Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Placing spiking layers only at chip boundaries gives ANN-level accuracy with up to 15.2x less latency and 5.3x less energy.

desk verdict A sensible partitioning idea with honest but assumption-heavy simulation; the efficiency headline needs sensitivity analysis before it can be trusted. read the letter →

arxiv 2501.08645 v2 pith:OWPDGUN5 submitted 2025-01-15 cs.AR

Joshua Nardone , Ruijie Zhu , Joseph Callenes , Mohammed E. Elbtity , Ramtin Zand , Jason Eshraghian This is my paper · ORCID
classification cs.AR
keywords heterogeneousneuralnetworksspikingdie-to-diecommunicationnetwork-on-chipenergyefficiencylatencyreductionlearnablesparsityspike-basedencoding
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 proposes heterogeneous neural networks (HNNs) that split a deep network so that dense ANN layers run inside each chip while spiking layers sit at the die-to-die boundary, converting activations into sparse spike trains before they cross the interconnect. The claim is that this simple partition, putting spikes only where bandwidth is tight, preserves ANN-level accuracy while capturing SNN-style communication savings. Across language modeling (Enwik8), CIFAR100, and ImageNet-1K, the authors report up to 15.2x lower inference latency and 5.3x better energy efficiency than pure ANN baselines, with HNN accuracy matching or slightly beating the ANN on two of three tasks. If correct, the long-standing SNN scalability gap can be sidestepped by using spikes as a communication code rather than as the sole computing substrate.

What carries the argument

The load-bearing mechanism is the cross-layer packet (CLP) converter that sits between ANN and SNN cores. It turns each 8-bit activation $a_i$ into a deterministic rate-coded spike train $s_i(t)=1$ if $t < \lfloor a_i/T \rfloor$ over a $T$-tick window for die-to-die transmission, and on the receiving side accumulates spikes and scales the count back to an activation with $a_i = \lfloor (2^b-1)/T \cdot \sum_{t=0}^{T-1} s_i(t) \rfloor$. This gives a learned, data-dependent sparsification of boundary traffic, encouraged by the regularizer $L = L_{CE} + \lambda \sum_i s_i$, and it is paired with an EMIO interconnect and a NoC simulator whose energy model counts MACs, ACCs, and routed packets.

What would settle it

Measure on a fabricated 65nm or equivalent test chip the actual per-packet die-to-die energy and latency through the EMIO and SerDes, plus the energy of an SNN accumulate versus an ANN MAC, and re-run the same RWKV, MS-ResNet, and EfficientNet mappings with those measured values. If die-to-die movement is not roughly an order of magnitude costlier than a MAC, or if ACC energy is not about 0.06x a MAC, the claimed HNN speedups and efficiency gains will not reproduce.

Watch

Extended reading notes

Core claim

The central discovery is that spiking layers do not need to run everywhere to deliver their benefits; confining them to bandwidth-limited chip boundaries is enough. An HNN partitions a network into ANN-dominated interior cores and SNN peripheral cores, uses a cross-layer packet converter to translate dense 8-bit activations into rate-coded spike trains and back, and learns sparsity through a regularization term that penalizes excessive spiking. On Enwik8 the HNN reaches 2.57 perplexity versus 2.66 for the ANN, on CIFAR100 it reaches 78.86% versus 78.65%, and on ImageNet it reaches 74.78% versus 75.48%, while the pure SNN trails at 67.50%. The efficiency gains, up to 15.2x speedup and 5.3x energy improvement, come from replacing dense cross-chip traffic with sparse spikes, and they grow as model size, bit precision, or the number of chip boundaries increases.

Load-bearing premise

The headline gains rest on the simulator's assumed relative energy costs, namely that a spike accumulate costs only 0.06x a multiply-accumulate and that die-to-die data movement costs 10x a MAC and 224x a core-to-core hop, so if real silicon shows different ratios, the 5.3x and 15.2x numbers shrink.

