Pith. sign in

REVIEW 4 major objections 5 minor 36 references

SimPhony: A Device-Circuit-Architecture Cross-Layer Modeling and Simulation Framework for Heterogeneous Electronic-Photonic AI System

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SimPhony is an open-source cross-layer simulator that claims end-to-end latency, energy, and area estimates for heterogeneous electronic-photonic AI systems, from device models to chip floorplan.

desk verdict A solid open-source framework for EPIC AI simulation with a genuinely new PTC representation; validation is mostly against the authors' prior sims, so treat accuracy numbers as provisional. read the letter →

arxiv 2411.13715 v1 pith:5PKTVPTZ submitted 2024-11-20 physics.optics cs.AIcs.ARcs.ETcs.LG

Ziang Yin , Meng Zhang , Amir Begovic , Rena Huang , Jeff Zhang , Jiaqi Gu This is my paper · ORCID
classification physics.opticscs.AIcs.ARcs.ETcs.LG
keywords electronic-photonicintegratedcircuitsphotonictensorcorecross-layersimulationdataflowmappinglayout-awareareaestimationdata-awareenergymodelinglinkbudgetanalysisAIaccelerator
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

SimPhony is a cross-layer modeling and simulation framework for heterogeneous electronic-photonic AI systems. It claims to estimate end-to-end system latency, data-aware energy, and layout-aware chip area for arbitrary photonic tensor core designs, not just array-style accelerators. The paper argues that existing simulators cannot represent mesh-style or time-multiplexed photonic cores, ignore real workload values in power modeling, and sum device footprints instead of using actual layouts. SimPhony's value, if true, is a common open platform where device, circuit, and architecture researchers can evaluate and fairly compare EPIC AI hardware before fabrication.

What carries the argument

The core object is the hierarchical netlist built from a minimal building block called a node, with directed two-pin nets capturing optical signal flow and scaling rules that replicate the node into a full multi-core architecture. The netlist is converted into a weighted directed acyclic graph whose longest path gives the critical insertion loss and whose topological order guides the layout-aware floorplanner. This representation carries the whole argument: it unifies PTC topologies, enables parametric architecture construction, supplies the link-budget and area analyses, and lets device library power models be applied per actual operand values.

What would settle it

Take an EPIC accelerator description in SimPhony and replace the unlimited-interconnect assumption with a cycle-accurate electrical and optical interconnect contention model for the same workload; if the contention-aware total latency exceeds SimPhony's estimate by more than the stated memory and reconfiguration margins on a bandwidth-heavy attention task, the end-to-end latency claim is falsified. Alternatively, compare SimPhony's auto-generated floorplan area against fabricated chip layouts for a diverse set of PTC netlists, not just the single node example in the paper.

Watch

Extended reading notes

Core claim

The central claim is that a single hierarchical netlist representation with directed two-pin nets and user-defined scaling rules can uniformly describe diverse photonic tensor core topologies, including array-style TeMPO cores and Clements-style MZI meshes. From that representation, SimPhony automatically derives critical-path insertion loss, link budget and laser power, signal-flow-aware floorplan area, and data-dependent device power that reflects actual operand values. The paper validates the simulator against area and energy breakdowns from TeMPO and power and area breakdowns from Lightening-Transformer, and shows that layout-unaware area estimation understates node area by 72 percent while data-aware phase-shifter energy drops by about 60 percent versus data-unaware estimates.

Load-bearing premise

The load-bearing premise is that memory bandwidth is the only data-movement bottleneck, with on-chip and cross-chiplet interconnects assumed to always provide enough bandwidth; if interconnect bandwidth is actually limiting, SimPhony's latency and energy estimates will be too optimistic.

Editorial extensions

If this is right

  • Designers can sweep architecture parameters such as number of wavelengths, tensor bitwidth, tile count, and core size and read out energy, area, and latency trade-offs before any fabrication; the paper demonstrates such sweeps for TeMPO-style cores.
  • Different photonic tensor core designs can be compared on the same memory hierarchy and layout assumptions, making reported area and energy numbers from different papers more directly comparable.
  • Heterogeneous accelerators that map different neural network layers to different photonic sub-architectures, such as SCATTER convolutions plus MZI-mesh linear layers, can be simulated with a shared memory system.
  • Data-dependent energy modeling makes pruning and power-gating effects visible in system-level energy, so algorithmic sparsity and device-level power savings can be co-optimized in simulation.

