Pith. sign in

REVIEW 4 major objections 5 minor 54 references

eMamba: Efficient Acceleration Framework for Mamba Models in Edge Computing

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read eMamba is an end-to-end framework for deploying Mamba models on edge FPGAs and 22nm ASICs, claiming 4.95-5.62x lower latency and 2.22-9.95x higher throughput than CNN and ViT accelerators while using far less area and energy.

desk verdict Real integration effort for edge Mamba acceleration, but the headline speedups rely on unfairly matched baselines; worth refereeing with artifact and comparison fixes. read the letter →

arxiv 2508.10370 v1 pith:YPBYIFR4 submitted 2025-08-14 cs.LG cs.AI

Jiyong Kim , Jaeho Lee , Jiahao Lin , Alish Kanani , Miao Sun , Umit Y. Ogras , Jaehyun Park This is my paper
classification cs.LGcs.AI
keywords MambastatespacemodelsedgecomputingFPGAaccelerationASICdesignquantizationneuralarchitecturesearchhardware-awareapproximation
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 argues that Mamba, a state-space architecture that processes sequences in linear time, can be made deployable on resource-constrained edge hardware by replacing exact layer normalization and non-linear activations with hardware-friendly approximations and by quantizing the entire pipeline to 8-bit with a specially higher-precision path for the SSM's recurrent state. The result is an accelerator that, the authors report, processes a pose-estimation frame 5.62x faster than a quantized CNN accelerator and 4.95x faster than a ViT accelerator on the same FPGA, while using a fraction of the area and energy on a 22nm ASIC. Accuracy stays comparable to or better than the ViT baselines while using up to 19.9x fewer parameters, and perplexity on WikiText2 remains flat as sequence length grows from 512 to 8,192. The reason to care: if these results hold, edge devices can run Mamba models at real-time speeds on small, low-energy silicon.

What carries the argument

The load-bearing mechanism is the pairing of range normalization with scale-aware quantization of the SSM recurrence. Range normalization replaces layer normalization's square-root and variance computations with max/min comparators, using learnable scale and shift parameters, which removes the dominant latency bottleneck in the pipelined Mamba block. The quantizer keeps the hidden state at higher precision (INT24 for the multiply-accumulate, INT17 after right-shifting by the scale of the transition matrix) while everything else runs at INT8 with power-of-two scales that compile to bit shifts, so recurrent precision does not compound across time steps. Around these, piecewise-linear approxima

What would settle it

Take the same CNN and ViT baselines used in the paper, apply eMamba's exact approximation set (range normalization, piecewise-linear SiLU and exponentials, scale-aware INT8 quantization with the higher-precision SSM state) without changing their architectures, and re-measure latency, throughput, and area on the same ZCU102 FPGA and 22nm flow. If the latency and throughput ratios fall below roughly 2x, the reported edge advantage is an artifact of how the baselines were configured rather than of Mamba's structure.

Watch

Extended reading notes

Core claim

The paper presents eMamba as a complete design flow for turning a Mamba state-space model into a deployed edge accelerator. At its center are three approximations: range normalization replaces layer normalization with comparator logic; SiLU and the exponentials used in SSM discretization are replaced with piecewise-linear fits bounded to a few percent error; and the network is fully quantized to INT8 except the SSM hidden state, which is computed at INT24, divided by the state-transition scale factor, and stored at INT17 so precision does not accumulate across the sequence. On the MARS pose-estimation task the resulting design processes a frame in 1,643 cycles at 100 MHz, which the authors r

Load-bearing premise

The claimed speed and efficiency advantages rest on the CNN and ViT baselines being implemented with comparable care: the ViT design uses 246% of the FPGA's LUTs (so it does not fit the target device) and the CNN baseline is tuned only to match eMamba's LUT usage rather than its own latency, so a fair re-optimization of the baselines could shrink the reported gains.