Editorial extensions

If this is right

  • HNNs beat both pure ANNs and pure SNNs on the accuracy-efficiency frontier across language and vision benchmarks.
  • The speedup and energy gains scale superlinearly with model size, bit precision, and the number of die-to-die crossings, making the approach more attractive for larger distributed models.
  • Static, non-temporal inputs can be handled by rate-encoding them over a $T$-step window without sacrificing accuracy, bringing SNN communication benefits to conventional deep learning datasets.
  • Spiking layers in an HNN can act as a regularizer, yielding slightly better generalization than the ANN alone on Enwik8 and CIFAR100.
  • The architecture is composable with existing ANN and SNN core designs, so the boundary-spiking recipe can be adapted to alternative accelerator cores.

Reading between the lines

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

  • Inference: The same boundary-spiking recipe could be applied to other communication bottlenecks inside a chip, such as between memory partitions or across chiplets in a 2.5D package, wherever link energy dominates compute energy.
  • Inference: A direct hardware measurement of the MAC/ACC and die-to-die energy ratios on a real 65nm or smaller process would be the fastest way to validate or refute the 5.3x and 15.2x figures, since the simulator's constants drive the result.
  • Inference: The rate-coding window size $T$ creates a tunable trade-off between communication precision and traffic volume; shrinking $T$ further would test whether accuracy degrades gracefully, potentially enabling even larger speedups.
  • Inference: Because the sparsity penalty is learned per layer, the method could be extended to automatically decide which layers should be spiking, rather than fixing the partition at chip boundaries.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a heterogeneous neural network (HNN) design in which spiking layers are confined to die-to-die boundaries while dense ANN layers occupy the chip interior. The claimed benefit is that spike-based, learnably sparse communication reduces inter-chip traffic without the accuracy loss of full SNN models. The authors describe an 8x8 NoC architecture with EMIO boundary blocks and CLP packet converters, implement a custom cycle/energy simulator, and evaluate RWKV, MS-ResNet18, and EfficientNet-B4 on Enwik8, CIFAR100, and ImageNet-1K. They report accuracy on par with or better than ANN baselines and up to 15.2x latency reduction and 5.3x energy efficiency gains over ANN baselines.

Significance. If the efficiency claims held under realistic assumptions, this would be a useful design point: it gives a principled reason to place SNN layers at bandwidth bottlenecks rather than replacing entire networks, and it provides a concrete architecture plus a simulator extension. The paper also makes a good-faith effort to report per-layer sparsity, latency sweeps, and energy breakdowns, and the RTL-based EMIO latency estimate is a concrete artifact. However, the headline 15.2x/5.3x numbers currently follow from two assumed energy ratios plus an asymmetric comparison with ANN baselines that do not exploit zero-skipping. The contribution is therefore conditionally significant: the concept is plausible, but the supporting evidence is not yet load-bearing.

