Pith. sign in

REVIEW 4 major objections 6 minor 25 references

A Low-Power Sparse Convolution Accelerator with Idle-First-Task-Assignment for Edge Vision

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A 16nm sparse convolution accelerator combines bitmap compression with an idle-first task scheduler to keep processing elements busy, achieving 6.5x and 2.8x speedups over dense accelerators on sparse VGG16 and MobileNetV2 while consuming o

desk verdict Novel sparse-scheduling mechanism and sensible DWConv dataflow, but the headline speedups come from an unvalidated in-house simulator and the abstract's 'validated' claim is unsupported by any measured silicon data. read the letter →

arxiv 2607.26835 v2 pith:AA5LWA7Y submitted 2026-07-29 cs.AR

Jingyue Zhuge , Johannes Partzsch , Christian Mayr , Christian Mayr (1 , 3) ((1) Chair of Highly-Parallel VLSI-Systems , Neuro-Microelectronics , TUD Dresden University of Technology , (2) Center for Scalable Data Analytics
show 2 more authors
Artificial Intelligence (ScaDS.AI Dresden) (3) Centre for Tactile Internet with Human-in-the-loop (CeTI))
This is my paper · ORCID
classification cs.AR
keywords sparseconvolutionacceleratoredgeAIloadbalancingdynamicschedulingdepthwiseseparablebitmapcompressionlowpowerPEutilization
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 tries to show that a small, low-power sparse convolution accelerator for edge vision can overcome the two classic obstacles to sparse acceleration: memory/bandwidth overhead and processing-element (PE) load imbalance. It proposes a bitmap-based compression format to cut data movement, an Idle-First-Task-Assignment (IFTA) dynamic scheduler that hands the next task to whichever PE finishes first, and a dedicated dataflow for depthwise separable convolutions. On pruned ImageNet models, the fabricated 16nm chip reports 6.5x and 2.8x speedups over dense accelerators, about 1.5x and 1.7x over the SparTen sparse accelerator, and above 90% average PE utilization. If accurate, this would make high-resolution, low-power edge vision feasible in bandwidth- and power-constrained IoT settings.

What carries the argument

The load-bearing mechanism is Idle-First-Task-Assignment (IFTA), implemented with a FIFO in each PE row: weight vectors are broadcast to all row FIFOs, and a PE that completes its current kernel immediately pulls the next available kernel, while sparse weight vectors can be fanned out to multiple PEs at once. A per-row partial-sum buffer prevents write-back contention. For depthwise separable convolution, the accelerator swaps the input-activation and weight buffer roles so each PE row processes a full kernel-length weight vector against a sliding activation window, using the decoder's decode mode to expand sparse activations.

What would settle it

Run the same pruned VGG16 and MobileNetV2 workloads on real SparTen hardware (or its official simulator) and on the fabricated chip, using identical layer configurations and sparsity patterns; if the measured speedup over SparTen drops below the reported 1.5x/1.7x, the central claim is falsified. Alternatively, if the chip's own measured cycle counts deviate from the paper's simulator predictions by more than a small margin, the simulation-based speedups would not hold.

Watch

Extended reading notes

Core claim

The accelerator's central claim is that the combination of bitmap-based sparse compression, IFTA dynamic load balancing, and a DWConv-specialized dataflow lets a 64-MAC, 0.5mm² core sustain high multiplier utilization on sparse CNNs. IFTA assigns the next convolution kernel to a PE the moment it idles, rather than waiting for all PEs to finish, cutting idle time by roughly a quarter in the paper's example. With a dual-mode decoder that either compacts non-zero values (pass-through) or expands them into full vectors (decode), the design keeps cache bandwidth high and supports the sliding-window access pattern of depthwise convolutions. Measured on-chip power is 12–16 mW, and simulation agains

Load-bearing premise

The cycle-accurate simulator built for this paper faithfully reproduces the actual cycle-level performance of the dense TPU-like accelerator and of SparTen under the same configuration, even though it is not validated against published SparTen measurements.