Editorial extensions

If this is right

  • Mamba can serve as a practical backbone for edge vision tasks: on the MARS human-pose dataset, eMamba beats a CNN baseline in MAE with 63x fewer parameters in FP32 and matches the quantized CNN while running at 5.62x lower frame latency.
  • The framework scales to long-sequence language modeling: perplexity on WikiText2 stays between 94.7 and 95.6 (FP32) from sequence length 512 to 8,192, while RNN and LSTM perplexities climb sharply over the same range.
  • A fully 8-bit, BRAM-free FPGA implementation using 40-68% of the platform's LUTs leaves resources available for the rest of an edge system, indicating the accelerator is memory-scalable and cheap to integrate.
  • On a 22nm ASIC, one eMamba inference consumes 1.254 microjoules versus 60.99 microjoules for the ViT accelerator, meaning the design can sustain continuous inference on devices with tight energy budgets.

Reading between the lines

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

  • The reported 5.62x latency gain over the CNN baseline likely overstates Mamba's intrinsic advantage because the baseline was tuned only to match eMamba's LUT usage, not to minimize its own latency; re-optimizing the CNN with the same approximation-aware search would provide a cleaner comparison.
  • The ViT baseline consumes 246% of the FPGA's LUTs, so the 4.95x latency comparison pits eMamba against a design that does not fit the target device; a ViT reduced to fit might narrow the gap.
  • The same recipe — range normalization, piecewise-linear SiLU/exponentials, and scale-aware state quantization — could be applied to other state-space architectures such as S4, S5, and H3, potentially giving edge deployment to a whole family of sequence models.
  • Because the normalization layer dominates the small-scale latency profile (up to 10 compute units), the framework's gains may look different for larger Mamba models where the SSM's recurrent and matrix operations occupy a bigger share of the pipeline; a scaling study with larger D, E, and N would test this.
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 presents eMamba, an end-to-end hardware acceleration framework for deploying Mamba models on edge platforms. The design replaces layer normalization with a learnable range normalization, approximates SiLU and exponential/softplus with piecewise-linear functions, performs approximation-aware NAS, and uses INT8 quantization with a scale-aware quantized SSM layer. The authors implement the pipeline on an AMD ZCU102 FPGA and in GF 22 nm ASIC technology, and evaluate it on Fashion-MNIST, CIFAR-10, MARS, and WikiText2. The main claims are that eMamba matches or beats ViT/CNN accuracy with far fewer parameters, and that on the MARS workload it achieves 5.62x lower latency and 9.95x higher throughput than a CNN accelerator and 4.95x lower latency and 2.22x higher throughput than a ViT accelerator, with substantially smaller area and energy.

Significance. If the reported results hold, the paper makes a useful contribution: it demonstrates a concrete, silicon-validated path for running SSM-based models on resource-constrained hardware, and it combines algorithmic approximation, NAS, quantization, FPGA prototyping, and ASIC synthesis into one framework. The work also includes a reasonably detailed ablation study showing which approximations contribute the most on the target FPGA. The comparison against a real CNN accelerator (FINN) and a ViT accelerator is valuable in principle. However, the significance of the headline speedup claims depends on the baselines being comparably optimized and deployable, and that is currently not the case; the issues below need to be resolved before the strong quantitative claims are accepted.