major comments (4)
  1. [Section 4.4] The energy model's dominant constants are assumed, not measured: SNN accumulation is set to 0.06x the energy of a MAC, and die-to-die movement is set to 10x a MAC and 224x a core-to-core packet per hop. Since the HNN advantage is realized precisely by moving boundary computation from MACs and dense packets to ACCs and spike packets, these two ratios are the main levers for the reported 5.3x energy and 15.2x latency gains. No sensitivity analysis over either constant is provided. The authors should report a sensitivity sweep over both ratios, identify the breakeven points at which HNN loses its advantage, and justify the chosen values with measurements or published data at the same 65nm/200MHz operating point.
  2. [Section 5.1] The comparison to ANN baselines is asymmetric: the paper states that 'zero-skipping is not implemented in the ANN cores.' The ANN baselines therefore count every MAC and every dense packet, while the HNN benefits from activation sparsity in its spiking layers. This conflates the architecture's benefit with the benefit of activation sparsity itself. A fair latency/energy comparison should either implement zero-skipping in the ANN cores or include an ANN variant that prunes/skips inactive activations, and then report how much of the 15.2x/5.3x remains.
  3. [Section 5.1, Table 4] The accuracy deltas between HNN and ANN are small (Enwik8 PPL 2.57 vs 2.66; CIFAR100 78.86% vs 78.65%; ImageNet 74.78% vs 75.48%), and no repeated-seed variance or error bars are reported. These differences are within the range of typical run-to-run noise for such training setups. The claim that HNNs are 'at least on par with' ANNs needs mean-plus-variance statistics, number of seeds, and, ideally, a statistical significance test; otherwise the accuracy parity claim is not established.
  4. [Sections 4.2 and 5.2] The reported latency and energy results are simulator outputs driven by assumed workload parameters, but the paper does not report the final realized sparsity or the actual packet counts for the selected operating points. The text assumes 90% input sparsity and T=8 timesteps for SNN input encoding, and the sparsity sweep is used to select the reported operating point, yet Table 4 and the latency/energy results do not state the final sparsity per model, per layer, or the number of die-to-die packets in the ANN vs HNN cases. Without these numbers, the reader cannot trace the claimed speedup and energy gain back to concrete traffic reductions, so the 15.2x and 5.3x figures should be treated as conditional on unreported workload details.
minor comments (5)
  1. [Section 5.2] The text says 'As shown in Fig. 13, the HNN exhibited speedups...' but Figure 13 is labeled as normalized energy efficiency; the speedup results appear to be in Figure 11. Please correct the cross-reference.
  2. [Figure 12] The energy-per-inference figure would benefit from explicit axis labels and units; the current caption states 'Energy Consumption (J) per Inference', but the axes are not labeled in the plot.
  3. [Equation (2)] The rate-coding rule in Eq. (2) can generate more than T spikes when a_i is large, because the condition t < floor(a_i / T) is not bounded by T. Please clarify the normalization or define a_i/T modulo the window.
  4. [Section 4.2] Reference [21] is formatted inconsistently ('Chen Kun-Chih and Wang ting Yi'); also, a brief statement about code/artifact availability for the custom simulator would aid reproducibility.
  5. [Section 6] The concluding section says the paper 'concludes with related works and an outline of future directions', but related work appears in Section 2; this sentence should be reworded.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central efficiency and latency claims are outputs of a simulation whose cost ratios are externally sourced assumptions, and the accuracy claims are independent benchmark results.

full rationale

The paper's derivation chain is not circular. The core architectural claim is that confining spiking layers to bandwidth-limited die-to-die boundaries reduces communication overhead while preserving accuracy. Accuracy is established by actual training runs on Enwik8, CIFAR100, and ImageNet-1K, compared against ANN and SNN baselines. The energy and latency numbers are produced by a custom NoC simulator, and its key constants are adopted from external sources, not fitted to the paper's own conclusions: Section 4.4 states that 'SNN inference consumes approximately 0.06× the energy of a MAC operation' and that 'die-to-die data movement consumes nearly 10× more energy than a MAC operation, and 224× that of a core-to-core packet per hop,' citing ORION 2.0, TrueNorth, and prior hardware-aware studies. These are modeling assumptions with stated provenance, so the resulting efficiency gains are conditional model outputs rather than conclusions that reduce to their own premises by definition. The statement in Section 5.3 that 'This efficiency arises from the computational cost reduction inherent in SNN layers' explains the mechanism of the model rather than introducing a new circular derivation. The sparsity sweep in Section 5.1 is a standard design-space exploration used to select a Pareto-optimal operating point, not a fitted parameter being renamed as a prediction. Self-citations, including Refs. [9], [39], [51], and [52], support background claims about LIF dynamics, hybrid SNN/ANN modes, and SNN language models, but none of these citations is load-bearing for the paper's central result. No equation in the paper is equivalent to its own input by construction, and no fitted quantity is presented as a prediction. The efficiency claims may be sensitive to the assumed cost ratios, but that is a correctness or robustness concern, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central efficiency claims depend on assumed energy ratios and chosen sparsity levels; the accuracy claims depend on standard training but lack statistical detail.

