Pith. sign in

REVIEW 2 major objections 6 minor 7 cited by

An Efficient Sparse Kernel Generator for O(3)-Equivariant Deep Networks

T0 review · 2 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A code generator that compiles the Clebsch–Gordan tensor product into fast GPU kernels, beating both open-source and NVIDIA's closed-source libraries for equivariant neural networks.

desk verdict Solid, honestly benchmarked kernel-engineering paper whose only real sin is an abstract that overstates the end-to-end win over cuEquivariance; the kernel-level results and the fusion technique are real and worth referee time. read the letter →

arxiv 2501.13986 v4 pith:MESP3L7G submitted 2025-01-23 cs.LG cs.AI

Vivek Bharadwaj , Austin Glover , Aydin Buluc , James Demmel This is my paper · ORCID
classification cs.LGcs.AI
keywords O(3)-equivariantneuralnetworksClebsch-GordantensorproductGPUkernelgenerationsparsecontractiongraphconvolutionfusionjust-in-timecompilationinteratomicpotentials
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

The paper claims that the Clebsch–Gordan (CG) tensor product, the main bottleneck in rotation-equivariant graph neural networks, can be made dramatically faster by generating specialized GPU kernels at model compile time. The authors break the tensor product into subkernels whose operands fit entirely in registers, schedule computation through limited shared memory using the known sparse block structure, and fuse the product with graph convolution. They report up to 6.2x inference-time speedup for a chemistry foundation model in FP64 precision.

What carries the argument

The central machinery is the compile-time generation of a schedule over subkernels B and C, which contract a block of the sparse CG tensor with a cached segment of the input vectors and then apply a tile of the weight matrix. Each GPU warp owns a private partition of shared memory and stages x, y, and W into registers, so the inner loop over nonzero CG coefficients becomes a long independent instruction stream. Shared memory is reused across subkernels, and the fused graph-convolution variant keeps a per-node accumulator, replacing edge-level scatter writes with node-level writes.

What would settle it

Run the generated kernels on a representative equivariant model with, say, irreps higher than 4 or batch sizes exceeding the register cache design, and measure whether the forward and backward throughput still exceeds cuEquivariance by the reported margins; if the speedups vanish on those configurations, the claimed generality of the method is falsified.

Watch

Extended reading notes

Core claim

The paper introduces a sparse kernel generator that compiles the entire CG tensor product, including its block-sparse structure and weight matrices, into a single fused GPU kernel per model configuration. By assigning each input triple to a warp, caching operands in registers, unrolling the nonzero tensor entries into straight-line arithmetic, and partitioning shared memory through compile-time static analysis, the kernels skip all explicit zeros and avoid most global memory traffic. The paper reports forward-pass throughput improvements up to 9.2x over e3nn and up to 1.3x over NVIDIA's cuEquivariance on representative model configurations, with comparable or faster backward and second-derivative kernels.

Load-bearing premise

The kernels assume current equivariant model sizes, where x, y, and W segments fit into registers and the sparse CG blocks fit into shared memory; much larger irreps or feature vectors would require more phases and could erase the speedups.

Editorial extensions

If this is right

  • Equivariant chemistry models such as Nequip and MACE can run molecular dynamics inference several times faster without changing model weights or outputs.
  • The same CG tensor product can be fused with graph convolution, eliminating the memory overhead of duplicating node features and reducing global writes from edge count to node count.
  • Higher-order derivative kernels for force-field training can be assembled from existing forward/backward kernels, avoiding custom double-backward implementations.
  • The JIT compilation approach extends to AMD GPUs through the same generated C++/HIP backend, widening the range of hardware that can train equivariant models.
  • Because results match e3nn up to floating-point roundoff, the kernels can be dropped into existing models as a drop-in replacement.