Editorial extensions

If this is right

  • If IFTA works as reported, PE load imbalance—often cited as the main reason sparse accelerators underperform—can be largely neutralized with a simple FIFO-based scheduler, making sparse convolution viable on small edge-scale arrays.
  • The bitmap format's dual-mode decoder suggests a practical trade-off: pass-through mode for dense layers and decode mode for depthwise layers, keeping both bandwidth and computation efficient without needing multiple sparse formats.
  • High PE utilization at 12–16 mW on MobileNetV2-type networks would strengthen the case for on-device, patch-wise inference that preserves input resolution while staying within IoT power envelopes.
  • The DWConv-specific dataflow addresses a known gap in sparse accelerators and, if correct, should generalize to other lightweight architectures that rely on depthwise separable layers.

Reading between the lines

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

  • The IFTA principle is not tied to this chip's specific bitmap format; a similar idle-first queue could be layered onto other sparse accelerators or even dense systolic arrays with data-dependent latencies, potentially improving their utilization too.
  • The paper's speedup claims rest entirely on the accuracy of its in-house cycle-accurate simulator for the dense and SparTen baselines; a direct comparison against measured SparTen hardware would be the most decisive check.
  • The reported energy efficiency (5435–8152 frames/J on MobileNetV2) suggests that for very low frame-rate monitoring, the bottleneck may shift from compute to sensor and memory access, so future work could target input-sensor co-design.
  • The sparse weight distributions in VGG16 and MobileNetV2 differ markedly (68% vs 35% sparsity), and the accelerator's advantage over SparTen is larger on the sparser model, implying IFTA's benefit scales with irregularity—an easily testable prediction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a 16 nm sparse convolution accelerator for edge vision, combining bitmap-based compression, an Idle-First-Task-Assignment (IFTA) scheduling scheme to mitigate PE load imbalance, and a dedicated dataflow for depthwise separable convolutions. The authors report cycle-accurate simulations comparing their design against a TPU-like dense accelerator and SparTen on ImageNet-pruned VGG16 and MobileNetV2, claiming about 6.5x/2.8x speedups over the dense baseline, 1.5x/1.7x over SparTen, and average PE utilization above 90%. They also report a 0.5 mm^2 core area and 12-16 mW power, and state in the abstract that the chip was fabricated and validated in a 16 nm process.

Significance. If substantiated, the IFTA scheduling mechanism and the combined bitmap/DWConv dataflow would be a useful contribution to low-power sparse CNN inference. The architectural idea is plausible, and the comparison is framed against a strong sparse baseline (SparTen) rather than only against a dense array. The authors also provide a complete RTL design and tape-out statement, which is more than a purely algorithmic paper. However, the quantitative case rests entirely on an in-house cycle-accurate simulator, and the manuscript contains no measured silicon data to support the 'fabricated and validated' wording. The strengths are the clarity of the architecture description and the explicit focus on depthwise separable convolutions, but the evidence base is currently too thin for the strength of the claims.