free parameters (6)
  • SNN-to-MAC energy ratio = 0.06x
    Assumed in the energy model (Section 4.4) based on Dampfhoffer et al. [6]; SNN accumulation is set to 6% of MAC energy, which directly determines much of the reported energy efficiency.
  • Die-to-die energy ratio = 10x MAC, 224x core-to-core packet per hop
    Assumed from TrueNorth and ORION 2.0 (Section 4.4); sets the cost of boundary packet movement and is a primary driver of the latency and energy comparisons.
  • Timestep window T = 8
    Rate-encoded spike trains use T=8 timesteps (Section 4.2), chosen following SpinalFlow; affects packet counts and operation counts for SNN and HNN layers.
  • Latency per MAC/ACC cycle = 1 cycle each
    Equations (6) and (7) assume one-cycle latency for both MAC and ACC operations, which makes ACC-versus-MAC comparisons favor SNN layers.
  • Sparsity regularization weight lambda = not reported
    Equation (10) uses a penalty active only above a desired sparsity; lambda and the target sparsity are chosen per model via the sweep in Section 5.1, and the final reported numbers use the selected Pareto-optimal point.
  • SNN input sparsity = 90%
    The simulator assumes 90% sparse inputs for SNN models (Section 4.2), which reduces ACC counts and therefore favors SNN and HNN energy estimates.
assumptions (4)
  • domain assumption ORION 2.0 energy model for the Intel 65nm 80-core chip scales correctly to the proposed 8x8 NoC at 1.0V and 200MHz.
    Invoked in Section 4.2 to compute PE, memory, and router energy; if the scaling is inaccurate, the absolute and relative energy numbers shift.
  • domain assumption TrueNorth-based 224x die-to-die versus core-to-core energy ratio holds for a synchronous 200MHz EMIO with 8-to-1 multiplexing.
    Section 4.4 borrows this ratio from TrueNorth's asynchronous, 32-to-1 design; the paper acknowledges the EMIO differs in clocking and multiplexing but still uses the same energy factor.
  • domain assumption Zero-skipping of spike traffic is available in SNN cores and absent in ANN cores.
    The simulator counts operations proportional to active spikes for SNN and HNN layers and full MACs for ANN layers; this asymmetry is the source of the efficiency gap (Sections 4.2 and 5.3).
  • domain assumption Static inputs can be rate-encoded with T=8 without losing task-relevant information.
    Section 4.2 uses T=8 for all SNN and HNN static-data evaluations; accuracy results depend on this encoding being sufficient for the tasks.
invented entities (2)
  • EMIO (Extended Mux I/O) block
    purpose: Die-to-die packet multiplexing and serialization at the chip boundary.
    Only a synthesized latency of 76 cycles is reported (Section 3.4); no power, area, or measured bandwidth numbers are given, so the claimed boundary efficiency lacks external confirmation.
  • CLP (Cross-layer Packet) converter
    purpose: Converts between activation-encoded ANN packets and rate-encoded spike trains in both directions.
    Described logically in Section 3.5 and Figure 4, but no synthesis or silicon results are provided; its cost is not separately measured in the simulator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding." pith.science (2026). https://pith.science/paper/OWPDGUN5

@misc{pith2026250108645,
  author       = {Pith},
  title        = {Pith review of: Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWPDGUN5}},
  note         = {Machine review of arXiv:2501.08645}
}
read the original abstract