Reading between the lines

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

  • The same compile-time sparse-unrolling technique could accelerate other structured equivariant operations, such as Gaunt tensor products or spherical-harmonic products, by treating their known coefficient tensors as the sparse pattern.
  • If the generator also emitted tensor-core-friendly variants, the register-level arithmetic streams could be packed into MMA instructions with low-precision formats, likely extending the speedups to training runs on larger models.
  • The schedule heuristics could be replaced with a proper optimization pass or learned scheduler that balances shared memory, register pressure, and phase overlap, which might close the remaining gap with cuEquivariance on backward passes and MACE end-to-end.
  • The performance claims were validated on a specific batch of benchmarks and a particular GPU generation; broader adoption would depend on whether the speedups hold across a wider range of irreps, batch sizes, and newer hardware.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The manuscript introduces OpenEquivariance, a JIT kernel generator for the Clebsch–Gordan (CG) tensor product used in O(3)-equivariant graph neural networks. The generator exploits the compile-time-known sparsity of the CG tensor, emits register-resident subkernels, schedules computation phases to reduce shared-memory pressure, fuses the tensor product with graph convolution using sparse-matrix techniques, and provides forward, backward, and double-backward kernels targeting NVIDIA and AMD GPUs. Performance is evaluated against e3nn v0.5.6 and NVIDIA cuEquivariance v0.4.0 on A100, A5000, and MI250x, both as isolated tensor products and inside Nequip and MACE, with correctness checked against e3nn up to floating-point roundoff.

Significance. If the reported kernel-level speedups are reproducible, this is a practically valuable contribution: it is open-source, cross-vendor, correctness-checked against an external package, and it addresses a recognized bottleneck in equivariant deep learning. The isolated-kernel benchmarks showing large wins over e3nn and competitive performance with cuEquivariance on many configurations are useful, and the graph-fusion design is well motivated. However, the paper's own full-model MACE benchmark contradicts the headline comparative claim against cuEquivariance, and the internal inconsistency between the reported maximum speedup over e3nn and the abstract's ">10x" claim must be resolved. The stress-test concern that kernel-level wins may not transfer to full models is confirmed by the paper's own Figure 10.

major comments (2)
  1. [Abstract, §1, §4.6, Fig. 10] The unqualified claim that the kernels provide "significant speedups over the best existing open and closed-source implementations" (Abstract) and "up to 1.3x speedup over NVIDIA's closed-source cuEquivariance package" is contradicted by the manuscript's own end-to-end experiment: §4.6 and Figure 10 show that the MACE implementation reaches only 0.73x the simulation speed of cuEquivariance v0.4.0 in both FP32 and FP64, and even the hybrid model that borrows cuE's optimized layers still trails cuE. The kernel-level benchmarks do support the more restricted claims (large gains over e3nn, modest gains over cuE on isolated tensor products), so the abstract and introduction should be revised to state those restrictions explicitly rather than making an unqualified comparison against the best closed-source implementation.
  2. [§4.1, Fig. 5 caption, Abstract] The abstract's claim of ">10x speedup over the widely-used e3nn package" is not supported by the paper's own numerical summary: §4.1 reports a maximum FP32 forward-pass speedup over e3nn of 9.2x, while Figure 5's caption claims "more than 10x improvement" for Nequip-benzene. The text and figure must be reconciled, and the abstract adjusted to match the actual measured maximum.
minor comments (6)
  1. [Throughout] The name "Clebsch-Gordon" should be spelled "Clebsch-Gordan" throughout the paper.
  2. [§1, bullet list] The bullet "F used Graph Convolution" contains a typo and should read "Fused Graph Convolution."
  3. [§3.4] The double-backward identity in Equations (3.4) and (3.5) is central to the second-derivative contribution, but the text only says "we can show" and does not provide the derivation. Given that the identity is asserted as novel, a short derivation or appendix is needed; the formulas appear correct, but the proof should be available to readers.
  4. [§4.3] The roofline paragraph contains a duplicated phrase: "contain contain a significant fraction of non fused-multiply-add (FMA) instructions" should be corrected.
  5. [§4.6] The word "speeedup" appears in the MACE benchmark paragraph and should be corrected to "speedup."
  6. [Abstract, §4.6] The abstract's statement of "up to 6.2x inference-time speedup for the MACE chemistry foundation model over the original unoptimized version" should specify what the baseline is; in Figure 10 the relevant comparison appears to be against e3nn or cuE-old, and this should be stated precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all performance claims are measured against external baselines and the derivative identities follow from the product and chain rules.