Reading between the lines

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

  • Editorial extension: If the data-aware energy model is correct, then energy comparisons against EPIC accelerators that ignore actual operand values are likely to overstate phase-shifter and modulator energy for pruned or sparse workloads, making sparsity-aware scheduling a directly testable optimization in this simulator.
  • Editorial extension: The unlimited-interconnect assumption means the framework cannot yet tell where optical broadcast or electrical interconnect bandwidth becomes the bottleneck; adding a contention-aware interconnect model would likely change latency estimates for broadcast-heavy transformer workloads.
  • Editorial extension: Because the netlist representation already auto-derives device counts and critical paths, it could be reused as a front end for automatic control-signal scheduling or physical-design closure, not just performance estimation.
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 SimPhony, an open-source cross-layer simulation framework for heterogeneous electronic-photonic AI systems. SimPhony provides a hierarchical netlist-based architecture representation, optics-specific dataflow modeling, bandwidth-adaptive memory hierarchy analysis, link budget analysis, data-aware energy estimation, and layout-aware chip area estimation. The framework is validated against two prior architecture simulation studies, TeMPO and Lightening-Transformer, and is demonstrated on use cases including wavelength and bitwidth sweeps, layout-aware area estimation, data-dependent energy modeling, and heterogeneous layer-to-architecture mapping.

Significance. If the accuracy claims are sustained, SimPhony would fill a genuine gap: a flexible, open, extensible simulator spanning device, circuit, and architecture levels for EPIC AI hardware. The paper's contributions include a unified PTC representation, photonics-specific dataflow handling, and a layout-aware area estimator, which are useful tools for the community. However, the validation is currently against simulations from overlapping authors and includes a fitted area scaling factor, so the significance is conditional on addressing these validation concerns.

major comments (4)
  1. [Section III-C3] The assumption that on-chip and cross-chiplet interconnects provide sufficient data transaction bandwidth is load-bearing for the claimed latency and energy accuracy, but it is not validated. The paper states 'we focus on memory bandwidth analysis and assume on-chip and cross-chiplet interconnects provide sufficient data transaction bandwidth' and cites [31] as evidence, yet that reference is a specific optical interconnect for ML training and does not establish that all EPIC architectures are interconnect-unconstrained. In an interconnect-limited system, SimPhony would underestimate data-movement latency and energy. A concrete improvement would be to add an interconnect bandwidth model or, at minimum, a sensitivity study that reports the range of interconnect bandwidths over which the memory-only assumption is valid.
  2. [Section IV-A] The validation is partially circular: both reference architectures, TeMPO [17] and Lightening-Transformer [4], come from overlapping authorship with the SimPhony team, and the device library and reference numbers likely share the same origin, so agreement is partly by construction. For Lightening-Transformer, the paper states 'SimPhony accurately reproduces the chip area when appropriate scaling factors are applied,' which indicates a fitted adjustment rather than a parameter-free prediction. This undermines the claim that layout-aware area estimation is predictive without calibration. The power comparison also uses different memory technology nodes (CACTI-45nm vs. PCACTI-14nm), making the match less informative. The authors should validate against independent measurements or a simulator with different assumptions, and they should report the exact scaling factors and justify them from physical layout constraints.
  3. [Section IV-A] The paper's central claim includes accurate performance (latency) modeling, but the validation section reports only area, energy, and power comparisons; no latency validation is presented. The latency model in Section III-C2 includes several penalties (range-restricted PTCs, reconfiguration latency) that could significantly affect results, yet none are checked against any reference. Without any latency comparison, the claim of 'accurate performance analysis' is unsubstantiated. The authors should either add a latency validation study or explicitly scope the accuracy claim to energy, area, and power.
  4. [Section III-B] The claim of supporting 'arbitrary PTC topologies' is not demonstrated. The validation covers TeMPO and MZI meshes, and the use cases are limited to TeMPO, SCATTER, and MZI meshes, all of which are either from the same research group or well-known examples. To support the generality claim, the authors should demonstrate the framework on a PTC topology that is not represented in the prior work of the authors, for example an FFT-based ONN or a WDM weight bank, and show that the netlist-based scaling rules and link budget analysis work without ad hoc adjustments.
minor comments (5)
  1. [Title] The title contains spacing errors: 'Sim ulation', 'Pho tonic', and 'Sy stem' should be corrected to 'Simulation', 'Photonic', and 'System'.
  2. [Section III-C3] The word 'searchs' should be 'searches'.
  3. [Figures 7 and 8] Figures 7 and 8 appear to be duplicated in the manuscript; please ensure only one copy of each figure is included.
  4. [Section IV-A] The sentence 'we validate our simulation results' should begin with a capital letter after the section heading.
  5. [References] Reference [14], the SCATTER paper, is cited as appearing at ICCAD 2024 but the URL points to an arXiv preprint; the reference should be formatted consistently.
Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claims of SimPhony rest on the accuracy of the device library, the assumption that interconnect bandwidth is not a bottleneck, the generality of the proposed floorplan estimator, and the applicability of CACTI memory models and the prior link-budget formula. None of these are validated against full-system silicon measurements in this paper, so they are treated as assumptions here.

