VUSA: Virtually Upscaled Systolic Array Architecture to Exploit Unstructured Sparsity in AI Acceleration

Alberto Garcia-Ortiz; Lennart Bamberg; Shereef Helal

REVIEW 4 major objections 4 minor 17 references

VUSA: Virtually Upscaled Systolic Array Architecture to Exploit Unstructured Sparsity in AI Acceleration

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read VUSA virtually grows a systolic array when weights are sparse, performing larger matrix multiplications with the same number of physical MAC units.

desk verdict A genuinely new virtual-growth trick for sparse systolic arrays, undermined by inconsistent headline numbers but worth a serious referee. read the letter →

arxiv 2506.01166 v1 pith:L27GE3N5 submitted 2025-06-01 cs.AR cs.AIcs.LG

Shereef Helal , Alberto Garcia-Ortiz , Lennart Bamberg This is my paper

classification cs.ARcs.AIcs.LG

keywords systolicarrayunstructuredsparsityweight-stationaryDNNacceleratorvirtualupscalingedgeAIMACunitpruning

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 that a systolic array can be built with far fewer multiply-accumulate (MAC) units than its apparent size, and use unstructured weight sparsity to 'virtually grow' to a larger array on the fly. In a weight-stationary array, zero weights need no computation, so the authors separate the data-flow registers (SPEs) from the arithmetic units and let each MAC connect to a small window of positions in its row. When every row has at most A non-zero weights, the same physical MACs perform the work of an N-by-M array, roughly doubling throughput at high sparsity. Because a dense fallback window always exists, the architecture still runs any DNN, even with no sparsity. The authors synthesize the design in a commercial 16-nm technology and report 37% area and 68% power savings at the same peak performance as a standard 3-by-6 array.

What carries the argument

The key mechanism is the decomposition of each PE into an SPE (pipeline registers only) and a MAC unit that can be dynamically connected to one of a small range of SPEs in the same row. Each row of the VUSA contains M SPEs and A MACs, and MAC j may connect only to SPE j through j+M-A; this constrained connectivity keeps wiring small while the authors claim it covers all possible sparse-row patterns with up to A non-zeros. The virtual-growth probability, $P_{\text{sparse}\rightarrow\text{standard}} = \left(\sum_{i=0}^{A} \binom{M}{i} P_1^i (1-P_1)^{M-i}\right)^{N}$, quantifies how often the array behaves as a full N-by-M array, and the dense N-by-A window is the guaranteed fallback that preserves general applicability.

What would settle it

For a fixed M and A, enumerate every row pattern of M bits with at most A ones and test whether each can be assigned to MAC units j=0..A-1 with the non-zero at a position within [j, j+M-A]. Any row pattern that cannot be mapped disproves the assumption and would lower the virtual-growth probabilities and the reported savings.

Watch

Extended reading notes

Core claim

The central claim is that decomposing each processing element into a data-flow pipeline stage (SPE) and an optional MAC unit makes a systolic array's effective dimensions a function of the current weight sparsity. A VUSA with parameters (N, M, A) has N rows of M SPEs but only A MAC units per row; each MAC can be shifted to any of M-A+1 adjacent SPE positions. If a row contains A or fewer non-zero weights, those weights are loaded into the A MACs and the row computes as if it were M columns wide, so an N-by-M matrix product is executed by only N-by-A physical MACs. The authors derive the probability of such virtual growth from a binomial model of non-zero weights, and on two pruned DNNs (ResNet-18 at 85% sparsity and MobileNetV1 at 75%) they report the array acting as a full 3-by-6 array for the large majority of the load, yielding 18-27% better performance per area and 45-56% better performance per power than a dense 3-by-6 array.

Load-bearing premise

The paper's efficiency gains rest on the assumption that any row with at most A non-zero weights among M positions can always be routed to the A available MAC units using the limited shifter connections; this is asserted but not proved.