major comments (4)
  1. [Section III, 'Evaluation and Results' (simulator description)] The headline results—6.5x/2.8x speedups over the dense accelerator, 1.5x/1.7x over SparTen, and >90% PE utilization—are produced entirely by the authors' in-house cycle-accurate simulator. The sentence 'The simulator can faithfully capture any idling caused by memory access and load imbalance' is an assertion, not a demonstration. No validation against published SparTen measurements or cycle counts, no sensitivity analysis, and no detailed baseline configuration are provided. If the simulator undercharges SparTen's load-balancing overhead or overcharges the dense baseline's idling, the main comparative claims are inflated. Please add simulator validation (e.g., reproduce a published SparTen workload/trace) and a sensitivity analysis over memory and FIFO parameters.
  2. [Abstract; Section III; Table II] The abstract states the accelerator is 'fabricated and validated in a 16 nm process,' but the body only says 'we also performed a complete RTL design and taped out using TSMC's 16nm technology' (Section III). There is no measured silicon data: no test setup, no die photo, no measured power/area, and no measurement methodology. Table II lists power (12–16 mW) and energy efficiency without saying whether these are measured or simulated. Please either include actual silicon measurements or rewrite the abstract, conclusion, and Table II to state clearly that power and area are post-layout/simulation estimates from a tape-out design.
  3. [Section III, 'All three accelerators use the same configuration'] Baseline fairness is not verifiable. Table I describes only the proposed accelerator; the dense TPU-like systolic array and SparTen configurations are not specified. SparTen's published memory hierarchy, PE organization, tile sizes, queue depths, and scheduling policy would need to be reproduced exactly for a cycle-accurate comparison. Please provide complete baseline parameters (including SRAM word width, bank count, FIFO depths, and schedule policy) and, if possible, the simulator source or a detailed cycle-level trace for a representative layer.
  4. [Section III-1, pruning setup] The networks are pruned to 68% (VGG16) and 35% (MobileNetV2) sparsity 'while maintaining accuracy,' but no accuracy numbers are given. Since the speedup comparison is meaningful only at comparable accuracy, report top-1/top-5 accuracy before and after pruning for both networks, as well as the distribution of sparsity across layers, especially for MobileNetV2's depthwise layers.
minor comments (6)
  1. [Introduction, first paragraph] Typo: 'ieffectively' should be 'effectively'.
  2. [Section II-1, paragraph after Fig. 2] The paragraph contains a run-on sentence ('When the sparsity exceeds 90%, the metadata of Bitmap already occupies more memory than the non-zero values, This means that...'). Also, clarify why Bitmap is selected even though Fig. 2 shows RLC/CSR/COO having smaller footprints at high sparsity; state the operating sparsity range of the evaluated networks.
  3. [Section II-2, Fig. 3] The claimed 'approximately 25%' reduction in execution time is not derived from the figure. State the assumed work distribution and the calculation behind this percentage.
  4. [Figure 5] Add axis labels, legends, and exact speedup values in the caption. Currently the reader cannot recover the 6.5x/2.8x/1.5x/1.7x numbers from the plot.
  5. [Table II] Clearly annotate which entries are measured, post-layout simulated, or RTL-simulated. Also specify the operating frequency used for the 12–16 mW power range, since the clock frequency is listed as 200–400 MHz.
  6. [Section II-3, depthwise convolution dataflow] Define K (kernel height/width) and stride explicitly before saying 'very small vector length' and before describing the sliding-window offset behavior.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: speedups are computed against externally specified baselines (TPU-like dense array and SparTen), not derived from fitted parameters or from self-citations.

full rationale

The paper's central speedup claims (6.5x/2.8x over a dense accelerator, 1.5x/1.7x over SparTen) are produced by a cycle-accurate simulator built by the authors. This raises legitimate concerns about baseline fidelity and validation, but those are correctness/empirical-support issues, not circularity. The paper does not fit any parameter to the target results and then rename that fit as a prediction; no equation defines a predicted quantity as equal to an input by construction. The comparison baselines (TPU-like systolic array [22] and SparTen [6]) are external published architectures, and the paper states that 'All three accelerators use the same configuration,' a claim that could be inaccurate but is not self-referential. The tile_co = 4*NPE and bitmap length choices are design decisions to favor the proposed method, not fitted values that mathematically force the reported speedups. There are no self-citations in the reference list, and no load-bearing argument reduces to prior work by the same authors. The absence of measured silicon data and lack of simulator validation against published SparTen/TPU measurements are serious evidentiary weaknesses, but under the circularity framework they do not constitute a derivation that reduces to its own inputs. Therefore no circular step is identified.

