pith. sign in

arxiv: 2605.31500 · v1 · pith:ZIAKOAF7new · submitted 2026-05-29 · 💻 cs.LG · cs.AI

On Efficient Scaling of GNNs via IO-Aware Layers Implementations

Daria Fomina , Daniil Krasylnikov , Alexey Boykov , Andrey Dolgovyazov , Vyacheslav Zhdanovskiy , Fedor Velikonivtsev This is my paper

Pith reviewed 2026-06-28 22:51 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords Graph Neural NetworksGPU kernelsSparse matrix multiplicationAttention layersMemory optimizationScalabilityGATGraph Transformer
0
0 comments X

The pith

Custom GPU kernels for GNN attention and aggregation layers cut memory traffic by avoiding edge-wise intermediates.

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

The paper claims that popular GNN frameworks like DGL create excessive memory traffic by materializing edge data for complex layers such as attention. It develops specialized GPU kernels for three families of operations—SpMM convolutions, reductions, and attention—that improve locality and reduce data movement while staying robust across graph structures. These changes produce measured speedups and memory savings on real workloads. A reader would care because the approach directly targets the I/O bottleneck that limits scaling GNNs to larger graphs without changing the model math.

Core claim

The authors show that GNN layers fall into SpMM-based, reduction-based, and attention-based kernel families, then implement fused GPU kernels for each that reduce data movement and improve locality; the attention kernels reach up to 3.9× speedup for Graph Transformer and 8.5× for GATv2 with up to 76× lower peak memory, degree-aware reductions reach 10×, and cached cuSPARSE reaches 8× over DGL.

What carries the argument

IO-aware fused kernels for the three families (SpMM convolutions, reduction aggregations, attention layers) that avoid materializing edge-wise intermediates.

If this is right

  • Graph reordering improves performance more for neighbor-parallel gather kernels than for feature-parallel designs.
  • Tensor Core block-sparse variants of the attention kernels yield additional speedups on locally dense graphs.
  • The kernels function as drop-in replacements that support reproducible hardware-aware acceleration.
  • Properly cached cuSPARSE can outperform custom SpMM baselines in most evaluated cases.

Where Pith is reading between the lines

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

  • The same I/O-centric redesign could be applied to other sparse tensor operations outside GNNs.
  • Lower memory footprints may allow training or inference on graphs that currently exceed GPU capacity.
  • The median speedups suggest the gains are consistent rather than limited to a few outlier graphs.

Load-bearing premise

Standard frameworks routinely materialize edge-wise intermediates that dominate memory traffic for complex GNN layers.

What would settle it

Run the released kernels against DGL on the same large graph and layer and measure whether peak memory and wall-clock time match the reported reductions.

Figures

Figures reproduced from arXiv: 2605.31500 by Alexey Boykov, Andrey Dolgovyazov, Daniil Krasylnikov, Daria Fomina, Fedor Velikonivtsev, Vyacheslav Zhdanovskiy.

Figure 1
Figure 1. Figure 1: Reordering gap ∆s on selected datasets (dot-aggr). ∆s > 0 means reordering helps neighbor-parallel (gather) more. We compare scalar vs. 16B vectorized loads (vec); boxes aggregate over partition sizes. improve forward throughput on some graphs (e.g., 2.69× on tolokers-2 and 1.50× on pokec-regions) but is not universally faster in backward due to scatter/atomic accumulation across tiles (e.g., 0.77× on ogbn… view at source ↗
Figure 2
Figure 2. Figure 2: Reordering gap ∆s = sneighbor − sfeature on all datasets (dot-aggr). ∆s > 0 means reordering helps neighbor-parallel (gather) more; ∆s < 0 favors feature-parallel (coalesced). We sweep D ∈ {32, 64, 128, 256, 512} with scalar vs. 16B vectorized loads (vec); boxes aggregate over partition sizes. CSR matrix descriptor (cusparseSpMatDescr t) that references the CSR structure and the (possibly normalized) value… view at source ↗
read the original abstract