Editorial extensions

If this is right

The same silicon area can deliver the throughput of a larger dense array whenever the weight rows are sparse enough, so pruned edge models run faster on a smaller accelerator.
A single VUSA can serve both sparse and dense DNNs without reconfiguring the programming model, making it a candidate for general-purpose edge AI.
At equal peak performance, a VUSA-based accelerator is smaller and consumes less power, relaxing area and thermal constraints for edge deployment.
The closed-form probability formula lets designers predict expected speedup from a model's sparsity rate before synthesis, supporting architecture-level design-space exploration.

Reading between the lines

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

The same SPE/MAC decomposition could be applied vertically as well as horizontally, potentially enabling two-dimensional virtual growth and amplifying gains for matrices sparse in both dimensions.
The constrained shifter connectivity may not cover every sparse row pattern; if counterexamples exist, the gap would show up as reduced gains for adversarial pruning masks that cluster non-zeros in certain positions, so a VUSA's worst-case performance is distribution-dependent in a way the binomial model does not capture.
Because the paper's gains grow with sparsity, the architecture would pair naturally with aggressive magnitude pruning; a testable extension is to train a DNN with a row-wise sparsity budget of exactly A non-zeros per row, guaranteeing full virtual growth for every layer.
The area and power overhead of the shifter network grows with M-A, so there is likely an optimal virtual-growth factor per technology node; sweeping M/A in synthesis would locate it.

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.

Desk Editor's Note

read the letter

Read VUSA. The central idea is new as far as the cited literature goes: split the PE into a data-flow-only SPE and an optional MAC, use fewer MACs than SPE columns, and let each MAC select among a small window of adjacent SPEs. When every row of the weight matrix has at most A nonzeros, the array runs as an N×M systolic array with only N×A physical MACs. That's a clean, practical trick for unstructured weight sparsity, and I don't see it in SCNN/EIE/Cnvlutin or the structured-sparsity work they cite. The binomial growth analysis (Eq. 1-4) is simple but appropriate, and the two SparseZoo benchmarks with load-split statistics are a reasonable first evaluation. The 16-nm synthesis numbers give concrete area/power data.

The soft spots are real but not fatal. First, the headline numbers are a mess. The abstract claims 37% area and 68% power savings; Table I shows the standard 3×6 is 37% larger in area and 68% larger in power, meaning VUSA saves 27% area and 40% power relative to that baseline. The conclusion then says 55% area and 25% power efficiency gains, which matches neither the abstract nor the tables (Tables II and III say 18–27% area efficiency and 45–56% power efficiency over the 3×6 baseline). That needs a consistent set of numbers and a clear statement of the comparison base. Second, the routing claim in Section III-C—that every set of at most A nonzeros among M positions can be assigned to MACs with each MAC j restricted to SPEs j..j+M-A—is asserted without proof. I tried to break it and couldn't; a right-to-left greedy assignment appears to work, because the intervals have the right covering property. So I think it's true, but the paper should state and prove it, and ideally sketch the mux structure. Third, the 'same peak performance' phrasing is misleading: VUSA's physical peak is N×A; it matches N×M only when sparsity cooperates. Fourth, they compare only against dense systolic arrays, not prior sparse accelerators, and the evaluated geometry (3×6, 9 MACs) is tiny.

Overall the core idea holds up. The paper is rough around the edges and would need careful revision, but it deserves a serious referee rather than a desk reject. Who's it for? People working on sparse DNN accelerators, especially systolic-array variants. I'd probably cite it if I were writing in that space, and I'd bring it to reading group as a good example of an incremental idea with a clear flaw in reporting.

Recommendation: send it to peer review; ask for major revision with corrected numbers and a proof of the routing lemma.