free parameters (2)
  • Area scaling factor for Lightening-Transformer validation = not specified
    Section IV-A states SimPhony reproduces LT's chip area 'when appropriate scaling factors are applied'; this is a post hoc parameter tuned to match the reference, making the area agreement partly fitted.
  • Memory technology node = CACTI-45nm (vs LT's PCACTI-14nm)
    Section III-C5 and IV-A: the memory energy and area model uses CACTI-45nm, a choice made by the authors; the paper attributes the 40% power deviation to this node difference, so it is a free modeling input affecting the central comparison.
assumptions (5)
  • domain assumption Device power, area, and latency values in SimPhony-DevLib (from Lumerical HEAT simulations or experimental measurements) are representative of real fabrication processes.
    Section III-A: the entire energy and area analysis rests on these device values, but the paper does not provide the raw data or uncertainty bounds.
  • domain assumption On-chip and cross-chiplet interconnects provide sufficient data transaction bandwidth, so only the modeled memory hierarchy constrains data movement.
    Explicitly stated in Section III-C3; if this fails, the simulated latency and energy omit a potentially dominant cost.
  • domain assumption The row-based signal-flow-aware floorplan (longest-device row width, hiding other devices beneath it, topological order placement) yields near-accurate real layout area for arbitrary PTC topologies.
    Section III-C6: the floorplan is validated against a single real layout (Figure 6), and generalization to other topologies is assumed.
  • domain assumption CACTI-simulated memory access energies and multi-block SRAM bandwidth scaling are accurate for the modeled technology nodes.
    Section III-C3 and C5: memory modeling is a major component of the energy and power estimate, yet it caused a large validation deviation when the node was changed.
  • domain assumption Equation (1) for laser power, derived from prior work [3],[4], correctly captures the required source power given PD sensitivity and extinction ratio.
    Section III-C4: the link budget formula is adopted from earlier photonic accelerator papers and is not re-derived or independently validated here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SimPhony: A Device-Circuit-Architecture Cross-Layer Modeling and Simulation Framework for Heterogeneous Electronic-Photonic AI System." pith.science (2026). https://pith.science/paper/5PKTVPTZ

@misc{pith2026241113715,
  author       = {Pith},
  title        = {Pith review of: SimPhony: A Device-Circuit-Architecture Cross-Layer Modeling and Simulation Framework for Heterogeneous Electronic-Photonic AI System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PKTVPTZ}},
  note         = {Machine review of arXiv:2411.13715}
}
read the original abstract

