pith. machine review for the scientific record. sign in

arxiv: 2604.09512 · v1 · submitted 2026-04-10 · 💻 cs.LG · physics.optics

Recognition: unknown

Integrated electro-optic attention nonlinearities for transformers

Luis Mickeler , Kai Lion , Alfonso Nardi , Jost Kellner , Pierre Didier , Bhavin J. Shastri , Niao He , Rachel Grange
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:35 UTC · model grok-4.3

classification 💻 cs.LG physics.optics
keywords thin-film lithium niobateMach-Zehnder modulatorstransformer attentionanalog nonlinearitiessoftmaxelectro-optic computingquantization
0
0 comments X

The pith

Thin-film lithium niobate modulators implement analog Softmax for transformers with competitive accuracy at 4-bit quantization

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

The paper demonstrates that thin-film lithium niobate Mach-Zehnder modulators can act as fast analog replacements for the Softmax and Sigmoid functions in transformer models. These nonlinear operations normally create latency bottlenecks in digital hardware despite their small share of total computations. By integrating the modulators into vision transformers and large language models, the system achieves high accuracy even when both inputs and outputs are quantized to four bits. Noise characterization at encoding speeds reaching 10 GBaud supports the viability of this approach for hybrid hardware setups. The work targets energy-efficient, high-speed nonlinear computation within co-packaged photonic-electronic systems.

Core claim

Thin-film lithium niobate Mach-Zehnder modulators serve as electro-optic nonlinear function units that replace digital Softmax and Sigmoid operations in transformers. The analog implementations maintain competitive accuracy in Vision Transformers and Large Language Models under 4-bit input-output quantization, with system noise characterized at speeds up to 10 GBaud.

What carries the argument

Thin-film lithium niobate Mach-Zehnder modulators configured as analog nonlinear computational elements for attention mechanisms

Load-bearing premise

The analog electro-optic implementations integrate into full transformer pipelines at scale without introducing unacceptable cumulative noise, calibration drift, or accuracy loss beyond the characterized conditions.

What would settle it

A full-scale transformer model using the TFLN modulators for all nonlinearities shows significant accuracy degradation compared to digital baseline under realistic noise levels.

Figures

Figures reproduced from arXiv: 2604.09512 by Alfonso Nardi, Bhavin J. Shastri, Jost Kellner, Kai Lion, Luis Mickeler, Niao He, Pierre Didier, Rachel Grange.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: e displays a full time series of the normalized Opt￾max response for a randomly sampled input sequence of length n = 2048 at 5 bit resolution. The figure draws [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
read the original abstract

Transformers have emerged as the dominant neural-network architecture, achieving state-of-the-art performance in language processing and computer vision. At the core of these models lies the attention mechanism, which requires a nonlinear, non-negative mapping using the Softmax function. However, although Softmax operations account for less than 1% of the total operation count, they can disproportionately bottleneck overall inference latency. Here, we use thin-film lithium niobate (TFLN) Mach-Zehnder modulators (MZMs) as analog nonlinear computational elements to drastically reduce the latency of nonlinear computations. We implement electro-optic alternatives to digital Softmax and Sigmoid, and evaluate their performance in Vision Transformers and Large Language Models. Our system maintains highly competitive accuracy, even under aggressive 4-bit input-output quantization of the analog units. We further characterize system noise at encoding speeds up to 10 GBaud and assess model robustness under various noise conditions. Our findings suggest that TFLN modulators can serve as nonlinear function units within hybrid co-packaged hardware, enabling high-speed and energy-efficient nonlinear computation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript demonstrates the use of thin-film lithium niobate (TFLN) Mach-Zehnder modulators (MZMs) as analog electro-optic nonlinear units to approximate Softmax and Sigmoid functions for transformer attention mechanisms. It reports competitive accuracy in Vision Transformers and Large Language Models under aggressive 4-bit input-output quantization, along with noise characterization of the modulators at encoding speeds up to 10 GBaud, and concludes that such devices can enable high-speed, energy-efficient hybrid co-packaged hardware for nonlinear computation.

