Pith. sign in

REVIEW 3 major objections 5 minor 61 references

TriADA: Massively Parallel Trilinear Matrix-by-Tensor Multiply-Add Algorithm and Device Architecture for the Acceleration of 3D Discrete Transformations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TriADA claims that a 3D grid of simple cells computes any 3D discrete orthogonal transform in a number of steps equal to the sum of its side lengths.

desk verdict Correct algebra, clean dataflow, but the 'linear time' headline is abstract cycles and the paper admits it; deserves a careful referee, not a desk reject. read the letter →

arxiv 2506.22818 v1 pith:MA2224NW submitted 2025-06-28 cs.DC cs.AIcs.ARcs.ETeess.SP

Stanislav Sedukhin (1) , Yoichi Tomioka (1) , Kazuya Matsumoto (1) , Yuichi Okuyama (1) ((1) The University of Aizu , Japan) This is my paper
classification cs.DCcs.AIcs.ARcs.ETeess.SP
keywords trilineardiscretetransformsouter-productGEMMmatrix-by-tensormultiplication3DmeshcellulararchitecturedecoupledstreamingmemorysparseprocessingmassivelyparallelacceleratorDXT
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

TriADA is a fused algorithm-and-architecture proposal for computing three-dimensional separable discrete orthogonal transforms — DFT, DCT, DHT, Walsh–Hadamard, and the more general three-mode matrix-by-tensor product — by re-expressing each stage as a sum of outer products, or rank-1 updates, rather than inner products. The paper claims that on an $N_1\times N_2\times N_3$ mesh of simple multiply-add cells, one per tensor element, all $(N_1N_2N_3)(N_1+N_2+N_3)$ multiply-adds of the transform can be executed in $N_1+N_2+N_3$ time-steps with 100% efficiency for dense data. The practical stakes are that the approach removes the power-of-two restriction of recursive fast transforms, accepts cuboid tensors directly, and keeps all tensor data stationary in the array while coefficient vectors stream in from three decoupled memories. An elastic sparse mode skips zero-valued operands, which the paper argues saves communication and arithmetic energy and shortens accumulation chains. If the underlying one-time-step-per-outer-product assumption holds, the architecture offers a path to extreme strong scaling for tensor-heavy HPC and AI workloads without the memory-bounded serialization of fast-transform algorithms.

What carries the argument

The carrying object is the outer-product (rank-1 update) notation of GEMM: an $N_1\times N_3$ intermediate matrix is accumulated as the sum of $N_3$ products of an $N_1$-vector and an $N_3$-vector, instead of as $N_1\cdot N_3$ independent inner products. TriADA maps three such sums onto a three-dimensional crossover mesh of compute-storage-communication cells in which each cell holds one element of the input, intermediate, or output tensor and updates it in place; coefficient vectors are multicast along horizontal, lateral, and frontal bus planes from three decoupled active streaming memories, and a tag bit on the diagonal of the square coefficient matrix selects the pivot cell that supplies the other operand vector. The local rule is data-driven — wait for two operands, multiply-add — with no dependence on cell coordinates, which is what makes the claimed time count linear and the array reusable for any problem shape $N_s\le P_s$.

What would settle it

A cycle-accurate simulator or physical prototype could settle the claim by measuring the time and energy of one rank-1 update as $N$ grows: if the critical path through the crossover multicast buses or the MAC update scales with vector length or with the array's bisection width, the total time for a 3D transform will grow faster than $N_1+N_2+N_3$, and the 100%-efficiency statement fails. A second check is to measure whether skipping zero-valued operands in the ESOP mode actually reduces the global time-step period or only the number of local updates.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a trilinear (3D) discrete orthogonal transformation of an $N_1\times N_2\times N_3$ tensor can be computed as three chained stages of rank-1 updates, with the tensor residing throughout in a 3D grid of $N_1\times N_2\times N_3$ cells and the three square coefficient matrices streamed in from decoupled active memories. Each stage performs its summation along one tensor mode by multicasting a coefficient vector along one set of buses and having a tagged 'pivot' cell per slice broadcast the matching data vector along the orthogonal buses; all cells then update their local element in place. Since one outer product is assumed to complete in a single time-step regardless of the lengths of the vectors, stage $s$ takes $N_s$ time-steps and the whole transformation finishes in $N_1+N_2+N_3$ time-steps. The architecture is called isomorphic to the algorithm because the three 4D iteration spaces of the stages are mapped onto the same 3D processing/storage/communication space by collapsing the summation direction onto the time axis, making cell activity coordinate-free and independent of problem size. The paper further claims an elastic sparse outer-product method that skips zero-valued operands, reducing energy and improving numerical accuracy by shortening the accumulation chains.

Load-bearing premise

The architecture's linear time count assumes that an outer-product step of any vector length completes in a single time-step, meaning coefficient vectors can be multicast to all relevant cells within that step with no wire-delay or energy cost that grows with array size.