Electronic-photonic integrated circuits (EPICs) offer transformative potential for next-generation high-performance AI but require interdisciplinary advances across devices, circuits, architecture, and design automation. The complexity of hybrid systems makes it challenging even for domain experts to understand distinct behaviors and interactions across design stack. The lack of a flexible, accurate, fast, and easy-to-use EPIC AI system simulation framework significantly limits the exploration of hardware innovations and system evaluations on common benchmarks. To address this gap, we propose SimPhony, a cross-layer modeling and simulation framework for heterogeneous electronic-photonic AI systems. SimPhony offers a platform that enables (1) generic, extensible hardware topology representation that supports heterogeneous multi-core architectures with diverse photonic tensor core designs; (2) optics-specific dataflow modeling with unique multi-dimensional parallelism and reuse beyond spatial/temporal dimensions; (3) data-aware energy modeling with realistic device responses, layout-aware area estimation, link budget analysis, and bandwidth-adaptive memory modeling; and (4) seamless integration with model training framework for hardware/software co-simulation. By providing a unified, versatile, and high-fidelity simulation platform, SimPhony enables researchers to innovate and evaluate EPIC AI hardware across multiple domains, facilitating the next leap in emerging AI hardware. We open-source our codes at https://github.com/ScopeX-ASU/SimPhony

Figures

Figures reproduced from arXiv: 2411.13715 by the authors.

Figure 1
Figure 1. Overview of our proposed SimPhony framework. modeling [3], [15], [16], [28], [29], unaware of the numeri￾cal values in actual workloads. CimLoop is a recent analog NN simulator; its photonic version [24] has showcased one architecture Albireo [5]. However, it focuses on architectural parameters (e.g., memory and dataflow), and lacks support for flexible PTC construction with photonics-specific dataflow and paralleli… view at source ↗
Figure 2
Figure 2. (a) Node-level circuit topology definition and (b) correspond￾ing weighted direct acyclic graph (DAG) representation. MZM A DAC A MZM B DAC B MZM B DAC B MZM B DAC B MZM B DAC B Node ADC Integrator ADC Integrator ADC Integrator ADC Integrator ADC Integrator ADC Integrator ADC Integrator ADC Integrator i0 i1 i9 i2 i3 i4 i5 i6 i7 i8 Parametric Photonic Tensor Core Construction Instances Scaling Rules Laser i0 Coupling… view at source ↗
Figure 3
Figure 3. Case studies of constructing parametric multi-core accelerator architecture (a) dynamic array-style TeMPO [17] and (b) static mesh￾style MZI arrays [1]. The circuits can be generated by scaling up the smallest node with a user-defined scaling rule. The critical path with the highest insertion loss can be auto-derived. device information is provided in detail to support accurate sim￾ulations of area, power, latency, … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Mapping blocking GEMM to output-stationary PTC with spatial/spectral parallelism and hierarchical accumulation. 2) Photonics-Specific Dataflow and Latency Analysis: Besides the support for standard dataflow for GEMM, e.g., weight/input/output stationary, here, we empha…
Figure 5
Figure 5. Figure 5: SimPhony-Sim supports accurate, data-dependent energy modeling with simulated/measured device power data. i2 i1 i0 i4 Netlist i3 i1 i3 i0 i2 i3 i4 level 1 level 2 level 3 level 1 i0 i1 level 2 level 3 i2 i3 i4 device spacing prior method: sum of device footprint= 1270.…
Figure 6
Figure 6. Figure 6: Proposed signal flow-aware row-based floorplan for accurate layout-aware area estimation. P CACTI-simulated access energy per bit emem, i.e., Emem = mem∼{HBM,GLB,LB,RF } ememDmem. For computing cost, SimPhony-Sim supports data￾dependent energy analysis with accurate de…
Figure 7
Figure 7. Figure 7: SimPhony results validated on a (280×28)×(28×280) GEMM task w/ TeMPO [17]. (a) area (b) energy breakdown. 0 0.2 0.4 0.6 0.8 1 Area (mm2) Crossing Integrator TIA Node MMI Y Branch MZM DAC ADC Area Validation Compare to Old simulation 0 50 100 150 Energy (pJ) Laser PS PD…
Figure 10
Figure 10. Figure 10: (a) TeMPO [17] area breakdown with layout awareness. (b) SCATTER [14] energy breakdown with data awareness. B. SimPhony Use Cases To show the capability of SimPhony, we study multiple design examples by sweeping PTC architectural parameters in SimPhony to gain design …
Figure 11
Figure 11. Figure 11: Layer energy breakdown with heterogeneous layer map￾ping of VGG-8(CIFAR10) [36]. Convolutions are mapped to SCAT￾TER [14], and Linear layers are mapped to MZI meshes [1]. 4) Heterogeneous Mapping: Lastly, we show SimPhony’s capability for heterogeneous architecture mo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 34 canonical work pages

