Pith. sign in

REVIEW 2 major objections 4 minor 24 references

A SIMT-style sparse traversal with per-PE address generation delivers the strongest energy and speed scaling with weight sparsity in neuromorphic cores, at near-zero area cost, while lockstep variants keep throughput fixed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:15 UTC pith:ZRYWYJWB

load-bearing objection Genuinely controlled three-way RTL comparison of SIMD/Sparse-SIMD/SIMT neuromorphic cores; the sparsity-tax story holds up, and the paper deserves a serious referee—just don't over-read the exact break-even. the 2 major comments →

arxiv 2607.22790 v1 pith:ZRYWYJWB submitted 2026-07-24 cs.AR

The Sparsity Tax: Weight Sparsity Trade-offs in Event-Driven SIMD and SIMT Neuromorphic Cores

Mattias Westerink , Sameed Sohail , Berend-Jan van der Zwaag , Sabih Gerez , Amirreza Yousefzadeh This is my paper
classification cs.AR
keywords sparsity taxweight sparsitySIMDSIMTevent-drivenneuromorphicrun-length codingenergy efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces the 'sparsity tax' — the control, metadata, and memory overheads that appear when weight sparsity breaks lockstep SIMD execution — and quantifies it by comparing three accelerators built into one neuromorphic core: dense SIMD, bitmap-gated Sparse-SIMD, and SIMT with per-PE address generation and run-length-coded weights. The central result is that SRAM dominates core area, so giving each lane its own address generator costs almost nothing in area, yet it changes how sparsity pays off: SIMD and Sparse-SIMD keep near-constant throughput, Sparse-SIMD only saves energy by disabling lanes, and SIMT skips zero-weight updates entirely, so both energy and latency scale with sparsity. The paper concludes that above roughly 50% unstructured weight sparsity, SIMT-style traversal is the preferred architecture, with sublinear gains bounded by metadata reads, load imbalance, and sparsity-independent phases. A sympathetic reader would care because it turns an architectural debate about handling irregular sparsity into a concrete, context-dependent trade-off with a quantitative break-even.

Core claim

The paper's central claim is that the way a neuromorphic core handles weight sparsity hinges on whether it gates work or skips work. In lockstep SIMD, zero-weight updates still execute; in Sparse-SIMD, a bitmap disables individual lanes and their memory accesses, saving energy but not execution time and adding bitmap-fetch overhead; in SIMT, each processing element has its own address generation unit and the weights are stored in 4-bit run-length coding, so each lane jumps directly to the next non-zero weight, skipping both the computation and the memory traffic of zero updates. Across post-training pruning levels on a VGG-16/CIFAR-10 workload, the paper reports that total core area changes

What carries the argument

The 'sparsity tax' is the paper's unifying concept: the overheads in control, metadata, and memory activity that appear when weight sparsity makes updates irregular and breaks lockstep execution. The argument is carried by three concrete microarchitectures sharing one SRAM-dominated core: (1) dense SIMD with a shared address generation unit, (2) bitmap-gated Sparse-SIMD, where a per-lane clock-enable register derived from a stored bitmap disables PEs and SRAM rows, and (3) SIMT, where each PE has its own AGU and weights are stored as non-zero values interleaved with 4-bit run-length codes. The RLC format is the key mechanism: it both compresses the weight matrix at high sparsity and gives ea

Load-bearing premise

The quantitative head-to-head depends on modeling each layer by one average activation event count, without simulating the temporal and spatial correlation of real event streams; if real traffic is bursty, the relative cycle and energy curves for SIMT could shift.

What would settle it

Run the three cores on real recorded event streams (e.g., from a DVS camera) at matched weight sparsity levels, instead of average per-layer event counts; if the SIMT latency or energy advantage over Sparse-SIMD at >50% sparsity reverses or disappears under bursty or clustered activity, the central claim would be refuted. A cheaper check: measure a single layer at 50% sparsity twice — once with non-zeros spread evenly, once clustered into a few PEs — and see if SIMT's speedup flips sign.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Above roughly 50% unstructured weight sparsity, SIMT-style sparse traversal dominates both energy and latency, while area stays flat because SRAM dominates.
  • Sparse-SIMD is a viable middle ground only when the storage format cannot be changed and only energy (not latency) matters, since throughput remains constant.
  • The qualitative 'gate vs skip' distinction should generalize across precisions and encodings, though the exact break-even point shifts with weight width, metadata width, and sparsity structure.
  • Because core area is dominated by memory, adding per-PE address generation is effectively free in this regime, lowering the barrier to adopting SIMT-style designs.
  • Designers can mix variants per layer: dense or lightly sparse layers run on the cheap SIMD stream, while high-sparsity layers are handed to the SIMT datapath.