Editorial extensions

If this is right

  • A single $N_1\times N_2\times N_3$ TriADA array computes the forward or inverse 3D DXT in $N_1+N_2+N_3$ time-steps, matching the paper's 100%-efficiency claim for dense data.
  • Tensor dimensions can be arbitrary and unequal, so cuboid problem sizes that arise in molecular dynamics and deep learning are handled without padding to powers of two.
  • The same architecture supports the more general three-mode matrix-by-tensor product, including Tucker-style compression and expansion, when the synchronization rule is extended to rectangular coefficient matrices.
  • A TriADA network with $P_1\times P_2\times P_3$ cells can be reused for any $N_1\times N_2\times N_3$ problem with $N_s\le P_s$, which the paper frames as a step toward wafer-scale tensor computing.
  • With ESOP, sparse tensors skip zero-valued operands entirely, so the paper's position is that energy drops and accuracy improves as sparsity increases.

Reading between the lines

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

  • The one-time-step outer-product assumption is where the linear-time claim meets physics: if broadcast or MAC cost grows with vector length, TriADA's advantage becomes a constant-factor gain rather than a complexity-class change, while the data-stationary structure remains valuable.
  • The tag-based coordinate-free cell rule could generalize beyond orthogonal transforms to any tensor contraction whose iteration space collapses onto a 3D grid, which the paper hints at but does not develop.
  • A concrete test for the sparse claims would be to run ESOP on ReLU-activated neural-network layers and measure end-to-end energy per inference; skipping zeros only pays if the waiting and tag-checking overhead stays below the cost of the avoided updates.
  • For the GEMT case with non-square matrices, the paper explicitly leaves the synchronization policy open; designing such a rule would unlock general tensor-decomposition workloads, not just orthogonal transforms.
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

3 major / 5 minor

Summary. The paper proposes TriADA, a co-designed algorithm and accelerator architecture for separable 3D discrete orthogonal transforms (3D-DXTs) and the more general 3-mode matrix-by-tensor multiplication (3D-GEMT). The algorithm expresses each of three transform stages as a sequence of outer-product rank-1 updates (Eqs. (6.1)-(6.3)) and maps these updates onto a 3D grid of multiply-accumulate cells connected by horizontal, lateral, and frontal buses. The central claim is that all N1*N2*N3*(N1+N2+N3) MAC operations can be executed in N1+N2+N3 time-steps with 100% efficiency on N1*N2*N3 cells. A sparse-data extension, Elastic Sparse Outer-Product Processing (ESOP), is also proposed. The paper is a design proposal with no physical implementation, simulation, or measured timing or energy results.

Significance. If the hardware timing assumptions were validated, the algebraic decomposition and the three-stage dataflow would be a clean and general way to accelerate separable 3D transforms without power-of-two constraints, and the mapping onto a regular 3D network is conceptually attractive. The algebraic decomposition itself is mathematically correct, and the dataflow description is detailed enough to reproduce the abstract cycle-level behavior. The main claimed advantage, however, depends on an unvalidated model in which each rank-1 update of arbitrary vector length costs one time-step and global multicast is free. Since the paper explicitly declines to justify this model, the practical significance of the headline linear-time and 100%-efficiency claims is not yet established.

major comments (3)
  1. [§3.2 and §5.4] The central claim, stated in §5.4 as executing all (N1N2N3)(N1+N2+N3) MAC operations in the linear number of N1+N2+N3 time-steps with 100% efficiency, rests entirely on the Section 3.2 assumption that each outer-product rank-1 update of a vector of arbitrary length is executed in one time-step independent of vector length. The paper explicitly says that a discussion of how realistic this assumption is 'out of the scope of this paper'. This is not a peripheral issue: on a physical die, the coefficient and data vectors must traverse or fan out across O(N) cells, so the per-step critical path grows with N and total physical latency scales at least as N(N1+N2+N3). The paper should either supply a physical timing model (wire delay, fanout, repeater insertion, pipelining) or rephrase the claim as an abstract cycle count and add explicit total-latency and energy estimates.
  2. [§5.4] The '100% efficiency' figure counts MAC operations per abstract cycle, but it does not account for the time needed to deliver operands. In Stage I, for example, a green cell's x value must propagate along a horizontal bus spanning up to N3 cells, and the coefficient value must be multicast to a planar face of the Tensor Core with fanout N1N2, all within the same time-step. With O(N)-length wires, the clock period grows at least linearly with N (or quadratically without repeaters), so 100% efficiency in abstract cycles does not imply 100% efficiency in physical time or energy. The paper needs to define the efficiency metric precisely and analyze the per-step critical path before making this claim.
  3. [§6] The ESOP section claims that skipping zero-valued operands improves accuracy, stability, and energy efficiency, but it does not define what counts as 'zero' (exact zero vs. insignificant value), gives no error analysis, and provides no energy model or measurement. Skipping exact-zero MACs is harmless, but if 'zero' means a small nonzero value, discarding the update changes the result and the effect on rounding error is not monotone in the number of skipped updates. The paper should formalize the sparsity model, define the zero threshold, and provide supporting analysis or experimental evidence for the accuracy and energy claims.
minor comments (5)
  1. [§3.2] The paper cites a document generated by 'Google Gemini' as support for the claim that outer-product notation is most promising for hardware acceleration; this is not a peer-reviewed technical source and should be replaced with established literature or removed.
  2. [§2.1] There is a garbled sentence in the text around Eq. (1): 'computing another grid of a N1×N2×N3 data or three-mode tensor ... X = J ...x k1,k2,k3 K' is incomplete and should be rewritten.
  3. [Figures 2-4] The figures are dense and some labels, such as c(3)_in and the green/orange cell coloring, are not fully defined in the captions; the captions should include a legend and define the bus directions.
  4. [§4] The text says a time-scheduling function is defined for each stage, but the actual scheduling equations are not given explicitly; adding them would make the conflict-free mapping reproducible.
  5. [§7] There are several typos, including 'ortogonal' and 'intellegent' in the conclusion, and the paper uses both 'cubical' and 'cuboid' without distinction.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the algebraic decomposition is self-contained, and the linear time-step count is an explicitly stated cost-model consequence rather than a fitted or self-cited prediction.

