM Transactions on Architecture and Code Optimization, 2024

    Yu Feng, Weikai Lin, Zihan Liu, Jingwen Leng, Minyi Guo, Han Zhao, Xiaofeng Hou, Jieru Zhao, and Yuhao Zhu. Potamoi: Accelerating neural rendering via a unified streaming architecture.ACM Transactions on Architecture and Code Optimization, 2024

    2024

  17. [25]

    Cicero: Address- ing algorithmic and architectural bottlenecks in neural rendering by radiance warping and memory optimizations.arXiv preprint arXiv:2404.11852, 2024

    Yu Feng, Zihan Liu, Jingwen Leng, Minyi Guo, and Yuhao Zhu. Cicero: Address- ing algorithmic and architectural bottlenecks in neural rendering by radiance warping and memory optimizations.arXiv preprint arXiv:2404.11852, 2024

    2024 arXiv

  18. [26]

    Asv: Accelerated stereo vision system

    Yu Feng, Paul Whatmough, and Yuhao Zhu. Asv: Accelerated stereo vision system. InProceedings of the 52nd Annual IEEE/ACM International Symposium 13 Yu Feng, Weikai Lin, Yuge Cheng, Zihan Liu, Jingwen Leng, Minyi Guo, Chen Chen, Shixuan Sun, and Yuhao Zhu on Microarchitecture, ...

    2019

  19. [27]

    Gen-nerf: Efficient and generalizable neural radiance fields via algorithm-hardware co-design

    Yonggan Fu, Zhifan Ye, Jiayi Yuan, Shunyao Zhang, Sixu Li, Haoran You, and Yingyan Lin. Gen-nerf: Efficient and generalizable neural radiance fields via algorithm-hardware co-design. InProceedings of the 50th Annual International Symposium on Computer Architecture, pages 1–12, 2023

    2023

  20. [28]

    Dynamic warp formation and scheduling for efficient gpu control flow

    Wilson WL Fung, Ivan Sham, George Yuan, and Tor M Aamodt. Dynamic warp formation and scheduling for efficient gpu control flow. In40th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO 2007), pages 407–420. IEEE, 2007

    2007

  21. [29]

    Nerf: Neural radiance field in 3d vision, a comprehensive review.arXiv preprint arXiv:2210.00379, 2022

    Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review.arXiv preprint arXiv:2210.00379, 2022

    2022 arXiv

  22. [30]

    Tetris: Scalable and efficient neural network acceleration with 3d memory

    Mingyu Gao, Jing Pu, Xuan Yang, Mark Horowitz, and Christos Kozyrakis. Tetris: Scalable and efficient neural network acceleration with 3d memory. InPro- ceedings of the 22nd ACM International Conference on Architectural Support for Programming Languages and Operating Systems, 2017

    2017

  23. [31]

    Megatrack: monochrome egocentric articulated hand-tracking for virtual reality

    Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG), 39(4):87–1, 2020

    2020

  24. [32]

    Deep blending for free-viewpoint image-based rendering

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    2018

  25. [33]

    Baking neural radiance fields for real-time view synthesis

    Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875–5884, 2021

    2021

  26. [34]

    Head movement amplitude and velocity during a common visual task.Investigative Ophthalmology & Visual Science, 43(13):4668–4668, 2002

    PL Hendicott, B Brown, KL Schmid, and S Fisher. Head movement amplitude and velocity during a common visual task.Investigative Ophthalmology & Visual Science, 43(13):4668–4668, 2002

    2002

  27. [35]

    Efficientnerf efficient neural radiance fields

    Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902–12911, 2022

    2022

  28. [36]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. InACM SIGGRAPH 2024 conference papers, pages 1–11, 2024

    2024

  29. [37]

    Neuman: Neural human radiance field from a single video

    Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. InEuropean Conference on Computer Vision, pages 402–418. Springer, 2022

    2022

  30. [38]

    Parallel multiple-bounce irradiance caching

    Nathaniel L Jones and Christoph F Reinhart. Parallel multiple-bounce irradiance caching. InComputer Graphics Forum, volume 35, pages 57–66. Wiley Online Library, 2016

    2016

  31. [39]