Where Pith is reading between the lines

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

  • The paper evaluates each layer using an input-event count from the average activation sparsity; if real event streams are bursty or spatially correlated, per-PE load balance would change, and the SIMT advantage at moderate sparsity could shrink or grow — this is the most direct testable extension.
  • The 4-bit RLC encoding is one point in a sparse-format design space; block-structured or grouped sparsity would reduce metadata reads and lane divergence, likely pushing the SIMT break-even point below 50%.
  • The same gate-vs-skip dichotomy applies beyond neuromorphic cores — any sparse accelerator that must choose between gating lanes and letting lanes self-address will face the same sparsity tax trade-off.
  • The near-zero area delta suggests that adding per-lane address generation is cheap in memory-dominated systems, so future cores could include both paths and dispatch per layer based on measured sparsity.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper compares three accelerator variants integrated into a single NEORV32-based neuromorphic core: a lockstep SIMD accelerator, a bitmap-gated Sparse-SIMD variant, and a SIMT-style coprocessor with per-PE address generation and 4-bit run-length-coded sparse weights. The evaluation uses RTL simulation, synthesis in GF22FDX+, and activity-driven energy estimation on a pruned VGG-16/CIFAR-10 workload. The main claims are that total core area is essentially unchanged across variants because SRAM dominates, that SIMD and Sparse-SIMD throughput is sparsity-independent, and that SIMT provides the strongest energy scaling and substantial speedups at high weight sparsity, albeit sublinear because of metadata reads, load imbalance, and sparsity-independent phases. The paper concludes that above roughly 50% unstructured weight sparsity, SIMT-style sparse traversal becomes preferable, while Sparse-SIMD remains useful only when the storage format cannot be changed.

Significance. If the findings hold, the paper gives hardware designers a concrete, silicon-backed criterion for choosing between lockstep gating and per-PE sparse traversal in event-driven cores. The study's strengths are the apples-to-apples integration of the three variants in one core, the RTL-to-gates flow with a stated 15% energy-accuracy bound, and the public release of the hardware code. The architectural trends are mechanically consistent with the designs: gating cannot reduce workload, whereas skipping zeros in a compressed representation can. The reported area insensitivity is well supported by the SRAM-dominated area table. The main fragility is the workload abstraction, which affects the quantitative break-even point rather than the qualitative direction of the results.

major comments (2)
  1. [III-E and Fig. 4 caption] The workload model reduces each layer to a single scalar: 'an input-event count derived from the average activation sparsity across CIFAR-10 images.' This ignores which input neurons are active and how they are correlated in space and time. The SIMT cycle count is set by the slowest PE when PEs reconverge (Section II-C), so per-PE load imbalance and effective RLC skip lengths depend on the distribution of active input neurons and their corresponding weight rows, not merely on the total event count. Real CIFAR-10 activations are spatially correlated (edges, objects, backgrounds), and bursty temporal activity could change how events are batched. The claimed 50% break-even and the speedup magnitudes in Figs. 4 and 6 may shift under more realistic event distributions. I request a sensitivity analysis or a bound: e.g., simulate with actual event traces from a few CIFAR-10 images, or inject ad
  2. [IV and Figs. 4/6] The paper extrapolates a design rule ('When the weight sparsity exceeds 50%... its advantages become evident') from measurements on only two layers: one dense layer and one convolutional layer (block2 conv1). Table II shows that activation sparsity in VGG-16 ranges from 1.6% to 98%, so a single average event count per layer may not capture the relative weight of different layers in end-to-end inference. The 50% threshold is presented as a general guideline, but the evidence is layer-specific and workload-specific. I recommend reporting the per-layer break-even sparsity across all convolutional and dense layers, or at least a range, and providing end-to-end energy/latency estimates using actual per-layer event counts. Without this, the quantitative threshold is not supported beyond the two shown layers, even though the qualitative claim may hold.
minor comments (4)
  1. [III-D] The statement 'The overhead of the RLC encoding at 0% sparsity is 106.25%' is correct given the 4-bit RLC value per 4-bit weight plus 64 kb of indices, but it may confuse readers because the RLC overhead at 0% sparsity includes the cost of encoding zero runs. Consider clarifying that the 106.25% is relative to the dense 1024 kb representation.
  2. [Table IV] The row for Sparse-SIMD8 reads '7 2.1' without units/separators; it should be '7 pJ, 2.1 ns' to be consistent with the rest of the table. Also, the header 'Energy per neuron update (pJ)' for our accelerators appears to include only the accelerator, while Loihi/NeuronFlow may include different scopes; the cautionary note is appreciated but a per-column footnote would help.
  3. [III-C] The notation 'KµM² = 10³ µM²' is awkward (likely kµm²). Please use standard units (kµm²) to avoid confusion.
  4. [I] The term 'sparsity tax' is used descriptively but never defined as a formal cost metric. Since the paper is about quantifying it, consider defining it operationally (e.g., the additional cycles/energy per useful MAC relative to an oracle sparse execution) early in the introduction.