full rationale

The paper's central derivation, Eqs. (4)-(6), expresses a three-mode GEMT as three stages of rank-1 outer-product updates. This is a valid algebraic identity and is derived directly from the tensor product definition, with no fitted parameters and no dependence on a circularly defined result. The headline 'linear N1+N2+N3 time-steps' is explicitly conditional: Section 3.2 states that if each outer-product vector operation is assumed to execute in one time-step regardless of vector length, then the OP formulation needs a linear number of time-steps, and Section 5.2 repeats the assumption ('assuming that all sends, receives, and updates are implemented in the Tensor Core with N1xN2xN3 cells, simultaneously, in one time-step'). The nontrivial content is the reduction from (N1N2N3)^2 MACs to N1N2N3(N1+N2+N3) MACs and the scheduling of the rank-1 updates, which are presented in the paper rather than imported from a prediction. The weakness is that physical global-broadcast wire delay and the realism of a single-cycle rank-1 update are explicitly left 'out of the scope of this paper'; that is a validation and correctness-risk limitation, not circularity. Self-citations to Sedukhin (2012) and Sedukhin and Sedukhin (1994) for index-space mapping are not load-bearing, because the mapping is stated in Eqs. (7) and the architecture is described locally. No fitted-input-called-prediction, uniqueness-imported-from-authors, or ansatz-smuggled-via-citation pattern is present. The score of 2 reflects minor self-citations and the central claim's dependence on an unvalidated but clearly disclosed timing model, not a circular derivation.

Assumptions & free parameters 0 free parameters · 5 assumptions · 4 invented entities

No numeric fitting occurs; the contribution rests on architectural postulates. The derived claims are conditional on the one-time-step-per-rank-1-update model, free global broadcast, and local cell storage, none of which are validated by measurement.

assumptions (5)
  • ad hoc to paper Each outer-product of an N-length vector and an M-length vector is executed in a single time-step, independent of N and M.
    This is the foundational performance model in Section 3.2 and is what makes the linear time-step count in Section 5.4 possible; it ignores wire delay, memory bandwidth, and pipelining constraints.
  • domain assumption The coefficient matrices are square and orthogonal (or unitary), with diagonal tags used to activate pivot columns.
    Section 2 defines C as square and orthogonal; Section 5.2 relies on diagonal tags, which the paper says only works for square matrices, limiting generalization to GEMT.
  • ad hoc to paper A global crossover mesh of operand buses can multicast a vector to all cells in one time-step with negligible energy and latency.
    Section 5.1 describes the mesh and axon-like buses but provides no electrical or timing analysis for global broadcast; this assumption is load-bearing for both the time-step count and ESOP energy claims.
  • domain assumption Each cell stores one tensor element and performs one MAC per time-step, with local memory sufficient for the input, intermediate, and output tensors.
    Section 5.1: 'Each cell in the Tensor Core stores in local memory the corresponding element x... as well as the elements ...'; capacity and access are not modeled.
  • ad hoc to paper Zero-valued operands can be detected and skipped without overhead, and skipping updates preserves or improves floating-point accuracy.
    Section 6 states ESOP avoids zero-operand updates and improves accuracy; no accounting for detection overhead or for the fact that skipping updates changes summation order, which may or may not improve error bounds.
invented entities (4)
  • TriADA Tensor Core
    purpose: N1xN2xN3 grid of cells storing input, intermediate, and output tensors
    No fabricated chip or simulation; the architecture is described at block-diagram level in Section 5.
  • Decoupled Active Streaming Memories (DASM)/Actuators
    purpose: Store and stream coefficient-matrix vectors into the Tensor Core faces
    Section 5.1 introduces three actuators; no implementation details or performance characterization.
  • Elastic Sparse Outer-Product (ESOP) processing
    purpose: Skip zero-valued vector elements to save computation and communication
    Described in Section 6 without quantitative evaluation of energy or accuracy.
  • Tag-based coordinate-free synchronization
    purpose: Diagonal tags in streamed coefficient vectors select pivot columns per time-step
    Section 5.2 proposes the tag mechanism; only works for square matrices, and no hardware mechanism for tags is specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TriADA: Massively Parallel Trilinear Matrix-by-Tensor Multiply-Add Algorithm and Device Architecture for the Acceleration of 3D Discrete Transformations." pith.science (2026). https://pith.science/paper/MA2224NW