    Adapt- ing cache partitioning algorithms to pseudo-lru replacement policies

    Kamil Kędzierski, Miquel Moreto, Francisco J Cazorla, and Mateo Valero. Adapt- ing cache partitioning algorithms to pseudo-lru replacement policies. In2010 IEEE International Symposium on Parallel & Distributed Processing (IPDPS), pages 1–12. IEEE, 2010

    2010

  32. [40]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):1–14, 2023

    2023

  33. [41]

    A hierarchical 3d gaussian repre- sentation for real-time rendering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian repre- sentation for real-time rendering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    2024

  34. [42]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4), 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4), 2017

    2017

  35. [43]

    Radiance caching for efficient global illumination computation.IEEE Transactions on Visualization and Computer Graphics, 11(5):550–561, 2005

    Jaroslav Krivánek, Pascal Gautron, Sumanta Pattanaik, and Kadi Bouatouch. Radiance caching for efficient global illumination computation.IEEE Transactions on Visualization and Computer Graphics, 11(5):550–561, 2005

    2005

  36. [44]

    Compact 3d gaussian representation for radiance field.arXiv preprint arXiv:2311.13681, 2023

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field.arXiv preprint arXiv:2311.13681, 2023

    2023 arXiv

  37. [45]

    Neurex: A case for neural rendering acceleration

    Junseo Lee, Kwanseok Choi, Jungi Lee, Seokwon Lee, Joonho Whangbo, and Jaewoong Sim. Neurex: A case for neural rendering acceleration. InProceedings of the 50th Annual International Symposium on Computer Architecture, pages 1–13, 2023

    2023

  38. [46]

    Vr-pipe: Streamlining hardware graphics pipeline for volume rendering.arXiv preprint arXiv:2502.17078, 2025

    Junseo Lee, Jaisung Kim, Junyong Park, and Jaewoong Sim. Vr-pipe: Streamlining hardware graphics pipeline for volume rendering.arXiv preprint arXiv:2502.17078, 2025

    2025 arXiv

  39. [47]

    Gscore: Efficient radiance field rendering via architectural support for 3d gaussian splat- ting

    Junseo Lee, Seokwon Lee, Jungi Lee, Junyong Park, and Jaewoong Sim. Gscore: Efficient radiance field rendering via architectural support for 3d gaussian splat- ting. InProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and O...

    2024

  40. [48]

    Rt-nerf: Real- time on-device neural radiance fields towards immersive ar/vr rendering

    Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real- time on-device neural radiance fields towards immersive ar/vr rendering. In Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, pages 1–9, 2022

    2022

  41. [49]

    Instant-3d: Instant neural radiance field training towards on-device ar/vr 3d reconstruction

    Sixu Li, Chaojian Li, Wenbo Zhu, Boyang Yu, Yang Zhao, Cheng Wan, Haoran You, Huihong Shi, and Yingyan Lin. Instant-3d: Instant neural radiance field training towards on-device ar/vr 3d reconstruction. InProceedings of the 50th Annual International Symposium on Computer Archit...

    2023

  42. [50]

    Envidr: Implicit differentiable renderer with neural envi- ronment lighting.arXiv preprint arXiv:2303.13022, 2023

    Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural envi- ronment lighting.arXiv preprint arXiv:2303.13022, 2023

    2023 arXiv

  43. [51]

    Metasapiens: Real-time neural rendering with efficiency-aware pruning and accelerated foveated rendering

    Weikai Lin, Yu Feng, and Yuhao Zhu. Metasapiens: Real-time neural rendering with efficiency-aware pruning and accelerated foveated rendering. InProceedings of the 30th ACM International Conference on Architectural Support for Program- ming Languages and Operating Systems, Volu...