Efficient communication is central to both biological and artificial intelligence (AI) systems. In biological brains, the challenge of long-range communication across regions is addressed through sparse, spike-based signaling, minimizing energy and latency. Conversely, modern AI workloads are increasingly constrained by bandwidth, leading to bottlenecks that hamper scalability and efficiency. Inspired by the brain's ability to execute dynamic and complex local computations coupled with sparse inter-neuron communication, we propose heterogeneous neural networks that combine spiking neural networks (SNNs) and artificial neural networks (ANNs) at bandwidth-limited regions, such as chip boundaries, where spike-based communication reduces data transfer overhead. Within each chip, dense ANN computations maintain high throughput, accuracy, and robustness. While SNNs have struggled to algorithmically scale, our approach surmounts this long-standing challenge through algorithm-architecture co-design where learnable sparsity is employed for die-to-die communication by confining spiking layers to specific partitions. This composable design combines high ANN performance with low-bandwidth SNN efficiency. Evaluations on language processing and computer vision exhibit up to 5.3x energy efficiency gains and 15.2x latency reductions, surpassing both purely spiking and non-spiking models. As model size grows, improvements scale accordingly. By targeting the inter-chip communication bottleneck with biologically inspired methods, this approach presents a promising path to more efficient AI systems.

Figures

Figures reproduced from arXiv: 2501.08645 by the authors.

Figure 1
Figure 1. An overview. Artificial neurons are integrated with spiking neurons, where [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. 2-D Mesh NoC Hybrid Hardware Accelerator Overview. A high-level overview of the hybrid approach to solve the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. High Level Interconnect (single port). Flow of I/O [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Cross-Layer Activation-to-Spiking and Spiking-to [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Model simulation workflow. Visualizing how ANN, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Activation sparsity sweep for each model. Latency improves with more sparsity. Model performance (perplex [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Activation heatmap across layers. HNNs are only sparsified at the spiking boundary layers, and partitioned based [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Comparison of performance across different datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Latency per Inference Speedup (×) for Enwik8, CIFAR100, & ImageNet-1K inputs on corresponding RWKV, MS-ResNet18, & EfficientNet-B4 model architectures using base parameters at 8-bit precision, 256 neuron grouping, & 8 core chip NoC dimensions. 6 Conclusion As deep lea…
Figure 11
Figure 11. Figure 11: Normalized speed-up w.r.t. ANN as a function of bit-width, NoC dimensions, and grouping. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Energy Consumption (J) per Inference for Enwik8, [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Normalized energy efficiency w.r.t ANN as a function bit-width, NoC dimensions, and grouping. [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 44 canonical work pages

  1. [1]

    Malyaban Bal and Abhronil Sengupta. 2024. Rethinking spiking neural networks as state space models. arXiv e-prints (2024), arXiv–2406

    work page 2024

  2. [2]

    Muya Chang, Ashwin Sanjay Lele, Samuel D Spetalnick, Brian Crafton, Shota Konno, Zishen Wan, Ashwin Bhat, Win-San Khwa, Yu-Der Chih, Meng-Fan Chang, et al. 2023. A 73.53 TOPS/W 14.74 TOPS heterogeneous RRAM in-memory and SRAM near-memory SoC for hybrid frame and event-based target tracking. In 2023 IEEE International Solid-State Circuits Conference (ISSCC...

    work page 2023

  3. [3]

    Kun-Chih Chen, Masoumeh Ebrahimi, Ting-Yi Wang, and Yuch-Chi Yang. 2020. A NoC-based simulator for design and evaluation of deep neural networks. Mi- croprocessors and Microsystems 77 (2020) (2020)

    work page 2020

  4. [4]

    Yu-Hsin Chen, Tushar Krishna, and Joel S. Emer. 2019. NeuronLink: An Efficient Chip-to-Chip Interconnect for Neuromorphic Computing. In Proceedings of the International Symposium on Low Power Electronics and Design . 1–6. https://doi. org/10.1145/3341325.3342030

    work page arXiv 2019

  5. [5]

    Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. 2019. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems 9, 2 (2019), 292–308

    work page 2019

  6. [6]