full rationale

The paper's central deliverables are (i) a JIT kernel generator for CG tensor products, (ii) fused graph-convolution kernels, and (iii) forward/backward/double-backward kernels. None of these is defined in terms of the claimed speedups. Correctness is checked against e3nn as an independent implementation ('We tested correctness against e3nn to ensure that our kernels produce identical results, up to floating point roundoff'), and performance is measured against two external packages, e3nn and NVIDIA cuEquivariance, on A100, A5000, and MI250x hardware. The register-caching design and phase scheduler are engineering heuristics whose parameters (warp counts, shared-memory partitions) are chosen statically at model compile time, not fitted to the benchmark outcomes; even if tuning were involved, it would not convert a measured performance claim into an analytical prediction that reduces to its inputs. The double-backward identity in Equations (3.4)-(3.5) is explicitly justified by 'repeatedly applying the product and chain rules to the formulas for a, b, and C in Section 3.3,' so it is a derivation, not an assumption or a renamed input. The paper itself discloses that end-to-end MACE simulation reaches only 0.73x the speed of cuEquivariance v0.4.0 with fused convolution (Section 4.6, Figure 10), which qualifies the abstract's unqualified 'significant speedups' claim; that discrepancy is a correctness/overclaim concern, not a circularity. No load-bearing self-citation, fitted-input-called-prediction, or uniqueness-imported-from-authors pattern is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on GPU architecture assumptions (warp-level execution, shared memory, register capacity) and on the correctness of the double-backward identity. There are no fitted physical parameters; the only hand-tuned choices are kernel launch heuristics.

free parameters (1)
  • Kernel launch configuration (warps per CTA, number of CTAs, shared memory per block) = not reported; chosen by heuristics
    Section 4.2 notes 'we could improve our heuristic selection of the warp count per block, the number of blocks, and the shared memory allotted to each block.' The speedup numbers depend on these hand-tuned choices, but they are not fitted physical constants.
assumptions (4)
  • domain assumption CG tensor product is a computationally significant bottleneck in equivariant models
    Motivates Section 1; the paper's benchmarks assume this bottleneck.
  • domain assumption A single GPU warp can process one (x, y, W) triple with coalesced memory access and no CTA-level synchronization
    Algorithm 1 assigns each input triple to one warp; this is the core parallelization premise.
  • standard math Eq. (3.5) correctly reconstructs double-backward gradients from seven forward/backward kernel evaluations
    Section 3.4 states it 'can be shown' but does not show the derivation; correctness is load-bearing for second-derivative speedups.
  • domain assumption Typical equivariant models have irreps with ℓ ≤ 4
    Section 3.2 asserts this for Nequip and MACE; needed for register-resident operands.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Efficient Sparse Kernel Generator for O(3)-Equivariant Deep Networks." pith.science (2026). https://pith.science/paper/MESP3L7G

@misc{pith2026250113986,
  author       = {Pith},
  title        = {Pith review of: An Efficient Sparse Kernel Generator for O(3)-Equivariant Deep Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MESP3L7G}},
  note         = {Machine review of arXiv:2501.13986}
}
read the original abstract

Rotation equivariant graph neural networks, i.e. networks designed to guarantee certain geometric relations between their inputs and outputs, yield state of the art performance on spatial deep learning tasks. They exhibit high data efficiency during training and significantly reduced inference time for interatomic potential calculations compared to classical approaches. Key to these models is the Clebsch-Gordon (CG) tensor product, a kernel that contracts two dense feature vectors with a highly-structured sparse tensor to produce a dense output vector. The operation, which may be repeated millions of times for typical equivariant models, is a costly and inefficient bottleneck. We introduce a GPU sparse kernel generator for the CG tensor product that provides significant speedups over the best existing open and closed-source implementations. Our implementation achieves high performance by carefully managing the limited GPU shared memory through static analysis at model compile-time, minimizing reads and writes to global memory. We break the tensor product into a series of smaller kernels with operands that fit entirely into registers, enabling us to emit long arithmetic instruction streams that maximize instruction-level parallelism. By fusing the CG tensor product with a subsequent graph convolution, we reduce both intermediate storage and global memory traffic over naive approaches that duplicate input data. We also provide optimized kernels for the gradient of the CG tensor product and a novel identity for the higher partial derivatives required to predict interatomic forces. Our kernels offer up to 1.3x speedup over NVIDIA's closed-source cuEquivariance package, as well as 10x speedup over the widely-used e3nn package. In FP64 precision, we offer up to 6.2x inference-time speedup for the MACE chemistry foundation model over the original unoptimized version.

