Pith. sign in

REVIEW 2 major objections 3 minor 63 references

Beyond Peak TOPS/W: A System-Level Perspective on Hybrid Digital, Analogue and Neuromorphic Computing

T0 review · 2 major / 3 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 Hybrid digital–analogue AI accelerators earn their place only when they deliver a measurable whole-task energy or latency gain over a strong digital baseline, not when they post high peak TOPS/W.

desk verdict A clear, honest perspective that codifies the emerging consensus against peak TOPS/W; the central criterion is underdetermined, but as a synthesis it earns a serious referee. read the letter →

arxiv 2608.03514 v1 pith:Q6IO725O submitted 2026-08-04 cs.AR

Eiman Kanjo , Varuna De Silva This is my paper
classification cs.AR
keywords hybriddigital-analoguecomputinganaloguein-memoryphotonicneuromorphicTOPS/Wsystem-levelbenchmarkingenergyefficiencyedgeAI
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

The paper sets out to change how energy efficiency of AI hardware is judged. Its central proposition is that analogue, photonic and neuromorphic accelerators should not be treated as replacements for digital computing, but as specialised engines that earn a place only when they deliver a measurable end-to-end improvement over a strong, optimised digital baseline on a well-matched workload. It argues that peak TOPS/W at the core cannot arbitrate between substrates, because conversion, data movement, calibration, control and host processing can erase a physical core's advantage. A sympathetic reader would care because the same logic applies to procurement, benchmarking and research priorities: if accepted, evaluation shifts from isolated accelerator metrics to deployed-task metrics, and hybrid systems are judged by whether their physical specialisation survives the full system boundary.

What carries the argument

The load-bearing machinery is an accounting boundary around the whole task, captured by the energy decomposition E_total = E_sensing + E_compute + E_memory + E_control + E_radio + E_actuation + E_idle + E_calibration, and by the eight-coordinate benchmark bundle B. For analogue in-memory cores the relevant decomposition includes the array plus output conversion, input driving, buffering, control, routing and calibration; for photonic systems it includes light generation, modulation, detection and conversion; for neuromorphic systems it includes baseline power, event routing, state updates and host I/O. These identities do not predict a winner; they define what must be measured before a physi

What would settle it

A concrete check: take one representative edge workload, such as keyword spotting or event-camera object detection, run it on a fully integrated hybrid system and on an optimised digital implementation at matched accuracy and latency, and measure full task energy including conversion, calibration amortised over a deployment window, memory traffic, idle power and host processing. If the hybrid system shows no measurable task-energy or latency advantage whenever the digital baseline is improved by the same team, the paper's decision rule loses its power to arbitrate between digital and hybrid de

Watch

Extended reading notes

Core claim

The central claim is a criterion, not a new device: the value of any physical computing technology—analogue in-memory arrays, photonic processors, neuromorphic chips—is a property of the whole deployed system, not of the core. A hybrid architecture is justified exactly when a physical substrate provides a measurable end-to-end benefit in task energy or latency relative to a strong digital baseline for workloads whose structure matches the substrate, once input encoding, output conversion, routing, buffering, calibration, correction and digital control are counted. In practice this means reporting three levels of results—the specialised core, the complete hybrid system, and an optimised digit

Load-bearing premise

The criterion only works if a fair, reproducible 'strong digital baseline' can be defined at matched task accuracy, latency and operating conditions; compiler, precision, batch-size and measurement-boundary choices all change that baseline, so the whole comparison is only as solid as the chosen baseline.

Editorial extensions

If this is right

  • Hardware studies should report at least three numbers: the specialised core, the complete hybrid system, and an optimised digital baseline at matched task accuracy and latency.
  • Energy and latency claims that exclude conversion, calibration, host processing, memory movement and idle power cannot be compared across papers as if they measured the same thing.
  • Hybrid adoption is likely to be incremental and workload-specific, starting with repeated matrix operations, sparse temporal streams and always-on sensing rather than general-purpose replacement.
  • Digital orchestration—scheduling, memory management, calibration, correction and fallback—is a required part of any hybrid design that wants its physical advantage to survive the full system boundary.