  1. [31]

    Sip-ml: high-bandwidth optical network interconnects for machine learning training,

    M. Khani, M. Ghobadi, M. Alizadeh, Z. Zhu, M. Glick, K. Bergman, A. Vahdat, B. Klenk, and E. Ebrahimi, “Sip-ml: high-bandwidth optical network interconnects for machine learning training,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference , ser. SIGCOMM ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 657–675. [Online]. Availabl...

    work page arXiv 2021

  2. [17]

    Tempo: Efficient time-multiplexed dynamic photonic tensor core for edge ai with compact slow-light electro-optic modulator,

    M. Zhang, D. Yin, N. Gangi, A. Begovi ´c, A. Chen, Z. R. Huang, and J. Gu, “Tempo: Efficient time-multiplexed dynamic photonic tensor core for edge ai with compact slow-light electro-optic modulator,” J. Appl. Phys, 2024

    work page 2024

  3. [4]

    Lightening-transformer: A dynamically-operated photonic tensor core for energy-efficient transformer accelerator,

    H. Zhu, J. Gu, H. Wang, Z. Jiang, Z. Zhang, R. Tang, C. Feng, S. Han et al., “Lightening-transformer: A dynamically-operated photonic tensor core for energy-efficient transformer accelerator,” in Proc. HPCA, 2024

    work page 2024

  4. [1]

    Deep learning with coherent nanophotonic circuits,

    Y . Shen, N. C. Harris, S. Skirlo et al. , “Deep learning with coherent nanophotonic circuits,” Nature Photonics, 2017

    work page 2017

  5. [2]

    Parallel convolutional processing using an integrated photonic tensor core,

    J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. L. Gallo, X. Fu, A. Lukashchuk, A. Raja, J. Liu, D. Wright, A. Sebas- tian, T. Kippenberg, W. Pernice, and H. Bhaskaran, “Parallel convolutional processing using an integrated photonic tensor core,” Nature, 2021

    work page 2021

  6. [3]

    A compact butterfly- style silicon photonic–electronic neural chip for hardware-efficient deep learning,

    C. Feng, J. Gu, H. Zhu, Z. Ying, Z. Zhao et al. , “A compact butterfly- style silicon photonic–electronic neural chip for hardware-efficient deep learning,” ACS Photonics, vol. 9, no. 12, pp. 3906–3916, 2022

    work page 2022

  7. [5]

    Albireo: Energy- efficient acceleration of convolutional neural networks via silicon pho- tonics,

    K. Shiflett, A. Karanth, R. Bunescu, and A. Louri, “Albireo: Energy- efficient acceleration of convolutional neural networks via silicon pho- tonics,” in Proc. ISCA, 2021, pp. 860–873

    work page 2021

  8. [6]

    Photonics for Artificial Intelligence and Neuromorphic Computing,

    B. J. Shastri, A. N. Tait et al., “Photonics for Artificial Intelligence and Neuromorphic Computing,” Nature Photonics, 2021

    work page 2021

Show all 36 references
  1. [7]

    Large-scale pho- tonic chiplet taichi empowers 160-tops/w artificial general intelligence,

    Z. Xu, T. Zhou, M. Ma, C. Deng, Q. Dai, and L. Fang, “Large-scale pho- tonic chiplet taichi empowers 160-tops/w artificial general intelligence,” Science, vol. 384, no. 6692, pp. 202–209, 2024

    2024

  2. [8]

    Photonic-electronic integrated circuits for high-performance computing and ai accelerators,

    S. Ning, H. Zhu, C. Feng, J. Gu, Z. Jiang, Z. Ying, J. Midkiff, S. Jain, M. H. Hlaing, D. Z. Pan, and R. T. Chen, “Photonic-electronic integrated circuits for high-performance computing and ai accelerators,” Journal of Lightwave Technology, pp. 1–26, 2024

