Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Hybrid Photonic-digital Accelerator for Attention Mechanism

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read HyAtten claims that sorting attention outputs by magnitude lets 4-bit converters handle 85% of signals while a small digital die takes the rest, achieving 9.8x performance and 2.2x energy efficiency per area over baseline.

desk verdict A real architectural idea for the photonic ADC bottleneck, but the digital fallback path is underspecified and the 9.8x claim is not yet credible. read the letter →

arxiv 2501.11286 v1 pith:IO2BR65E submitted 2025-01-20 cs.AR

Huize Li , Dan Chen , Tulika Mitra This is my paper · ORCID
classification cs.AR
keywords photonicacceleratorattentionmechanismTransformeranalog-to-digitalconvertersignalconversionoverheadhybridphotonic-digitallow-resolutionADCdata-dependentrouting
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

HyAtten is a photonic Transformer accelerator built around a simple observation: in attention GEMMs, more than 85% of the analog currents produced by the photonic arrays fit inside the range of a 4-bit analog-to-digital converter, while the remaining high-magnitude outputs carry most of the accuracy. The paper's claim is that the ADC bottleneck that dominates area and latency in prior photonic accelerators can be removed by splitting the signal path by magnitude: convert the 85% with many low-resolution ADCs, and route the 15% high-resolution outputs to a small digital die that recomputes those dot products. If this split works, photonic accelerators get most of the speed of low-resolution conversion without paying the accuracy cost of clipping high-resolution signals. The paper reports 9.8x performance per unit area and 2.2x energy efficiency per unit area over the Lightening-Transformer baseline, with accuracy loss below 0.3%.

What carries the argument

The load-bearing component is the analog comparator plus coordinate register attached to each DPTC array, the photonic multiply-accumulate array inherited from the baseline design. The comparator marks any photonic current that exceeds the 4-bit ADCs' full-scale range, the coordinate register records where it occurred, and the memory controller uses those coordinates to fetch the corresponding inputs and send them to the digital PE. This converts a monolithic ADC-resolution problem into a data-dependent partition of outputs: only the over-range tail consumes digital resources, making it possible to replace one shared 8-bit ADC per 32x32 array with 32 parallel 4-bit ADCs per 64x64 array.

What would settle it

Simulate or prototype the digital tail with a cycle-accurate model: count the exact number of high-resolution dot products per DPTC array, the size of the operand fetch for each, and the MAU cycles needed, and compare that path's latency and energy against the savings from using 32 4-bit ADCs instead of one 8-bit ADC; if the digital path's overhead is comparable to the ADC savings, the reported 9.8x and 2.2x gains collapse.

Watch

Extended reading notes

Core claim

The central claim is that signal conversion overhead, not photonic computation, is the binding constraint, and that it can be partitioned rather than eliminated. HyAtten uses an analog comparator to classify each DPTC output current as low-resolution (within a 4-bit ADC full-scale range) or high-resolution; the former go to banks of 32 low-resolution ADCs, the latter are recorded by coordinate and their digital input values are sent to a digital processing element for recomputation. Over BERT-base on GLUE and DeiT-T on ImageNet, HyAtten reports 9.8x higher performance per square millimeter and 2.2x higher energy efficiency per square millimeter than the baseline photonic accelerator, and it slightly exceeds the baseline's accuracy (for example, 85.89 versus 85.77 on MRPC) because the high-resolution path avoids ADC noise.

Load-bearing premise

The design assumes the digital die handles the over-range 15% of outputs with negligible latency, energy, and area, using only the data routed to it, but the paper never specifies how many operands, cycles, or joules that recomputation path actually consumes.

Editorial extensions

If this is right

  • Replacing one high-resolution shared ADC with many low-resolution ADCs removes the conversion latency that made photonic arrays idle, so throughput per unit area rises without an area penalty.
  • The digital die's area and power stay small because it processes only the over-range fraction, reported as 4.6% of area and 0.8% of power in the 32-tile configuration.
  • Accuracy is preserved because high-resolution signals are computed in the noise-free digital domain rather than clipped or quantized; the paper reports under 0.3% loss versus GPU and about 0.2% better than the photonic baseline.
  • Sequence-length scaling keeps per-matrix overhead stable, while scaling the number of tiles is eventually limited by HBM transfer bandwidth rather than photonic compute.