Figures

Figures reproduced from arXiv: 2501.13986 by the authors.

Figure 1
Figure 1. The CG tensor product, which contracts a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Three examples of coefficient tensors depicted [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Fundamental subkernels that compose to implement the CG tensor product in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Applications of the CG tensor product. The simplest and most general use case (A) calls the kernel repeatedly with distinct x, y, and W inputs. Interatomic potential models embed the operation in a graph convolution (B), where the tensor product combines node features …
Figure 5
Figure 5. Figure 5: Throughput of CG tensor products (batch size 50K), kernel B configurations without SpMM kernel fusion. On difficult configurations like Nequip-benzene with massive output vector lengths, we exhibit more than 10x improvement over e3nn. included any overhead in the measu…
Figure 7
Figure 7. Figure 7: Throughput of second derivative kernels (batch size 20K) for chemistry / protein models. tuning these hyperparameters as future work. 4.3 Roofline Analysis We conducted a roofline analysis [35] by profiling our forward / backward pass implementations on varied input co…
Figure 8
Figure 8. Figure 8: Roofline analysis for input configurations [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Speedup of convolution kernels over cuE [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Simulation speed of MACE for varying kernel [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Virtues and Vices of Equivariant Transformers

    hep-ph 2026-08 conditional novelty 7.0 of 10

    Lorentz-equivariant transformers outperform standard transformers for jet and flavor tagging whenever geometric 4-vector features dominate, and a 48M pretrained equivariant model matches far larger foundation models o...

  2. Transferable Implicit Solvent Machine Learning Potential for Drugs and Proteins Approaching Ab Initio Accuracy

    physics.chem-ph 2026-07 conditional novelty 7.0 of 10

    TWIN, a MACE-based implicit-solvent MLP trained solely on ab initio and experimental data, transfers across drugs, peptides and proteins with near-DFT accuracy at ~100 imes lower cost.

  3. The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products

    cs.LG 2025-06 conditional novelty 6.0 of 10

    The reported speedups of Gaunt and matrix tensor products over the full Clebsch-Gordan tensor product come from reduced expressivity, and the only true per-expressivity speedup comes from fast spherical harmonic transforms.

  4. Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems

    physics.chem-ph 2025-05 conditional novelty 6.0 of 10

    InstaDeep's mlip library ports MACE, NequIP, and ViSNet to JAX with a JAX-MD backend, ships SPICE2-trained organics models, reports faster MD steps than its own Torch routes, and proposes a faster gated MACE variant i...

  5. High-performance training and inference for deep equivariant interatomic potentials

    physics.comp-ph 2025-04 conditional novelty 6.0 of 10

    A redesigned NequIP framework using torch.compile, AOT Inductor compilation, and a custom Triton tensor product kernel accelerates Allegro MD inference by up to 18x and enables distributed training on large datasets.

  6. Fast and Accurate Foundation Models for Equivariant Machine-Learned Interatomic Potentials

    physics.comp-ph 2026-07 accept novelty 5.0 of 10

    Fast NequIP/Allegro foundation MLIPs reach leading MD inference speeds and strong benchmark accuracy; materials-discovery gains need better chemical diversity and consistent transition-metal energy surfaces.