    2024

  3. [9]

    Femto- joule per MAC Neuromorphic Photonics: An Energy and Technology Roadmap,

    A. R. Totovi ´c, G. Dabos, N. Passalis, A. Tefas, and N. Pleros, “Femto- joule per MAC Neuromorphic Photonics: An Energy and Technology Roadmap,” IEEE Journal of Selected Topics in Quantum Electronics , vol. 26, no. 5, pp. 1–15, 2020

    2020

  4. [10]

    Towards Hardware-Efficient Optical Neural Networks: Beyond FFT Architecture via Joint Learnability,

    J. Gu, Z. Zhao, C. Feng et al. , “Towards Hardware-Efficient Optical Neural Networks: Beyond FFT Architecture via Joint Learnability,” IEEE TCAD, 2020

    2020

  5. [11]

    FLOPS: Effi- cient On-Chip Learning for Optical Neural Networks Through Stochastic Zeroth-Order Optimization,

    J. Gu, Z. Zhao, C. Feng, W. Li, R. T. Chen, and D. Z. Pan, “FLOPS: Effi- cient On-Chip Learning for Optical Neural Networks Through Stochastic Zeroth-Order Optimization,” in Proc. DAC, 2020

    2020

  6. [12]

    ROQ: A noise-aware quantization scheme towards robust optical neural networks with low-bit controls,

    J. Gu, Z. Zhao, C. Feng, H. Zhu, R. T. Chen, and D. Z. Pan, “ROQ: A noise-aware quantization scheme towards robust optical neural networks with low-bit controls,” in Proc. DATE, 2020

    2020

  7. [13]

    Pruning Coherent Integrated Photonic Neural Networks Using the Lottery Ticket Hypothesis,

    S. Banerjee, M. Nikdast, S. Pasricha, and K. Chakrabarty, “Pruning Coherent Integrated Photonic Neural Networks Using the Lottery Ticket Hypothesis,” in 2022 IEEE Comput. Soc. Annu. Symp. VLSI ISVLSI , Jul. 2022, pp. 128–133

    2022

  8. [14]

    Scatter: Algorithm-circuit co-sparse photonic accelerator with thermal- tolerant, power-efficient in-situ light redistribution,

    Z. Yin, N. Gangi, M. Zhang, J. Zhang, R. Huang, and J. Gu, “Scatter: Algorithm-circuit co-sparse photonic accelerator with thermal- tolerant, power-efficient in-situ light redistribution,” in International Conference on Computer-Aided Design (ICCAD) , 2024. [Online]. Available...

    2024 arXiv

  9. [15]

    Holylight: A nanophotonic accelerator for deep learning in data centers,

    W. Liu, W. Liu, Y . Ye, Q. Lou, Y . Xie, and L. Jiang, “Holylight: A nanophotonic accelerator for deep learning in data centers,” in Proc. DATE, 2019

    2019

  10. [16]

    LightBulb: A Photonic- Nonvolatile-Memory-based Accelerator for Binarized Convolutional Neu- ral Networks,

    F. Zokaee, Q. Lou, N. Youngblood et al. , “LightBulb: A Photonic- Nonvolatile-Memory-based Accelerator for Binarized Convolutional Neu- ral Networks,” in Proc. DATE, 2020

    2020

  11. [18]

    PIXEL: Photonic Neural Network Accelerator,

    K. Shiflett, D. Wright, A. Karanth, and A. Louri, “PIXEL: Photonic Neural Network Accelerator,” in Proc. HPCA, 2020, pp. 474–487

    2020

  12. [19]

    Mirage: An rns- based photonic accelerator for dnn training,

    C. Demirkiran, G. Yang, D. Bunandar, and A. Joshi, “Mirage: An rns- based photonic accelerator for dnn training,” in Proc. ISCA , 2024, pp. 73–87

    2024

  13. [20]

    Neuromorphic photonic networks using silicon photonic weight banks,

    A. N. Tait, T. F. de Lima, E. Zhou et al. , “Neuromorphic photonic networks using silicon photonic weight banks,” Sci. Rep., 2017

    2017

  14. [21]