Reading between the lines

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

  • The 15% high-resolution fraction is an empirical property of these quantized models; on distributions with heavier tails (longer sequences, unquantized activations) the digital tail could grow, and the design's advantage would shrink unless the comparator threshold or digital parallelism is adapted.
  • The comparator-and-coordinate mechanism is effectively a data-dependent routing table, which suggests it could be combined with output-stationary scheduling to prefetch digital operands, something the paper does not explore.
  • The same magnitude-splitting idea could transfer to other analog accelerators whose ADC resolution is the bottleneck, such as ReRAM crossbars, where an over-range comparator would play the same role.
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 / 7 minor

Summary. The manuscript proposes HyAtten, a hybrid photonic-digital accelerator for the attention mechanism. It makes the empirical observation that over 85% of analog signals in the Lightening-Transformer baseline can be converted by 4-bit ADCs, and it uses this observation to replace high-resolution ADCs with many low-resolution ADCs while offloading the remaining less-than-15% of signals that would overflow the low-resolution converters to a small digital die. The paper claims 9.8x performance/area and 2.2x energy-efficiency/area over the LT-B baseline with under 0.3% accuracy loss, based on a modified version of the Lightening-Transformer simulator, and reports area/power breakdowns and scalability trends.

Significance. If the claims hold, the work contributes a useful architectural idea: magnitude-based signal classification can reduce the ADC bottleneck in photonic accelerators without a full high-resolution conversion path. The empirical signal-distribution study (Figure 2) is a valuable input for the community, and the paper is not circular in the sense that the headline speedup is computed from component counts and energy models rather than fitted to a target. The paper also provides concrete area, power, and accuracy numbers for realistic Transformer workloads. However, the central mechanism for handling the high-resolution tail is under-specified, and the evaluation does not currently demonstrate that its latency, data movement, and energy costs are negligible; the headline comparison is also confounded by changes in ADC count and array size.

major comments (4)
  1. [Section III-A, Section III-B (Fig. 3 steps 4-6); Table I] The digital recompute path is not specified at the level needed to verify the claimed 9.8x speedup. The comparator in Figure 3(e) logs only coordinates, so the correct overflowed dot product cannot be recovered from a converted scalar; the digital PE must recompute it from the original operands (a row of Q and a column of K for QK^T; a row of S and a column of V for SV). The text's phrase 'digital values of the high-resolution signals' is ambiguous about whether only scalar results or full operand vectors are moved to the digital die. With one 0.014 mm2, 8.2 mW MAU per tile (Table I), the added MAC work for a 15% overflow rate is about 0.15 * N^2 * d per GEMM, which is roughly 3.6e8 MACs for BERT-base QK^T alone before S*V, and the operand-fetch traffic could be of the same order. The paper should specify the data volume, operand reuse, cycle count, and serialization behavior of this path, and the simulator should account for it explicitly.
  2. [Section II-C, Section IV-B; Fig. 2(b), Fig. 5] The 15% high-resolution fraction is measured on the LT-B baseline (Figure 2(b)) and then reused as the workload of the digital die in HyAtten. HyAtten uses 64x64 DPTC arrays with 4-bit input quantization and 32 4-bit ADCs per array, whereas the distribution in Figure 2 was obtained on the baseline configuration; the overflow fraction could differ. The evaluation should re-measure or bound this fraction under the HyAtten configuration, and the sensitivity of the headline speedup to the overflow rate should be reported.
  3. [Section IV-A, Section IV-B; Table I, Figure 5] The performance comparison confounds the proposed signal-classification idea with a large increase in ADC parallelism. LT-B shares one ADC per 32x32 DPTC array, while HyAtten uses 32 ADCs per 64x64 DPTC array; the text also says one 8-bit ADC is replaced by 16 4-bit ADCs, which is inconsistent with the 32 ADCs in Table I. To attribute the 9.8x result to the hybrid digital offload, the authors should compare against LT-B variants with matched ADC count and array size, or decompose the speedup into the contribution from additional ADCs and the contribution from the digital die.
  4. [Section III-B (Fig. 3 step 7), Section III-A (softmax unit)] The softmax operation consumes the collaboratively generated score matrix S, but the paper does not explain how the softmax unit combines partial results from the photonic and digital dies. In particular, an overflowed entry of S that is recomputed digitally must still participate in the exponent and normalization for its row, and the dataflow for passing those entries and the partial row sums is not described. This omission affects both correctness and the latency model of the attention pipeline.
minor comments (7)
  1. [Abstract] The phrase 'with minimize signal conversion overhead' should read 'with minimal signal conversion overhead.'
  2. [Section IV-B] There is a typo in 'dispalys' that should be 'displays.'
  3. [Section IV-A] The simulator is described only as a modified Python-based simulator from [34]; no artifact or detailed list of modifications is given, which limits reproducibility.
  4. [Figure 2(b)] The axis label 'Signals Ratio' is unclear; it should specify that this is the proportion of signals within the ADC's full-scale range, and the normalization should be stated.
  5. [Table I] The HyAtten total row '17.38 39.9W Numbers 1' is not a normal table entry; the totals should be presented as separate rows with clear units.
  6. [Table II] The caption and text refer to a GPU comparison, but the table row is labeled 'Original'; these labels should be harmonized.
  7. [References] Reference [3] is cited for the MAU, softmax, and accumulator SPICE simulations, but the cited paper is about ferrite core losses; please verify that this is the intended source.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 9.8x and 2.2x claims are evaluated from component-level area/energy/latency models, and the 85/15 signal split is an empirical profiling observation, not a fitted target.