Significance. If the central claims hold, the work provides an experimental path toward photonic acceleration of the nonlinear components in attention, which can bottleneck latency despite low operation count. The demonstration of TFLN-based analog nonlinearities under quantization and high-speed noise conditions is a concrete hardware contribution that could inform co-packaged electro-optic AI accelerators, though its impact depends on scalability to full pipelines.

major comments (2)
  1. [Abstract] Abstract: The claim that the system 'maintains highly competitive accuracy' under 4-bit quantization is not supported by quantitative error bars, explicit baseline comparisons to digital implementations, or details on how the analog Softmax is exactly realized, leaving the central claim only partially substantiated.
  2. [Evaluation] Evaluation in ViTs and LLMs: The reported results appear limited to isolated modulator characterizations or small test conditions; no end-to-end pipeline measurements or noise-propagation analysis across multi-layer attention blocks, residual connections, and layer stacking are provided, which is load-bearing for the claim that TFLN units can be integrated into full transformers without unacceptable cumulative errors.
minor comments (1)
  1. [Abstract] The abstract would benefit from a concise statement of the exact functional mapping implemented by the TFLN MZM (e.g., how input voltage maps to the nonlinear output) to clarify the electro-optic Softmax approximation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We have revised the manuscript to address the concerns about quantitative support in the abstract and the scope of the evaluations, adding clarifications, comparisons, and analyses while maintaining an accurate representation of our experimental contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that the system 'maintains highly competitive accuracy' under 4-bit quantization is not supported by quantitative error bars, explicit baseline comparisons to digital implementations, or details on how the analog Softmax is exactly realized, leaving the central claim only partially substantiated.

    Authors: We agree that the abstract would benefit from stronger quantitative backing. In the revised manuscript, we have added error bars from repeated experimental runs to the accuracy results. We have also included explicit side-by-side comparisons to digital 4-bit Softmax baselines under identical quantization, showing our electro-optic approach remains within 1-2% accuracy. The methods section has been expanded with the precise electro-optic mapping (MZM transfer function approximation to Softmax) and calibration procedure used. revision: yes

  2. Referee: [Evaluation] Evaluation in ViTs and LLMs: The reported results appear limited to isolated modulator characterizations or small test conditions; no end-to-end pipeline measurements or noise-propagation analysis across multi-layer attention blocks, residual connections, and layer stacking are provided, which is load-bearing for the claim that TFLN units can be integrated into full transformers without unacceptable cumulative errors.

    Authors: We appreciate the emphasis on cumulative effects. Our work experimentally characterizes the individual TFLN nonlinear units and evaluates them within attention blocks under measured noise conditions up to 10 GBaud. In the revision, we have added a noise-propagation simulation that models error accumulation across stacked attention layers and residual connections, confirming robustness within the reported accuracy margins. Full end-to-end hardware measurements of a complete multi-layer pipeline exceed the scope of the current device-level demonstration and are identified as future work in the updated discussion. revision: partial

Circularity Check

0 steps flagged

No circularity: experimental hardware demonstration is self-contained

full rationale

The paper reports direct experimental measurements of TFLN MZM-based analog nonlinear units implementing electro-optic Softmax/Sigmoid approximations, with accuracy evaluated on ViTs and LLMs under 4-bit I/O quantization and noise characterized up to 10 GBaud. No mathematical derivation chain, fitted parameters renamed as predictions, or self-citation load-bearing steps exist; claims rest on hardware test results rather than any self-referential equations or uniqueness theorems. The work is therefore self-contained against external benchmarks with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work is an experimental demonstration relying on standard electro-optic device physics and transformer model definitions; no new free parameters, axioms, or invented entities are introduced in the abstract.

pith-pipeline@v0.9.0 · 5504 in / 1149 out tokens · 83570 ms · 2026-05-10T17:35:18.901100+00:00 · methodology

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

52 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    Attention is All you Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, L ukasz Kaiser, and Illia Polosukhin. Attention is All you Need. InProc. Neural Information Processing Systems (NeurIPS), 2017

    2017

  2. [2]

    Available: http://dx.doi.org/10.1038/s41586-025-09422-z

    Daya et al Guo. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.Nature, 645 (8081):633–638, 2025. doi:10.1038/s41586-025-09422-z

    work page doi:10.1038/s41586-025-09422-z 2025

  3. [3]

    An im- age is worth 16x16 words: Transformers for image recog- nition at scale

    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. An im- age is worth 16x16 words: Transformers for image recog- nition at scale. InProc. Int. Conf. on Learning Repre- sentations (ICLR), 2021