Graph Neural Networks (GNNs) are bottlenecked by sparse, irregular memory access. Popular frameworks such as DGL and PyTorch Geometric support general message passing, but complex layers often materialize edge-wise intermediates, increasing memory traffic and limiting scalability on large graphs. We take an I/O- and arithmetic-intensity--centric view and show that widely used layers fall into three kernel families: SpMM-based convolutions, reduction-based aggregations, and attention-based layers (GATv2/Graph Transformer). For each family, we develop GPU kernels that reduce data movement, improve locality, and remain robust across realistic graphs. We also study graph reordering and find that its impact depends on the kernel mapping: it benefits neighbor-parallel (gather-dominated) kernels more consistently than feature-parallel designs. Empirically, our fused attention kernels reach up to $\textbf{3.9}\times$ speedup for Graph Transformer (median $\textbf{1.6}\times$), with Tensor Core (block-sparse) variants up to $\textbf{7.3}\times$ on locally dense graphs; for GATv2 we reach up to $\textbf{8.5}\times$ speedup (median $\textbf{2.0}\times$) while reducing peak memory by up to $\textbf{76}\times$ (median $\textbf{6}\times$). Our degree-aware reduction kernels achieve up to $\textbf{10}\times$ speedup (median $\textbf{2.6}\times$). For SpMM-based layers, properly cached cuSPARSE achieves up to $\textbf{8}\times$ speedup over DGL and outperforms evaluated custom baselines in the majority of evaluations. We release our implementations as drop-in replacements to support reproducible, hardware-aware GNN acceleration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript develops IO-aware GPU kernels for three GNN layer families (SpMM convolutions, reduction aggregations, attention layers including GATv2 and Graph Transformer) to reduce memory traffic from edge-wise intermediates. It reports empirical speedups of up to 3.9× (median 1.6×) for Graph Transformer, 8.5× (median 2.0×) for GATv2 with up to 76× memory reduction, 10× (median 2.6×) for reductions, and 8× for properly cached cuSPARSE over DGL; it also examines graph reordering effects and releases the kernels as drop-in replacements.

Significance. If the performance claims are reproducible and the speedups are attributable to the IO-aware redesign rather than implementation artifacts, the work would offer practical, hardware-aware accelerations for scaling attention-based GNNs on large graphs. The release of implementations is a clear strength for reproducibility.

major comments (3)
  1. [Abstract] Abstract: the central motivation asserts that 'complex layers often materialize edge-wise intermediates, increasing memory traffic and limiting scalability' in DGL/PyG, yet no profiling data, byte-movement breakdown, or source-level comparison of GATv2/Graph Transformer implementations is supplied to establish this as the dominant cost versus launch overhead or SpMM tiling.
  2. [Experimental results] Experimental results (throughout): speedups and memory reductions are stated as maxima and medians (e.g., 3.9×, 8.5×, 76×) with no error bars, number of runs, dataset statistics, or explicit baseline configuration details (hardware, DGL/PyG versions, compilation flags), making it impossible to judge whether selection or measurement artifacts affect the claims.
  3. [Results on graph reordering] Results on graph reordering: the claim that reordering 'benefits neighbor-parallel kernels more consistently than feature-parallel designs' is presented without quantitative tables or figures showing the interaction between reordering and the three kernel families, which is needed to support the kernel-mapping dependence.
minor comments (1)
  1. [Abstract] The abstract and text use 'median' and 'up to' without clarifying the distribution of graphs or layers over which the median is computed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which has helped identify areas where the manuscript can be strengthened for clarity and reproducibility. We have revised the paper to address each major comment as described below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central motivation asserts that 'complex layers often materialize edge-wise intermediates, increasing memory traffic and limiting scalability' in DGL/PyG, yet no profiling data, byte-movement breakdown, or source-level comparison of GATv2/Graph Transformer implementations is supplied to establish this as the dominant cost versus launch overhead or SpMM tiling.