full rationale

I find no circular step in the derivation chain. The paper's central observation (over 85% of signals fit in 4-bit ADC range, Section II-C) is obtained from experiments on the Lightening-Transformer baseline and is then used as a design input; this is an empirical co-design loop, not a prediction that reduces to its inputs by construction. The reported 9.8x performance-per-area and 2.2x energy-efficiency-per-area improvements are derived from Table I component counts, area tables, and power/latency models in the modified simulator, not fitted to reproduce those headline numbers. The digital-die path for the remaining ~15% of signals is indeed underspecified, but underspecification is a correctness/completeness risk, not circularity: the paper never defines the digital recompute in terms of the final speedup. Self-citations appear in related work ([9], [11]) and in the softmax lookup-table implementation ([10]); none is load-bearing for the main technical claim, and the paper does not invoke any uniqueness theorem or hidden ansatz from prior work. The accuracy comparison (Table II) is an independent evaluation against GPU and LT-B results rather than a restatement of the design assumptions. Therefore the derivation is self-contained enough that no equation-level circularity can be quoted.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The evaluation rests on assumptions about photonic behavior, simulator fidelity, the stability of the 85% signal split, and the digital recomputation path. These are listed as axioms; none are proven by measurement in this paper.

free parameters (4)
  • ADC resolution / comparator threshold = 4 bits
    Hand-chosen boundary between low- and high-resolution signals; it determines the reported 85% and 15% shares and the size of the digital fallback. Section II-C selects it after measuring accuracy at 2, 4, and 8 bits.
  • Input quantization bit-width for weight and activation matrices = 4 bits
    Applied to Q and K to cut conversion and compute load; adopted from [34] and [5]. It is a configurable design choice that shapes the signal distribution entering the ADCs.
  • Number of 4-bit ADCs per 64x64 DPTC array = 32
    Architectural configuration chosen to lower conversion latency; its area and power are summed in Table I. The number is not swept, and the paper does not show why 32 is sufficient for the 9.8x claim.
  • ADC area/power scaling model = Murmann survey [16] scaling
    Section IV-A uses the survey to claim a 5-bit ADC needs twice the area of a 4-bit ADC; this scaling is an external model, not a fitted measurement in this paper.
assumptions (5)
  • domain assumption The DDot and DPTC photonic units behave exactly as modeled in Lightening-Transformer [34].
    Section II-B reproduces the coherent interference and DPTC crossbar model from [34] as the basis for all GEMM throughput and energy numbers; this paper adds no photonic-device test.
  • domain assumption The modified Python simulator from [34] is cycle-accurate and the authors' modifications are faithful.
    Section IV-A says they 'modified an existing Python-based simulator' to evaluate latency, power, area, and energy, but no simulator source, configuration, or validation against silicon is provided.
  • domain assumption The 85%/15% signal-resolution split measured on Lightening-Transformer remains valid for HyAtten's 4-bit operating point.
    Section II-C derives the design from that split, and Section IV-B uses it to size the digital die without reporting a re-measurement of the split for HyAtten with its own noise-aware training.