Reading between the lines

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

  • The authors do not say this explicitly, but the same task-level test applies to purely digital accelerators: a digital chip should also be judged against a strong optimised baseline, so 'hybrid versus digital' becomes one case of a general end-to-end evaluation standard.
  • The bundle B has no scalar score and no stated way to trade off its coordinates; a natural next step the paper leaves open is an application-specific cost function (for example, energy-delay product or cost per completed task) to make decisions under the bundle.
  • The paper's measurement-boundary argument suggests a testable extension: benchmark reports could include a short machine-readable 'boundary manifest' listing which terms from the energy decomposition are included, which would let readers redistribute results across definitions without rerunning experiments.
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

2 major / 3 minor

Summary. This perspective paper argues that hybrid digital-analogue, photonic, and neuromorphic computing should be understood as specialized engines hosted and orchestrated by a digital system, and that their value must be demonstrated through measurable end-to-end benefits relative to an optimised digital baseline at matched task accuracy and operating conditions. The authors review analogue in-memory, photonic, and neuromorphic hardware, characterize system-level overheads (conversion, calibration, routing, host processing), and propose a multi-metric benchmark bundle B={E_task, L_task, A_task, M_move, C_cal, R_field, P_prog, C_deploy} to replace core-level TOPS/W claims. The paper is a synthesis and position statement; it does not present new experimental measurements, and its equations are explicitly scoped as accounting boundaries rather than predictive models.

Significance. If adopted, the proposed system-level evaluation framework would materially improve the quality of cross-technology comparisons in AI accelerator research, since current peak-TOPS/W metrics are, as the paper shows, not comparable across different measurement boundaries. The manuscript consolidates a broad literature and provides a useful taxonomy of hybrid design patterns (sensor-first front-end, controller/accelerator split, coarse/fine split, correction feedback, verification/fallback) together with a clear checklist of overheads that should be included in fair evaluations. Its strengths include the explicit recognition of the digital baseline, the system-level energy accounting equations, and the recommendation to report a multi-coordinate bundle rather than a single scalar. The main weakness is that the central 'measurable end-to-end benefit' criterion is not yet operationalizable because the baseline is underspecified and the B-bundle lacks a comparison rule; this limits the framework's immediate actionability but does not undermine the paper's broader argument.

major comments (2)
  1. [§5.4 (condition 4) and §7.1 (Eq. 21)] The central claim is underdetermined. Section 7.1 explicitly says the B-bundle 'is not a scalar score: its coordinates have different units and should not be collapsed into arbitrary weights,' but no comparison relation or decision rule is provided. Consequently, the condition 'end-to-end system advantage' in §5.4 cannot be operationally applied whenever a hybrid system improves some B-coordinates and worsens others (e.g., lower E_task but higher C_cal or worse R_field). This is load-bearing because the paper's central proposition defines value through this condition. Please specify a concrete protocol: at matched task performance, report all B coordinates for both the hybrid system and the optimised digital baseline, and define a decision rule—for example, Pareto dominance with stated tolerances, a transparent multi-criteria analysis with weights justified by the application, or a pre-r
  2. [§2.1 and §5.4] The 'strong digital baseline' is not reproducible. The text repeatedly states that comparisons should be against 'an optimised digital implementation rather than a general-purpose or outdated baseline,' but does not define how that baseline is constructed: same process node, same numerical precision, same batch size, same compiler/runtime, same measurement boundary? A weak baseline would make nearly any hybrid system appear beneficial; an unrealistically strong baseline would make all hybrids fail. Since the central claim turns on this comparison, provide explicit baseline construction guidelines—for instance, using a state-of-the-art digital accelerator on the same process node, with matched task accuracy and operating conditions, and reporting the baseline with the same B-vector coordinates. Reference [41] (NeuroBench) could serve as a starting point for this protocol.
minor comments (3)
  1. [§7.1, Eq. (21)] The equation 'B={E task,L task,A task,M move,C cal,R field,P prog,C deploy}' has missing spaces after commas; add spaces for readability.
  2. [Reference [29]] The reference 'W. B. et al., “Programmable photonic circuits,” Nature' is incomplete; the author list should be expanded to 'W. Bogaerts et al.' (or the actual author names).
  3. [§7.1, Table 1] The table effectively demonstrates that published numbers in the literature are not comparable, but the paper does not show how the proposed B-bundle would be used in a concrete comparison. A worked example, even stylized, would strengthen the actionability of the framework.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the paper is a framework/position argument, not a derivation whose output is equivalent to its input.

