Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Lookup Table-based Multiplication-free All-digital DNN Accelerator Featuring Self-Synchronous Pipeline Accumulation

T0 review · 4 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 This paper claims an all-digital, clock-free MADDNESS macro reaches 174 TOPS/W and 2.01 TOPS/mm2 in 22nm post-layout simulation, beating an analog baseline 2.5x and 5x at identical DNN accuracy.

desk verdict A clever all-digital MADDNESS macro with attractive simulated numbers, but the lack of any functional simulation means the efficiency claims rest on an unsupported assumption that the self-timed pipeline actually computes the right dot products. read the letter →

arxiv 2506.16800 v1 pith:LJQZQBYM submitted 2025-06-20 cs.AR

Hiroto Tagata , Takashi Sato , Hiromitsu Awano This is my paper · ORCID
classification cs.AR
keywords approximatematrixmultiplicationMADDNESSlookuptablecompute-in-memoryself-synchronouspipelinebinarydecisiontree10T-SRAMDNNaccelerator
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

This paper claims a compact, all-digital accelerator for the MADDNESS scheme, which replaces neural-network multiply-accumulate with product-quantization table lookups. The central claim is that a self-synchronous, clock-free pipeline combined with a binary-decision-tree encoder and a 10T-SRAM lookup table reaches 174 TOPS/W energy efficiency and 2.01 TOPS/mm2 area efficiency in post-layout 22nm simulation. Those numbers are reported as 2.5x and 5x improvements over a prior analog MADDNESS accelerator, while keeping the same 92.6% CIFAR-10 accuracy on ResNet9 as the all-digital Stella Nera design. If correct, the work would show that multiplication-free DNN inference can be done digitally with better energy and area than analog approaches, without PVT calibration. The paper is a circuit-design claim; its accuracy figure is inherited from prior work rather than demonstrated functionally.

What carries the argument

The load-bearing structure is the compute block and its self-synchronous handshake. Each block contains an encoder built as a four-level binary decision tree of 15 dual-rail dynamic logic comparators that classifies a subvector into one of 16 prototypes by activating only the comparators on the selected path, and a decoder made of a two-port 10T-SRAM array storing precomputed prototype-weight dot products. The decoder uses column-level read-completion detection (RCD) trees to sense when the bitlines have fully discharged, so the carry-save adder latches only after the read is done; the RCD outputs feed a four-phase request-acknowledge pipeline that passes results between blocks without a global clock. This asynchronous dataflow is what makes the macro PVT-invariant and lets its speed and energy scale with supply voltage.

What would settle it

Run a post-layout simulation that feeds real activation subvectors and trained thresholds through the macro and compares each compute block's accumulated output to the exact lookup-table sum; any mismatch or handshake stall across process corners would show the macro is not a working 174 TOPS/W accelerator.

Watch

Extended reading notes

Core claim

The paper's claim is that a fully digital MADDNESS accelerator—one that replaces multiply-accumulate with product-quantization table lookups—can be made compact and efficient by encoding inputs with a dataflow-driven binary decision tree of dynamic comparators, storing precomputed dot products in a 16-row by 8-column two-port 10T-SRAM array read without a sense amplifier, and chaining compute blocks with a four-phase self-synchronous pipeline instead of a global clock. Post-layout HSPICE simulation in a commercial 22nm process gives 174 TOPS/W and 2.01 TOPS/mm2 at 0.5V with $N_{\mathrm{dec}}=16$, $N_S=32$; the paper reports this as 2.5x energy and 5x area improvement over the analog accelerator [21], and 4.0x energy improvement over the prior all-digital design [22]. The same ResNet9 CIFAR-10 accuracy of 92.6% as [22] is reported, because the decision-tree mapping and lookup contents are the same algorithm.

Load-bearing premise

The whole efficiency result applies only if the macro actually computes the right approximate dot products, but the paper never runs input vectors through the circuit to check the outputs.

