Pith. sign in

REVIEW 2 major objections 6 minor 25 references

COBRA: Algorithm-Architecture Co-optimized Binary Transformer Accelerator for Edge Inference

T0 review · 2 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read COBRA is an algorithm-architecture co-designed accelerator that runs binary transformers on edge FPGAs, claiming 3,894.7 GOPS throughput, 448.7 GOPS/W, a 311x energy-efficiency gain over a GPU, and 3.5x throughput over the prior binary…

desk verdict A genuinely interesting FPGA accelerator for binary Transformers with a clever 1-bit RBMM unit, but the 'negligible accuracy loss' claim is calibrated per-task and overstated, and the 3.5x speedup number doesn't match the paper's own table. read the letter →

arxiv 2504.16269 v2 pith:BL5YLXM7 submitted 2025-04-22 cs.AR cs.LG

Ye Qiao , Zhiheng Chen , Yian Wang , Yifan Zhang , Yunzhe Deng , Sitao Huang This is my paper · ORCID
classification cs.ARcs.LG
keywords binarytransformerFPGAacceleratoredgeinferencehardware-softwareco-design1-bitmatrixmultiplicationquantizationattentionmechanismenergyefficiency
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

COBRA is a co-designed algorithm and hardware accelerator that runs binarized BERT-style transformers on edge FPGAs. The paper claims that by replacing softmax with a threshold-only shifted polarized softmax (SPS) and by building a genuine 1-bit matrix engine that works on -1, 0, and +1 values, the accelerator reaches 3,894.7 GOPS and 448.7 GOPS/W on a ZCU102 board. That is a 311x energy-efficiency gain over a GPU baseline and 3.5x the throughput of the prior best binary transformer accelerator, with a relative average accuracy drop of 1.8% on GLUE compared with BiT. The claim matters because it suggests extreme low-bit transformers, normally thought too inaccurate and too awkward to accelerate, can be both accurate enough and fast enough for local edge inference.

What carries the argument

The central object is the RBMM engine, a real 1-bit binary matrix-multiplication unit. Its vector dot-product identity is $2\,\mathrm{popcount}(\mathrm{XNOR}(a_m,b_p))-N$ for signed $(-1,1)$ data and $2\,\mathrm{popcount}(a_m\,\mathrm{AND}\,b_p)-N+\delta_m$ for unsigned $(0,1)$ data, where $\delta_m$ counts the zeros in the unsigned datapack. Popcount counts the number of `1` bits. This one engine serves both binarization schemes used by binary transformers. The second mechanism is SPS, $SPS(z)=1$ if $z\ge\lambda_{i,k}$ else $0$, a threshold-only attention function that removes softmax's exponentials and the separate rounding-clipping step. Together they reduce attention and feedforward computation to AND/XNOR, comparisons, and popcounts.

What would settle it

Run COBRA's SPS attention on a GLUE task excluded from the threshold search (or on a domain-shifted dataset such as a different sentiment corpus) without any calibration data or fine-tuning, and compare against BiT's softmax attention; if the accuracy gap exceeds a few points, the claim that SPS causes only negligible degradation would be falsified.

Watch

Extended reading notes

Core claim

The paper claims that the two obstacles to edge binary transformers—softmax's floating-point cost and the lack of a true binary multiply unit—can both be removed. Its RBMM engine treats -1 and 0 as the same hardware bit and derives the vector dot product from popcounts of XNOR and AND operations, fusing quantization and ReLU into the comparison. Its SPS replaces softmax plus rounding/clipping with a per-head threshold comparison, with thresholds searched on a 10% calibration sample to match BiT's attention maps, followed by fine-tuning. On BERT-base, the co-designed system runs the full model with only a small accuracy reduction while delivering the reported throughput and energy efficiency. The key claim is that real 1-bit matrix multiplication plus threshold-only attention is sufficient for practical edge inference.

Load-bearing premise

The accuracy claim rests on having a small calibration set (10% of the target task) to fit SPS thresholds and then fine-tuning the model to imitate BiT's softmax attention; on a new edge task without such calibration, the accuracy loss could be larger than the reported 1.8%.

Editorial extensions

If this is right

  • Binarized BERT-base can be served on a mid-range edge FPGA (ZCU102) at 3,894.7 GOPS, so on-device NLP inference at this scale does not require a GPU.
  • Removing softmax in favor of SPS eliminates the main latency and area bottleneck: the ablation shows a 564x throughput gain against a softmax unit, so attention stops being the limiting factor.
  • One RBMM engine with six operation modes covers all matrix multiplications in a transformer (Q/K/V, attention scores, context, and both FFN layers), so area and power stay low.
  • The same architectural choices work on a lower-power KV260 board with fewer processing elements, supporting deployment on more constrained edge devices.
  • Because the engine is fully pipelined with an initiation interval of one cycle, throughput scales with PE count, which is why Npe=32 on ZCU102 gives the best measured result.