Referee Report

4 major / 4 minor

Summary. The paper proposes VUSA, a systolic-array architecture in which each row has M data-flow elements (SPEs) but only A < M physical MAC units, so that rows with at most A nonzero weights can be processed with the data flow of an N × M array. The authors derive a binomial probability model for full virtual growth (Eqs. 1–4), synthesize a 3 × 6 VUSA with 9 MACs in a commercial 16-nm technology, and evaluate it on pruned ResNet-18 and MobileNetV1 using a cycle-accurate simulator plus empirically measured weight distributions. The central claims are that VUSA achieves area and power efficiency gains over standard dense systolic arrays while preserving the standard programming model.

Significance. If the central routing claim is established, VUSA is a conceptually simple and potentially useful way to exploit unstructured weight sparsity in a weight-stationary systolic array without changing the programming model. The paper has clear strengths: it presents a closed-form, parameter-free probability model for virtual growth; it includes a concrete, reproducible physical-design study in 16-nm technology; and it evaluates on real pruned models rather than synthetic workloads only. However, the significance is currently conditional on an unproven combinatorial routing assertion and on the accuracy of the reported efficiency numbers, which are internally inconsistent across the abstract, tables, and conclusion.

major comments (4)

[Section III-C] The claim that every set of at most A nonzero weight positions among M SPEs can be routed with MAC j connected only to SPEs [j, j+M−A] is asserted without proof. This is load-bearing because Eq. (4) and the load-split percentages in Section V-D assume that every qualifying row can be processed in a single N × M window. Please provide a proof (for example, by Hall's theorem or an explicit greedy matching algorithm) or a counterexample. If some sparse-row patterns require a smaller window, the virtual-growth probabilities and the simulated area/power efficiency gains in Tables II and III would be optimistic. In addition, the paper does not specify the switching network, register placement, or multiplexer structure that realizes the claimed limited-range shifting; the synthesis results in Table I should include the area and power of this configuration logic and of the reconfiguration control.
[Abstract; Section V; Section VII] The headline numbers are inconsistent. The abstract reports 37% area and 68% power savings, which matches Table I only if these are raw resource savings relative to a standard 3 × 6 array, but Section V-D reports application-level efficiency gains of only 18–27% in area efficiency and 45–56% in power efficiency (Tables II and III). The conclusion's statement that the technique improves 'area and power efficiencies by around 55% and 25%, respectively' matches neither set of numbers and appears to swap the two quantities. Please state explicitly which comparison is being made (raw component savings versus workload-level efficiency) and correct the conclusion so that it is consistent with the tables.
[Section V-C] The methodology for computing VUSA's cycle counts is underspecified. The paper says that 'the percentages of the VUSA performing as 3×6, 3×5, 3×4, and 3×3 systolic arrays' are determined from the weight filters, but it does not describe the algorithm used to partition the model's matrix multiplications into these windows, nor how the MAC-to-SPE configuration changes between windows are modeled. Reconfiguration latency and energy are apparently not included in the reported cycles or power. Without this detail, the efficiency gains in Tables II and III are not fully reproducible and may be optimistic.
[Section III-C; Table I] The phrase 'same peak-performance' is ambiguous and potentially misleading. A 9-MAC VUSA cannot sustain the same peak multiply-accumulate throughput as an 18-MAC standard array on dense data; it can match the throughput of a larger array only when the workload is sparse enough that all MAC units are usefully occupied. The paper should define the performance model explicitly (for example, whether peak throughput means useful MACs per cycle on a given sparse matrix, or the dense peak of the physical hardware) and state how this definition relates to the cycle counts from SCALE-Sim.

minor comments (4)