    2021

  4. [4]

    Theory, analysis, and best practices for sigmoid self-attention

    Jason Ramapuram, Federico Danieli, Eeshan Dhekane, Floris Weers, Dan Busbridge, Pierre Ablin, Tatiana Likhomanenko, Jagrit Digani, Zijin Gu, Amitis Shidani, and Russ Webb. Theory, analysis, and best practices for sigmoid self-attention. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  5. [5]

    cosformer: Rethinking softmax in attention

    Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In Proc. Int. Conf. on Learning Representations (ICLR), 2022

    2022

  6. [6]

    Rodriguez Condia, Juan-David Guerrero- Balaguera, Edwar J

    Josie E. Rodriguez Condia, Juan-David Guerrero- Balaguera, Edwar J. Pati˜ no N´ u˜ nez, Robert Limas, and Matteo Sonza Reorda. Investigating and Reducing the Architectural Impact of Transient Faults in Special Func- tion Units for GPUs.Journal of Electronic Testing, 40 (2):215–228, 2024. doi:10.1007/s10836-024-06107-9

    work page doi:10.1007/s10836-024-06107-9 2024

  7. [7]

    Flashattention- 3: fast and accurate attention with asynchrony and low- precision

    Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, and Tri Dao. Flashattention- 3: fast and accurate attention with asynchrony and low- precision. InProc. Neural Information Processing Sys- tems (NeurIPS), 2024

    2024

  8. [8]

    arrive remote, wait local

    Ted Zadouri, Markus Hoehnerbach, Jay Shah, Timmy Liu, Vijay Thakkar, and Tri Dao. Flashattention-4: Al- gorithm and kernel pipelining co-design for asymmetric hardware scaling. arXiv preprint, 2026. URLhttps: //arxiv.org/abs/2603.05451

    work page arXiv 2026

  9. [9]

    Scaling Laws for Neural Language Models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint, 2020. URL https://arxiv.org/abs/2001.08361

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  10. [10]

    Training compute-optimal large language models

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, DDL Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. InProc. Neural Information Processing Systems (NeurIPS), 2022

    2022

  11. [11]

    Stevens, Rangharajan Venkatesan, Steve Dai, Brucek Khailany, and Anand Raghunathan

    Jacob R. Stevens, Rangharajan Venkatesan, Steve Dai, Brucek Khailany, and Anand Raghunathan. Softer- max: Hardware/software co-design of an efficient soft- max for transformers. InProceedings of the 58th Annual ACM/IEEE Design Automation Conference, 2022

    2022

  12. [12]

    Sigmoid Self-Attention has Lower Sample Complexity than Softmax Self- Attention: A Mixture-of-Experts Perspective,

    Fanqi Yan, Huy Nguyen, Pedram Akbarian, Nhat Ho, and Alessandro Rinaldo. Sigmoid self-attention has lower sample complexity than softmax self-attention: A mixture-of-experts perspective. arXiv preprint, 2025. URLhttps://arxiv.org/abs/2502.00281

    work page arXiv 2025

  13. [13]

    URLhttps://docs.pytorch.org/docs/stable/ generated/torch.nn.attention.SDPBackend.html

    SDPBackend — PyTorch 2.11 documentation, 2026. URLhttps://docs.pytorch.org/docs/stable/ generated/torch.nn.attention.SDPBackend.html

    2026

  14. [14]

    Language models are unsupervised multitask learners, 2019

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners, 2019. URLhttps:// cdn.openai.com/better-language-models/language_ models_are_unsupervised_multitask_learners.pdf

    2019

  15. [15]

    H100 tensor core gpu, 2022

    NVIDIA. H100 tensor core gpu, 2022. URLhttps:// resources.nvidia.com/en-us-hopper-architecture/ nvidia-h100-tensor-c?ncid=no-ncid

    2022

  16. [16]

    PyTorch: An Impera- tive Style, High-Performance Deep Learning Library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, and Natalia Gimelshein. PyTorch: An Impera- tive Style, High-Performance Deep Learning Library. In Proc. Neural Information Processing Systems (NeurIPS), 2019

    2019

  17. [17]

    Fu, Stefano Ermon, Atri Rudra, and Christopher R´ e

    Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher R´ e. FlashAttention: Fast and memory- efficient exact attention with IO-awareness. InProc. Neu- ral Information Processing Systems (NeurIPS), 2022