major comments (4)
  1. [Sections 5.4 and 5.5, Table 6, Figure 9] The headline performance claims rest on baselines that are not comparably optimized. The ViT accelerator consumes 246% of the ZCU102 LUTs (Section 5.5), so it cannot be deployed on the target FPGA at all; reporting its latency and throughput as if it were a deployable baseline is misleading. The CNN baseline is configured by incrementally adjusting the FINN throughput setting only until LUT utilization is comparable to reconfigurable eMamba (71.0% vs 68.5%), not until latency or throughput is optimized under that resource budget. Meanwhile eMamba benefits from approximation-aware NAS, range normalization, piecewise SiLU/exp approximations, and custom pipelining. Thus the claimed 5.62x/4.95x latency and 9.95x/2.22x throughput gains conflate architectural efficiency with unequal design effort. The authors should either re-optimize the baselines under the same resource and optimization cons
  2. [Section 5.4, Table 6] The throughput metric is presented inconsistently. The text states that throughput is the number of inferences per second in a pipelined manner, but Table 6 labels the column as 'Mb/s'. For eMamba, 1,643 cycles at 100 MHz corresponds to about 60,900 inferences per second, not 263 Mb/s. If 'Mb/s' is intended as a data rate, the conversion from inferences to bits is not given; if 'inferences/s' is intended, the label is wrong. The ratios may be unchanged, but as reported the metric cannot be interpreted and the claim '263 Mb/s' is unsupported. Please correct the unit and provide the conversion if a data-rate metric is intended.
  3. [Section 5.2, Table 3] The accuracy claim is not uniformly supported under INT8, which is the precision used in the hardware implementation. On CIFAR-10, INT8 eMamba achieves 72.6% vs 75.6% for INT8 ViT, a 3-point gap, whereas the FP32 versions are nearly tied. The text attributes this to 'sensitivity to channel-wise statistical variation under quantization' but provides no analysis or mitigation. Since the hardware results use 8-bit quantization, the claim of 'comparable accuracy' needs to be tempered or the CIFAR-10 INT8 degradation must be explained and, ideally, addressed. In addition, no variance, seeds, or multiple-run statistics are reported for any accuracy, RMSE, or MAE result; given the small datasets and the approximation/NAS pipeline, single-point measurements are not sufficient to support the quantitative comparisons.
  4. [Section 5.6, Table 7] The ASIC comparison is reported only against the ViT baseline; there is no GF 22 nm area, power, or energy result for the CNN or naïve Mamba baselines. This is not a fatal issue by itself, but it means the '4.77x smaller area, 9.84x lower power, 48.6x lower energy' claims are narrower than the abstract suggests, because the CNN baseline is absent from the ASIC comparison. Please clarify whether the CNN comparator was omitted because of fit/effort issues, and state which baselines are included in each hardware comparison.
minor comments (5)
  1. [Section 5.6] The sentence 'We also implement ViT and the proposed eMamba designs using GF 22FDX FDSOI 22 nm technology to obtain detailed area, area, and power comparisons' has a duplicated word 'area' and should read 'area, power, and energy comparisons'.
  2. [Section 4.1] Typo: 'devision' should be 'division'.
  3. [References] References [23] and [24] are the same MARCA paper; one reference should be removed or they should be cross-referenced properly.
  4. [Contributions list, Section 1] The third contribution says 'comparable accuracy to CNN and ViT with 63x and 1.63x parameter reduction', but the abstract and Section 5 report 1.63-19.9x parameter reductions for the vision datasets. Please clarify that 63x is the reduction relative to the CNN baseline and 1.63x is relative to the ViT baseline, and make the notation consistent.
  5. [Section 5.4] The throughput gain of 5.32x over naïve Mamba is listed in Table 6 but not mentioned in the abstract; please ensure the abstract and Section 5.4 use consistent metrics and numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; central results are measured hardware/accuracy outcomes rather than derivations from their own inputs, with baseline-choice and self-citation issues noted as non-circular validity risks.

full rationale

The claimed speedups and accuracy comparisons come from FPGA/ASIC implementations and standard evaluations, not from a derivation that reduces to its assumptions. Approximations (range normalization, piecewise SiLU/exp, softplus) are stated as independent functional substitutions and validated by ablation; the NAS selects hyperparameters from a Pareto front rather than imposing the reported parameter reductions; quantization is standard symmetric INT8 with explicit scale handling. No equation in Sections 4.1-4.6 is constructed from the quantity it is later said to predict. The CNN/ViT comparisons raise fairness questions -- the FINN CNN throughput configuration was adjusted only to match eMamba LUT utilization (Section 5.4), and the ViT accelerator uses 246% of ZCU102 LUTs (Section 5.5) -- but these are threats to external validity, not circularity, because eMamba's measured latency/throughput is not definitionally forced by those baseline choices. The citation to MARS [3] is a self-citation of a corresponding author's dataset, but it is an external, previously published benchmark, so it is independent support rather than load-bearing circular reasoning.