full rationale

The paper does not fit parameters, generate quantitative predictions, or import a uniqueness theorem from prior work; its central claim is a normative evaluation standard. Section 1 states that hybrid architectures are valuable only when they deliver measurable end-to-end benefits relative to a strong digital baseline, and Section 5.4 restates this same criterion (workload compatibility, sufficient activity, acceptable accuracy, end-to-end system advantage). This is a repeated assertion of a proposed criterion rather than a reduction of a derived result to its own inputs. The energy decompositions (Eqs. 5, 11, 16, 19, 20) are explicitly introduced as accounting boundaries rather than predictive models, so no fitted input is later relabelled as a prediction. The evaluation bundle B in Eq. (21) is explicitly non-scalar and the paper refuses to collapse it into arbitrary weights; this makes the proposed benchmark underdetermined, which is a testability weakness, not circularity. The authors cite several of their own works (e.g., [18], [20], [37], [43]), but these function as illustrative examples or background reviews, and the central argument does not reduce to them. No equation or claimed result is shown to be equivalent, by construction, to a definition or to a self-citation. Therefore no circular step meeting the required quote-and-reduction standard can be identified.

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

The paper introduces no free parameters and no invented entities. Its equations are explicitly framed as accounting identities (not predictive models), so the main background assumptions are the standard device models it borrows from the literature and the additivity of its energy decompositions.

assumptions (4)
  • domain assumption Ideal crossbar model (Eq. 1): column current equals sum of G_ij * V_i, assuming linear ohmic summation.
    The paper itself flags this as ideal and later adds non-ideality terms (Eq. 3), so the linear model is an assumption used to define AIMC operation (Section 2.2.1).
  • domain assumption Photonic transformations are represented as linear complex matrices W_opt (Eq. 6), with intensity detection in Eq. 7.
    The argument that photonic cores suit linear algebra assumes ideal linear optical networks and ignores nonlinearities, which the paper acknowledges in the 'Nonlinear operations' part of Section 2.2.2.
  • domain assumption Leaky integrate-and-fire (LIF) neuron model (Eq. 12-15) as the representative neuromorphic abstraction.
    The paper notes that LIF is one possible abstraction and others exist (Section 3.1), but its activity-energy discussion (Eq. 16) uses this model as the baseline.
  • domain assumption The energy decompositions (Eqs. 5, 11, 16, 20) are additive and complete enough to capture the relevant system costs.
    The paper repeatedly calls these 'accounting boundaries rather than universal device models', meaning it assumes that the listed terms cover the dominant energy contributions and that they can be treated as separable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Peak TOPS/W: A System-Level Perspective on Hybrid Digital, Analogue and Neuromorphic Computing." pith.science (2026). https://pith.science/paper/Q6IO725O

@misc{pith2026260803514,
  author       = {Pith},
  title        = {Pith review of: Beyond Peak TOPS/W: A System-Level Perspective on Hybrid Digital, Analogue and Neuromorphic Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q6IO725O}},
  note         = {Machine review of arXiv:2608.03514}
}
read the original abstract

The digital revolution, which progressively replaced analogue methods with digital circuits, has entered a new phase as AI expands across cloud infrastructure, mobile networks, wearables and physical systems, including drones and robots. Digital computing remains the general-purpose foundation of this expansion: it supports heterogeneous, on-device and decentralised AI through programmable control, mature software and decades of accumulated engineering infrastructure. Yet as energy and data-movement constraints become more significant, that same foundation is increasingly being extended rather than replaced by selected analogue and physical principles that it can host, configure and verify. Photonic, in-memory and neuromorphic architectures offer routes to reducing data movement and accelerating matrix-intensive and event-driven processing, not as alternatives to digital infrastructure but as specialised engines operating within it. This paper argues that hybrid digital--analogue computing represents a credible pathway towards more energy-efficient AI systems: one in which physical substrates earn an expanding role only where they deliver a measurable system-level advantage, under digital orchestration that manages integration, uncertainty and fallback. It examines the architectural principles, workload suitability, energy accounting, software requirements, limitations and open challenges associated with this transition, and argues that future progress should be evaluated through deployed-system metrics rather than isolated peak tera operations per second per watt (TOPS/W) claims.