    Manon Dampfhoffer, Thomas Mesquida, Alexandre Valentian, and Lorena Anghel

    work page

  7. [7]

    Mike Davies, Alex Neckar, Kenneth Lai, Steve Goetz, Tobi Lin, Vikram Chandra, Matt Chou, Sylvain Cardin, and Steve Esser. 2021. Loihi2: A neuromorphic research chip combining spike-based learning and data-driven plasticity. arXiv preprint arXiv:2108.03229 (2021)

    work page Pith review arXiv 2021

  8. [8]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition . Ieee, 248–255

    2009

Show all 55 references
  1. [9]

    Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D Lu. 2023. Training spiking neural networks using lessons from deep learning. Proc. IEEE (2023). 10 Learnable Sparsification of Die-to-Die Commun...

    2023

  2. [10]

    Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. 2021. Deep residual learning in spiking neural networks. Ad- vances in Neural Information Processing Systems 34 (2021), 21056–21069

    2021

  3. [11]

    Charlotte Frenkel and Giacomo Indiveri. 2022. ReckOn: A 28nm sub-mm2 task- agnostic spiking recurrent neural network processor enabling on-chip learning over second-long timescales. In 2022 IEEE International Solid-State Circuits Con- ference (ISSCC), Vol. 65. IEEE, 1–3

    2022

  4. [12]

    Charlotte Frenkel, Martin Lefebvre, Jean-Didier Legat, and David Bol. 2019. A 0.086-mm2 12.7-pJ/SOP 65k-Synapse 256-Neuron Online-Learning Digital Spik- ing Neuromorphic Processor in 28-nm CMOS. IEEE Transactions on Biomedical Circuits and Systems, Vol. 13, NO. 1, February 2019 (2019)

    2019

  5. [13]

    Mingyu Gao, Jing Pu, Xuan Yang, Mark Horowitz, and Christos Kozyrakis. 2017. Tetris: Scalable and efficient neural network acceleration with 3d memory. In Proceedings of the Twenty-Second International Conference on Architectural Support for Programming Languages and Operating...

    2017

  6. [14]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)

    2023 arXiv

  7. [15]

    John L Hennessy and David A Patterson. 2019. A new golden age for computer architecture. Commun. ACM 62, 2 (2019), 48–60

    2019

  8. [16]

    Yatin Hoskote, Sriram Vangal, Arvind Singh, Nitin Borkar, and Shekhar Borkar

  9. [17]

    Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. 2021. Advancing Spiking Neural Networks towards Deep Residual Learning. arXiv preprint arXiv:2112.08954 (2021)

    2021 arXiv

  10. [18]

    Andrew Kahng, Bin Li, Li-Shiuan Peh, and Kambiz Samadi. 2009. ORION 2.0: A Fast and Accurate NoC Power and Area Model for Early-Stage Design Space Exploration. 2009 Design, Automation & Test in Europe Conference & Exhibition (2009)

    2009

  11. [19]

    Kim, Joonhyung K

    Jonghoon K. Kim, Joonhyung K. Kim, and Jongsun P. Kim. 2018. RANC: Re- configurable Architecture for Neuromorphic Computing. In Proceedings of the International Conference on Reconfigurable Computing and FPGAs . 1–6. https: //doi.org/10.1109/ReConFig.2018.8641707

    2018

  12. [20]

    Alex Krizhevsky and Geoffrey Hinton. 2009. Learning multiple layers of features from tiny images. Technical Report. University of Toronto, Toronto, ON, Canada

    2009

  13. [21]

    Chen Kun-Chih and Wang ting Yi. 2018. NN-Noxim: High-Level Cycle-Accurate NoC-based Neural Networks Simulator. IEEE Access 8 (2018), 108527–108539. https://doi.org/10.1109/ACCESS.2018.3002082

    2018

  14. [22]