    Experimental realization of any discrete unitary operator,

    M. Reck, A. Zeilinger, H. Bernstein et al., “Experimental realization of any discrete unitary operator,” Physical review letters, 1994

    1994

  15. [22]

    Optimal Design for Universal Multiport Interferometers,

    W. R. Clements, P. C. Humphreys, B. J. Metcalf et al., “Optimal Design for Universal Multiport Interferometers,” Optica, 2018

    2018

  16. [23]

    11 TOPS photonic convolutional accelerator for optical neural networks,

    X. Xu, M. Tan, B. Corcoran, J. Wu, A. Boes, T. G. Nguyen, S. T. Chu, B. E. Little, D. G. Hicks, R. Morandotti, A. Mitchell, and D. J. Moss, “11 TOPS photonic convolutional accelerator for optical neural networks,” Nature, 2021

    2021

  17. [24]

    Architecture-level modeling of photonic deep neural network accelerators,

    T. Andrulis, G. I. Chaudhry, V . M. Suriyakumar, J. S. Emer, and V . Sze, “Architecture-level modeling of photonic deep neural network accelerators,” arXiv preprint arXiv:2405.07266 , 2024

    2024 arXiv

  18. [25]

    Towards area-efficient optical neural networks: an FFT-based architecture,

    J. Gu, Z. Zhao, C. Feng et al. , “Towards area-efficient optical neural networks: an FFT-based architecture,” in Proc. ASPDAC, 2020

    2020

  19. [26]

    M3ICRO: Machine Learning-Enabled Compact Photonic Tensor Core based on PRogrammable Multi-Operand Multimode Interference,

    J. Gu, H. Zhu, C. Feng, Z. Jiang, R. T. Chen, and D. Z. Pan, “M3ICRO: Machine Learning-Enabled Compact Photonic Tensor Core based on PRogrammable Multi-Operand Multimode Interference,” APL Machine Learning, 2023

    2023

  20. [27]

    Photonic Tensor Cores for Machine Learning,

    M. Miscuglio and V . J. Sorger, “Photonic Tensor Cores for Machine Learning,” Applied Physics Review , 2020

    2020

  21. [28]

    Adept: Automatic differentiable design of photonic tensor cores,

    J. Gu, H. Zhu, C. Feng, Z. Jiang, M. Liu, S. Zhang, R. T. Chen, and D. Z. Pan, “Adept: Automatic differentiable design of photonic tensor cores,” in Proc. DAC, 2022

    2022

  22. [29]

    O-has: Optical hardware accelerator search for boosting both acceleration performance and devel- opment speed,

    M. Li, Z. Yu, Y . Zhang, Y . Fu, and Y . Lin, “O-has: Optical hardware accelerator search for boosting both acceleration performance and devel- opment speed,” in Proc. ICCAD, 2021

    2021

  23. [30]

    L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization,

    J. Gu, H. Zhu, C. Feng, Z. Jiang, R. T. Chen, and D. Z. Pan, “L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization,” in Proc. NeurIPS, 2021

    2021

  24. [32]

    Cacti 7: New tools for interconnect exploration in innovative off-chip memories,

    R. Balasubramonian, A. B. Kahng, N. Muralimanohar, A. Shafiee, and V . Srinivas, “Cacti 7: New tools for interconnect exploration in innovative off-chip memories,” ACM Trans. Archit. Code Optim. , vol. 14, no. 2, jun 2017

    2017

  25. [33]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2019. [Online]. Available: https://arxiv.org/abs/1810.04805

    2019 arXiv

  26. [34]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in Proc. CVPR , 2009, pp. 248–255

    2009

  27. [35]

    Fincacti: Architectural analysis and modeling of caches with deeply-scaled finfet devices,

    A. Shafaei, Y . Wang, X. Lin, and M. Pedram, “Fincacti: Architectural analysis and modeling of caches with deeply-scaled finfet devices,” in 2014 IEEE Computer Society Annual Symposium on VLSI , 2014, pp. 290–295

    2014

  28. [36]

    A sparse-group lasso,

    N. Simon, J. Friedman, T. Hastie, and R. Tibshirani, “A sparse-group lasso,” Journal of Computational and Graphical Statistics , 2013

    2013

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.