[Section III-C] There is a typo: 'M ACnumber r' should read 'MAC number r'.
[Section V-B] Figure 7 is difficult to interpret: the caption '3x6: 100%' and '3x6: 50%, 3x3: 50%' is not explained in the text, and the binary matrix layout is not clearly tied to the positions of nonzero weights in the VUSA rows.
[Section IV] Equations (1)–(4) assume row-wise independence of weight sparsity. The authors are careful to use actual distributions later, but Figure 6 should be labeled as an idealized illustration rather than a prediction for real workloads, since real pruned models may deviate substantially from the binomial model.
[Tables II and III] The normalized efficiency columns would be clearer if the text stated explicitly that every design is normalized to the standard 3 × 6 array, not to its own best-case metric; this is evident from the numbers but worth stating to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VUSA's gains are derived from workload sparsity via a parameter-free binomial model and independent synthesis/simulation data.

full rationale

The paper's derivation chain is self-contained. The theoretical gain model (Eqs. 1-4) computes P_sparse_to_standard as a binomial CDF with P1 taken from the workload's non-zero weight probability; no output quantity is used to define P1, and no fitted parameter is renamed as a prediction. The experimental efficiency numbers (Tables I-III, Figures 8-9) come from physical synthesis in a commercial 16-nm technology and cycle counts from SCALE-Sim combined with weight sparsity statistics measured from pre-trained SparseZoo models; these are independent of the claimed outcomes. The architecture's fallback rule 'down to N x A, at which the conditions are guaranteed to be satisfied' makes the 'supports any sparsity' statement true by construction, but the paper presents this as an architectural property rather than as a derived empirical prediction, so it is not circular. Citations [7] and [8] are background on sparse DNN processing and are not load-bearing; there is no imported uniqueness theorem or ansatz. The Section III-C routing assertion (each MAC r connects only to SPEs [r, ..., r+M-A] and that this suffices for every sparse-row pattern) is asserted without proof and is load-bearing for Eq. (4) and the load-split percentages in Tables II-III; however, this is a correctness/completeness risk, not a circularity, since the claim is not defined in terms of the predicted efficiency and is not justified by self-citation. Therefore the paper receives no circularity score.

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

The central claim rests on four domain assumptions: the sliding-window mapping is always feasible, sparsity positions are independent, SCALE-Sim cycles transfer to VUSA, and SparseZoo models are representative. The mapping assumption is the most fragile because it is asserted without proof and directly supports the growth probabilities and benchmark gains.

assumptions (4)

domain assumption Every row with A or fewer non-zero weights in M positions can always be mapped onto the A MACs using only the allowed sliding-window connections (MAC j can reach SPEs j through j+M-A).
Stated in Section III-C without proof; the paper's theoretical gain probability Eq. (4) and the load-split calculations in Section V depend on it.
domain assumption Unstructured weight sparsity positions are independent and identically distributed Bernoulli(P1), so row non-zero counts are binomial.
Section IV explicitly assumes no cross-correlation; real pruning distributions can be correlated across rows and columns, which would change P_sparse_to_standard.
domain assumption SCALE-Sim accurately models cycle counts of standard weight-stationary systolic arrays, and these cycles transfer directly to VUSA's virtual windows.
Used in Sections V-C and V-D to compute performance and efficiency; no validation of the transfer is given.
domain assumption SparseZoo pretrained model weights have unstructured sparsity representative of edge-AI workloads.
Benchmarks in Section V-D rely on this; the two models cover limited diversity and sparsity patterns.

how reviews work

0 comments

Cite this review

Pith. "Pith review of VUSA: Virtually Upscaled Systolic Array Architecture to Exploit Unstructured Sparsity in AI Acceleration." pith.science (2026). https://pith.science/paper/L27GE3N5

@misc{pith2026250601166,
  author       = {Pith},
  title        = {Pith review of: VUSA: Virtually Upscaled Systolic Array Architecture to Exploit Unstructured Sparsity in AI Acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L27GE3N5}},
  note         = {Machine review of arXiv:2506.01166}
}

read the original abstract

Leveraging high degrees of unstructured sparsity is a promising approach to enhance the efficiency of deep neural network DNN accelerators - particularly important for emerging Edge-AI applications. We introduce VUSA, a systolic-array architecture that virtually grows based on the present sparsity to perform larger matrix multiplications with the same number of physical multiply-accumulate MAC units. The proposed architecture achieves saving by 37% and 68% in area and power efficiency, respectively, at the same peak-performance, compared to a baseline systolic array architecture in a commercial 16-nm technology. Still, the proposed architecture supports acceleration for any DNN with any sparsity - even no sparsity at all. Thus, the proposed architecture is application-independent, making it viable for general-purpose AI acceleration.