Circularity Check

0 steps flagged

No significant circularity: the central results are empirical measurements from independent RTL implementations, not derived from their inputs by construction.

full rationale

The paper's core quantitative claims—area parity, near-constant SIMD/Sparse-SIMD throughput, Sparse-SIMD's limited energy slope, and SIMT's stronger energy/cycle scaling at high sparsity—are reported outcomes of an RTL-to-gates flow (Cadence Genus, GF22FDX+), functional simulation (Cadence Xcelium), and activity-driven energy estimation (Cadence Joules) on three concrete accelerator implementations. There are no fitted parameters that are then renamed as predictions, and no equation defines a claimed result in terms of its own inputs. The workload model in Section III-E that uses 'an input-event count derived from the average activation sparsity across CIFAR-10 images' is a modeling simplification, and the skeptic's concern that spatial/temporal event structure could shift SIMT load imbalance is a legitimate generalization risk. However, that assumption is not circular: the measured cycle counts, energies, and areas are not algebraic consequences of the average count, and the paper does not use that count as a fitted parameter to force the observed trends. Self-citations (e.g., the SENECA loop-controller reference in Section II-A and prior event-driven neuromorphic works) provide context and architectural background, but they are not the load-bearing evidence for the central comparison, which stands on the independently described and publicly accessible RTL implementations. No self-citation chain, imported uniqueness theorem, or ansatz-by-citation constrains the conclusion. The paper also explicitly qualifies the quantitative break-even as context-specific, further indicating that its claims are not definitional. Therefore no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on a small number of disclosed workload and tool assumptions, not on fitted models or new physical entities. The fixed design choices (8 PEs, 4-bit RLC, average event counts) are the main carriers of context dependence; none of them are hidden.

free parameters (4)
  • Input event count per layer = average CIFAR-10 activation sparsity
    All latency and energy numbers are reported per this average event count (Section III-E); high variance in real event streams would change absolute and possibly relative results.
  • PE count = 8
    Fixed architectural design point (Table III); per-PE AGU overhead and load-balance effects scale with PE count.
  • RLC metadata width = 4 bits
    Chosen to match INT4 weights (Section II-C); wider indices or alternative sparse formats would shift the SIMT break-even point.
  • Sparsity levels = 0-95% post-training magnitude pruning
    Pruning method and thresholds chosen for the architectural study; the authors state advanced pruning would not change architectural trends (Section III-B).
axioms (4)
  • domain assumption Post-training magnitude pruning preserves architectural trends compared to retraining-aware pruning.
    Stated in Section III-B; all sparsity-level results assume this.
  • domain assumption Cadence Joules activity-driven energy estimates are accurate within 15% of signoff.
    Section III-A; all energy conclusions rely on this tool accuracy.
  • domain assumption Average activation sparsity over CIFAR-10 is a sufficient workload proxy for event-driven comparison.
    Section III-E; full event streams are not simulated.
  • domain assumption 4-bit RLC encoding is a representative sparse format for the architectural question.
    Section II-C; the authors acknowledge other formats would change metadata and address-generation behavior but claim the central question is unchanged.

pith-pipeline@v1.3.0-alltime-deepseek · 11915 in / 13643 out tokens · 109358 ms · 2026-08-01T05:15:29.640625+00:00 · methodology

0 comments
read the original abstract

Event-driven neuromorphic inference exploits activation sparsity by updating neuron state only on spikes. However, weight sparsity introduces irregular gather-style updates that undermine lockstep Single Instruction Multiple Data (SIMD) execution. We call the resulting overheads in control, metadata, and memory activity the sparsity tax. This paper quantifies that tax by comparing three closely related accelerators integrated into one neuromorphic core: (i) baseline lockstep SIMD, (ii) bitmap-gated Sparse-SIMD that selectively disables lanes without compressing weights, and (iii) a Single Instruction Multiple Threads (SIMT) style design with per-PE address generation and run-length coded sparse weights. Using an RTL-to-gates flow in GF22FDX+ and activity-driven energy estimation, we evaluate event-driven neural network inference across varying post-training pruning levels. Results show that total core area changes are insignificant because SRAM dominates area, while performance and energy strongly depend on how sparsity is handled: SIMD and Sparse-SIMD exhibit near-constant throughput, Sparse-SIMD achieves limited energy savings due to bitmap and dense-storage overheads, and SIMT provides the strongest energy scaling and substantial speedups at high sparsity, albeit with sublinear gains due to metadata reads, load imbalance, and sparsity-independent phases. The hardware code for the proposed architectures and experiments is publicly accessible for research purposes.