    2022

  18. [18]

    VEXP: A Low-Cost RISC-V ISA Extension for Ac- celerated Softmax Computation in Transformers

    Run Wang, Gamze Islamoglu, Andrea Belano, and et al. VEXP: A Low-Cost RISC-V ISA Extension for Ac- celerated Softmax Computation in Transformers. In 2025 IEEE 32nd Symposium on Computer Arithmetic (ARITH), 2025

    2025

  19. [19]

    Understanding the Performance Horizon of the Latest ML Workloads with NonGEMM Workloads

    Rachid Karami, Sheng-Chun Kao, and Hyoukjun Kwon. Understanding the Performance Horizon of the Latest ML Workloads with NonGEMM Workloads. In2025 IEEE International Symposium on Performance Analy- sis of Systems and Software (ISPASS), May 2025

    2025

  20. [20]

    SOLE: Hardware-Software Co-design of Softmax and LayerNorm for Efficient Transformer In- ference

    Wenxun Wang, Shuchang Zhou, Wenyu Sun, Peiqin Sun, and Yongpan Liu. SOLE: Hardware-Software Co-design of Softmax and LayerNorm for Efficient Transformer In- ference. In2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), 2023

    2023

  21. [21]

    Hyft: A Reconfig- urable Softmax Accelerator with Hybrid Numeric For- mat for both Training and Inference

    Tianhua Xia and Sai Qian Zhang. Hyft: A Reconfig- urable Softmax Accelerator with Hybrid Numeric For- mat for both Training and Inference. InProceedings of the 29th ACM/IEEE International Symposium on Low Power Electronics and Design, 2024

    2024

  22. [22]

    FlashAttention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. InProc. Int. Conf. on Learning Representations (ICLR), 2024

    2024

  23. [23]

    Schraudolph

    Nicol N. Schraudolph. A Fast, Compact Approximation of the Exponential Function.Neural Computation, 1999

    1999

  24. [24]

    NN- LUT: neural approximation of non-linear operations for efficient transformer inference

    Joonsang Yu, Junki Park, Seongmin Park, Minsoo Kim, Sihwa Lee, Dong Hyun Lee, and Jungwook Choi. NN- LUT: neural approximation of non-linear operations for efficient transformer inference. InProceedings of the 59th ACM/IEEE Design Automation Conference. Association for Computing Machinery, 2022

    2022

  25. [25]

    PEANO-ViT: Power- Efficient Approximations of Non-Linearities in Vision Transformers

    Mohammad Erfan Sadeghi, Arash Fayyazi, Seyedarmin Azizi, and Massoud Pedram. PEANO-ViT: Power- Efficient Approximations of Non-Linearities in Vision Transformers. InProceedings of the 29th ACM/IEEE International Symposium on Low Power Electronics and Design. ACM, 2024

    2024

  26. [26]

    Analog in-memory computing atten- tion mechanism for fast and energy-efficient large lan- guage models.Nature Computational Science, 5(9):813– 824, 2025

    Nathan Leroux, Paul-Philipp Manea, Chirag Sudarshan, Jan Finkbeiner, Sebastian Siegel, John Paul Strachan, 12 and Emre Neftci. Analog in-memory computing atten- tion mechanism for fast and energy-efficient large lan- guage models.Nature Computational Science, 5(9):813– 824, 2025. doi:10.1038/s43588-025-00854-1

    work page doi:10.1038/s43588-025-00854-1 2025

  27. [27]

    Opto- electronic nonlinear Softmax operator based on diffrac- tive neural networks.Optics Express, 32(15):26458, 2024

    Ziyu Zhan, Hao Wang, Qiang Liu, and Xing Fu. Opto- electronic nonlinear Softmax operator based on diffrac- tive neural networks.Optics Express, 32(15):26458, 2024. doi:10.1364/OE.527843

    work page doi:10.1364/oe.527843 2024

  28. [28]

    SOFTONIC: A Photonic Design Approach to Softmax Activation for High-Speed Fully Analog AI Acceleration

    Priyabrata Dash, Anxiao Jiang, and Dharanidhar Dang. SOFTONIC: A Photonic Design Approach to Softmax Activation for High-Speed Fully Analog AI Acceleration. InProceedings of the Great Lakes Symposium on VLSI 2025, 2025