Editorial extensions

If this is right

  • If the reported simulation numbers hold, edge devices can run ResNet9-class CNNs at 174 TOPS/W and 2.01 TOPS/mm2, putting multiplication-free inference within reach of battery-powered sensors.
  • Because the pipeline is self-synchronous, the macro's operating speed shifts automatically with process corner and supply voltage, so the same layout can be reused across PVT conditions without re-tuning a clock.
  • Increasing $N_{\mathrm{dec}}$ from 4 to 16 improves area efficiency by about 43% and energy efficiency by about 4%, with diminishing returns beyond 16, guiding how the macro should be configured.
  • The encoder consumes roughly 95% less energy than the prior all-digital design because it removes threshold memory reads and internal registers, so most remaining energy sits in the SRAM array and carry-save adder.
  • The reported accuracy equals the Stella Nera algorithm's 92.6% CIFAR-10 on ResNet9, so the accuracy cost of the lookup approximation is no worse than the prior digital MADDNESS design.

Reading between the lines

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

  • Beyond the paper's own demonstration, the 92.6% CIFAR-10 number is inherited from the Stella Nera algorithm rather than produced by a functional simulation of this macro, so end-to-end accuracy of the physical circuit remains untested until real vectors are pushed through the netlist.
  • Because all figures come from post-layout simulation at 25°C, leakage and dynamic-comparator behavior at elevated temperature or near-threshold voltage could change both energy and correctness; a temperature sweep would be the natural next check.
  • The claimed 5x area-efficiency advantage over the analog baseline depends on process-normalization assumptions, including scaling only the digital portions of the 65nm comparator, so the comparison is sensitive to how scaling is done.
  • The macro's latency is data-dependent because comparator discharge paths vary with input values, so hard-real-time deployments would need a worst-case latency budget or an output buffering scheme.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper presents a fully digital MADDNESS-based DNN accelerator macro that replaces matrix multiplications with table lookups. The encoder is a binary decision tree implemented with dual-rail dynamic comparators, and the decoder uses a two-port 10T-SRAM array as a lookup table with a carry-save accumulator, all organized in a globally asynchronous, self-synchronous pipeline. Post-layout HSPICE simulation in a commercial 22 nm bulk-CMOS process is used to report energy and area efficiency: up to 174 TOPS/W at 0.5 V, 2.01 TOPS/mm2 at 0.5 V, and 11.34 TOPS/mm2 at 0.8 V, with claimed improvements of 2.5x and 5x over the analog accelerator [21]. The paper states that CIFAR-10 accuracy with ResNet9 is identical to that of Stella Nera [22] (92.6%) and claims PVT-invariant operation. No functional simulation with input/output vectors or comparison against a golden model is reported.

Significance. If the functional correctness of the self-synchronous accumulator were demonstrated, this work would be a useful contribution: it is a fully digital, clock-free design that avoids analog-to-digital conversion overhead and is argued to be robust to PVT variation. The post-layout simulation methodology, the energy/latency/area breakdowns, and the parameter sweep over decoder count are strengths, and the design is described in sufficient detail to be reimplemented. However, the significance is currently limited: all power and area numbers are obtained under the assumption that the macro computes the correct approximate dot products, and the reported accuracy is inherited from [22] rather than measured on this circuit. The absence of functional verification and the questionable comparison assumptions prevent the central efficiency claims from being fully substantiated.