Figures

Figures reproduced from arXiv: 2608.03514 by the authors.

Figure 1
Figure 1. Digital computing as the system-level baseline. and quantisation-aware fine-tuning can reduce the cost of adapting and storing model parameters. Quantisation Aware - Low Rank Adaptation (QA-LoRA), for example, combines low-rank adaptation with quantisation-aware training to support low-bit deployment [8]. T-MAC reformulates low-bit matrix multiplication as table-lookup operations and demonstrates how software–hardwa… view at source ↗
Figure 2
Figure 2. Analogue in-memory computing: physical operation and system overheads. Conversion and peripheral overhead An AIMC tile requires more than the conductance array. It may include input drivers or digital-to-analogue converters, row and column selection circuits, sample-and-hold stages, current sensing, analogue-to-digital converters, multiplexers, buffers, routing and digital control. Depending on precision, throughput… view at source ↗
Figure 3
Figure 3. System-level operation of photonic computing. Electronic inputs modulate optical carriers that pass through a configurable interferometric network to perform a linear transformation. Photodetection, data conversion and digital nonlinear processing remain necessary. Optical loss, fabrication and phase errors, thermal drift and conversion overhead affect the realised system-level performance. Yˆ = f(Qb[D (WoptEin)]), … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: System-level view of neuromorphic computing. Sparse input events are encoded and routed through a network of stateful neurons, where membrane states integrate, leak, cross a threshold and reset. Sparse activity can reduce computation and communication, but encoding, st…
Figure 5
Figure 5. Figure 5: Representative technologies and circuit roles associated with physical–digital integration. Inputs may be supplied by a digital system or arrive directly as sensor, electrical or optical signals. Physical processing may occur after digital encoding or before digitisati…
Figure 6
Figure 6. Figure 6: Relationship between digital, physical and neuromorphic computing. Neuromorphic systems may use digital, analogue or mixed-signal hardware. Mixed-signal describes the integration of physical signal processing with digital conversion, control or correction. Hybrid compu…
Figure 7
Figure 7. Figure 7: Analogue acceleration as one hybrid design pattern. A digital host prepares and schedules the workload, delegates suitable operations to an analogue accelerator and retains unsupported processing, orchestration and output verification. These layers define functional re…
Figure 8
Figure 8. Figure 8: Common hybrid digital–analogue design patterns. Physical processing may occur before digitisation, operate as a digitally controlled accelerator, generate an approximate result for digital refinement, receive digital correction feedback or produce outputs that are digi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 52 canonical work pages

  1. [41]

    The neurobench framework for benchmarking neuromorphic computing algorithms and systems,

    J. Yiket al., “The neurobench framework for benchmarking neuromorphic computing algorithms and systems,”Nature Communications, vol. 16, p. 1545, 2025

    work page 2025

  2. [1]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” inNIPS Deep Learning and Representation Learning Workshop, 2015

    2015

  3. [2]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,”International Journal of Computer Vision, vol. 129, pp. 1789–1819, 2021

    work page 2021

  4. [3]

    Domain-specific hardware accelerators,

    W. J. Dally, Y. Turakhia, and S. Han, “Domain-specific hardware accelerators,”Communications of the ACM, vol. 63, no. 7, pp. 48–57, 2020

    work page 2020

  5. [4]

    Computing’s energy problem (and what we can do about it),

    M. Horowitz, “Computing’s energy problem (and what we can do about it),” inProceedings of the IEEE International Solid-State Circuits Conference, 2014, pp. 10–14

    work page 2014

  6. [5]

    Efficient processing of deep neural networks: A tutorial and survey,

    V. Sze, Y.-H. Chen, T.-J. Yang, and J. S. Emer, “Efficient processing of deep neural networks: A tutorial and survey,”Proceedings of the IEEE, vol. 105, no. 12, pp. 2295–2329, 2017

    work page 2017

  7. [6]

    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,” inProceedings of the International Conference on Learning Representations (ICLR), 2016

    work page 2016

  8. [7]