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

The central claims rest on several hand-chosen approximation thresholds (SiLU segment count and error bound), data-derived quantization scales, task-specific bit-width choices, and NAS-selected hyperparameters. These are not derived from first principles; they are tuned per application. The main structural assumptions are the correctness of the Mamba discretization from prior work, the accuracy-preservation of the piecewise approximations, and the stability of calibration-time distributions.

free parameters (4)
  • Per-layer quantization scale factors S (zero-point Z=0) = not enumerated; powers of two
    Determined from data distribution profiling in Section 4.6; scale factors directly set fixed-point rounding and shift amounts.
  • SiLU approximation: 17 segments, 3% max error = 17 linear segments over [-7,7]
    Hand-chosen in Section 4.2; controls the accuracy/hardware trade-off and is part of the central approximation claim.
  • SSM hidden-state bit-width / re-quantization shift = INT24 computed, right-shift by 7 -> INT17 for MARS
    Selected to prevent overflow in recurrent accumulation (Section 4.6); task-specific tuning.
  • NAS-selected architecture hyperparameters = e.g., MARS D=20, E=2, M=2, N=8; other tasks in Table 2
    Chosen from the Pareto front in approximation-aware NAS (Section 4.5), balancing validation RMSE and parameter count.
assumptions (4)
  • domain assumption The Mamba SSM discretization and architecture from Gu and Dao [17] are correct and are used as the algorithmic substrate.
    Equation 4 and Section 4.3 rely on the selective SSM definitions without re-derivation.
  • ad hoc to paper A 17-segment piecewise linear SiLU and piecewise exponential/softplus approximations with stated max errors preserve task accuracy within reported bounds.
    Empirically validated on four datasets; no formal guarantee connects approximation error to final accuracy.
  • domain assumption Calibration-time data distributions are representative of inference-time activations, so per-layer scale factors and clipping remain valid.
    Section 4.6 profiles Mamba layers to set quantization parameters; assumes no distribution shift at deployment.
  • ad hoc to paper Range normalization with learnable gamma and beta (Eq. 2) is a sufficient replacement for layer normalization in Mamba.
    Ablation shows improvement over I-ViT's integer layer norm, but the comparison to exact layer norm under INT8 is not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of eMamba: Efficient Acceleration Framework for Mamba Models in Edge Computing." pith.science (2026). https://pith.science/paper/YPBYIFR4

@misc{pith2026250810370,
  author       = {Pith},
  title        = {Pith review of: eMamba: Efficient Acceleration Framework for Mamba Models in Edge Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YPBYIFR4}},
  note         = {Machine review of arXiv:2508.10370}
}
abstract

State Space Model (SSM)-based machine learning architectures have recently gained significant attention for processing sequential data. Mamba, a recent sequence-to-sequence SSM, offers competitive accuracy with superior computational efficiency compared to state-of-the-art transformer models. While this advantage makes Mamba particularly promising for resource-constrained edge devices, no hardware acceleration frameworks are currently optimized for deploying it in such environments. This paper presents eMamba, a comprehensive end-to-end hardware acceleration framework explicitly designed for deploying Mamba models on edge platforms. eMamba maximizes computational efficiency by replacing complex normalization layers with lightweight hardware-aware alternatives and approximating expensive operations, such as SiLU activation and exponentiation, considering the target applications. Then, it performs an approximation-aware neural architecture search (NAS) to tune the learnable parameters used during approximation. Evaluations with Fashion-MNIST, CIFAR-10, and MARS, an open-source human pose estimation dataset, show eMamba achieves comparable accuracy to state-of-the-art techniques using 1.63-19.9$\times$ fewer parameters. In addition, it generalizes well to large-scale natural language tasks, demonstrating stable perplexity across varying sequence lengths on the WikiText2 dataset. We also quantize and implement the entire eMamba pipeline on an AMD ZCU102 FPGA and ASIC using GlobalFoundries (GF) 22 nm technology. Experimental results show 4.95-5.62$\times$ lower latency and 2.22-9.95$\times$ higher throughput, with 4.77$\times$ smaller area, 9.84$\times$ lower power, and 48.6$\times$ lower energy consumption than baseline solutions while maintaining competitive accuracy.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 32 canonical work pages

  1. [1]

    2007.Fundamentals of electric circuits

    Charles K Alexander, Matthew NO Sadiku, and Matthew Sadiku. 2007.Fundamentals of electric circuits. McGraw-Hill Higher Education Boston, MA, USA

    work page 2007

  2. [2]