  • domain assumption The digital die can recompute the overflow outputs using only the transferred high-resolution signal data with a single small MAU.
    Section III-A/B describes the digital path but does not quantify operand fetching, recomputation cycle count, or energy for the 15% tail; the claim that overhead is minimal is asserted, not demonstrated.
  • domain assumption 4-bit ADCs can be substituted for higher-resolution ADCs with area/power from the Murmann survey [16] and no architecture-level side effects.
    Section IV-A scales ADC power and area based on [16] and replaces one 8-bit ADC with multiple 4-bit ADCs; no measured silicon data from this design is available.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Photonic-digital Accelerator for Attention Mechanism." pith.science (2026). https://pith.science/paper/IO2BR65E

@misc{pith2026250111286,
  author       = {Pith},
  title        = {Pith review of: Hybrid Photonic-digital Accelerator for Attention Mechanism},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IO2BR65E}},
  note         = {Machine review of arXiv:2501.11286}
}
read the original abstract

The wide adoption and substantial computational resource requirements of attention-based Transformers have spurred the demand for efficient hardware accelerators. Unlike digital-based accelerators, there is growing interest in exploring photonics due to its high energy efficiency and ultra-fast processing speeds. However, the significant signal conversion overhead limits the performance of photonic-based accelerators. In this work, we propose HyAtten, a photonic-based attention accelerator with minimize signal conversion overhead. HyAtten incorporates a signal comparator to classify signals into two categories based on whether they can be processed by low-resolution converters. HyAtten integrates low-resolution converters to process all low-resolution signals, thereby boosting the parallelism of photonic computing. For signals requiring high-resolution conversion, HyAtten uses digital circuits instead of signal converters to reduce area and latency overhead. Compared to state-of-the-art photonic-based Transformer accelerator, HyAtten achieves 9.8X performance/area and 2.2X energy-efficiency/area improvement.

Figures

Figures reproduced from arXiv: 2501.11286 by the authors.

Figure 2
Figure 2. (a) The model accuracy when employing different ADC [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Architecture and dataflow of HyAtten utilized five datasets, CoLA, MNLI, MRPC, RTE, and SQuAD, from GLUE [25], running on a BERT-based model. For high￾resolution output signals that exceed the ADC’s full-scale measurement range, the conversion output was capped at the ADC’s maximum value. The accuracy results (ratio between the number of correctly predicted samples by the total number of samples), shown in [PITH_FU… view at source ↗
Figure 4
Figure 4. GEMM operations on multiple photonic Tiles [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: (a) HyAtten speedups compared to CPU and GPU, and [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: (a) Sequence length scalability, and (b) Tiles scalability [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MDTransformer: A Hardware-Software Co-Design of Mode-Division Photonic Transformer Accelerator with Inverse-Designed Coherent Crossbar

    cs.AR 2026-07 reject novelty 6.0 of 10

    A mode-division photonic tensor core with inverse-designed coherent couplers runs transformer inference from a single 1550 nm laser, reporting 40.4% area, 63.6% power, and 40.6% energy savings over LT-class accelerato...