  7. Quick ViTs: Speeding up Vision Transformers through Equivariance

    cs.CV 2025-05 conditional novelty 5.0 of 10

    D8-equivariant linear layers give ViTs a 5.33x FLOP reduction and 8x parameter reduction per layer, and hybrid octic ViTs match or slightly exceed ImageNet-1K accuracy while using about 40% fewer FLOPs.

Reference graph

Works this paper leans on

43 extracted references · 33 canonical work pages · cited by 7 Pith papers

  1. [1]

    Cormorant: covariant molecular neural net- works

    Brandon Anderson, Truong-Son Hy, and Risi Kon- dor. Cormorant: covariant molecular neural net- works. In Proceedings of the 33rd International Conference on Neural Information Processing Sys- tems, Red Hook, NY, USA, 2019. Curran Asso- ciates Inc

    work page 2019

  2. [2]

    Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Sys- tems, 2022

    work page 2022

  3. [3]

    Elena, D´ avid P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, D´ avid P. Kov´ acs, Janosh Riebe- sell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Samuel M. Blau, Vlad C˘ arare, James P. Darby, Sandip De, Flaviano Della Pia, Volker L. Deringer, Rokas Elijoˇ sius, Zakariya El-Machachi, Fa...

    arXiv 2024

  4. [4]

    Mailoa, Mordechai Ko- rnbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Ko- rnbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic poten- tials. Nature Communications , 13(1):2453, May

    work page

  5. [5]

    Brown- ing, and Michele Ceriotti

    Filippo Bigi, Guillaume Fraux, Nicholas J. Brown- ing, and Michele Ceriotti. Fast evaluation of spher- ical harmonics with sphericart. J. Chem. Phys. , 159:064802, 2023

    work page 2023

  6. [6]

    NVIDIA A100 tensor core GPU architecture

    NVIDIA Corporation. NVIDIA A100 tensor core GPU architecture. Technical report, NVIDIA,

    work page

  7. [7]

    Jaakkola

    Gabriele Corso, Hannes St¨ ark, Bowen Jing, Regina Barzilay, and Tommi S. Jaakkola. Diffdock: Dif- fusion steps, twists, and turns for molecular dock- ing. In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  8. [8]

    SE(3)-transformers: 3d roto- translation equivariant attention networks

    Fabian Fuchs, Daniel Worrall, Volker Fischer, Copyright © 2025 by SIAM Unauthorized reproduction of this article is prohibited and Max Welling. SE(3)-transformers: 3d roto- translation equivariant attention networks. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Bal- can, and H. Lin, editors, Advances in Neural In- formation Processing Systems , volume 3...

    work page 2025

Show all 43 references
  1. [9]

    e3nn: Euclidean neural networks, 2022

    Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks, 2022. URL https://arxiv.org/ abs/2207.09453

    2022 arXiv

  2. [10]

    Euclidean neural networks: e3nn, April 2022

    Mario Geiger, Tess Smidt, Alby M., Benjamin Kurt Miller, Wouter Boomsma, Bradley Dice, Kos- tiantyn Lapchevskyi, Maurice Weiler, Micha l Tyszkiewicz, Simon Batzner, Dylan Madisetti, Martin Uhrin, Jes Frellsen, Nuri Jung, Sophia San- born, Mingjian Wen, Josh Rackers, Marcel Rød...

    2022

  3. [11]

    Accelerate drug and material discovery with new math library NVIDIA cuEquivariance, 11 2024

    Mario Geiger, Emine Kucukbenli, Becca Zand- stein, and Kyle Tretina. Accelerate drug and material discovery with new math library NVIDIA cuEquivariance, 11 2024. URL https: //developer.nvidia.com/blog/accelerate- drug-and-material-discovery-with-new- math-library-nvidia-cuequi...

    2024

  4. [12]

    Helal, Jan Laukemann, Fabio Chec- coni, Jesmin Jahan Tithi, Teresa Ranadive, Fab- rizio Petrini, and Jeewhan Choi

    Ahmed E. Helal, Jan Laukemann, Fabio Chec- coni, Jesmin Jahan Tithi, Teresa Ranadive, Fab- rizio Petrini, and Jeewhan Choi. ALTO: adap- tive linearized storage of sparse tensors. In Pro- ceedings of the 35th ACM International Confer- ence on Supercomputing , ICS ’21, page 404–...