    AMD Inc. 2023. ZCU102 Evaluation Board User Guide (UG1182). https://docs.amd.com/v/u/en-US/ug1182-zcu102- eval-bd

    work page 2023

  3. [3]

    Sizhe An and Umit Y. Ogras. 2021. MARS: mmWave-based Assistive Rehabilitation System for Smart Healthcare.ACM Transactions on Embedded Computing Systems20, 5s (Oct. 2021), 1–22. https://doi.org/10.1145/3477003

    work page doi:10.1145/3477003 2021

  4. [4]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer Normalization.arXiv preprint arXiv:1607.06450 (2016)

    arXiv 2016

  5. [5]

    Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry. 2018. Scalable Methods for 8-Bit Training of Neural Networks. InAdvances in Neural Information Processing Systems, Vol. 31. Curran Associates, Inc

    work page 2018

  6. [6]

    2011.Modern control systems

    Richard C Dorf Robert H Bishop. 2011.Modern control systems

    work page 2011

  7. [7]

    Michaela Blott, Thomas B Preußer, Nicholas J Fraser, Giulio Gambardella, Kenneth O’brien, Yaman Umuroglu, Miriam Leeser, and Kees Vissers. 2018. FINN-R: An end-to-end deep-learning framework for fast exploration of quantized neural networks.ACM Transactions on Reconfigurable Technology and Systems (TRETS)11, 3 (2018), 1–23

    work page 2018

  8. [8]

    Jerry Chee, Yaohui Cai, Volodymyr Kuleshov, and Christopher M De Sa. 2023. Quip: 2-bit quantization of large language models with guarantees.Advances in Neural Information Processing Systems36 (2023), 4396–4429

    2023

Show all 54 references
  1. [9]

    Yiran Chen, Yuan Xie, Linghao Song, Fan Chen, and Tianqi Tang. 2020. A survey of accelerator architectures for deep neural networks.Engineering6, 3 (2020), 264–274

    2020

  2. [10]

    Yu-Hsin Chen, Joel Emer, and Vivienne Sze. 2016. Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks.ACM SIGARCH Computer Architecture News44, 3 (June 2016), 367–379. https: //doi.org/10.1145/3007787.3001177

    2016

  3. [11]

    Hung-Yueh Chiang, Chi-Chih Chang, Natalia Frumkin, Kai-Chiang Wu, and Diana Marculescu. 2024. Quamba: A post-training quantization recipe for selective state space models.arXiv preprint arXiv:2410.13229(2024). eMamba : Efficient Acceleration Framework for Mamba Models in Edge ...

    2024 arXiv

  4. [12]

    Sobelman

    Kyubaik Choi and Gerald E. Sobelman. 2022. An Efficient CNN Accelerator for Low-Cost Edge Systems.ACM Trans. Embed. Comput. Syst.21, 4 (Aug. 2022), 44:1–44:20. https://doi.org/10.1145/3539224

    2022 doi

  5. [13]

    Alex De Vries. 2023. The growing energy footprint of artificial intelligence.Joule7, 10 (2023), 2191–2194

    2023

  6. [14]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognit...

    2021

  7. [15]

    Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. 2022. Hungry hungry hippos: Towards language modeling with state space models.arXiv preprint arXiv:2212.14052(2022)

    2022 arXiv

  8. [16]

    Xue Geng, Jie Lin, Bin Zhao, Anmin Kong, Mohamed M Sabry Aly, and Vijay Chandrasekhar. 2019. Hardware-aware softmax approximation for deep neural networks. InComputer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Sele...