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

The central speedup figure rests on several hand-picked hardware parameters (tile_co multiplier, FIFO depth, bitmap length) and on the fidelity of the in-house simulator for both the proposed design and the baselines. No free parameters are fitted to produce the central result in a mathematical sense, but the configuration choices and sparse workloads are not swept, so the claimed numbers should be read as point estimates under those choices.

free parameters (5)
  • tile_co multiplier = 4
    Section II-3: 'we set the number of weight vectors to be four times the number of PEs in each row' to maximize the advantages of IFTA; chosen by hand, not derived.
  • FIFO depth = equal to number of PEs per row (8)
    Section II-2: 'we configure its depth to be consistent with the number of PEs in each row'; a design choice that determines how tasks buffer.
  • Bitmap length per PE = 32
    Table I lists bitmap length per PE as 32; this controls metadata overhead and non-zero matching efficiency, with no derivation shown.
  • Model sparsity after pruning = VGG16 68%, MobileNetV2 35%
    Section III: sparsity levels come from applying the pruning method of [23] and fine-tuning on ImageNet; the speedup numbers depend directly on these chosen sparsity levels.
  • Operating clock frequency = 200-400 MHz
    Table II gives a frequency range; power and energy-efficiency numbers depend on the operating point, and no measurement setup or voltage domain is specified.
assumptions (4)
  • domain assumption The cycle-accurate simulator faithfully models the dense TPU-like accelerator, SparTen, and the proposed accelerator under the same configuration, including memory-access idling and load-imbalance idling.
    Section III states 'The simulator can faithfully capture any idling caused by memory access and load imbalance.' This premise is load-bearing for all speedup claims and is not independently validated in the paper.
  • domain assumption Bitmap format has the least metadata overhead over the sparsity ranges actually found in CNNs, and sparsity greater than 99% is exceptionally rare in CNNs.
    Section II-1 justifies the format choice using Fig. 2 and the statement that 'the boundary case of sparsity greater than 99% is exceptionally rare in CNNs', without a cited survey of CNN sparsity distributions.
  • domain assumption The FIFO broadcast and row PSUM buffers prevent simultaneous weight-access conflicts and PSUM write-back from becoming bottlenecks.
    Section II-2 describes the data distribution strategy and asserts it prevents PSUM write-back from becoming a bottleneck; no proof or measured verification is given.
  • domain assumption The DWConv dataflow maps all MobileNetV2 depthwise layers with negligible overhead.
    Section II-3 describes the swapped-buffer DWConv dataflow, but the evaluation reports only whole-network speedups and does not provide a per-layer sensitivity analysis for depthwise layers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Low-Power Sparse Convolution Accelerator with Idle-First-Task-Assignment for Edge Vision." pith.science (2026). https://pith.science/paper/AA5LWA7Y

@misc{pith2026260726835,
  author       = {Pith},
  title        = {Pith review of: A Low-Power Sparse Convolution Accelerator with Idle-First-Task-Assignment for Edge Vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AA5LWA7Y}},
  note         = {Machine review of arXiv:2607.26835}
}
abstract

In recent years, edge-vision monitoring systems for applications such as smart animal husbandry have faced strict tripartite constraints: maintaining input resolution under extremely limited transmission bandwidth and strict power budgets. Conventional dense convolutional neural networks (CNNs) cannot satisfy the resource limits of such constrained IoT nodes. To address this challenge, this paper presents a low-power sparse convolution accelerator for edge devices, fabricated and validated in a 16 nm process. First, the accelerator adopts a bitmap-based format for compression in both data transmission and computation, effectively reducing memory and bandwidth overhead. Second, to mitigate load imbalance in sparse computation, an Idle-First-Task-Assignment (IFTA) dynamic scheduling strategy is proposed, significantly reducing processing-element (PE) idle time and improving multiplier utilization. In addition, a dedicated dataflow is designed to support and accelerate depthwise separable convolution (DWConv), which is widely used in lightweight networks. Experimental results show that the chip occupies only 0.5~mm$^2$ core area and consumes as little as 12--16~mW. On ImageNet, for sparse VGG16 and MobileNetV2, the proposed accelerator achieves 6.5$\times$ and 2.8$\times$ speedups, respectively, over traditional dense accelerators, and also delivers significant performance gains over the existing sparse accelerator.

Figures

Figures reproduced from arXiv: 2607.26835 by the authors.