    Eren Kurshan and Paul Franzon. 2024. Towards 3D AI Hardware: Fine-Grain Hardware Characterization of 3D Stacks for Heterogeneous System Integration & AI Systems. arXiv preprint arXiv:2409.10539 (2024)

    2024 arXiv

  15. [23]

    Simon B Laughlin and Terrence J Sejnowski. 2003. Communication in neuronal networks. Science 301, 5641 (2003), 1870–1874

    2003

  16. [24]

    Donghyun Lee, Guoqi Li, Hongsik Jeong, and Dong-Hyeok Lim. 2022. QAD-SNN: Quantization-aware Distillation on Spiking Neural Network. Korea Artificial Intelligence Conference (KAIC) (2022). 11 , , Figure 13: Normalized energy efficiency w.r.t ANN as a function bit-width, NoC di...

    2022

  17. [25]

    Jong Chern Lee, Jihwan Kim, Kyung Whan Kim, Young Jun Ku, Dae Suk Kim, Chunseok Jeong, Tae Sik Yun, Hongjung Kim, Ho Sung Cho, Sangmuk Oh, et al

  18. [26]

    Edgar Lemaire, Loïc Cordone, Andrea Castagnetti, Pierre-Emmanuel Novac, Jonathan Courtois, and Benoît Miramond. 2022. An analytical estimation of spiking neural networks energy efficiency. In International Conference on Neural Information Processing. Springer, 574–587

    2022

  19. [27]

    Faqiang Liu, Hao Zheng, Songchen Ma, Weihao Zhang, Xue Liu, Yansong Chua, Luping Shi, and Rong Zhao. 2024. Advancing brain-inspired computing with hybrid neural networks. National Science Review 11, 5 (2024), nwae066

    2024

  20. [28]

    Anni Lu, Xiaochen Peng, Yandong Luo, Shanshi Huang, and Shimeng Yu. 2021. A runtime reconfigurable design of compute-in-memory based hardware acceler- ator. In 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE) . IEEE, 932–937

    2021

  21. [29]

    Matt Mahoney. 2011. Large text compression benchmark

    2011

  22. [30]

    Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Naka- mura, et al. 2014. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 3...

    2014

  23. [31]

    Merolla, John V

    Paul A. Merolla, John V. Arthur, Rodrigo Alvarez-Icaza, Andrew S. Cassidy, Jun Sawada, Filipp Akopyan, Bryan L. Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, Bernard Brezzo, Ivan Vo, Steven K. Esser, Rathinakumar Appuswamy, Brian Taba, Arnon Amir, Myron D. Flickner, William ...

    2014

  24. [32]

    Surya Narayanan, Karl Taht, Rajeev Balasubramonian, Edouard Giacomin, and Pierre-Emmanuel Gaillardon. 2020. SpinalFlow: An architecture and dataflow tailored for spiking neural networks. In 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA) . IEE...

    2020

  25. [33]

    Eustace Painkras, Luis A Plana, Jim Garside, Steve Temple, Francesco Galluppi, Cameron Patterson, David R Lester, Andrew D Brown, and Steve B Furber. 2013. SpiNNaker: A 1-W 18-core system-on-chip for massively-parallel neural network simulation. IEEE Journal of Solid-State Cir...

    2013

  26. [34]

    Dylan Patel and Gerald Wong. 2023. GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE. https://www.semianalysis.com/p/gpt- 4-architecture-infrastructure. Accessed: April 15, 2025

    2023

  27. [35]

    Jens E Pedersen, Steven Abreu, Matthias Jobst, Gregor Lenz, Vittorio Fra, Fe- lix Christian Bauer, Dylan Richard Muir, Peng Zhou, Bernhard Vogginger, Kade Heckel, et al. 2024. Neuromorphic intermediate representation: a unified instruc- tion set for interoperable brain-inspire...