major comments (4)
  1. [Section IV, Table II] No functional simulation is reported anywhere in the paper. There are no input vectors, encoder output addresses, LUT read values, accumulator outputs, or comparisons with expected results from a software model. The statement in Section IV that the CIFAR-10 accuracy is "identical to that reported in [22]" is not supported by any measurement on the proposed circuit. Every TOPS/W and TOPS/mm2 figure assumes the macro performs the intended approximate dot product. Please add a circuit-level functional verification using randomized input subvectors and a golden model of the BDT encoder and LUT accumulation, and report the resulting accuracy or bit-error rate.
  2. [Section III-C, Fig. 5] The self-timed completion detection does not establish that the accumulated values are settled. The RCD_col signal is generated from the full-adder outputs through a NAND/NOR network, so it indicates that those outputs have toggled, not that the carry-save addition has reached a stable final state. The GE pulse is generated after a "brief delay" whose data- and PVT-independence is not demonstrated; a late-arriving carry input or a glitchy FA output could cause the latch to capture an incorrect value and release the next pipeline stage. The same issue applies to the DLC encoder in Section III-B, where the next-level comparator is activated on the basis of a possibly unsettled comparison. Please provide transient simulations for worst-case data patterns across all corners and supply voltages, or a formal analysis of why the handshake guarantees settled data.
  3. [Section IV, Table II] The comparison against [21] is not apples-to-apples. Footnote 5 of Table II states that the decoder energy for [21] does not include the accumulator, whereas the proposed macro includes it, which inflates the claimed energy-efficiency improvement. In addition, the area of [21] is scaled to 22 nm only for the "digital parts" using a 65-to-22 nm factor, while energy is not scaled at all; this mixed scaling is not justified. The claimed 2.5x energy and 5x area improvements depend on these assumptions. Please report the raw unscaled numbers, state explicitly which portions of [21] were scaled and why, and rerun the comparison with accumulator overhead included in [21].
  4. [Section IV, Fig. 6] The claim that the circuit is "PVT-invariant" is stronger than the evidence. Figure 6 sweeps process corners and supply voltage, but only at a single temperature (25 °C), and all data come from post-layout simulation rather than silicon measurement. The self-synchronous design may mitigate some variation, but the paper should either add temperature and supply-variation simulations or soften the claim to "simulation-based robustness across process corners and supply voltage" in the abstract and conclusion.
minor comments (5)
  1. [Section III-B, Fig. 4(C)] The DLC truth table lists the "t = x" case, but the text only describes comparisons that are determined by a single digit. Please clarify how the equal case is resolved by lower-order digits and whether the equal row can occur in practice.
  2. [Section III-C] The memory is called a "two-port 10T-SRAM" but is described as having an independent read port and a separate write port. Please clarify why the bitcell is two-port rather than single-port with a separate read wordline.
  3. [Fig. 6] The figure caption contains stray characters ("TTG Average Conventional works" and repeated "G"), which makes the legend difficult to read. Please clean up the caption and the figure labels.
  4. [Table I and throughout] The decoder-count parameter is written as "Ndec" in Table I and "N_dec" in the text. Please unify the notation.
  5. [References] Reference [22] is an arXiv preprint rather than a peer-reviewed publication. If a formal version exists, please cite it; otherwise, note that the accuracy comparison is against a preprint.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the hardware efficiency figures are independent post-layout simulation measurements, and the accuracy claim is inherited from an external paper rather than derived from this paper's own inputs.

full rationale

The paper's central hardware claims (174 TOPS/W, 2.01 TOPS/mm2, energy and area breakdowns) are obtained from post-layout HSPICE simulation of a layout in a commercial 22nm process, not derived from the target result or from any fitted parameter. There are no self-citations: the comparison baselines [21] and [22] are authored by independent groups, and the BDT-based encoding is explicitly attributed to the external MADDNESS literature ([16], [22]) rather than presented as this paper's own prior result. The CIFAR-10 accuracy of 92.6% is quoted from Stella Nera [22] and does not function as a derivation of the proposed macro's efficiency; it is an external algorithmic benchmark. The absence of functional simulation for the self-timed pipeline is a legitimate correctness and verification concern, but it is not a circularity: the reported power, area, and latency numbers would indeed be invalid if the macro does not compute the intended dot products, yet that possibility is a missing validation, not a logical equivalence between inputs and outputs. No equation, parameter, or claim in the paper reduces by construction to another claim or to a self-citation, so the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. The central performance figures depend on hand-chosen design parameters (N_dec, N_S, VDD) and on inherited results from prior work: the MADDNESS accuracy and the BDT thresholds/codebooks from [22]. The soundness of the efficiency numbers also rests on the unverified assumption that the circuit functions correctly.