    Authors: We agree that explicit empirical support strengthens the motivation. In the revised manuscript we add a new profiling subsection (with Nsight Compute traces) that quantifies byte movement for DGL/PyG GATv2 and Graph Transformer on representative graphs; the data show edge-wise intermediate materialization accounts for the majority of memory traffic relative to launch overhead or SpMM tiling. revision: yes

  2. Referee: [Experimental results] Experimental results (throughout): speedups and memory reductions are stated as maxima and medians (e.g., 3.9×, 8.5×, 76×) with no error bars, number of runs, dataset statistics, or explicit baseline configuration details (hardware, DGL/PyG versions, compilation flags), making it impossible to judge whether selection or measurement artifacts affect the claims.

    Authors: The referee correctly notes insufficient experimental detail. We have revised the experimental section and appendix to report: standard-deviation error bars over 5 runs, exact run counts, full per-graph statistics (nodes/edges/features), and complete baseline configurations (A100 80 GB, DGL 1.1.0, PyG 2.3.0, CUDA 12.0, -O3 flags). All timing excludes host-device transfer. revision: yes

  3. Referee: [Results on graph reordering] Results on graph reordering: the claim that reordering 'benefits neighbor-parallel kernels more consistently than feature-parallel designs' is presented without quantitative tables or figures showing the interaction between reordering and the three kernel families, which is needed to support the kernel-mapping dependence.

    Authors: We acknowledge the request for explicit quantification. The revised manuscript adds Figure 7 and Table 4 that report per-kernel-family speedups (with/without METIS reordering) across eight graphs; the numbers confirm more consistent gains for neighbor-parallel kernels (median 1.4×) versus feature-parallel attention kernels (median 1.1×). revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical kernel benchmarks with no derivation chain

full rationale

The manuscript describes development of custom GPU kernels for SpMM, reductions, and attention layers in GNNs, followed by empirical timing and memory measurements on graphs. No equations, fitted parameters, predictions, or uniqueness theorems appear in the provided text. All performance claims (speedups, memory reductions) are presented as direct measurement outcomes rather than derived quantities. The opening motivation about DGL/PyG materialization is an unverified assumption but does not participate in any self-referential derivation or reduction to inputs. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are introduced; the contribution consists of concrete kernel implementations and empirical measurements.

pith-pipeline@v0.9.1-grok · 5879 in / 1169 out tokens · 20486 ms · 2026-06-28T22:51:09.029878+00:00 · methodology

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

35 extracted references · 8 canonical work pages

  1. [1]

    URL https://link

    doi: 10.1103/RevModPhys.74.47. URL https://link. aps.org/doi/10.1103/RevModPhys.74.47. Bazhenov, G., Platonov, O., and Prokhorenkova, L. Graph- Land: Evaluating graph machine learning models on di- verse industrial data.Advances in Neural Information Processing Systems,

    work page doi:10.1103/revmodphys.74.47

  2. [2]

    Brody, S., Alon, U., and Yahav, E

    URL https: //arxiv.org/abs/2308.12093. Brody, S., Alon, U., and Yahav, E. How attentive are graph attention networks?,

    arXiv

  3. [3]

    org/abs/2105.14491

    URL https://arxiv. org/abs/2105.14491. Chen, T., Moreau, T., Jiang, Z., Zheng, L., Yan, E., Cowan, M., Shen, H., Wang, L., Hu, Y ., Ceze, L., Guestrin, C., and Krishnamurthy, A. Tvm: An automated end- to-end optimizing compiler for deep learning,

    Pith/arXiv arXiv

  4. [4]

    Chen, Z., Yan, M., Zhu, M., Deng, L., Li, G., Li, S., and Xie, Y

    URL https://arxiv.org/abs/1802.04799. Chen, Z., Yan, M., Zhu, M., Deng, L., Li, G., Li, S., and Xie, Y . fusegnn: accelerating graph convolutional neu- ral network training on gpgpu. InProceedings of the 39th International Conference on Computer-Aided De- sign, ICCAD ’20, New York, NY , USA,