    SparseGPT: Massive language models can be accurately pruned in one-shot,

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “SparseGPT: Massive language models can be accurately pruned in one-shot,” inProceedings of the 40th International Conference on Machine Learning (ICML), 2023

    work page 2023

Show all 63 references
  1. [8]

    QA-LoRA: Quantization-aware low-rank adaptation of large language models,

    Y. Xu, L. Xie, X. Gu, X. Chen, H. Chang, H. Zhang, Z. Chen, X. Zhang, and Q. Tian, “QA-LoRA: Quantization-aware low-rank adaptation of large language models,”arXiv preprint arXiv:2309.14717, 2023

    2023 arXiv

  2. [9]

    T-MAC: CPU renaissance via table lookup for low-bit LLM deployment on edge,

    J. Wei, S. Cao, T. Cao, L. Ma, L. Wang, Y. Zhang, and M. Yang, “T-MAC: CPU renaissance via table lookup for low-bit LLM deployment on edge,”arXiv preprint arXiv:2407.00088, 2024

    2024 arXiv

  3. [10]

    Mr. wolf: An energy-precision scalable parallel ultra low power SoC for IoT edge processing,

    A. Pullini, F. Conti, D. Rossi, I. Loi, A. Di Mauro, and L. Benini, “Mr. wolf: An energy-precision scalable parallel ultra low power SoC for IoT edge processing,”IEEE Journal of Solid-State Circuits, vol. 54, no. 7, pp. 1970–1981, 2019

    1970

  4. [11]

    MCUNet: Tiny deep learning on IoT devices,

    J. Lin, W.-M. Chen, Y. Lin, J. Cohn, C. Gan, and S. Han, “MCUNet: Tiny deep learning on IoT devices,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 11711–11722

    2020

  5. [12]

    MCUNetV2: Memory-efficient patch-based inference for tiny deep learning,

    J. Lin, W.-M. Chen, H. Cai, C. Gan, and S. Han, “MCUNetV2: Memory-efficient patch-based inference for tiny deep learning,” inAdvances in Neural Information Processing Systems, 2021

    2021

  6. [13]

    Combining multiple tiny machine learning models for multimodal context-aware stress recognition on constrained microcontrollers,

    M. Gibbs, K. Woodward, and E. Kanjo, “Combining multiple tiny machine learning models for multimodal context-aware stress recognition on constrained microcontrollers,”IEEE Micro, vol. 44, no. 3, pp. 54–62, 2023

    2023

  7. [14]

    TinyOL: TinyML with online learning on microcontrollers,

    H. Ren, D. Anicic, and T. Runkler, “TinyOL: TinyML with online learning on microcontrollers,” inProceedings of the International Joint Conference on Neural Networks (IJCNN), 2021, pp. 1–8

    2021

  8. [15]

    Review of movement sensor applications in livestock animal activity recognition: Communications, data collection practices, and edge-ai solutions,

    B. C. Patrick, E. Kanjo, and O. Kaiwartya, “Review of movement sensor applications in livestock animal activity recognition: Communications, data collection practices, and edge-ai solutions,” Smart Agricultural Technology, vol. 14, p. 100412, 2026

    2026

  9. [16]

    A multicore and edge tpu-accelerated multimodal tinyml system for livestock behavior recognition,

    Q. Zhang and E. Kanjo, “A multicore and edge tpu-accelerated multimodal tinyml system for livestock behavior recognition,”IEEE Transactions on Mobile Computing, vol. 24, no. 10, pp. 5412–5425, 2025. 22

    2025

  10. [17]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural Networks, vol. 113, pp. 54–71, 2019

    2019

  11. [18]

    A hybrid edge classifier: Combining tinyml-optimised cnn with rram-cmos acam for energy-efficient inference,

    K. Woodward, E. Kanjo, G. Papandroulidakis, S. Agwa, and T. Prodromakis, “A hybrid edge classifier: Combining tinyml-optimised cnn with rram-cmos acam for energy-efficient inference,” IEEE Transactions on Knowledge and Data Engineering, vol. 38, no. 4, pp. 2122–2135, 2026

    2026

  12. [19]

    Internetless low-cost sensing system for real-time livestock monitoring,

    B. Patrick, T. Johnson, and E. Kanjo, “Internetless low-cost sensing system for real-time livestock monitoring,”IEEE Wireless Communications Letters, vol. 13, no. 5, pp. 1284–1288, 2024