Figures

Figures reproduced from arXiv: 2506.01166 by the authors.

**Figure 4.** Single row of the proposed VUSA architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

**Figure 3.** Decomposing the PE to a data-flow and a compute block. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 6.** Figure 6: Growth probability of VUSA with N = 3, M = 6, and A = 3 TABLE I SYNTHESIS RESULTS IN A COMMERCIAL 16-NM TECHNOLOGY AT 1 GHz Design #MACs Area (Normalized) Power (Normalized) Standard 3×3 9 0.69 0.86 Standard 3×4 12 0.91 1.15 Standard 3×5 15 1.14 1.41 Standard 3×6 18 1.…

**Figure 7.** Figure 7: Example 50% sparsity distribution B. Effect of Sparsity on Performance The number of zero weights and their positions in the different layers of a given application determines the number of jobs to be executed on the VUSA structure. If the weights of a given load are a…

**Figure 8.** Figure 8: Area efficiency at different pruning rates. [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Power efficiency at different pruning rates. [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages

[1]

A survey of convolutional neural networks: analysis, applications, and prospects,

Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks: analysis, applications, and prospects,” IEEE transac- tions on neural networks and learning systems , 2021

2021
[2]

In-datacenter performance analysis of a tensor pro- cessing unit,

N. P. Jouppi et al., “In-datacenter performance analysis of a tensor pro- cessing unit,” in Proceedings of the 44th annual international symposium on computer architecture, 2017, pp. 1–12

work page 2017
[3]

An accelerator for sparse convolutional neural networks leveraging systolic general matrix- matrix multiplication,

M. Soltaniyeh, R. P. Martin, and S. Nagarakatte, “An accelerator for sparse convolutional neural networks leveraging systolic general matrix- matrix multiplication,” ACM Transactions on Architecture and Code Optimization (TACO), vol. 19, no. 3, pp. 1–26, 2022

2022
[4]

Sense: Model-hardware codesign for accelerating sparse cnns on systolic arrays,

W. Sun et al., “Sense: Model-hardware codesign for accelerating sparse cnns on systolic arrays,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems , vol. 31, no. 4, pp. 470–483, 2023

work page 2023
[5]

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

arXiv 2015
[6]

Eie: Efficient inference engine on compressed deep neural network,

S. Han et al., “Eie: Efficient inference engine on compressed deep neural network,” ACM SIGARCH Computer Architecture News , vol. 44, no. 3, pp. 243–254, 2016

work page 2016
[7]

Synapse compression for event-based convolutional-neural-network ac- celerators,

L. Bamberg, A. Pourtaherian, L. Waeijen, A. Chahar, and O. Moreira, “Synapse compression for event-based convolutional-neural-network ac- celerators,” IEEE Transactions on Parallel and Distributed Systems , vol. 34, no. 4, pp. 1227–1240, 2023

work page 2023
[8]

Exploiting neural-network statistics for low-power DNN inference,

L. Bamberg, A. Najafi, and A. Garcia-Ortiz, “Exploiting neural-network statistics for low-power DNN inference,” IEEE Open Journal of Circuits and Systems, 2024

work page 2024

Show all 17 references

[9]

Why systolic architectures?