free parameters (3)
  • N_dec (decoders per compute block) = 16 (recommended; 4, 8, 32 also evaluated)
    The headline 174 TOPS/W and 2.01 TOPS/mm2 figures assume N_dec=16, N_S=32 at 0.5V. N_dec=32 gives 174.9 TOPS/W and the same area efficiency, but the paper recommends 16 for robustness to local variation.
  • N_S (pipeline stages) = 32
    The main evaluation uses N_S=32, corresponding to 32 input channels processed concurrently. This affects area, energy, and latency; other values are not reported.
  • Supply voltage VDD = 0.5V (max energy efficiency), 0.8V (nominal)
    174 TOPS/W is the 0.5V result; the 0.8V result is 75.1 TOPS/W and 11.34 TOPS/mm2. The claimed 2.5x/5x improvements are specifically at 0.5V.
assumptions (3)
  • domain assumption MADDNESS with a 4-level binary decision tree (16 prototypes) and 8-bit integer inputs achieves 92.6% accuracy on CIFAR-10 with ResNet9.
    This accuracy is taken from the Stella Nera paper [22]; the present paper does not retrain or re-evaluate the algorithm and simply reuses the trained thresholds and codebooks.
  • domain assumption The proposed digital circuits (dynamic comparators, 10T-SRAM read, CSA accumulation, asynchronous handshake) compute results identical to the software MADDNESS model.
    No functional simulation is provided; the paper only reports energy, area, and latency. Correctness is implicitly assumed from the digital logic design.
  • ad hoc to paper Cross-process comparison by scaling area with (65/22)^2 for [21] and (14/22)^2 for [22], while leaving energy unscaled, is a fair basis for the claimed improvements.
    Such scaling is common but can overstate improvements when the baseline is analog and its analog portions are not scaled. The paper applies scaling only to digital parts of [21].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lookup Table-based Multiplication-free All-digital DNN Accelerator Featuring Self-Synchronous Pipeline Accumulation." pith.science (2026). https://pith.science/paper/LJQZQBYM

@misc{pith2026250616800,
  author       = {Pith},
  title        = {Pith review of: Lookup Table-based Multiplication-free All-digital DNN Accelerator Featuring Self-Synchronous Pipeline Accumulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJQZQBYM}},
  note         = {Machine review of arXiv:2506.16800}
}
read the original abstract

Deep neural networks (DNNs) have been widely applied in our society, yet reducing power consumption due to large-scale matrix computations remains a critical challenge. MADDNESS is a known approach to improving energy efficiency by substituting matrix multiplication with table lookup operations. Previous research has employed large analog computing circuits to convert inputs into LUT addresses, which presents challenges to area efficiency and computational accuracy. This paper proposes a novel MADDNESS-based all-digital accelerator featuring a self-synchronous pipeline accumulator, resulting in a compact, energy-efficient, and PVT-invariant computation. Post-layout simulation using a commercial 22nm process showed that 2.5 times higher energy efficiency (174 TOPS/W) and 5 times higher area efficiency (2.01 TOPS/mm2) can be achieved compared to the conventional accelerator.

Figures

Figures reproduced from arXiv: 2506.16800 by the authors.

Figure 1
Figure 1. Overview of the MADDNESS algorithm. The input vec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall circuit of the proposed LUT-based DNN accel [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Proposed decision tree-based encoder. (A) Overall schematic of the proposed encoder circuit. (B), (C) Schematic and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: Energy and area efficiency of the proposed circuit [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: (A) Energy, (B) latency, and (C) area breakdown based [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 13 canonical work pages

  1. [21]

    Lookup Table-Based Computing-in-Memory Macro Approximating Dot Products Without Multiplications for Energy-Efficient CNN Inference,

    H. Fuketa, “Lookup Table-Based Computing-in-Memory Macro Approximating Dot Products Without Multiplications for Energy-Efficient CNN Inference,”IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 70, no. 10, pp. 3954–3963, 2023.DOI: 10.1109/TCSI.2023.3290151

    arXiv 2023

  2. [22]

    Stella Nera: A Differentiable Maddness-Based Hardware Accelerator for Efficient Approximate Matrix Multiplication

    J. Sch ¨onleber, L. Cavigelli, R. Andri, M. Perotti, and L. Benini,Stella Nera: Achieving 161 TOp/s/W with Multiplier- free DNN Acceleration based on Approximate Matrix Multipli- cation, 2023. arXiv: 2311.10207[cs.AR]

    work page Pith review arXiv 2023

  3. [1]