    2024

  13. [20]

    Node learning: A framework for adaptive, decentralised and collaborative network edge ai,

    E. Kanjo and M. Aslanov, “Node learning: A framework for adaptive, decentralised and collaborative network edge ai,”arXiv preprint arXiv:2602.16814, 2026

    2026

  14. [21]

    Enabling fast deep learning on tiny energy-harvesting IoT devices,

    S. Islam, J. Deng, S. Zhou, C. Pan, C. Ding, and M. Xie, “Enabling fast deep learning on tiny energy-harvesting IoT devices,”arXiv preprint arXiv:2105.01003, 2021

    2021 arXiv

  15. [22]

    Communication-efficient learning of deep networks from decentralized data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017, pp. 1273–1282

    2017

  16. [23]

    Memory devices and applications for in-memory computing,

    A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, and E. Eleftheriou, “Memory devices and applications for in-memory computing,”Nature Nanotechnology, vol. 15, pp. 529–544, 2020

    2020

  17. [24]

    Mixed-precision in-memory computing,

    M. Le Gallo, A. Sebastian, G. Cherubini, H. Giefers, T. Tuma, and E. Eleftheriou, “Mixed-precision in-memory computing,”Nature Electronics, vol. 1, no. 9, pp. 511–519, 2018

    2018

  18. [25]

    An optimal design for universal multiport interferometers,

    W. R. Clements, P. C. Humphreys, B. J. Metcalf, W. S. Kolthammer, and I. A. Walmsley, “An optimal design for universal multiport interferometers,”Optica, vol. 3, no. 12, pp. 1460–1465, 2016

    2016

  19. [26]

    The future of electronics based on memristive systems,

    M. A. Zidan, J. P. Strachan, and W. D. Lu, “The future of electronics based on memristive systems,”Nature Electronics, vol. 1, pp. 22–29, 2018

    2018

  20. [27]

    A phase-change memory model for neuromorphic computing,

    S. R. Nandakumar, M. Le Gallo, I. Boybat, B. Rajendran, A. Sebastian, and E. Eleftheriou, “A phase-change memory model for neuromorphic computing,”Journal of Applied Physics, vol. 124, no. 15, p. 152135, 2018

    2018

  21. [28]

    Optical neural networks: progress and challenges,

    T. Fu, J. Zhang, R. Sun, Y. Huang, W. Xu, S. Yang, Z. Zhu, and H. Chen, “Optical neural networks: progress and challenges,”Light: Science & Applications, vol. 13, p. 263, 2024

    2024

  22. [29]

    Programmable photonic circuits,

    W. B. et al., “Programmable photonic circuits,”Nature, vol. 586, no. 7830, pp. 520–530, October 2020

    2020

  23. [30]

    Photonic neural networks: Principles, progress, and prospects toward streaming processes,

    G. Cong and S. Namiki, “Photonic neural networks: Principles, progress, and prospects toward streaming processes,”J. Lightwave Technol., vol. 44, no. 3, pp. 821–844, Feb 2026. [Online]. Available: https://opg.optica.org/jlt/abstract.cfm?URI=jlt-44-3-821

    2026

  24. [31]

    Parallel convolutional processing using an integrated photonic tensor core,

    J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. Le Gallo, X. Fu, A. Lukashchuk, A. S. Raja, J. Liu, C. D. Wright, A. Sebastian, T. J. Kippenberg, W. H. P. Pernice, and H. Bhaskaran, “Parallel convolutional processing using an integrated photonic tenso...

    2021

  25. [32]

    An integrated large-scale photonic accelerator with ultralow latency,

    S. Hua, E. Divita, S. Yu, B. Peng, C. Roques-Carmes, Z. Su, Z. Chen, Y. Bai, J. Zou, Y. Zhu, Y. Xu, C.-k. C. Lu, Y. Di, H. Chen, L. Jiang, L. Wang, L. Ou, C. Zhang, J. Chen, W. Zhang, H. Zhu, W. Kuang, L. Wang, H. Meng, M. Steinman, and Y. Shen, “An integrated large-scale phot...