@misc{pith2026250622818,
  author       = {Pith},
  title        = {Pith review of: TriADA: Massively Parallel Trilinear Matrix-by-Tensor Multiply-Add Algorithm and Device Architecture for the Acceleration of 3D Discrete Transformations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MA2224NW}},
  note         = {Machine review of arXiv:2506.22818}
}
read the original abstract

Multilinear transformations are key in high-performance computing (HPC) and artificial intelligence (AI) workloads, where data is represented as tensors. However, their high computational and memory demands, which grow with dimensionality, often slow down critical tasks. Moreover, scaling computation by enlarging the number of parallel processing units substantially increases energy consumption, limiting widespread adoption, especially for sparse data, which is common in HPC and AI applications. This paper introduces the Trilinear Algorithm and isomorphic to algorithm Device Architecture (TriADA) to address these challenges with the following innovations: (1) a massively parallel, low-rank algorithm for computing a family of trilinear (3D) discrete orthogonal transformations (3D-DXTs), which is a special case of the more general 3-mode matrix-by-tensor multiplication (3D-GEMT); (2) a new outer-product-based GEMM kernel with decoupled streaming active memory, specially designed to accelerate 3D-GEMT operation; (3) an isomorphic to the proposed algorithm, fully distributed 3D network of mesh interconnected processing elements or cells with a coordinate-free, data-driven local processing activity, which is independent of problem size; (4) an elastic sparse outer-product (ESOP) method that avoids unnecessary computing and communication operations with zero-valued operands, thereby enhancing energy efficiency, computational accuracy, and stability. TriADA is capable of performing a variety of trilinear transformations with hypercubic arithmetic complexity in a linear number of time-steps. The massively parallel, scalable, and energy-efficient architecture of TriADA is ideal for accelerating multilinear tensor operations, which are the most demanding parts of AI and HPC workloads.

Figures

Figures reproduced from arXiv: 2506.22818 by the authors.

Figure 1
Figure 1. Partitions of a 3-mode tensor into the set of slices or matrices (shown in gray) and corresponding two-mode [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Stage I: computing and data movement on each time-step [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Stage II: computing and data movement on each of [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Stage III: computing and data movement on each of [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Diagram of cell’s activity to manage a data sparsity. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 31 canonical work pages

  1. [8]

    IEEE Micro 42, 5 (2022), 34–40

    Compiling for the IBM Matrix Engine for Enterprise Workloads. IEEE Micro 42, 5 (2022), 34–40. https://doi.org/10.1109/MM.2022.3176529 Lieven De Lathauwer, Bart De Moor, and Joos Vandewalle

    arXiv 2022

  2. [12]

    Computing in Science & Engineering 2, 1 (2000), 22–23

    Guest Editors Introduction to the top 10 algorithms. Computing in Science & Engineering 2, 1 (2000), 22–23. https://doi.org/10.1109/MCISE.2000.814652 Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri

    arXiv 2000

  3. [13]

    Neurocomput

    Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomput. 503, C (Sept. 2022), 92–108. https://doi.org/10.1016/ j.neucom.2022.06.111 Sanghamitra Dutta, Viveck Cadambe, and Pulkit Grover

    work page 2022

  4. [15]

    In2021 IEEE International Symposium on Circuits and Systems (ISCAS)

    Accelerating 3D Convolutional Neural Networks Using 3D Fast Fourier Transform. In2021 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, Daegu, Korea, 1–5. https://doi.org/10.1109/ISCAS51556.2021.9401765 G.C Fox, S.W Otto, and A.J.G Hey

    arXiv 2021

  5. [21]

    arXiv:2502.04563 [cs.LG] https://arxiv.org/abs/2502.04563 Dan Hendrycks and Kevin Gimpel

    WaferLLM: A Wafer-Scale LLM Inference System. arXiv:2502.04563 [cs.LG] https://arxiv.org/abs/2502.04563 Dan Hendrycks and Kevin Gimpel

    arXiv

  6. [23]

    Sparsity in Deep Learning: Pruning and growth for efficient inference and training in neural networks. arXiv:2102.00554 [cs.LG] https: //arxiv.org/abs/2102.00554 Yang Hu, Xinhan Lin, Huizheng Wang, Zhen He, Xingmao Yu, Jiahao Zhang, Qize Yang, Zheng Xu, Sihan Guan, Jiahao Fang, Haoran Shang, Xinru Tang, Xu Dai, Shaojun Wei, and Shouyi Yin

    arXiv

  7. [24]

    IEEE Circuits and Systems Magazine 24, 1 (2024), 52–81

    Wafer-Scale Computing: Advancements, Challenges, and Future Perspectives [Feature]. IEEE Circuits and Systems Magazine 24, 1 (2024), 52–81. https://doi.org/10.1109/MCAS.2024.3349669 16 TriADA: Algorithm and Device Architecture for Acceleration of 3D Transforms A PREPRINT Paul Hübner, Andong Hu, Ivy Peng, and Stefano Markidis

    arXiv 2024

  8. [25]