    2025

  29. [29]

    Photonic exponen- tial approximation via cascaded tfln microring resonators toward softmax

    Hyoseok Park and Yeonsang Park. Photonic exponen- tial approximation via cascaded tfln microring resonators toward softmax. arXiv preprint, 2026. URLhttps: //arxiv.org/abs/2603.12934

    work page arXiv 2026

  30. [30]

    A case study on the performance metrics of in- tegrated photonic computing

    Frank Br¨ uckerhoff-Pl¨ uckelmann, Jelle Dijkstra, Julian B¨ uchel, Bottyan Batkai, Falk Ebert, Luis Mickeler, Urs Egger, Abu Sebastian, Wolfram Pernice, and Ghazi Sar- wat Syed. A case study on the performance metrics of in- tegrated photonic computing. arXiv preprint, 2025. URL https://arxiv.org/abs/2511.00186

    work page arXiv 2025

  31. [31]

    Daniel Brunner et. al. Roadmap on neuromorphic pho- tonics. arXiv preprint, 2025. URLhttps://arxiv.org/ abs/2501.07917

    work page arXiv 2025

  32. [32]

    240 gbps high-efficiency opti- cal interconnection with tfln transmitter and ge-pd re- ceiver.Opt

    Zhipei Wang, Xuanhao Wang, Jinwen Song, Xu Wang, Aoxue Wang, Shuai Yuan, Xiao Hu, Fangchen Hu, Hai- wen Cai, and Wei Chu. 240 gbps high-efficiency opti- cal interconnection with tfln transmitter and ge-pd re- ceiver.Opt. Lett., 50(21):6469–6472, Nov 2025. doi: 10.1364/OL.575339

    work page doi:10.1364/ol.575339 2025

  33. [33]

    Chapman, Giovanni Finco, Tristan Kut- tner, Andreas Maeder, and Rachel Grange

    Alessandra Sabatti, Jost Kellner, Fabian Kaufmann, Robert J. Chapman, Giovanni Finco, Tristan Kut- tner, Andreas Maeder, and Rachel Grange. Extremely high extinction ratio electro-optic modulator via fre- quency upconversion to visible wavelengths.Optics Let- ters, 49(14):3870–3873, 2024. doi:10.1364/OL.525733. URLhttps://opg.optica.org/ol/abstract.cfm?ur...

    work page doi:10.1364/ol.525733 2024

  34. [34]

    Redeposition-free inductively-coupled plasma etching of lithium niobate for integrated pho- tonics.Nanophotonics, 12(8):1601–1611, 2023

    Fabian Kaufmann, Giovanni Finco, Andreas Maeder, and Rachel Grange. Redeposition-free inductively-coupled plasma etching of lithium niobate for integrated pho- tonics.Nanophotonics, 12(8):1601–1611, 2023. doi: doi:10.1515/nanoph-2022-0676

    work page doi:10.1515/nanoph-2022-0676 2023

  35. [35]

    Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. Mnist handwritten digit database, 2010. URL http://yann.lecun.com/exdb/mnist/

    2010

  36. [36]

    Learning multiple layers of fea- tures from tiny images

    Alex Krizhevsky. Learning multiple layers of fea- tures from tiny images. Technical report, University of Toronto, 2009. URLhttps://www.cs.toronto.edu/ ~kriz/cifar.html

    2009

  37. [37]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis- sacco, Bo Wu, and Andrew Y. Ng. Reading digits in nat- ural images with unsupervised feature learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011

    2011

  38. [38]

    Fineweb-edu: the finest collection of educational content, 2024

    Anton Lozhkov, Loubna Ben Allal, Leandro von Werra, and Thomas Wolf. Fineweb-edu: the finest collection of educational content, 2024. URLhttps://huggingface. co/datasets/HuggingFaceFW/fineweb-edu

    2024

  39. [39]

    I-bert: Integer-only bert quan- tization.Proc

    Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W Ma- honey, and Kurt Keutzer. I-bert: Integer-only bert quan- tization.Proc. Int. Conf. on Machine Learning (ICML), 2021

    2021

  40. [40]

    Softmax Bias Correction for Quan- tized Generative Models

    Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax Bias Correction for Quan- tized Generative Models. In2023 IEEE/CVF Interna- tional Conference on Computer Vision Workshops (IC- CVW), 2023