Figures

Figures reproduced from arXiv: 2607.22790 by Amirreza Yousefzadeh, Berend-Jan van der Zwaag, Mattias Westerink, Sabih Gerez, Sameed Sohail.

Figure 1
Figure 1. Figure 1: A) A layer of a neural network. When one input neuron activates [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the Processing Elements Processing Element: Processing elements operate in INT16 and INT4 formats, with a register file structured into 8×16- bit and 8×4-bit registers as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Memory Footprint of “VGG-16 dense layer” Weights for SIMD (Dense [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of Sparsity on Execution Time (in clock cycles) for SIMD, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of Sparsity on energy consumption for SIMD, Sparse-SIMD [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 4 linked inside Pith

  1. [1]

    Optimizing event-based neural networks on digital neuromorphic architecture: a comprehensive design space exploration,

    Y . Xu, K. Shidqi, G.-J. van Schaik, R. Bilgic, A. Dobrita, S. Wang, R. Meijer, P. Nembhani, C. Arjmand, P. Martinelloet al., “Optimizing event-based neural networks on digital neuromorphic architecture: a comprehensive design space exploration,”Frontiers in Neuroscience, vol. 18, p. 1335422, 2024

    2024

  2. [2]

    µbrain: An event-driven and fully synthesizable architecture for spiking neural networks,

    J. Stuijt, M. Sifalakis, A. Yousefzadeh, and F. Corradi, “µbrain: An event-driven and fully synthesizable architecture for spiking neural networks,”Frontiers in neuroscience, vol. 15, p. 664208, 2021

    2021

  3. [3]

    Neuron- flow: a neuromorphic processor architecture for live ai applications,

    O. Moreira, A. Yousefzadeh, F. Chersi, G. Cinserin, R.-J. Zwartenkot, A. Kapoor, P. Qiao, P. Kievits, M. Khoei, L. Rouillardet al., “Neuron- flow: a neuromorphic processor architecture for live ai applications,” in 2020 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2020, pp. 840–845

    2020

  4. [4]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jainet al., “Loihi: A neuromorphic manycore processor with on-chip learning,”Ieee Micro, vol. 38, no. 1, pp. 82–99, 2018

    2018

  5. [5]

    Hybrid neural network, an efficient low- power digital hardware implementation of event-based artificial neural network,

    A. Yousefzadeh, G. Orchard, E. Stromatias, T. Serrano-Gotarredona, and B. Linares-Barranco, “Hybrid neural network, an efficient low- power digital hardware implementation of event-based artificial neural network,” in2018 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2018, pp. 1–5

    2018

  6. [6]

    Performance comparison of time- step-driven versus event-driven neural state update approaches in spin- naker,

    A. Yousefzadeh, M. Soto, T. Serrano-Gotarredona, F. Galluppi, L. Plana, S. Furber, and B. Linares-Barranco, “Performance comparison of time- step-driven versus event-driven neural state update approaches in spin- naker,” in2018 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2018, pp. 1–4

    2018

  7. [7]

    Efficient neuromorphic signal processing with loihi 2,

    G. Orchard, E. P. Frady, D. B. D. Rubin, S. Sanborn, S. B. Shrestha, F. T. Sommer, and M. Davies, “Efficient neuromorphic signal processing with loihi 2,” in2021 IEEE Workshop on Signal Processing Systems (SiPS). IEEE, 2021, pp. 254–259

    2021

  8. [8]

    The state of sparsity in deep neural networks,

    T. Gale, E. Elsen, and S. Hooker, “The state of sparsity in deep neural networks,”arXiv preprint arXiv:1902.09574, 2019

    Pith/arXiv arXiv 1902

  9. [9]

    Eie: Efficient inference engine on compressed deep neural network,

    S. Han, X. Liu, H. Mao, J. Pu, A. Pedram, M. A. Horowitz, and W. J. Dally, “Eie: Efficient inference engine on compressed deep neural network,”ACM SIGARCH Computer Architecture News, vol. 44, no. 3, pp. 243–254, 2016