    2021

  5. [13]

    Johansson, P.D

    J.R. Johansson, P.D. Nation, and F. Nori. QuTiP 2: A Python framework for the dynamics of open quantum systems. Computer Physics Communica- tions, 184(4):1234–1240, apr 2013. doi: 10 .1016/ j.cpc.2012.11.019

    2013

  6. [14]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin ˇZ ´ ıdek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Bal- lard, Andrew Cowie, Bernardino Romera-Paredes,...

    2021

  7. [15]

    W. Kahan. Pracniques: further remarks on re- ducing truncation errors. Commun. ACM , 8(1): 40, January 1965. ISSN 0001-0782. doi: 10 .1145/ 363707.363723

    1965

  8. [16]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics , 3(6):422–440, Jun 2021. ISSN 2522-5820. doi: 10 .1038/s42254-021-00314-5

    2021

  9. [17]

    e3nn.c, November 2024

    Teddy Koker. e3nn.c, November 2024. URL https: //github.com/teddykoker/e3nn.c

    2024

  10. [18]

    Higher-order equivariant neural networks for charge density prediction in materials

    Teddy Koker, Keegan Quigley, Eric Taw, Kevin Tibbetts, and Lin Li. Higher-order equivariant neural networks for charge density prediction in materials. npj Computational Materials , 10(1):161, Jul 2024. ISSN 2057-3960. doi: 10 .1038/s41524- 024-01343-1

    2024

  11. [19]

    Gelib, July

    Risi Kondor and Erik Henning Thiede. Gelib, July

  12. [20]

    Clebsch–Gordan nets: a fully Fourier space spher- ical convolutional neural network

    Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch–Gordan nets: a fully Fourier space spher- ical convolutional neural network. In Proceed- ings of the 32nd International Conference on Neu- ral Information Processing Systems, NIPS’18, page 10138–10147, Red Hook, NY, USA, 20...

    2018

  13. [21]

    Nation, J

    Neill Lambert, Eric Gigu` ere, Paul Menczel, Boxi Li, Patrick Hopf, Gerardo Su´ arez, Marc Gali, Jake Lishman, Rushiraj Gadhvi, Rochisha Agar- wal, Asier Galicia, Nathan Shammah, Paul D. Nation, J. R. Johansson, Shahnawaz Ahmed, Si- mon Cross, Alexander Pitchford, and Franco N...

    2024 arXiv

  14. [22]

    Scaling computational performance of spherical harmonics kernels with Triton

    Kin Long Kelvin Lee, Mikhail Galkin, and Santi- ago Miret. Scaling computational performance of spherical harmonics kernels with Triton. In AI for Accelerated Materials Design - Vienna 2024 , 2024. Copyright © 2025 by SIAM Unauthorized reproduction of this article is prohibited

    2024

  15. [23]

    Equiformer: Equiv- ariant graph attention transformer for 3D atomistic graphs

    Yi-Lun Liao and Tess Smidt. Equiformer: Equiv- ariant graph attention transformer for 3D atomistic graphs. In International Conference on Learning Representations, 2023

    2023

  16. [24]

    What is

    Lek-Heng Lim and Bradley J Nelson. What is... an equivariant neural network? Notices of the American Mathematical Society , 70(4):619–624, 4

  17. [25]

    Krish- napriyan

    Shengjie Luo, Tianlang Chen, and Aditi S. Krish- napriyan. Enabling efficient equivariant operations in the Fourier basis via gaunt tensor products. In The Twelfth International Conference on Learning Representations, 2024

    2024

  18. [26]

    Owen, Mordechai Ko- rnbluth, and Boris Kozinsky

    Albert Musaelian, Simon Batzner, Anders Johans- son, Lixin Sun, Cameron J. Owen, Mordechai Ko- rnbluth, and Boris Kozinsky. Learning local equiv- ariant representations for large-scale atomistic dy- namics. Nature Communications , 14(1):579, Feb

  19. [27]

    Lawrence Zitnick

    Saro Passaro and C. Lawrence Zitnick. Reducing SO(3) convolutions to SO(2) for efficient equiv- ariant GNNs. In Andreas Krause, Emma Brun- skill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference o...