    Pith/arXiv arXiv

  5. [5]

    ISBN 9781450380263

    Associa- tion for Computing Machinery. ISBN 9781450380263. doi: 10.1145/3400302.3415610. URL https://doi. org/10.1145/3400302.3415610. Dao, T. Flashattention-2: Faster attention with better par- allelism and work partitioning,

    work page doi:10.1145/3400302.3415610

  6. [6]

    Dao, T., Fu, D

    URL https: //arxiv.org/abs/2307.08691. Dao, T., Fu, D. Y ., Ermon, S., Rudra, A., and R´e, C. Flashat- tention: Fast and memory-efficient exact attention with io- awareness,

    Pith/arXiv arXiv

  7. [7]

    Fey, M., Hu, W., Huang, K., Lenssen, J

    URL https://arxiv.org/abs/ 2205.14135. Fey, M., Hu, W., Huang, K., Lenssen, J. E., Ranjan, R., Robinson, J., Ying, R., You, J., and Leskovec, J. Rela- tional deep learning: Graph representation learning on relational databases,

    Pith/arXiv arXiv

  8. [8]

    Gilmer, J., Schoenholz, S

    URL https:// arxiv.org/abs/2507.16991. Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry,

    arXiv

  9. [9]

    Hamilton, W., Ying, Z., and Leskovec, J

    URL https://arxiv.org/abs/ 1704.01212. Hamilton, W., Ying, Z., and Leskovec, J. Inductive repre- sentation learning on large graphs.Advances in Neural Information Processing Systems, 30,

    Pith/arXiv arXiv

  10. [10]

    Hu, W., Fey, M., Zitnik, M., Dong, Y ., Ren, H., Liu, B., Catasta, M., and Leskovec, J

    doi: 10.1145/ 3293883.3295712. Hu, W., Fey, M., Zitnik, M., Dong, Y ., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs,

    arXiv

  11. [11]

    Huang, G., Dai, G., Wang, Y ., and Yang, H

    URL https://arxiv.org/abs/2005.00687. Huang, G., Dai, G., Wang, Y ., and Yang, H. Ge-spmm: General-purpose sparse matrix-matrix multiplication on gpus for graph neural networks,

    arXiv 2005

  12. [12]

    Ivanov, A., Dryden, N., Ben-Nun, T., Li, S., and Hoefler, T

    URL https: //arxiv.org/abs/2007.03179. Ivanov, A., Dryden, N., Ben-Nun, T., Li, S., and Hoefler, T. Data movement is all you need: A case study on optimizing transformers,

    arXiv 2007

  13. [13]

    org/abs/2007.00072

    URL https://arxiv. org/abs/2007.00072. Karypis, G. and Kumar, V . Metis—a software package for partitioning unstructured graphs, partitioning meshes and computing fill-reducing ordering of sparse matrices. 01

    arXiv 2007

  14. [14]

    URL https://arxiv.org/abs/2403.17410. Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks,

    arXiv

  15. [15]

    URL https: //arxiv.org/abs/1609.02907. 9 On Efficient Scaling of GNNs via IO-Aware Layers Implementations Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton- Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., and Battaglia, P. Graphcast: ...