    2019

  9. [17]

    Albert Gu and Tri Dao. 2024. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. InFirst Conference on Language Modeling

    2024

  10. [18]

    Albert Gu, Karan Goel, and Christopher Re. 2022. Efficiently Modeling Long Sequences with Structured State Spaces. InInternational Conference on Learning Representations

    2022

  11. [19]

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. 2021. Combining recurrent, convolutional, and continuous-time models with linear state space layers.Advances in neural information processing systems34 (2021), 572–585

    2021

  12. [20]

    Albert Gu, Isys Johnson, Aman Timalsina, Atri Rudra, and Christopher Re. 2023. How to Train your HIPPO: State Space Models with Generalized Orthogonal Basis Projections. InInternational Conference on Learning Representations

    2023

  13. [21]

    Atsutake Kosuge, Mototsugu Hamada, and Tadahiro Kuroda. 2021. A 16 nJ/classification FPGA-based wired-logic DNN accelerator using fixed-weight non-linear neural net.IEEE Journal on Emerging and Selected Topics in Circuits and Systems11, 4 (2021), 751–761

    2021

  14. [22]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. (2009)

    2009

  15. [24]

    Jinhao Li, Shan Huang, Jiaming Xu, Jun Liu, Li Ding, Ningyi Xu, and Guohao Dai. 2024. Marca: Mamba accelerator with reconfigurable architecture.arXiv preprint arXiv:2409.11440(2024)

    2024 arXiv

  16. [25]

    Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. 2022. Q-vit: Accurate and fully quantized low-bit vision transformer.Advances in neural information processing systems35 (2022), 34451–34463

    2022

  17. [26]

    Zhikai Li and Qingyi Gu. 2023. I-vit: Integer-only quantization for efficient vision transformer inference. InProceedings of the IEEE/CVF International Conference on Computer Vision. 17065–17075

    2023

  18. [27]

    Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. 2021. Fq-vit: Post-training quantization for fully quantized vision transformer.arXiv preprint arXiv:2111.13824(2021)

    2021 arXiv

  19. [28]

    Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. 2021. Post-training quantization for vision transformer.Advances in Neural Information Processing Systems34 (2021), 28092–28103

    2021

  20. [29]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843(2016)

    2016 arXiv

  21. [30]

    Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernock`y, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model.. InInterspeech, Vol. 2. Makuhari, 1045–1048

    2010

  22. [31]

    Shashank Nag, Gourav Datta, Souvik Kundu, Nitin Chandrachoodan, and Peter A. Beerel. 2023. ViTA: A Vision Transformer Inference Accelerator for Edge Applications. In2023 IEEE International Symposium on Circuits and Systems (ISCAS). 1–5. https://doi.org/10.1109/ISCAS46773.2023....

    2023

  23. [32]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort

  24. [33]

    2023.Xilinx/brevitas

    Alessandro Pappalardo. 2023.Xilinx/brevitas. https://doi.org/10.5281/zenodo.3333552

    2023 doi

  25. [34]

    Junki Park, Hyunsung Yoon, Daehyun Ahn, Jungwook Choi, and Jae-Joon Kim. 2020. OPTIMUS: OPTImized matrix MUltiplication Structure for Transformer neural network accelerator.Proceedings of Machine Learning and Systems2 (March 2020), 363–378

    2020

  26. [35]

    Albert Reuther, Peter Michaleas, Michael Jones, Vijay Gadepally, Siddharth Samsi, and Jeremy Kepner. [n. d.]. AI and ML Accelerator Survey and Trends. In2022 IEEE High Performance Extreme Computing Conference (HPEC)(2022-09-19). 1–10. https://doi.org/10.1109/HPEC55821.2022.992...

    2022

  27. [36]

    David Rodriguez and Ram Krishnan. 2023. Learnable Image Transformations for Privacy Enhanced Deep Neural Networks. In2023 5th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA). IEEE, 64–73. 1:22 Jiyong Kim, Jaeho Lee...

    2023

  28. [37]

    Hasim Sak, Andrew W Senior, Françoise Beaufays, et al. 2014. Long short-term memory recurrent neural network architectures for large scale acoustic modeling.. InInterspeech, Vol. 2014. 338–342

    2014

  29. [38]

    Ananda Samajdar, Yuhao Zhu, Paul Whatmough, Matthew Mattina, and Tushar Krishna. 2018. Scale-sim: Systolic cnn accelerator simulator.arXiv preprint arXiv:1811.02883(2018)

    2018 arXiv

  30. [39]

    Shivanshu Shekhar, Tanishq Dubey, Koyel Mukherjee, Apoorv Saxena, Atharv Tyagi, and Nishanth Kotla. 2024. Towards optimizing the costs of llm usage.arXiv preprint arXiv:2402.01742(2024)