    ImageNet clas- sification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet clas- sification with deep convolutional neural networks,”Commun. ACM, vol. 60, no. 6, 84–90, May 2017.DOI: 10.1145/3065386

    doi:10.1145/3065386 2017

  4. [2]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778.DOI: 10.1109/ CVPR.2016.90

    work page 2016

  5. [3]

    Training Very Deep Networks,

    R. K. Srivastava, K. Greff, and J. Schmidhuber, “Training Very Deep Networks,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2015

    work page 2015

  6. [4]

    Steiner, M

    B. Steiner, M. Elhoushi, J. Kahn, and J. Hegarty,OLLA: Optimizing the Lifetime and Location of Arrays to Reduce the Memory Usage of Neural Networks, 2022. arXiv: 2210.12924 [cs.LG]

    arXiv 2022

  7. [5]

    Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolu- tional Neural Networks,

    Y .-H. Chen, T. Krishna, J. S. Emer, and V . Sze, “Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolu- tional Neural Networks,”IEEE J. Solid-State Circuits, vol. 52, no. 1, pp. 127–138, 2017.DOI: 10.1109/JSSC.2016.2616357

    arXiv 2017

  8. [6]

    In-Memory Computation of a Machine-Learning Classifier in a Standard 6T SRAM Array,

    J. Zhang, Z. Wang, and N. Verma, “In-Memory Computation of a Machine-Learning Classifier in a Standard 6T SRAM Array,” IEEE J. Solid-State Circuits, vol. 52, no. 4, pp. 915–924, 2017. DOI: 10.1109/JSSC.2016.2642198

    arXiv 2017

Show all 26 references
  1. [8]

    A Charge- Sharing Based 8T SRAM In-Memory Computing for Edge DNN Acceleration,

    K. Lee, S. Cheon, J. Jo, W. Choi, and J. Park, “A Charge- Sharing Based 8T SRAM In-Memory Computing for Edge DNN Acceleration,” inProc. ACM/IEEE Design Autom. Conf. (DAC), 2021, pp. 739–744.DOI: 10 . 1109 / DAC18074 . 2021 . 9586103

    2021

  2. [10]

    34.4 A 3nm, 32.5TOPS/W, 55.0TOPS/mm 2 and 3.78Mb/mm 2 Fully-Digital Compute-in-Memory Macro Supporting INT12×INT12 with a Parallel-MAC Architecture and Foundry 6T-SRAM Bit Cell,

    H. Fujiwara et al., “34.4 A 3nm, 32.5TOPS/W, 55.0TOPS/mm 2 and 3.78Mb/mm 2 Fully-Digital Compute-in-Memory Macro Supporting INT12×INT12 with a Parallel-MAC Architecture and Foundry 6T-SRAM Bit Cell,” inProc. IEEE Int. Solid- State Circuits Conf. (ISSCC), vol. 67, 2024, pp. 572...

    2024

  3. [11]

    A 1-16b Reconfigurable 80Kb 7T SRAM-Based Digital Near- Memory Computing Macro for Processing Neural Networks,

    H. Kim, J. Mu, C. Yu, T. T.-H. Kim, and B. Kim, “A 1-16b Reconfigurable 80Kb 7T SRAM-Based Digital Near- Memory Computing Macro for Processing Neural Networks,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 70, no. 4, pp. 1580–1590, 2023.DOI: 10.1109/TCSI.2022.3232648

    2023

  4. [12]

    1.1 Computing’s Energy Problem (and What We Can Do About It),

    M. Horowitz, “1.1 Computing’s Energy Problem (and What We Can Do About It),” inProc. IEEE Int. Solid-State Circuits Conf. (ISSCC), 2014, pp. 10–14.DOI: 10.1109/ISSCC.2014.6757323

    2014

  5. [13]

    Patient- Specific Seizure Prediction via Adder Network and Supervised Contrastive Learning,