    Pith/arXiv arXiv

  16. [16]

    doi: 10.1109/tpds.2020.3030548

    ISSN 2161-9883. doi: 10.1109/tpds.2020.3030548. URL http://dx.doi. org/10.1109/TPDS.2020.3030548. Li, Z. and Chandramowlishwaran, A. Fused3s: Fast sparse attention on tensor cores,

    work page doi:10.1109/tpds.2020.3030548 2020

  17. [17]

    org/abs/2505.08098

    URLhttps://arxiv. org/abs/2505.08098. Liu, J., Cai, Z., Chen, Z., and Wang, M. Df-gnn: Dynamic fusion framework for attention graph neural networks on gpus,

    arXiv

  18. [18]

    Matthias, F., Jinu, S., Akihiro, N., Rishi, P., Manan, S., Blaˇz, S., Bendias, R., Barghi, A., Vid, K., Zecheng, Z., Xinwei, H., E., L

    URL https://arxiv.org/abs/ 2411.16127. Matthias, F., Jinu, S., Akihiro, N., Rishi, P., Manan, S., Blaˇz, S., Bendias, R., Barghi, A., Vid, K., Zecheng, Z., Xinwei, H., E., L. J., and Jure, L. PyG 2.0: Scalable learn- ing on real world graphs. InTemporal Graph Learning Workshop @ KDD,

    arXiv

  19. [19]

    Min, E., Chen, R., Bian, Y ., Xu, T., Zhao, K., Huang, W., Zhao, P., Huang, J., Ananiadou, S., and Rong, Y

    URL https: //arxiv.org/abs/2409.11129. Min, E., Chen, R., Bian, Y ., Xu, T., Zhao, K., Huang, W., Zhao, P., Huang, J., Ananiadou, S., and Rong, Y . Trans- former for graphs: An overview from architecture per- spective.arXiv preprint arXiv:2202.08455,

    arXiv

  20. [20]

    URL https://doi.org/ 10.1137/S003614450342480

    1137/S003614450342480. URL https://doi.org/ 10.1137/S003614450342480. NVIDIA. Cuda programming guide. https://docs.nvidia.com/cuda/ cuda-programming-guide/. Accessed 2026-01-

    work page doi:10.1137/s003614450342480 2026

  21. [21]

    Peng, H., Xie, X., Shivdikar, K., Hasan, M., Zhao, J., Huang, S., Khan, O., Kaeli, D., and Ding, C

    URL https://arxiv.org/abs/1912.01703. Peng, H., Xie, X., Shivdikar, K., Hasan, M., Zhao, J., Huang, S., Khan, O., Kaeli, D., and Ding, C. Maxk-gnn: Ex- tremely fast gpu kernel design for accelerating graph neu- ral networks training. InProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Sys...

    Pith/arXiv arXiv 1912

  22. [22]

    doi: 10.1109/InPar.2012.6339602

    ISBN 978-1-4673-2632-2. doi: 10.1109/InPar.2012.6339602. Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., and Battaglia, P. W. Learning to simulate complex physics with graph networks,

    work page doi:10.1109/inpar.2012.6339602 2012

  23. [23]

    Shi, C., Wang, C., Lu, J., Zhong, B., and Tang, J

    URL https: //arxiv.org/abs/2407.08608. Shi, C., Wang, C., Lu, J., Zhong, B., and Tang, J. Pro- tein sequence and structure co-design with equivariant translation,

    Pith/arXiv arXiv

  24. [24]

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A

    URL https://arxiv.org/abs/2009.03509. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. At- tention is all you need.Advances in neural information processing systems, 30,

    arXiv 2009

  25. [25]

    Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y ., Xiao, T., He, T., Karypis, G., Li, J., and Zhang, Z

    URL https://arxiv.org/abs/1710.10903. Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y ., Xiao, T., He, T., Karypis, G., Li, J., and Zhang, Z. Deep graph library: A graph- centric, highly-performant package for graph neural net- works.arXiv preprint arXiv:1909.01315,

    Pith/arXiv arXiv 1909

  26. [26]

    10 On Efficient Scaling of GNNs via IO-Aware Layers Implementations Wei, N., Jiexiong, G., Yanzhi, W., Gagan, A., and Bin, R

    URL https://arxiv.org/ abs/2112.02052. 10 On Efficient Scaling of GNNs via IO-Aware Layers Implementations Wei, N., Jiexiong, G., Yanzhi, W., Gagan, A., and Bin, R. Dnnfusion: accelerating deep neural networks execution with advanced operator fusion. InProceedings of the 42nd ACM SIGPLAN International Conference on Pro- gramming Language Design and Implem...