    2023

  41. [41]

    Machine behaviour

    Nayem Al-Kayed, Charles St-Arnault, Hugh Morison, A. Aadhi, Chaoran Huang, Alexander N. Tait, David V. Plant, and Bhavin J. Shastri. Programmable 200 GOPS Hopfield-inspired photonic Ising machine.Nature, 648 (8094):576–584, December 2025. doi:10.1038/s41586- 025-09838-7

    work page doi:10.1038/s41586- 2025

  42. [42]

    Analog implementation of the softmax function

    Jacob Sillman. Analog implementation of the softmax function. arXiv preprint, 2023. URLhttps://arxiv. org/abs/2305.13649

    work page arXiv 2023

  43. [43]

    firwin — SciPy v1.17.0 Manual, 2026

    SciPy. firwin — SciPy v1.17.0 Manual, 2026. URLhttps://docs.scipy.org/doc/scipy/reference/ generated/scipy.signal.firwin.html

    2026

  44. [44]

    CTL 1550, 2026

    TOPTICA Photonics. CTL 1550, 2026. URLhttps://www.toptica.com/products/ narrow-linewidth-lasers/ctl

    2026

  45. [45]

    E36106A DC power sup- ply, 100V, 0.4A, 40W [Obsolete], 2026

    Keysight Technologies. E36106A DC power sup- ply, 100V, 0.4A, 40W [Obsolete], 2026. URL https://www.keysight.com/us/en/product/E36106A/ dc-power-supply-100v-0-4a-40w.html

    2026

  46. [46]

    Jiang.High Data Rate DMT SERDES Design

    Z. Jiang.High Data Rate DMT SERDES Design. PhD thesis, Carleton University, Ottawa, Ontario, 2022

    2022

  47. [47]

    A 128 gb/s, 11.2 mw single-ended pam4 linear tia with 2.7µarms input noise in 22 nm finfet cmos

    Saeid Daneshgar, Hao Li, Taehwan Kim, and Ganesh Balamurugan. A 128 gb/s, 11.2 mw single-ended pam4 linear tia with 2.7µarms input noise in 22 nm finfet cmos. IEEE Journal of Solid-State Circuits, 57(5):1397–1408,

  48. [48]

    AUTHOR CONTRIBUTIONS COMPETING INTEREST The authors declare no competing financial or nonfinan- cial interests

    doi:10.1109/JSSC.2022.3147467. AUTHOR CONTRIBUTIONS COMPETING INTEREST The authors declare no competing financial or nonfinan- cial interests. ACKNOWLEDGMENTS This work was supported by the Swiss National Science Foundation SNSF Consolidator Grant APIC (TMCG- 2 213713), by Sinergia LION (CRII5-216600), and by the European Union’s Horizon Europe research a...

    work page doi:10.1109/jssc.2022.3147467 2022

  49. [49]

    Setup and Measurement We perform high-speed measurements of a Mach-Zehnder modulator (MZM) response to compare the experiment with the simulations used within Transformer training. Fig. S1a illustrates the experimental setup. BothOpt- maxandOptmoidare designed to perform a nonlinear transform on a sequence of valuesx i withi∈[1, n]. To test this, a 5-bit ...

    2048

  50. [50]

    S1b displays the response of a uniformly-sampled 10 GBaud 4-bit sequences atV π at three stages: (1) DAC output, (2) RF amplifier output, and (3) photo- diode output

    Experimental Noise Fig. S1b displays the response of a uniformly-sampled 10 GBaud 4-bit sequences atV π at three stages: (1) DAC output, (2) RF amplifier output, and (3) photo- diode output. We observe additive noise introduced by the RF amplifier, which is most pronounced at the MZM quadrature point where the voltage-to-power transcon- ductance is maximi...

  51. [51]

    The input encoding range [x min, xmax], which maps onto the rising slope of the first MZM’s transfer function

  52. [52]

    The normalization range [z min, zmax], which maps the accumulated optical power onto the falling slope of the second MZM. To translate from the digital simulation domain w∈[w min, wmax] to the physical voltage inputV∈ [Vmin, Vmax] driving the MZM, we define a fixed, range- preserving affine transformation: V(w) =γw+δ,(S22) where the scaling factorγand the...

    work page arXiv 2048