    Y . Zhao, C. Li, X. Liu, R. Qian, R. Song, and X. Chen, “Patient- Specific Seizure Prediction via Adder Network and Supervised Contrastive Learning,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 30, pp. 1536–1547, 2022.DOI: 10 . 1109 / TNSRE . 2022 . 3180155

    2022

  6. [14]

    LAcc: Exploiting Lookup Table-Based Fast and Accurate Vector Multiplication in DRAM-Based CNN Accelerator,

    Q. Deng, Y . Zhang, M. Zhang, and J. Yang, “LAcc: Exploiting Lookup Table-Based Fast and Accurate Vector Multiplication in DRAM-Based CNN Accelerator,” inProc. ACM/IEEE Design Autom. Conf. (DAC), 2019, pp. 1–6

    2019

  7. [15]

    BiQGEMM: Matrix Multiplication with Lookup Table for Binary-Coding-Based Quantized DNNs,

    Y . Jeon, B. Park, S. J. Kwon, B. Kim, J. Yun, and D. Lee, “BiQGEMM: Matrix Multiplication with Lookup Table for Binary-Coding-Based Quantized DNNs,” inProc. Int. Conf. High Perform. Comput., Netw., Storage Anal. (SC20), 2020, pp. 1–14.DOI: 10.1109/SC41405.2020.00099

    2020 arXiv

  8. [16]

    Blalock and J

    D. Blalock and J. Guttag,Multiplying Matrices Without Multi- plying, 2021. arXiv: 2106.10860[cs.LG]

    2021 arXiv

  9. [17]

    J. Ran, R. Lin, J. C. L. Li, J. Zhou, and N. Wong,PECAN: A Product-Quantized Content Addressable Memory Network,

  10. [18]

    LUT-NN: Empower Efficient Neural Network Inference with Centroid Learning and Table Lookup,

    X. Tang et al., “LUT-NN: Empower Efficient Neural Network Inference with Centroid Learning and Table Lookup,” inProc. 29th Annu. Int. Conf. Mobile Computing Netw. (ACM MobiCom ’23), 2023.DOI: 10.1145/3570361.3613285

    2023

  11. [19]

    Product Quantization for Nearest Neighbor Search,

    H. J ´egou, M. Douze, and C. Schmid, “Product Quantization for Nearest Neighbor Search,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 1, pp. 117–128, 2011.DOI: 10.1109/TPAMI. 2010.57

    2011 doi

  12. [20]

    McCarter and N

    C. McCarter and N. Dronen,Look-ups are not (yet) all you need for deep learning inference, 2022. arXiv: 2207 . 05808 [cs.LG]

    2022

  13. [23]

    Gradient-based learning applied to document recognition,

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proc. of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.DOI: 10.1109/5.726791

    1998 doi

  14. [24]

    Pipeline Architecture,

    C. V . Ramamoorthy and H. F. Li, “Pipeline Architecture,”ACM Comput. Surv., vol. 9, no. 1, 61–102, Mar. 1977.DOI: 10.1145/ 356683.356687

    1977

  15. [25]

    Asynchronous design methodologies: an overview,

    S. Hauck, “Asynchronous design methodologies: an overview,” Proc. IEEE, vol. 83, no. 1, pp. 69–93, 1995.DOI: 10.1109/5. 362752

    1995 doi

  16. [26]

    A four-phase handshaking asynchronous static RAM design for self-timed systems,

    V . W.-Y . Sit, C.-S. Choy, and C.-F. Chan, “A four-phase handshaking asynchronous static RAM design for self-timed systems,”IEEE J. Solid-State Circuits, vol. 34, no. 1, pp. 90–96, 1999.DOI: 10.1109/4.736660

    1999 doi

  17. [27]

    A Replica Technique for Wordline and Sense Control in Low-Power SRAMs,

    B. Amrutur and M. Horowitz, “A Replica Technique for Wordline and Sense Control in Low-Power SRAMs,”IEEE J. Solid-State Circuits, vol. 33, no. 8, pp. 1208–1219, 1998.DOI: 10.1109/4.705359

    1998 doi

  18. [2022]

    arXiv: 2208.13571[cs.LG]

    arXiv

Pith tools

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