Reference graph

Works this paper leans on

35 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    SW AT: Scalable and Efficient Window Attention-based Transformers Acceleration on FPGAs,

    Z. Bai, P. Dangi, H. Li, and T. Mitra, “SW AT: Scalable and Efficient Window Attention-based Transformers Acceleration on FPGAs,” in Proceedings of the 61st ACM/IEEE Design Automation Conference , ser. DAC ’24. New York, NY , USA: Association for Computing Machinery,

    work page

  2. [2]

    Metanmp: Leveraging cartesian-like product to accelerate hgnns with near-memory processing,

    D. Chen, H. He, H. Jin, L. Zheng, Y . Huang, X. Shen, and X. Liao, “Metanmp: Leveraging cartesian-like product to accelerate hgnns with near-memory processing,” in Proceedings of the 50th Annual International Symposium on Computer Architecture , ser. ISCA ’23. New York, NY , USA: Association for Computing Machinery, 2023. [Online]. Available: https://doi....

    arXiv 2023

  3. [3]

    Improved Spice Simulation of Dynamic Core Losses for Ferrites With Nonuniform Field and Its Experimental Validation,

    F. Corti, A. Reatti, E. Cardeli, A. Faba, and H. Rimal, “Improved Spice Simulation of Dynamic Core Losses for Ferrites With Nonuniform Field and Its Experimental Validation,” IEEE Transactions on Industrial Electronics, vol. 68, no. 12, pp. 12 069–12 078, 2021

    work page 2021

  4. [4]

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

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    arXiv 2018

  5. [5]

    Learned step size quantization,

    S. K. Esser, J. L. McKinstry, D. Bablani, R. Appuswamy, and D. S. Modha, “Learned step size quantization,” arXiv preprint arXiv:1902.08153, 2019

    arXiv 1902

  6. [6]

    Parallel convolu- tional 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 et al., “Parallel convolu- tional processing using an integrated photonic tensor core,” Nature, vol. 589, no. 7840, pp. 52–58, 2021

    work page 2021

  7. [7]

    MeG2: In-Memory Acceleration for Genome Graphs Analysis,

    Y . Huang, L. Zheng, H. Liu, Z. Zhou, D. Chen, P. Yao, Q. Wang, X. Liao, and H. Jin, “MeG2: In-Memory Acceleration for Genome Graphs Analysis,” in 2023 60th ACM/IEEE Design Automation Conference (DAC), 2023, pp. 1–6

    work page 2023

  8. [8]

    Accel- erating Graph Convolutional Networks Using Crossbar-based Processing- In-Memory Architectures,

    Y . Huang, L. Zheng, P. Yao, Q. Wang, X. Liao, H. Jin, and J. Xue, “Accel- erating Graph Convolutional Networks Using Crossbar-based Processing- In-Memory Architectures,” in 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2022, pp. 1029–1042

    work page 2022

Show all 35 references
  1. [9]

    SADIMM: Accelerating Sparse Attention using DIMM-based Near-memory Processing,

    H. Li, D. Chen, and T. Mitra, “SADIMM: Accelerating Sparse Attention using DIMM-based Near-memory Processing,” IEEE Transactions on Computers, no. 01, pp. 1–12, 2024

    2024

  2. [10]

    CPSAA: Accelerating Sparse Attention Using Crossbar-Based Processing-In-Memory Architecture,

    H. Li, H. Jin, L. Zheng, X. Liao, Y . Huang, C. Liu, J. Xu, Z. Duan, D. Chen, and C. Gui, “CPSAA: Accelerating Sparse Attention Using Crossbar-Based Processing-In-Memory Architecture,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 43, no....

    2024

  3. [11]

    ASADI: Accelerating Sparse Attention Using Diagonal-based In-Situ Computing,

    H. Li, Z. Li, Z. Bai, and T. Mitra, “ASADI: Accelerating Sparse Attention Using Diagonal-based In-Situ Computing,” in 2024 IEEE International Symposium on High-Performance Computer Architecture (HPCA) , 2024, pp. 774–787

    2024

  4. [12]

    ReGNN: a ReRAM-based heterogeneous architecture for general graph neural networks,

    C. Liu, H. Liu, H. Jin, X. Liao, Y . Zhang, Z. Duan, J. Xu, and H. Li, “ReGNN: a ReRAM-based heterogeneous architecture for general graph neural networks,” in Proceedings of the 59th ACM/IEEE Design Automation Conference , ser. DAC ’22. New York, NY , USA: Association for Comp...