    2025 doi

  26. [33]

    Prospects of photonic computing for artificial intelligence,

    H. Zhou and J. Dong, “Prospects of photonic computing for artificial intelligence,”National Science Review, 02 2026. [Online]. Available: https://doi.org/10.1093/nsr/nwag089

    2026 doi

  27. [34]

    Photonic neural networks based on integrated silicon microresonators,

    S. Biasi, G. Donati, A. Lugnan, M. Mancinelli, E. Staffoli, and L. Pavesi, “Photonic neural networks based on integrated silicon microresonators,”arXiv preprint arXiv:2306.04779, 2023

    2023 arXiv

  28. [35]

    Neuromorphic silicon neuron circuits,

    G. Indiveri, B. Linares-Barranco, T. J. Hamilton, A. van Schaik, R. Etienne-Cummings, T. Delbruck, S.-C. Liu, P. Dudek, P. Häfliger, S. Renaudet al., “Neuromorphic silicon neuron circuits,”Frontiers in Neuroscience, vol. 5, p. 73, 2011

    2011

  29. [36]

    Neuromorphic computing at scale,

    D. Kudithipudiet al., “Neuromorphic computing at scale,”Nature, vol. 637, pp. 801–812, 2025

    2025

  30. [37]

    Advancements in neuromorphic computing for bio-inspired artificial vision: A review,

    S. A. Gabayre, M. Illeperuma, V. De-Silva, X. Shi, and S. E. Savel’ev, “Advancements in neuromorphic computing for bio-inspired artificial vision: A review,”Neurocomputing, vol. 653, p. 131221, 2025

    2025

  31. [38]

    Adaptive exponential integrate-and-fire model as an effective description of neuronal activity,

    R. Brette and W. Gerstner, “Adaptive exponential integrate-and-fire model as an effective description of neuronal activity,”Journal of Neurophysiology, vol. 94, no. 5, pp. 3637–3642, 2005

    2005

  32. [39]

    Resonate-and-fire neurons,

    E. M. Izhikevich, “Resonate-and-fire neurons,”Neural Networks, vol. 14, no. 6–7, pp. 883–894, 2001

    2001

  33. [40]

    Gerstner, W

    W. Gerstner, W. M. Kistler, R. Naud, and L. Paninski,Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition. Cambridge University Press, 2014

    2014

  34. [42]

    Benchmarking neuromorphic hardware and its energy expenditure,

    C. Ostrau, C. Klarhorst, M. Thies, and U. Rückert, “Benchmarking neuromorphic hardware and its energy expenditure,”Frontiers in Neuroscience, vol. 16, p. 873935, 2022

    2022

  35. [43]

    Novel directions for neuromorphic machine intelligence guided by functional connectivity: A review,

    M. Illeperuma, R. Pina, V. D. Silva, and X. Liu, “Novel directions for neuromorphic machine intelligence guided by functional connectivity: A review,”Machines, vol. 12, no. 8, p. 574, 2024

    2024

  36. [44]

    Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

    2019

  37. [45]

    Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing,

    J. E. Pedersenet al., “Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing,”Nature Communications, vol. 15, p. 8122, 2024

    2024

  38. [46]

    The SpiNNaker project,

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

    2014

  39. [47]

    SpiNNaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,

    H. A. Gonzalez, J. Huang, F. Kelber, K. K. Nazeer, T. Langer, C. Liu, M. Lohrmann, A. Rostami, M. Schöne, B. Vogginger, T. C. Wunderlich, Y. Yan, M. Akl, and C. Mayr, “SpiNNaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,” arXiv:2401....

    2024 arXiv

  40. [48]

    Loihi 2: An asynchronous neuromorphic processor for programmable spiking neural networks,

    G. Orchard, E. P. Frady, M. Davieset al., “Loihi 2: An asynchronous neuromorphic processor for programmable spiking neural networks,”Intel Technology Journal, vol. 25, no. 1, 2021

    2021

  41. [49]

    The BrainScaleS-2 accelerated neuromorphic system with hybrid plasticity,

    C. Pehle, S. Billaudelle, B. Cramer, J. Kaiser, K. Schreiber, Y. Stradmann, J. Weis, A. Leibfried, E. Müller, and J. Schemmel, “The BrainScaleS-2 accelerated neuromorphic system with hybrid plasticity,”Frontiers in Neuroscience, vol. 16, p. 795876, 2022