    Oranges: Evaluating the Apple Silicon M-Series SoCs for HPC Performance and Efficiency

    Apple vs. Oranges: Evaluating the Apple Silicon M-Series SoCs for HPC Performance and Efficiency. arXiv:2502.05317 [cs.AR] https://arxiv.org/abs/2502. 05317 Yuki Ikegaki, Toshiaki Miyazaki, and Stanislav G. Sedukhin

    arXiv

Show all 61 references
  1. [27]

    arXiv:2403.07953 [cs.LG] https://arxiv

    Enabling Unstructured Sparse Acceleration on Structured Sparse Accelerators. arXiv:2403.07953 [cs.LG] https://arxiv. org/abs/2403.07953 Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Nai, Nishant Patil, Suvinay Subra- manian, Andy Swing, Brian Tow...

    arXiv

  2. [28]

    arXiv:2304.01433 [cs.AR] https://arxiv.org/abs/2304.01433 Tamara G

    TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings. arXiv:2304.01433 [cs.AR] https://arxiv.org/abs/2304.01433 Tamara G. Kolda and Brett W. Bader

    arXiv

  3. [29]

    SIAM Rev

    Tensor Decompositions and Applications. SIAM Rev. 51, 3 (Aug. 2009), 455–500. https://doi.org/10.1137/07070111X H. Kung and C. Leiserson

    2009 doi

  4. [31]

    Acta Numerica 30 (May 2021), 555–764

    Tensors in computations. Acta Numerica 30 (May 2021), 555–764. https://doi.org/10. 1017/S0962492921000076 Sheng Lin, Ning Liu, Mahdi Nazemi, Hongjia Li, Caiwen Ding, Yanzhi Wang, and Massoud Pedram

    2021

  5. [34]

    IEEE Spectrum 61, 7 (2024), 22–27

    The Path to a 1-Trillion-Transistor GPU: AI’s Boom Demands New Chip Technology. IEEE Spectrum 61, 7 (2024), 22–27. https://doi.org/10.1109/MSPEC.2024.10589682 Xingyi Liu and Keshab K. Parhi

    2024

  6. [35]

    IEEE Circuits and Systems Magazine 23, 2 (2023), 8–28

    Tensor Decomposition for Model Reduction in Neural Networks: A Review. IEEE Circuits and Systems Magazine 23, 2 (2023), 8–28. https://doi.org/10.1109/MCAS.2023.3267921 Qingda Lu, Xiaoyang Gao, Sriram Krishnamoorthy, Gerald Baumgartner, J. Ramanujam, and P. Sadayappan

    2023

  7. [37]

    arXiv:2402.13499 [cs.AR] https://arxiv.org/abs/2402.13499 Nitin Malapally, Viacheslav Bolnykh, Estela Suarez, Paolo Carloni, Thomas Lippert, and Davide Mandelli

    Benchmarking and Dissecting the Nvidia Hopper GPU Architecture. arXiv:2402.13499 [cs.AR] https://arxiv.org/abs/2402.13499 Nitin Malapally, Viacheslav Bolnykh, Estela Suarez, Paolo Carloni, Thomas Lippert, and Davide Mandelli

    arXiv

  8. [38]

    http://arxiv.org/abs/ 2303.13337 arXiv:2303.13337 [physics]

    Scalability of 3D-DFT by block tensor-matrix multiplication on the JUWELS Cluster. http://arxiv.org/abs/ 2303.13337 arXiv:2303.13337 [physics]. Nitin Malapally, Viacheslav Bolnykh, Estela Suarez, Paolo Carloni, Thomas Lippert, and Davide Mandelli

    arXiv

  9. [39]

    3D DFT by block tensor-matrix multiplication via a modified Cannon’s algorithm: Implementation and scaling on distributed-memory clusters with fat tree networks. J. Parallel and Distrib. Comput. 193 (2024), 104945. https://doi.org/10.1016/j.jpdc.2024.104945 17 TriADA: Algorith...

    2024

  10. [40]

    In Proceedings of the 37th ACM International Conference on Supercomputing (Orlando, FL, USA) (ICS ’23)

    Wafer-Scale Fast Fourier Transforms. In Proceedings of the 37th ACM International Conference on Supercomputing (Orlando, FL, USA) (ICS ’23). Association for Computing Machinery, New York, NY , USA, 180–191. https://doi.org/10.1145/3577193.3593708 Konstantin F. Pilz, James Sand...

  11. [41]

    arXiv:2504.16026 [cs.CY] https://arxiv.org/abs/2504.16026 T

    Trends in AI Supercomputers. arXiv:2504.16026 [cs.CY] https://arxiv.org/abs/2504.16026 T. E. Pogue and N. Nicolici

    arXiv

  12. [42]

    IEEE Trans

    Fast Inner-Product Algorithms and Architectures for Deep Neural Network Accelerators. IEEE Trans. Comput.73, 02 (feb 2024), 495–509. https://doi.org/10.1109/TC.2023.3334140 Douglas N. Rutledge and Delphine Jouan-Rimbaud Bouveresse

    2024

  13. [44]

    University of Aizu, Japan

    3D Discrete Transforms with Cubical Data Decomposition on the IBM Blue Gene/Q. University of Aizu, Japan. https://u-aizu.ac.jp/files/page/research/techreport/ 2013-001.pdf Kylee Santos, Stan Moore, Tomas Oppelstrup, Amirali Sharifian, Ilya Sharapov, Aidan Thompson, Delyan Z. K...