    2022

  5. [13]

    A ReRAM-Based Processing-In-Memory Architecture for Hyperdimensional Computing,

    C. Liu, K. Wu, H. Liu, H. Jin, X. Liao, Z. Duan, J. Xu, H. Li, Y . Zhang, and J. Yang, “A ReRAM-Based Processing-In-Memory Architecture for Hyperdimensional Computing,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , pp. 1–1, 2024

    2024

  6. [14]

    A 10GS/s 8b 25fJ/c-s 2850um2 Two-Step Time-Domain ADC Using Delay-Tracking Pipelined- SAR TDC with 500fs Time Step in 14nm CMOS Technology,

    J. Liu, M. Hassanpourghadi, and M. S.-W. Chen, “A 10GS/s 8b 25fJ/c-s 2850um2 Two-Step Time-Domain ADC Using Delay-Tracking Pipelined- SAR TDC with 500fs Time Step in 14nm CMOS Technology,” in 2022 IEEE International Solid-State Circuits Conference (ISSCC) , vol. 65, 2022, pp. 160–162

    2022

  7. [15]

    Sanger: A Co-Design Framework for Enabling Sparse Attention using Reconfigurable Architecture,

    L. Lu, Y . Jin, H. Bi, Z. Luo, P. Li, T. Wang, and Y . Liang, “Sanger: A Co-Design Framework for Enabling Sparse Attention using Reconfigurable Architecture,” in MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture , ser. MICRO ’21. New York, NY , USA: A...

    2021

  8. [16]

    ADC Performance Survey 1997-2024,

    B. Murmann, “ADC Performance Survey 1997-2024,” [Online]. Avail- able: https://github.com/bmurmann/ADC-survey

    1997

  9. [17]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” International journal of computer vision, vol. 115, pp. 211–252, 2015

    2015

  10. [18]

    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

  11. [19]

    Photonics for artificial intelligence and neuromorphic computing,

    B. J. Shastri, A. N. Tait, T. Ferreira de Lima, W. H. Pernice, H. Bhaskaran, C. D. Wright, and P. R. Prucnal, “Photonics for artificial intelligence and neuromorphic computing,” Nature Photonics, vol. 15, no. 2, pp. 102–114, 2021

    2021

  12. [20]

    Deep learning with coherent nanophotonic circuits,

    Y . Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund et al., “Deep learning with coherent nanophotonic circuits,” Nature photonics , vol. 11, no. 7, pp. 441–446, 2017

    2017

  13. [21]

    Compact 4-bit all optical digital to analog converter based on photonic crystal ring resonators,

    T. Sridarshini, S. I. Gandhi, and V . J. U. Firthouse, “Compact 4-bit all optical digital to analog converter based on photonic crystal ring resonators,” Laser Physics, vol. 30, no. 11, p. 116206, 2020

    2020

  14. [22]

    CrossLight: A cross- layer optimized silicon photonic neural network accelerator,

    F. Sunny, A. Mirza, M. Nikdast, and S. Pasricha, “CrossLight: A cross- layer optimized silicon photonic neural network accelerator,” in 2021 58th ACM/IEEE Design Automation Conference (DAC) . IEEE, 2021, pp. 1069–1074

    2021

  15. [23]

    Neuromorphic photonic networks using silicon photonic weight banks,

    A. N. Tait, T. F. De Lima, E. Zhou, A. X. Wu, M. A. Nahmias, B. J. Shastri, and P. R. Prucnal, “Neuromorphic photonic networks using silicon photonic weight banks,” Scientific reports, vol. 7, no. 1, p. 7430, 2017

    2017

  16. [24]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou, “Training data-efficient image transformers & distillation through attention,” in Proceedings of the 38th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, M...

    2021

  17. [25]

    Glue: A multi-task benchmark and analysis platform for natural language understanding,

    A. Wang, “Glue: A multi-task benchmark and analysis platform for natural language understanding,” arXiv preprint arXiv:1804.07461, 2018

    2018 arXiv

  18. [26]

    High-Performance and Resource-Efficient Dynamic Memory Management in High-Level Synthesis,

    Q. Wang, L. Zheng, Z. An, H. Huang, H. Zhu, Y . Huang, P. Yao, X. Liao, and H. Jin, “High-Performance and Resource-Efficient Dynamic Memory Management in High-Level Synthesis,” in Proceedings of the 61st ACM/IEEE Design Automation Conference , ser. DAC ’24. New York, NY , USA:...