H.-T. Kung, “Why systolic architectures?” IEEE computer , vol. 15, no. 1, pp. 37–46, 1982

1982
[10]

Scale-sim: Systolic cnn accelerator simulator,

A. Samajdar, Y . Zhu, P. Whatmough, M. Mattina, and T. Kr- ishna, “Scale-sim: Systolic cnn accelerator simulator,” arXiv preprint arXiv:1811.02883, 2018

2018 arXiv
[11]

Cnvlutin: Ineffectual-neuron-free deep neural network computing,

J. Albericio, P. Judd, T. Hetherington, T. Aamodt, N. E. Jerger, and A. Moshovos, “Cnvlutin: Ineffectual-neuron-free deep neural network computing,” ACM SIGARCH Computer Architecture News , vol. 44, no. 3, pp. 1–13, 2016

2016
[12]

Eyeriss: An energy- efficient reconfigurable accelerator for deep convolutional neural net- works,

Y .-H. Chen, T. Krishna, J. S. Emer, and V . Sze, “Eyeriss: An energy- efficient reconfigurable accelerator for deep convolutional neural net- works,” IEEE journal of solid-state circuits, vol. 52, no. 1, pp. 127–138, 2016

2016
[13]

SparseZoo: Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes,

NeuralMagic, “SparseZoo: Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes,” accessed: 2023-12-21. [Online]. Available: https://sparsezoo. neuralmagic.com

2023
[14]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

2016
[15]

Mobilenets: Efficient convolutional neural net- works for mobile vision applications,

A. G. Howard et al. , “Mobilenets: Efficient convolutional neural net- works for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017

2017 arXiv
[16]

S2ta: Exploiting structured sparsity for energy-efficient mobile cnn acceleration,

Z.-G. Liu, P. N. Whatmough, Y . Zhu, and M. Mattina, “S2ta: Exploiting structured sparsity for energy-efficient mobile cnn acceleration,” in 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 2022, pp. 573–586

2022
[17]

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

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A survey of convolutional neural networks: analysis, applications, and prospects,

Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks: analysis, applications, and prospects,” IEEE transac- tions on neural networks and learning systems , 2021

2021

[2] [2]

In-datacenter performance analysis of a tensor pro- cessing unit,

N. P. Jouppi et al., “In-datacenter performance analysis of a tensor pro- cessing unit,” in Proceedings of the 44th annual international symposium on computer architecture, 2017, pp. 1–12

work page 2017

[3] [3]

An accelerator for sparse convolutional neural networks leveraging systolic general matrix- matrix multiplication,

M. Soltaniyeh, R. P. Martin, and S. Nagarakatte, “An accelerator for sparse convolutional neural networks leveraging systolic general matrix- matrix multiplication,” ACM Transactions on Architecture and Code Optimization (TACO), vol. 19, no. 3, pp. 1–26, 2022

2022

[4] [4]

Sense: Model-hardware codesign for accelerating sparse cnns on systolic arrays,

W. Sun et al., “Sense: Model-hardware codesign for accelerating sparse cnns on systolic arrays,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems , vol. 31, no. 4, pp. 470–483, 2023

work page 2023

[5] [5]

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

arXiv 2015

[6] [6]

Eie: Efficient inference engine on compressed deep neural network,

S. Han et al., “Eie: Efficient inference engine on compressed deep neural network,” ACM SIGARCH Computer Architecture News , vol. 44, no. 3, pp. 243–254, 2016

work page 2016

[7] [7]

Synapse compression for event-based convolutional-neural-network ac- celerators,

L. Bamberg, A. Pourtaherian, L. Waeijen, A. Chahar, and O. Moreira, “Synapse compression for event-based convolutional-neural-network ac- celerators,” IEEE Transactions on Parallel and Distributed Systems , vol. 34, no. 4, pp. 1227–1240, 2023

work page 2023

[8] [8]

Exploiting neural-network statistics for low-power DNN inference,

L. Bamberg, A. Najafi, and A. Garcia-Ortiz, “Exploiting neural-network statistics for low-power DNN inference,” IEEE Open Journal of Circuits and Systems, 2024

work page 2024