    2013

  14. [45]

    In SC24: International Conference for High Performance Computing, Networking, Storage and Analysis

    Breaking the Molecular Dynamics Timescale Barrier Using a Wafer-Scale System. In SC24: International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE, Atlanta, GA, USA, 1–13. https://doi.org/10.1109/SC41406.2024.00014 Gabin Schieffer, Daniel Ar...

    arXiv

  15. [46]

    In2024 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS)

    On the Rise of AMD Matrix Cores: Performance, Power Efficiency, and Programmability. In2024 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) . IEEE, Indianapolis, IN, USA, 132–143. https: //doi.org/10.1109/ISPASS61541.2024.00022 S.G. Sedukhin

  16. [47]

    In Proceedings of 4th Great Lakes Symposium on VLSI

    A new systolic architecture for pipeline prime factor DFT-algorithm. In Proceedings of 4th Great Lakes Symposium on VLSI. IEEE CS, Notre Dame, IN, USA, 40–45. https://doi.org/10.1109/GLSV.1994. 289998 Stanislav Sedukhin

    1994 doi

  17. [48]

    Technical Report Tech

    Co-design of Extremely Scalable Algorithms/Architecture for 3-Dimensional Linear Trans- forms. Technical Report Tech. Report 2012-001. University of Aizu, Japan. https://u-aizu.ac.jp/files/ page/research/techreport/2012-001.pdf Stanislav. Sedukhin, Toshiaki. Miyazaki, and Keni...

    2012

  18. [49]

    IEICE Transactions on Electronics E105.C, 6 (June 2022), 209–221

    In Search of the Performance- and Energy- Efficient CNN Accelerators. IEICE Transactions on Electronics E105.C, 6 (June 2022), 209–221. https: //doi.org/10.1587/transele.2021LHP0003 S. G. Sedukhin and I. S. Sedukhin

    2022 doi

  19. [50]

    In Parallel Processing: CONPAR 94 — VAPP VI, Bruno Buchberger and Jens V olkert (Eds.)

    Systematic approach and software tool for systolic design. In Parallel Processing: CONPAR 94 — VAPP VI, Bruno Buchberger and Jens V olkert (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 172–183. https://doi.org/10.1007/3-540-58430-7_16 Stanislav G. Sedukhin, Ahmed S. ...

    doi

  20. [54]

    arXiv:2109.08668 [cs.LG] https://arxiv.org/abs/2109

    Primer: Searching for Efficient Transformers for Language Modeling. arXiv:2109.08668 [cs.LG] https://arxiv.org/abs/2109. 08668 E.E. Swartzlander, B.K. Gilbert, and I.S. Reed

    arXiv

  21. [55]

    IEEE Trans

    Inner Product Computers. IEEE Trans. Comput. C-27, 1 (1978), 21–31. https://doi.org/10.1109/TC.1978.1674948 18 TriADA: Algorithm and Device Architecture for Acceleration of 3D Transforms A PREPRINT Earl E Swartzlander Jr

    1978

  22. [56]

    arXiv:2303.11525 [cs.LG] https://arxiv.org/abs/2303.11525 R

    Sparse-IFT: Sparse Iso-FLOP Transforma- tions for Maximizing Training Efficiency. arXiv:2303.11525 [cs.LG] https://arxiv.org/abs/2303.11525 R. A. Van De Geijn and J. Watts

    arXiv

  23. [57]

    Concurrency: Practice and Experience 9, 4 (April 1997), 255–274

    SUMMA: scalable universal matrix multiplication algorithm. Concurrency: Practice and Experience 9, 4 (April 1997), 255–274. Wikipedia contributors

    1997

  24. [58]

    https: //en.wikipedia.org/w/index.php?title=Advanced_Matrix_Extensions&oldid=1281191096

    Advanced Matrix Extensions — Wikipedia, The Free Encyclopedia. https: //en.wikipedia.org/w/index.php?title=Advanced_Matrix_Extensions&oldid=1281191096. [Online; accessed 3-May-2025]. Yannan Nellie Wu, Po-An Tsai, Saurav Muralidharan, Angshuman Parashar, Vivienne Sze, and Joel Emer

    2025

  25. [60]

    In 56th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’23)

    Tailors: Accelerating Sparse Tensor Algebra by Overbooking Buffer Capacity. In 56th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’23). ACM, Toronto, ON, Canada, 1347–1363. https://doi.org/10.1145/3613424.3623793 Ian En-Hsu Yen, Zhibin Xiao, and Dongkuan Xu

  26. [61]

    arXiv:2207.08006 [cs.AR] https://arxiv.org/abs/2207.08006 Guowei Zhang, Nithya Attaluri, Joel S

    S4: a High-sparsity, High-performance AI Accelerator. arXiv:2207.08006 [cs.AR] https://arxiv.org/abs/2207.08006 Guowei Zhang, Nithya Attaluri, Joel S. Emer, and Daniel Sanchez

    arXiv

  27. [62]

    In Proceedings of the 26th ACM International Conference on Architec- tural Support for Programming Languages and Operating Systems(Virtual, USA) (ASPLOS ’21)

    Gamma: leveraging Gustavson’s algorithm to accelerate sparse matrix multiplication. In Proceedings of the 26th ACM International Conference on Architec- tural Support for Programming Languages and Operating Systems(Virtual, USA) (ASPLOS ’21). Association for Computing Machiner...