    2024

  19. [27]

    A Data-Centric Accelerator for High-Performance Hyper- graph Processing,

    Q. Wang, L. Zheng, A. Hu, Y . Huang, P. Yao, C. Gui, X. Liao, H. Jin, and J. Xue, “A Data-Centric Accelerator for High-Performance Hyper- graph Processing,” in 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO), 2022, pp. 1326–1341

    2022

  20. [28]

    GraSU: A Fast Graph Update Library for FPGA-based Dynamic Graph Processing,

    Q. Wang, L. Zheng, Y . Huang, P. Yao, C. Gui, X. Liao, H. Jin, W. Jiang, and F. Mao, “GraSU: A Fast Graph Update Library for FPGA-based Dynamic Graph Processing,” in The 2021 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays , ser. FPGA ’21. New York, NY , US...

    2021

  21. [29]

    Hardware-Accelerated Hypergraph Processing with Chain- Driven Scheduling,

    Q. Wang, L. Zheng, J. Yuan, Y . Huang, P. Yao, C. Gui, A. Hu, X. Liao, and H. Jin, “Hardware-Accelerated Hypergraph Processing with Chain- Driven Scheduling,” in 2022 IEEE International Symposium on High- Performance Computer Architecture (HPCA) , 2022, pp. 184–198

    2022

  22. [30]

    A Conflict-free Scheduler for High-performance Graph Processing on Multi-pipeline FPGAs,

    Q. Wang, L. Zheng, J. Zhao, X. Liao, H. Jin, and J. Xue, “A Conflict-free Scheduler for High-performance Graph Processing on Multi-pipeline FPGAs,” ACM Trans. Archit. Code Optim. , vol. 17, no. 2, May 2020. [Online]. Available: https://doi.org/10.1145/3390523

    2020 doi

  23. [31]

    ReHarvest: An ADC Resource-Harvesting Crossbar Architecture for ReRAM-Based DNN Accelerators,

    J. Xu, H. Liu, Z. Duan, X. Liao, H. Jin, X. Yang, H. Li, C. Liu, F. Mao, and Y . Zhang, “ReHarvest: An ADC Resource-Harvesting Crossbar Architecture for ReRAM-Based DNN Accelerators,” ACM Trans. Archit. Code Optim. , vol. 21, no. 3, Sep. 2024. [Online]. Available: https://doi....

    2024 doi

  24. [32]

    ViTCoD: Vision Transformer Acceleration via Dedicated Algorithm and Accelerator Co-Design,

    H. You, Z. Sun, H. Shi, Z. Yu, Y . Zhao, Y . Zhang, C. Li, B. Li, and Y . Lin, “ViTCoD: Vision Transformer Acceleration via Dedicated Algorithm and Accelerator Co-Design,” in 2023 IEEE International Symposium on High- Performance Computer Architecture (HPCA) , 2023, pp. 273–286

    2023

  25. [33]

    TransPIM: A Memory- based Acceleration via Software-Hardware Co-Design for Transformer,

    M. Zhou, W. Xu, J. Kang, and T. Rosing, “TransPIM: A Memory- based Acceleration via Software-Hardware Co-Design for Transformer,” in 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2022, pp. 1071–1085

    2022

  26. [34]

    Lightening-Transformer: A Dynamically- Operated Optically-Interconnected Photonic Transformer Accelerator,

    H. Zhu, J. Gu, H. Wang, Z. Jiang, Z. Zhang, R. Tang, C. Feng, S. Han, R. T. Chen, and D. Z. Pan, “Lightening-Transformer: A Dynamically- Operated Optically-Interconnected Photonic Transformer Accelerator,” in 2024 IEEE International Symposium on High-Performance Computer Archi...

    2024

  27. [2024]

    Available: https://doi.org/10.1145/3649329.3658488

    [Online]. Available: https://doi.org/10.1145/3649329.3658488

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.