Figure 1
Figure 1. Architecture overview. (Cache), a dual-way bitmap decoder (Decoder), multi-level buffers for input activations, weights, and partial sums, and a two-dimensional processing unit array composed of multiple PE rows (PE ROW). Each PE contains a MAC unit, as well as buffers for input activations and weights. Each PE row contains multiple PEs, and each PE row has a FIFO and a row partial sum buffer [PITH_FULL_IMAGE:figur… view at source ↗
Figure 3
Figure 3. Load balancing comparison between (a)load imbalance and (b)our [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Data distribution To implement our IFTA method, we introduce a FIFO in each PE row. The weight buffer broadcasts data simultaneously to the FIFOs of all PE rows. This avoids simultaneous access to different weight data in the weight buffer by each row. The FIFO also serves as a buffer to balance the task processing speed among the PE rows, and we configure its depth to be consistent with the number of PEs in each ro… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Performance comparison of our accelerator, SparTen, and a dense [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

  1. [1]

    Memory-efficient patch-based inference for tiny deep learning,

    J. Lin, W.-M. Chen, H. Cai, C. Gan, and S. Han, “Memory-efficient patch-based inference for tiny deep learning,”Advances in Neural Information Processing Systems, vol. 34, pp. 2346–2358, 2021

    2021

  2. [2]

    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. [Online]. Available: http://arxiv.org/abs/1602.01528

    arXiv

  3. [3]

    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

  4. [4]

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

    Y .-H. Chen, T. Krishna, J. S. Emer, and V . Sze, “Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks,” vol. 52, no. 1, pp. 127–138. [Online]. Available: https://ieeexplore.ieee.org/document/7738524

    arXiv

  5. [5]

    Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices,

    Y .-H. Chen, T.-J. Yang, J. Emer, and V . Sze, “Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices,”IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 9, no. 2, pp. 292–308, 2019

    2019

  6. [6]

    Sparten: A sparse tensor accelerator for convolutional neural networks,

    A. Gondimalla, N. Chesnut, M. Thottethodi, and T. Vijaykumar, “Sparten: A sparse tensor accelerator for convolutional neural networks,” inProceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture, 2019, pp. 151–165

    2019

  7. [7]

    Procrustes: A Dataflow and Accelerator for Sparse Deep Neural Net- work Training,

    D. Yang, A. Ghasemazar, X. Ren, M. Golub, G. Lemieux, and M. Lis, “Procrustes: A Dataflow and Accelerator for Sparse Deep Neural Net- work Training,” Sep. 2020

    2020

  8. [8]

    Snap: An efficient sparse neural acceleration processor for unstructured sparse deep neural network inference,

    J.-F. Zhang, C.-E. Lee, C. Liu, Y . S. Shao, S. W. Keckler, and Z. Zhang, “Snap: An efficient sparse neural acceleration processor for unstructured sparse deep neural network inference,”IEEE Journal of Solid-State Circuits, vol. 56, no. 2, pp. 636–647, 2020

    2020

Show all 25 references
  1. [9]

    GoSPA: An Energy-efficient High-performance Globally Optimized SParse Convolutional Neural Network Accelerator,

    C. Deng, Y . Sui, S. Liao, X. Qian, and B. Yuan, “GoSPA: An Energy-efficient High-performance Globally Optimized SParse Convolutional Neural Network Accelerator,” in2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA), pp. 1110–1123. [Online]. Avai...

  2. [10]

    Paca: A pattern pruning algorithm and channel- fused high pe utilization accelerator for cnns,

    J. Wang, S. Yu, Z. Yuan, J. Yue, Z. Yuan, R. Liu, Y . Wang, H. Yang, X. Li, and Y . Liu, “Paca: A pattern pruning algorithm and channel- fused high pe utilization accelerator for cnns,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. ...