    2025

  44. [52]

    Bacon: Band-limited coordinate networks for multiscale scene representation

    David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16252–16262, 2022

    2022

  45. [53]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians

    Yang Liu, He Guan, Chuanchen Luo, Lue Fan, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. arXiv preprint arXiv:2404.01133, 2024

    2024 arXiv

  46. [54]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039–18048, 2024

    2024

  47. [55]

    Dynamic warp subdivision for integrated branch and memory divergence tolerance

    Jiayuan Meng, David Tarjan, and Kevin Skadron. Dynamic warp subdivision for integrated branch and memory divergence tolerance. InProceedings of the 37th annual international symposium on Computer architecture, pages 235–246, 2010

    2010

  48. [56]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  49. [57]

    Hardware acceleration of neural graphics

    Muhammad Husnain Mubarik, Ramakrishna Kanungo, Tobias Zirr, and Rakesh Kumar. Hardware acceleration of neural graphics. InProceedings of the 50th Annual International Symposium on Computer Architecture, pages 1–12, 2023

    2023

  50. [58]

    Real-time neural radiance caching for path tracing.arXiv preprint arXiv:2106.12372, 2021

    Thomas Müller, Fabrice Rousselle, Jan Novák, and Alexander Keller. Real-time neural radiance caching for path tracing.arXiv preprint arXiv:2106.12372, 2021

    2021 arXiv

  51. [59]

    Improving gpu performance via large warps and two-level warp scheduling

    Veynu Narasiman, Michael Shebanow, Chang Joo Lee, Rustam Miftakhutdinov, Onur Mutlu, and Yale N Patt. Improving gpu performance via large warps and two-level warp scheduling. InProceedings of the 44th Annual IEEE/ACM International Symposium on Microarchitecture, pages 308–317, 2011

    2011

  52. [60]

    Icarus: A specialized architecture for neural radiance fields rendering.ACM Transactions on Graphics (TOG), 41(6):1–14, 2022

    Chaolin Rao, Huangjie Yu, Haochuan Wan, Jindong Zhou, Yueyang Zheng, Minye Wu, Yu Ma, Anpei Chen, Binzhe Yuan, Pingqiang Zhou, et al. Icarus: A specialized architecture for neural radiance fields rendering.ACM Transactions on Graphics (TOG), 41(6):1–14, 2022

    2022

  53. [61]

    Re-rend: Real-time rendering of nerfs across devices

    Sara Rojas, Jesus Zarzar, Juan C Pérez, Artsiom Sanakoyeu, Ali Thabet, Albert Pumarola, and Bernard Ghanem. Re-rend: Real-time rendering of nerfs across devices. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3632–3641, 2023

    2023

  54. [62]

    Deepscaletool: A tool for the accurate estima- tion of technology scaling in the deep-submicron era

    Satyabrata Sarangi and Bevan Baas. Deepscaletool: A tool for the accurate estima- tion of technology scaling in the deep-submicron era. In2021 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2021

    2021

  55. [63]

    Pre-convolved radiance caching

    Daniel Scherzer, Chuong H Nguyen, Tobias Ritschel, and Hans-Peter Seidel. Pre-convolved radiance caching. InComputer Graphics Forum, volume 31, pages 1391–1397. Wiley Online Library, 2012

    2012

  56. [64]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  57. [65]

    Vr-dann: Real-time video recognition via decoder-assisted neural net- work acceleration

    Zhuoran Song, Feiyang Wu, Xueyuan Liu, Jing Ke, Naifeng Jing, and Xiaoyao Liang. Vr-dann: Real-time video recognition via decoder-assisted neural net- work acceleration. In2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 698–710. IEEE, 2020

  58. [66]

    Scaling equations for the accurate prediction of cmos device performance from 180 nm to 7 nm.Integration, 58:74–81, 2017

    Aaron Stillmaker and Bevan Baas. Scaling equations for the accurate prediction of cmos device performance from 180 nm to 7 nm.Integration, 58:74–81, 2017

    2017

  59. [67]

    Block-nerf: Scalable large scene neural view synthesis

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2022

  60. [68]

    Real-time radiance caching using chrominance compression.Journal of Computer Graphics Techniques Vol, 3(4), 2014

    K Vardis, G Papaioannou, and A Gkaravelis. Real-time radiance caching using chrominance compression.Journal of Computer Graphics Techniques Vol, 3(4), 2014

    2014

  61. [69]

    Effect of frame rate on user experience, performance, and simulator sickness in virtual reality.IEEE Transactions on Visualization and Computer Graphics, 29(5):2478–2488, 2023