    arXiv 2021

  27. [27]

    ISBN 9781450383912

    Associa- tion for Computing Machinery. ISBN 9781450383912. doi: 10.1145/3453483.3454083. URL https://doi. org/10.1145/3453483.3454083. Williams, S., Waterman, A., and Patterson, D. Roofline: an insightful visual performance model for multicore architectures.Commun. ACM, 52(4):65–76, April

    work page doi:10.1145/3453483.3454083

  28. [28]

    doi: 10.1145/1498765

    ISSN 0001-0782. doi: 10.1145/1498765. 1498785. URL https://doi.org/10.1145/ 1498765.1498785. Xiang, L., Asudeh, O., Sabin, G., Sukumaran-Rajam, A., and Sadayappan, P. cutespmm: Accelerating sparse-dense matrix multiplication using gpu tensor cores,

    work page doi:10.1145/1498765

  29. [29]

    Yan, J., Jiang, W., He, D., Wen, S., Li, Y ., Jin, H., and Shao, Z

    URL https://arxiv.org/abs/2504.06443. Yan, J., Jiang, W., He, D., Wen, S., Li, Y ., Jin, H., and Shao, Z. Rt-gnn: Accelerating sparse graph neural net- works by tensor-cuda kernel fusion.ACM Trans. Ar- chit. Code Optim., 22(1), March

    arXiv

  30. [30]

    doi: 10.1145/3702001

    ISSN 1544-3566. doi: 10.1145/3702001. URL https://doi.org/10. 1145/3702001. Yang, Z., Cohen, W., and Salakhudinov, R. Revisiting semi-supervised learning with graph embeddings. In International Conference on Machine Learning, pp. 40– 48,

    work page doi:10.1145/3702001

  31. [31]

    Zhang, L., Huang, J., Di, S., Matsuoka, S., and Wahib, M

    URLhttps://arxiv.org/abs/2110.09524. Zhang, L., Huang, J., Di, S., Matsuoka, S., and Wahib, M. Can tensor cores benefit memory-bound ker- nels? (no!),

    arXiv

  32. [32]

    11 On Efficient Scaling of GNNs via IO-Aware Layers Implementations A

    URL https://arxiv.org/abs/ 2502.16851. 11 On Efficient Scaling of GNNs via IO-Aware Layers Implementations A. Code availability and Reproducibility Experimental details are described in Section

    arXiv

  33. [33]

    cuSPARSE + ˜A⊤

    and (ii) widely used citation-network benchmarks (Cora/Citeseer/Pubmed) from the Planetoid setup (Yang et al., 2016), as well as (iii) OGB node-property graphs (ogbn-arxiv, ogbn-products) (Hu et al., 2021). Table 8 reports basic structural statistics computed on our processed graph representations. GraphLand datasets (Bazhenov et al., 2025).GraphLand is a...

    2016

  34. [34]

    The same scale factors are reused to combine partial output accumulators before applying the final normalization

    The merge first computes the global maximum across warps and then the globally normalized denominator by reweighting each warp’s local sum via exp(mw −m global). The same scale factors are reused to combine partial output accumulators before applying the final normalization. The kernel writes the final output O[i, h,:] and additionally stores the per-node...

    2048

  35. [35]

    · 8 516/µ ,(14) which for µ= 256 (fully dense tiles) and large D approaches ≈5× . When node features dominate (N HD≫nnz ), the edge-index terms vanish and the ratio approaches AIBS AICSR N HD≫nnz − − − − − − − →(40D+ 512)·6 (4D+ 2)·14 · 1 ρ ,(15) where ρ=µ/256∈(0,1] is the mean tile fill fraction. For large D this simplifies to ≈(6/14)·(40/4)/ρ≈4.3/ρ , co...

    arXiv 1911