    2024 arXiv

  31. [40]

    Yongming Shen, Michael Ferdman, and Peter Milder. 2017. Escher: A CNN Accelerator with Flexible Buffering to Minimize Off-Chip Transfer. In2017 IEEE 25th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM). 93–100. https://doi.org/10.1109/FCCM.2017.47

    2017 doi

  32. [41]

    Smith, Andrew Warrington, and Scott Linderman

    Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. 2023. Simplified State Space Layers for Sequence Modeling. InThe Eleventh International Conference on Learning Representations

    2023

  33. [42]

    Yaman Umuroglu, Nicholas J Fraser, Giulio Gambardella, Michaela Blott, Philip Leong, Magnus Jahre, and Kees Vissers

  34. [43]

    Aimee Van Wynsberghe. 2021. Sustainable AI: AI for sustainability and the sustainability of AI.AI and Ethics1, 3 (2021), 213–218

    2021

  35. [44]

    Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. 2024. Graph-mamba: Towards long-range graph sequence modeling with selective state spaces.arXiv preprint arXiv:2402.00789(2024)

    2024 arXiv

  36. [45]

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. [n. d.]. HAQ: Hardware-Aware Automated Quantization With Mixed Precision. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)(Long Beach, CA, USA, 2019-06). IEEE, 8604–8612. https://doi.org/10.11...

    2019

  37. [47]

    Renjie Wei, Songqiang Xu, Linfeng Zhong, Zebin Yang, Qingyu Guo, Yuan Wang, Runsheng Wang, and Meng Li. 2025. LightMamba: Efficient Mamba Acceleration on FPGA with Quantization and Hardware Co-design.arXiv preprint arXiv:2502.15260(2025)

    2025

  38. [48]

    Shuang Wu, Guoqi Li, Feng Chen, and Luping Shi. 2018. Training and Inference with Integers in Deep Neural Networks. InInternational Conference on Learning Representations

    2018

  39. [49]

    Xiao Wu, Shuang Liang, Meiqi Wang, and Zhongfeng Wang. 2023. ReAFM: A reconfigurable nonlinear activation function module for neural networks.IEEE Transactions on Circuits and Systems II: Express Briefs70, 7 (2023), 2660–2664

    2023

  40. [50]

    Yuchen Xia, Jiho Kim, Yuhan Chen, Haojie Ye, Souvik Kundu, Cong Callie Hao, and Nishil Talati. 2024. Understanding the performance and estimating the cost of llm fine-tuning. In2024 IEEE International Symposium on Workload Characterization (IISWC). IEEE, 210–223

    2024

  41. [51]

    Han Xiao, Kashif Rasul, and Roland Vollgraf. 2017. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747(2017)

    2017 arXiv

  42. [52]

    Zukang Xu, Yuxuan Yue, Xing Hu, Zhihang Yuan, Zixu Jiang, Zhixuan Chen, Jiangyong Yu, Chen Xu, Sifan Zhou, and Dawei Yang. 2025. MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods.arXiv preprint arXiv:2501.13484(2025)

    2025 arXiv

  43. [53]

    Haoran You, Zhanyi Sun, Huihong Shi, Zhongzhi Yu, Yang Zhao, Yongan Zhang, Chaojian Li, Baopu Li, and Yingyan Lin. 2023. Vitcod: Vision transformer acceleration via dedicated algorithm and accelerator co-design. In2023 IEEE International Symposium on High-Performance Computer ...

    2023

  44. [54]

    Haoran Zhu, Boyuan Chen, and Carter Yang. 2023. Understanding why vit trains badly on small datasets: An intuitive perspective.arXiv preprint arXiv:2302.03751(2023)

    2023 arXiv

  45. [2017]

    InProceedings of the 2017 ACM/SIGDA international symposium on field-programmable gate arrays

    Finn: A framework for fast, scalable binarized neural network inference. InProceedings of the 2017 ACM/SIGDA international symposium on field-programmable gate arrays. 65–74

    2017

  46. [2021]

    A white paper on neural network quantization.arXiv preprint arXiv:2106.08295(2021)

    2021 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.