    Jialin Wang, Rongkai Shi, Wenxuan Zheng, Weijie Xie, Dominic Kao, and Hai- Ning Liang. Effect of frame rate on user experience, performance, and simulator sickness in virtual reality.IEEE Transactions on Visualization and Computer Graphics, 29(5):2478–2488, 2023

    2023

  62. [70]

    Humannerf: Free-viewpoint rendering of moving people from monocular video

    Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humannerf: Free-viewpoint rendering of moving people from monocular video. InProceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210–16220, 2022

    2022

  63. [71]

    Recent advances in 3d gaussian splatting.arXiv preprint arXiv:2403.11134, 2024

    Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting.arXiv preprint arXiv:2403.11134, 2024

    2024 arXiv

  64. [72]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks.IEEE transactions 14 Lumina: Real-Time Mobile Neural Rendering by Exploiting Computational Redundancy on neural networks and learning systems, 32...

    2020

  65. [73]

    Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering

    Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. InEuropean conference on computer vision, pages 106–122. Springer, 2022

    2022

  66. [74]

    Neural supersampling for real-time rendering.ACM Transactions on Graphics (TOG), 39(4):142–1, 2020

    Lei Xiao, Salah Nouri, Matt Chapman, Alexander Fix, Douglas Lanman, and An- ton Kaplanyan. Neural supersampling for real-time rendering.ACM Transactions on Graphics (TOG), 39(4):142–1, 2020

    2020

  67. [75]

    Gs-slam: Dense visual slam with 3d gaussian splatting

    Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595–19604, 2024

    2024

  68. [76]

    Ganax: A unified mimd-simd acceleration for generative adversarial networks

    Amir Yazdanbakhsh, Kambiz Samadi, Nam Sung Kim, and Hadi Esmaeilzadeh. Ganax: A unified mimd-simd acceleration for generative adversarial networks. 2018

    2018

  69. [77]

    Intrinsicnerf: Learning intrinsic neural radiance fields for editable novel view synthesis

    Weicai Ye, Shuo Chen, Chong Bao, Hujun Bao, Marc Pollefeys, Zhaopeng Cui, and Guofeng Zhang. Intrinsicnerf: Learning intrinsic neural radiance fields for editable novel view synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 339–351, 2023

    2023

  70. [78]

    Exploiting frame similarity for efficient inference on edge devices

    Ziyu Ying, Shulin Zhao, Haibo Zhang, Cyan Subhra Mishra, Sandeepa Bhuyan, Mahmut T Kandemir, Anand Sivasubramaniam, and Chita R Das. Exploiting frame similarity for efficient inference on edge devices. In2022 IEEE 42nd International Conference on Distributed Computing Systems ...

  71. [79]

    Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation

    Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. InProceedings of the IEEE/CVF Conference on Computer Vision an...

    2022

  72. [80]

    Déja view: Spatio-temporal compute reuse for ‘energy-efficient 360 vr video streaming

    Shulin Zhao, Haibo Zhang, Sandeepa Bhuyan, Cyan Subhra Mishra, Ziyu Ying, Mahmut T Kandemir, Anand Sivasubramaniam, and Chita R Das. Déja view: Spatio-temporal compute reuse for ‘energy-efficient 360 vr video streaming. In 2020 ACM/IEEE 47th Annual International Symposium on C...

    2020

  73. [81]

    Holoar: On-the-fly optimization of 3d holographic processing for augmented reality

    Shulin Zhao, Haibo Zhang, Cyan Subhra Mishra, Sandeepa Bhuyan, Ziyu Ying, Mahmut Taylan Kandemir, Anand Sivasubramaniam, and Chita Das. Holoar: On-the-fly optimization of 3d holographic processing for augmented reality. In MICRO-54: 54th Annual IEEE/ACM International Symposium...

    2021

  74. [82]

    Euphrates: Algorithm-soc co-design for low-power mobile continuous vision.arXiv preprint arXiv:1803.11232, 2018

    Yuhao Zhu, Anand Samajdar, Matthew Mattina, and Paul Whatmough. Euphrates: Algorithm-soc co-design for low-power mobile continuous vision.arXiv preprint arXiv:1803.11232, 2018. 15

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.