    2022

  3. [11]

    An optimized dataflow based accelerator for sparse convolutional neural networks,

    X. Ding, G. Su, and J. Zhang, “An optimized dataflow based accelerator for sparse convolutional neural networks,” in2023 IEEE 15th Interna- tional Conference on ASIC (ASICON). IEEE, 2023, pp. 1–4

    2023

  4. [12]

    An efficient sparse neural network accelerator for low-cost edge systems,

    K. Choi and G. E. Sobelman, “An efficient sparse neural network accelerator for low-cost edge systems,” in2023 Congress in Computer Science, Computer Engineering, & Applied Computing (CSCE). IEEE, 2023, pp. 887–892

    2023

  5. [13]

    An efficient cnn accelerator achieving high pe utilization using a dense- /sparse-aware redundancy reduction method and data–index decoupling workflow,

    Y . Meng, C. Yang, S. Xiang, J. Wang, K. Mei, and L. Geng, “An efficient cnn accelerator achieving high pe utilization using a dense- /sparse-aware redundancy reduction method and data–index decoupling workflow,”IEEE Transactions on Very Large Scale Integration (VLSI) Systems,...

    2023

  6. [14]

    ISOSceles: Accelerating Sparse CNNs through Inter-Layer Pipelining,

    Y . Yang, J. S. Emer, and D. Sanchez, “ISOSceles: Accelerating Sparse CNNs through Inter-Layer Pipelining,” in2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pp. 598–610. [Online]. Available: https://ieeexplore.ieee.org/document/10071080

  7. [15]

    Eyelet: A cross-mesh noc-based fine-grained sparse cnn accelerator for spatio-temporal parallel computing optimization,

    B. Yao, L. Liu, Y . Peng, X. Peng, R. Xu, H. Liu, and H. Zhang, “Eyelet: A cross-mesh noc-based fine-grained sparse cnn accelerator for spatio-temporal parallel computing optimization,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 72, no. 4, pp. 1634–1647, 2024

    2024

  8. [16]

    Edge-side fine-grained sparse cnn accelerator with efficient dynamic pruning scheme,

    B. Wu, T. Yu, K. Chen, and W. Liu, “Edge-side fine-grained sparse cnn accelerator with efficient dynamic pruning scheme,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 71, no. 3, pp. 1285– 1298, 2024

    2024

  9. [17]

    An efficient sparse cnn inference accelerator with balanced intra-and inter-pe workload,

    J. Guo, T. Xu, Z. Wu, and H. Xiao, “An efficient sparse cnn inference accelerator with balanced intra-and inter-pe workload,”IEEE Transac- tions on Very Large Scale Integration (VLSI) Systems, vol. 33, no. 5, pp. 1278–1291, 2024

    2024

  10. [18]

    An fpga-based approach for compressing and accelerating depthwise separable convolution,

    R. Yang, Z. Chen, L. Hu, X. Cui, and Y . Guo, “An fpga-based approach for compressing and accelerating depthwise separable convolution,” IEEE Signal Processing Letters, vol. 31, pp. 2590–2594, 2024

    2024

  11. [19]

    An efficient fpga-based depth- wise separable convolutional neural network accelerator with hardware pruning,

    Z. Liu, Q. Liu, S. Yan, and R. C. Cheung, “An efficient fpga-based depth- wise separable convolutional neural network accelerator with hardware pruning,”ACM Transactions on Reconfigurable Technology and Systems, vol. 17, no. 1, pp. 1–20, 2024

    2024

  12. [20]

    Gupa: Group-wise uniform pruning accelerator for depthwise separable convo- lution,

    Y . Chen, M. Wabnitz, J. Lou, C. Lanius, and T. Gemmeke, “Gupa: Group-wise uniform pruning accelerator for depthwise separable convo- lution,” in2025 IEEE Symposium on Low-Power and High-Speed Chips and Systems (COOL CHIPS). IEEE, 2025, pp. 1–6

    2025

  13. [21]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520

    2018

  14. [22]

    In-datacenter performance analysis of a tensor processing unit,

    N. P. Jouppi, C. Young, N. Patil, D. Patterson, G. Agrawal, R. Bajwa, S. Bates, S. Bhatia, N. Boden, A. Borcherset al., “In-datacenter performance analysis of a tensor processing unit,” inProceedings of the 44th annual international symposium on computer architecture, 2017, pp. 1–12

    2017

  15. [23]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,

    S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,”arXiv preprint arXiv:1510.00149, 2015

    2015 arXiv

  16. [24]

    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

    2014 arXiv

  17. [25]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009, pp. 248–255

    2009

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.