    2023

  20. [28]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Van- derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Ma- chine learning in Python. Journal of Machine Lear...

    2011

  21. [29]

    Equivariant message passing for the prediction of tensorial properties and molec- ular spectra

    Kristof Sch¨ utt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molec- ular spectra. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning , volume 139 of P...

    2021

  22. [30]

    Tensor-matrix products with a compressed sparse tensor

    Shaden Smith and George Karypis. Tensor-matrix products with a compressed sparse tensor. In Pro- ceedings of the 5th Workshop on Irregular Appli- cations: Architectures and Algorithms , New York, NY, USA, 2015. Association for Computing Ma- chinery. ISBN 9781450340014. doi: 10...

    2015

  23. [31]

    doi: 10 .1038/s41467-023- 36329-y

    ISSN 2041-1723. doi: 10 .1038/s41467-023- 36329-y

  24. [32]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds, 2018

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds, 2018. URL https://arxiv.org/abs/ 1802.08219

    2018 arXiv

  25. [33]

    Unke and Hartmut Maennel

    Oliver T. Unke and Hartmut Maennel. E3x: E(3)- equivariant deep learning made easy, 2024. URL https://arxiv.org/abs/2401.07595

    2024 arXiv

  26. [34]

    3D steerable cnns: learning rotationally equivariant features in volumetric data

    Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3D steerable cnns: learning rotationally equivariant features in volumetric data. In Proceedings of the 32nd In- ternational Conference on Neural Information Pro- cessing Systems, NIPS’18, page 10402–10...

    2018

  27. [35]

    Roofline: an insightful visual perfor- mance model for multicore architectures.Commun

    Samuel Williams, Andrew Waterman, and David Patterson. Roofline: an insightful visual perfor- mance model for multicore architectures.Commun. ACM, 52(4):65–76, April 2009. ISSN 0001-0782. doi: 10 .1145/1498765.1498785

    2009

  28. [36]

    Descoteaux, Mit Ko- tak, Gabriel de Miranda Nascimento, Se´ an R

    Chuin Wei Tan, Marc L. Descoteaux, Mit Ko- tak, Gabriel de Miranda Nascimento, Se´ an R. Kavanagh, Laura Zichi, Menghang Wang, Aadit Saluja, Yizhong R. Hu, Tess Smidt, Anders Johans- son, William C. Witt, Boris Kozinsky, and Albert Musaelian. High-performance training and infe...

  29. [37]

    Carl Yang, Aydın Bulu¸ c, and John D. Owens. Design principles for sparse matrix multiplication on the gpu. In Euro-Par 2018: Parallel Process- ing: 24th International Conference on Parallel and Distributed Computing, Turin, Italy, August 27 - 31, 2018, Proceedings , page 672–...

    2018

  30. [42]

    The price of freedom: Exploring trade- offs between expressivity and computational effi- ciency in equivariant tensor products

    YuQing Xie, Ameya Daigavane, Mit Kotak, and Tess Smidt. The price of freedom: Exploring trade- offs between expressivity and computational effi- ciency in equivariant tensor products. In ICML 2024 Workshop on Geometry-grounded Represen- tation Learning and Generative Modeling , 6 2024

    2024

  31. [2020]

    URL https://images.nvidia.com/aem- dam/en-zz/Solutions/data-center/nvidia- ampere-architecture-whitepaper.pdf

  32. [2022]

    doi: 10 .1038/s41467-022- 29939-5

    ISSN 2041-1723. doi: 10 .1038/s41467-022- 29939-5

  33. [2023]

    doi: 10 .1090/ noti2666

    ISSN 0002-9920, 1088-9477. doi: 10 .1090/ noti2666

  34. [2024]

    URL https://github.com/risi-kondor/ GElib

  35. [2025]

    URL https://arxiv.org/abs/2504.16068

    arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.