  28. [1965]

    An Algorithm for the Machine Calculation of Complex Fourier Series. Math. Comp. 19, 90 (1965), 297–301. João P. L. de Carvalho, José E. Moreira, and José Nelson Amaral

    1965

  29. [1978]

    ACM Trans

    Two Fast Algorithms for Sparse Matrices: Multiplication and Permuted Transposition. ACM Trans. Math. Softw.4, 3 (Sept. 1978), 250–269. https://doi.org/10.1145/355791.355796 H. Kung

    1978

  30. [1982]

    1982), 37–46

    Why systolic architectures? Computer 15, 1 (Jan. 1982), 37–46. https://doi.org/10.1109/MC. 1982.1653825 Congjie He, Yeqi Huang, Pei Mu, Ziming Miao, Jilong Xue, Lingxiao Ma, Fan Yang, and Luo Mai

    1982

  31. [1985]

    In1985 IEEE 7th Symposium on Computer Arithmetic (ARITH)

    Design of a fast inner product processor. In1985 IEEE 7th Symposium on Computer Arithmetic (ARITH). IEEE, Urbana, Illinois, USA, 38–43. https://doi.org/10.1109/ARITH.1985.6158974 David R. So, Wojciech Ma´nke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V . Le

    1985

  32. [1987]

    Parallel Comput

    Matrix algorithms on a hypercube I: Matrix multiplication. Parallel Comput. 4, 1 (Feb. 1987), 17–31. https://doi.org/10.1016/0167-8191(87)90060-3 Google Gemini

    1987 doi

  33. [1994]

    IBM Journal of Research and Development 38, 6 (Nov

    A high-performance matrix-multiplication algorithm on a distributed-memory parallel computer, using overlapped communication. IBM Journal of Research and Development 38, 6 (Nov. 1994), 673–681. https://doi.org/10.1147/rd.386.0673 Shahanur Alam, Chris Yakopcic, Qing Wu, Mark Ba...

    1994 doi

  34. [1997]

    Concurrency: Practice and Experience 9, 5 (1997), 345–389

    A poly-algorithm for parallel dense matrix multiplication on two- dimensional process grid topologies. Concurrency: Practice and Experience 9, 5 (1997), 345–389. https: //doi.org/10.1002/(SICI)1096-9128(199705)9:5<345::AID-CPE258>3.0.CO;2-7 Qiong Li, Chao Fang, and Zhongfeng W...

    1997

  35. [2000]

    A Multilinear Singular Value Decomposition. SIAM J. Matrix Anal. Appl. 21, 4 (Jan. 2000), 1253–1278. https://doi.org/10.1137/S0895479896305696 Sheng Di, Jinyang Liu, Kai Zhao, Xin Liang, Robert Underwood, Zhaorui Zhang, Milan Shah, Yafan Huang, Jiajun Huang, Xiaodong Yu, Congr...

    2000 doi

  36. [2005]

    Synthesis of High-Performance Parallel Programs for a Class of ab Initio Quantum Chemistry Models. Proc. IEEE 93, 2 (2005), 276–292. https://doi.org/10.1109/JPROC.2004. 840311 Sathwika Bavikadi, Abhijitt Dhavlle, Amlan Ganguly, Anand Haridass, Hagar Hendy, Cory Merkel, Vijay J...

    2005 doi

  37. [2006]

    In ACM/IEEE SC 2006 Conference (SC’06)

    Scalable Algorithms for Molecular Dynamics Simulations on Commodity Clusters. In ACM/IEEE SC 2006 Conference (SC’06). IEEE, Tampa, FL, 43–43. https://doi.org/10.1109/SC.2006.54 Lynn Elliot Cannon

    2006 doi

  38. [2007]

    Chemometrics and Intelligent Laboratory Systems 85, 2 (2007), 170–178

    Multi-way analysis of outer product arrays using PARAFAC. Chemometrics and Intelligent Laboratory Systems 85, 2 (2007), 170–178. https://doi.org/ 10.1016/j.chemolab.2006.06.011 Tomoya Sakai and Stanislav Sedukhin

    2007 doi

  39. [2008]

    ACM Trans

    Anatomy of high-performance matrix multiplication. ACM Trans. Math. Software 34, 3 (May 2008), 1–25. https://doi.org/10.1145/1356052.1356053 Fred G. Gustavson

    2008

  40. [2009]

    Technical Report

    Workshop Report Future Directions in Tensor-Based Computation and Modeling . Technical Report. NDF. 20 pages. https: //doi.org/10.13140/2.1.4040.4807 R. C. Agarwal, F. G. Gustavson, and M. Zubair

  41. [2010]

    In 2010 39th International Conference on Parallel Processing Workshops

    Orbital Algorithms and Unified Array Processor for Computing 2D Separable Transforms. In 2010 39th International Conference on Parallel Processing Workshops. IEEE, San Diego, CA, USA, 127–134. https://doi.org/10.1109/ICPPW.2010.29 Nicholas D. Sidiropoulos, Lieven De Lathauwer,...