    2024

  28. [36]

    Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. 2023. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048 (2023)

    2023 arXiv

  29. [37]

    Michael Pfeiffer and Thomas Pfeil. 2018. Deep learning with spiking neurons: opportunities and challenges. Frontiers in neuroscience 12 (2018), 409662

    2018

  30. [38]

    Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. 2019. Towards spike- based machine intelligence with neuromorphic computing. Nature 575, 7784 (2019), 607–617

    2019

  31. [39]

    James Seekings, Peyton Chandarana, Mahsa Ardakani, MohammadReza Moham- madi, and Ramtin Zand. 2024. Towards Efficient Deployment of Hybrid SNNs on Neuromorphic and Edge AI Hardware. In 2024 International Conference on Neuro- morphic Systems (ICONS). 71–77. https://doi.org/10.1...

    2024

  32. [40]

    Sonali Singh, Anup Sarma, Nicholas Jao, Ashutosh Pattnaik, Sen Lu, Kezhou Yang, Abhronil Sengupta, Vijaykrishnan Narayanan, and Chita R Das. 2020. NEBULA: A neuromorphic spin-based ultra-low power architecture for SNNs and ANNs. In 2020 ACM/IEEE 47th Annual International Sympo...

    2020

  33. [41]

    Matei-Ioan Stan and Oliver Rhodes. 2024. Learning long sequences in spiking neural networks. Scientific Reports 14, 1 (2024), 21957

    2024

  34. [42]

    Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. 2023. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 6555– 6565

    2023

  35. [43]

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. 2017. Efficient processing of deep neural networks: A tutorial and survey. Proc. IEEE 105, 12 (2017), 2295–2329

    2017

  36. [44]

    Mingxing Tan and Quoc Le. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning . PMLR, 6105–6114

    2019

  37. [45]

    Mingxing Tan and Quoc Le. 2021. Efficientnetv2: Smaller models and faster training. In International conference on machine learning . PMLR, 10096–10106

    2021

  38. [46]

    Aaron Voelker, Ivana Kajić, and Chris Eliasmith. 2019. Legendre memory units: Continuous-time representation in recurrent neural networks.Advances in neural 12 Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding , , information processing systems 32 (2019)

    2019

  39. [47]

    Aaron R Voelker, Daniel Rasmussen, and Chris Eliasmith. 2020. A spike in performance: Training hybrid-spiking neural networks with quantized activation functions. arXiv preprint arXiv:2002.03553 (2020)

    2020 arXiv

  40. [48]

    Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. 2023. Spike-driven transformer. Advances in neural information processing systems 36 (2023), 64043–64058

    2023

  41. [49]

    Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. 2023. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence (2023)

    2023

  42. [50]

    Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. 2024. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020 (2024)

    2024 arXiv

  43. [51]

    Rui-Jie Zhu, Yu Zhang, Ethan Sifferman, Tyler Sheaves, Yiqiao Wang, Dustin Richmond, Peng Zhou, and Jason K Eshraghian. 2024. Scalable matmul-free language modeling. arXiv preprint arXiv:2406.02528 (2024)

    2024 arXiv

  44. [52]

    Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. 2024. Spikegpt: Gen- erative pre-trained language model with spiking neural networks. Transactions on Machine Learning Research (2024). 13

    2024

  45. [2007]

    IEEE Micro (2007)

    A 5-Ghz Mesh Interconnect for a Teraflops PRocessor. IEEE Micro (2007)

    2007

  46. [2016]

    In 2016 International SoC Design Conference (ISOCC)

    High bandwidth memory (HBM) with TSV technique. In 2016 International SoC Design Conference (ISOCC) . IEEE, 181–182

    2016

  47. [2023]

    IEEE Transactions on Emerging Topics in Computational Intelligence 7, 3 (2023), 731–741

    Are SNNs Really More Energy-Efficient Than ANNs? an In-Depth Hardware-Aware Study. IEEE Transactions on Emerging Topics in Computational Intelligence 7, 3 (2023), 731–741. https://doi.org/10.1109/TETCI.2022.3214509

    2023

Pith tools

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