    2022

  42. [50]

    DYNAP-SE2: A scalable multi-core dynamic neuromorphic asynchronous spiking neural network processor,

    O. Richter, C. Wu, A. M. Whatley, G. Köstinger, C. Nielsen, N. Qiao, and G. Indiveri, “DYNAP-SE2: A scalable multi-core dynamic neuromorphic asynchronous spiking neural network processor,”Neuromorphic Computing and Engineering, vol. 4, no. 1, p. 014003, 2024. 24

    2024

  43. [51]

    Event-based vision: A survey,

    G. Gallego, T. Delbrück, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis, and D. Scaramuzza, “Event-based vision: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 154–180, 2022

    2022

  44. [52]

    Roadmapping the next generation of silicon photonics,

    S. Shekharet al., “Roadmapping the next generation of silicon photonics,”Nature Communications, vol. 15, p. 751, 2024

    2024

  45. [53]

    Once-for-all: Train one network and specialize it for efficient deployment,

    H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, “Once-for-all: Train one network and specialize it for efficient deployment,” inProceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  46. [54]

    Hardware implementation of memristor-based artificial neural networks,

    F. Aguirre, A. Sebastian, M. L. Gallo, W. Song, T. Wang, J. J. Yanget al., “Hardware implementation of memristor-based artificial neural networks,”Nature Communications, vol. 15, p. 1974, 2024

    1974

  47. [55]

    A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference,

    M. L. Gallo, R. Khaddam-Aljameh, M. Stanisavljevic, A. Vasilopouloset al., “A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference,”Nature Electronics, vol. 6, no. 9, pp. 680–693, 2023

    2023

  48. [56]

    An analog-AI chip for energy-efficient speech recognition and transcription,

    S. Ambrogio, P. Narayanan, A. Okazaki, A. Fasoli, C. Mackinet al., “An analog-AI chip for energy-efficient speech recognition and transcription,”Nature, vol. 620, pp. 768–775, 2023

    2023

  49. [57]

    Fully row/column-parallel in-memory computing SRAM macro employing capacitor-based mixed-signal computation with 5-b inputs,

    J. Lee, H. Valavi, Y. Tang, and N. Verma, “Fully row/column-parallel in-memory computing SRAM macro employing capacitor-based mixed-signal computation with 5-b inputs,” in2021 Symposium on VLSI Circuits, 2021

    2021

  50. [58]

    An evolution in power management with mythic’s analog matrix processor,

    Mythic, “An evolution in power management with mythic’s analog matrix processor,” https: //mythic.ai/whats-new/an-evolution-in-power-management-with-mythics-analog-matrix-pro cessor/, 2021, company technical resource; accessed 31 July 2026

    2021

  51. [59]

    Innatera showcases neuromorphic processor for ambient intelligence at CES 2025,

    Innatera Nanosystems, “Innatera showcases neuromorphic processor for ambient intelligence at CES 2025,” https://www.innatera.com/newsroom/innatera-showcases-revolutionary-neuromo rphic-processor-enabling-next-generation-ambient-intelligence-at-ces-2025/, 2025, company product ...

    2025

  52. [60]

    AnalogML technology,

    Aspinity, “AnalogML technology,” https://www.aspinity.com/technology/, 2024, company technical resource; accessed 31 July 2026

    2024

  53. [61]

    A complete no-brainer: ReRAM for neuromorphic computing,

    G. Piccolboni, “A complete no-brainer: ReRAM for neuromorphic computing,” https://www. weebit-nano.com/wp-content/uploads/2024/06/IMW2024-Weebit-technology-IP-company-e mbedded-RRAM-ReRAM-NVM-for-semiconductors-replacing-Flash-memory-Neuromorphi c-Computing.pdf, Weebit Nano, 2...

    2024

  54. [62]

    Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators,

    M. J. Rasch, C. Mackin, M. L. Galloet al., “Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators,”Nature Communications, vol. 14, p. 5282, 2023

    2023

  55. [63]

    Reaxion: A hardware evaluation framework for analog computing,

    A. Lottarini, A. Vasan, J. Wawrzynek, and K. Bhatt, “Reaxion: A hardware evaluation framework for analog computing,” inProceedings of the IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2021. 25

    2021

Pith tools

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