    2010 doi

  42. [2011]

    IEICE Transactions on Information and Systems E94-D, 7 (2011), 1409–1418

    3D-DCT Processor and Its FPGA Implementation. IEICE Transactions on Information and Systems E94-D, 7 (2011), 1409–1418. https://doi.org/10.1587/ transinf.E94.D.1409 Geonhwa Jeong, Po-An Tsai, Abhimanyu R. Bambhaniya, Stephen W. Keckler, and Tushar Krishna

    2011

  43. [2012]

    Empirical performance model-driven data layout optimization and library call selection for tensor contraction expressions. J. Parallel and Distrib. Comput.72, 3 (2012), 338–352. https://doi.org/10.1016/j.jpdc.2011. 09.006 Weile Luo, Ruibo Fan, Zeyu Li, Dayou Du, Qiang Wang, an...

    2012 doi

  44. [2013]

    The Johns Hopkins University Press, Baltimore, MD

    Matrix Computations (4th ed.). The Johns Hopkins University Press, Baltimore, MD. https://doi.org/10.56021/9781421407944 Kazushige Goto and Robert A. Van De Geijn

    doi

  45. [2017]

    IEEE Transactions on Signal Processing 65, 13 (2017), 3551–3582

    Tensor Decomposition for Signal Processing and Machine Learning. IEEE Transactions on Signal Processing 65, 13 (2017), 3551–3582. https://doi.org/10.1109/TSP.2017.2690524 S. P. Smith and H. C. Torng

    2017

  46. [2018]

    In 2018 Design, Automation & Test in Europe Conference & Exhibition (DATE)

    FFT-based deep learning deployment in embedded systems. In 2018 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, Dresden, Germany, 1045–1050. https://doi.org/10.23919/DATE.2018.8342166 Ivan Lirkov

    2018

  47. [2019]

    Short-Dot

    “Short-Dot”: Computing Large Linear Transforms Distributedly Using Coded Short Dot Products. IEEE Transactions on Information Theory 65, 10 (2019), 6171–6193. https://doi.org/10.1109/TIT.2019.2927558 Chao Fang, Liulu He, Haonan Wang, Jinghe Wei, and Zhongfeng Wang

    2019

  48. [2020]

    Cybernetics and Information Technologies 20, 6 (Dec

    Performance Analysis of a Scalable Algorithm for 3D Linear Transforms on Supercomputer with Intel Processors/Co-Processors. Cybernetics and Information Technologies 20, 6 (Dec. 2020), 94–104. https: //doi.org/10.2478/cait-2020-0064 Mark Liu and H.-S. Philip Wong

    2020 doi

  49. [2021]

    Matrix Engines for High Performance Computing:A Paragon of Performance or Grasping at Straws? arXiv:2010.14373 [cs.DC] https://arxiv.org/abs/2010. 14373 J. Dongarra and F. Sullivan

    2010 arXiv

  50. [2022]

    IEEE Design & Test 39, 3 (2022), 91–116

    A Survey on Machine Learning Accelerators and Evolutionary Hardware Platforms. IEEE Design & Test 39, 3 (2022), 91–116. https: //doi.org/10.1109/MDAT.2022.3161126 Kevin J. Bowers, David E. Chow, Huafeng Xu, Ron O. Dror, Michael P. Eastwood, Brent A. Gregersen, John L. Klepeis,...

    2022

  51. [2023]

    arXiv:1606.08415 [cs.LG] https://arxiv.org/abs/1606.08415 Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste

    Gaussian Error Linear Units (GELUs). arXiv:1606.08415 [cs.LG] https://arxiv.org/abs/1606.08415 Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste

    arXiv

  52. [2024]

    Electronics 13, 15 (2024), 1–44

    Survey of Deep Learning Accelerators for Edge and Emerging Computing. Electronics 13, 15 (2024), 1–44. https: //doi.org/10.3390/electronics13152988 G. Baumgartner, A. Auer, D.E. Bernholdt, A. Bibireata, V . Choppella, D. Cociorva, Xiaoyang Gao, R.J. Harrison, S. Hirata, S. Kri...

    2024 doi

  53. [2025]

    A Survey on Error-Bounded Lossy Compression for Scientific Datasets. arXiv:2404.02840 [cs.DC] https://arxiv.org/abs/2404.02840 Jens Domke, Emil Vatai, Aleksandr Drozd, Peng Chen, Yosuke Oyama, Lingqi Zhang, Shweta Salaria, Daichi Mukunoki, Artur Podobas, Mohamed Wahib, and Sat...

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.