    2016

  10. [10]

    Cambricon-s: Addressing irregularity in sparse neural networks through a cooperative software/hardware approach,

    X. Zhou, Z. Du, Q. Guo, S. Liu, C. Liu, C. Wang, X. Zhou, L. Li, T. Chen, and Y . Chen, “Cambricon-s: Addressing irregularity in sparse neural networks through a cooperative software/hardware approach,” in 2018 51st Annual IEEE/ACM International Symposium on Microarchi- tecture (MICRO). IEEE, 2018, pp. 15–28

    2018

  11. [11]

    Scnn: An accelerator for compressed-sparse convolutional neural networks,

    A. Parashar, M. Rhu, A. Mukkara, A. Puglielli, R. Venkatesan, B. Khailany, J. Emer, S. W. Keckler, and W. J. Dally, “Scnn: An accelerator for compressed-sparse convolutional neural networks,”ACM SIGARCH computer architecture news, vol. 45, no. 2, pp. 27–40, 2017

    2017

  12. [12]

    Seneca: building a fully digital neuromorphic processor, design trade-offs and challenges,

    G. Tang, K. Vadivel, Y . Xu, R. Bilgic, K. Shidqi, P. Detterer, S. Traferro, M. Konijnenburg, M. Sifalakis, G.-J. van Schaiket al., “Seneca: building a fully digital neuromorphic processor, design trade-offs and challenges,” Frontiers in Neuroscience, vol. 17, p. 1187252, 2023

    2023

  13. [13]

    Spinnaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,

    H. A. Gonzalez, J. Huang, F. Kelber, K. K. Nazeer, T. Langer, C. Liu, M. Lohrmann, A. Rostami, M. Sch¨one, B. V oggingeret al., “Spinnaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,”arXiv preprint arXiv:2401.04491, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Spinalflow: An architecture and dataflow tailored for spiking neural networks,

    S. Narayanan, K. Taht, R. Balasubramonian, E. Giacomin, and P.- E. Gaillardon, “Spinalflow: An architecture and dataflow tailored for spiking neural networks,” in2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA). IEEE, 2020, pp. 349– 362

    2020

  15. [15]

    Vsa: Reconfigurable vector- wise spiking neural network accelerator,

    H.-H. Lien, C.-W. Hsu, and T.-S. Chang, “Vsa: Reconfigurable vector- wise spiking neural network accelerator,” in2021 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2021, pp. 1–5

    2021

  16. [16]

    Spare: Spiking neural network acceleration using rom-embedded rams as in-memory-computation prim- itives,

    A. Agrawal, A. Ankit, and K. Roy, “Spare: Spiking neural network acceleration using rom-embedded rams as in-memory-computation prim- itives,”IEEE Transactions on Computers, vol. 68, no. 8, pp. 1190–1200, 2018

    2018

  17. [17]

    Spiking transformer hardware accelerators in 3d integration,

    B. Xu, J. Hwang, P. Vanna-iampikul, S. K. Lim, and P. Li, “Spiking transformer hardware accelerators in 3d integration,” inProceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, 2024, pp. 1–9

    2024

  18. [18]

    Prune once for all: Sparse pre-trained language models,

    O. Zafrir, A. Larey, G. Boudoukh, H. Shen, and M. Wasserblat, “Prune once for all: Sparse pre-trained language models,”arXiv preprint arXiv:2111.05754, 2021

    Pith/arXiv arXiv 2021

  19. [19]

    Neorv32

    “Neorv32.” [Online]. Available: https://github.com/stnolting/neorv32

  20. [20]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  21. [21]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” 2009

    2009

  22. [22]

    A comprehensive review of network pruning based on pruning granularity and pruning time perspectives,

    K. Zhu, F. Hu, Y . Ding, W. Zhou, and R. Wang, “A comprehensive review of network pruning based on pruning granularity and pruning time perspectives,”Neurocomputing, p. 129382, 2025

    2025

  23. [23]

    Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip,

    F. Akopyan, J. Sawada, A. Cassidy, R. Alvarez-Icaza, J. Arthur, P. Merolla, N. Imam, Y . Nakamura, P. Datta, G.-J. Namet al., “Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip,”IEEE transactions on computer- aided design of integrated circuits and systems, vol. 34, no. 10, pp. 1537–1557, 2015

    2015

  24. [24]

    The spinnaker project,

    S. B. Furber, F. Galluppi, S. Temple, and L. A. Plana, “The spinnaker project,”Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014

    2014