Reading between the lines

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

  • Not explored in the paper: the same "don't-care" zero-count mechanism could encode structured sparsity, letting pruned or sparse binary weights be represented as zeros that the RBMM skips, combining binarization and pruning in one engine.
  • If SPS thresholds were derived from the model's own attention statistics rather than from a BiT teacher, the design could transfer to new tasks without any calibration set; this is testable and would extend the accuracy claim beyond GLUE.
  • The decomposition property (Eq. 8) suggests the single RBMM engine could be time-multiplexed to larger hidden sizes or longer sequences on the same FPGA, trading throughput for capacity.
  • The "negligible accuracy degradation" claim is specific to the 10%-calibration SPS setup; a harder test is zero-shot SPS with fixed thresholds, where the gap may widen.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. COBRA presents an FPGA accelerator for binarized BERT-style transformers. The algorithmic side replaces softmax in attention with Shifted Polarized Softmax (SPS), a per-head threshold comparison, and introduces a Real Binary Matrix Multiplication (RBMM) engine that computes dot products of 1-bit {-1,1} and {0,1} vectors with XNOR/AND, popcount, and a don't-care count. The hardware is implemented on KV260 and ZCU102; on ZCU102 it reports 3,894.7 GOPS, 448.7 GOPS/W, a 3.5x throughput improvement over BAT, and a 311x energy-efficiency improvement over a GPU baseline. On GLUE, the COBRA (Head) configuration averages 68.2 versus 71.0 for BiT, a 2.8-point average drop, with per-task drops up to 9.6 points on QNLI.

Significance. If the hardware measurements hold, COBRA is a meaningful advance in edge binary-transformer acceleration. The RBMM arithmetic in Eq. (7) is correct, the ablation in Table V cleanly isolates the SPS, 6:3-compressor popcount, and pipelining contributions, and the on-board results on two FPGAs with detailed resource utilization are valuable. The main caveat is that the accuracy claim is tied to a per-task calibration protocol: thresholds are fitted to match BiT's attention on 10% of each benchmark (Eq. (6)) and the model is then fine-tuned, so 'negligible accuracy degradation' is not yet established as a portable property of the architecture for uncalibrated edge tasks.

major comments (2)
  1. [III-A3 / Table I] The abstract and Section III-A4 claim 'negligible inference accuracy degradation' relative to BiT, but Table I shows per-task drops that are not negligible for deployed models: QNLI falls from 82.8 to 73.2 (-9.6 points), and MNLI-m/mm fall by about 4.3 points, even though the GLUE average drops only 2.8 points. Since the SPS thresholds are selected by minimizing distortion against BiT's attention maps on a 10% calibration sample of the same benchmark (Eq. (6)) and the model is then fine-tuned, this accuracy is a property of a per-task calibration procedure, not a portable property of the architecture. The text also refers to a 'less than 2% minor average performance drop' on the basis of the 98.2% relative-performance column, which hides the large per-task drops. The paper should either add a transfer or zero-calibration evaluation (e.g., thresholds fixed across tasks, or a new task with no calibration data), or explicitly scope the claim as 'after per-task threshold calibration and fine-tuning.'
  2. [III-A3 / evaluation protocol] The threshold-search calibration is not fully specified: the paper states that 10% of each benchmark is sampled to search the thresholds, but it does not say whether this sample comes from the training split or the evaluation split. If any evaluation data are used to set thresholds, Table I overstates generalization; if training data are used, that should be stated explicitly. The paper should also report multiple seeds or per-task variance, because a single-run 9.6-point QNLI drop makes the 'negligible degradation' claim fragile.
minor comments (6)
  1. [III-A3] Eq. (5) uses n both as the fixed matrix dimension and as the summation index; rename the index or the dimension to avoid confusion.
  2. [III-B1] The derivation of Eq. (7) is omitted with the phrase 'steps omitted for brevity'; given that the don't-care-count mechanism is a core novelty, include the derivation explicitly.
  3. [IV-B / Table II] Table II's layout is ambiguous: the Platform row and the Network row do not align, and it is unclear which GPU baseline (BiT or COBRA) is used for the 311x claim; reformat the table and state the baselines clearly in the text.
  4. [IV-B] The power numbers in Table II are central to the energy-efficiency claims, but the paper does not describe the power measurement methodology; add a sentence on whether power is measured on-board or estimated by the toolchain.
  5. [III-A4] The claim that row-wise thresholds do not offer meaningful improvements is reported without any supporting numbers; either include the row-wise results or soften the claim.
  6. [Throughout] There are several typos and inconsistent labels (e.g., 'serveral', 'originial', and 'low-power ZCU102' in Section IV-B, where only KV260 is the low-power board); a copyedit pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: SPS is a calibrated approximation whose accuracy is empirically evaluated, and the hardware results are measured independently.

full rationale

The paper's central hardware claims (RBMM throughput, energy efficiency, FPGA resource results) are measured on ZCU102 and KV260 and do not depend on the SPS threshold search. The SPS thresholds in Eq. (6) are fitted to minimize attention-map distortion relative to BiT's softmax-based attention, but the reported 'negligible inference accuracy degradation' is an empirical outcome after fine-tuning and evaluation on GLUE benchmarks, not a constructional consequence of the fit. The per-task calibration protocol is a legitimate generalization and robustness concern, especially given the QNLI drop, but it is not circularity. The attention-map similarity in Figure 3 is a fitting target rather than an independent prediction, and it is not the load-bearing evidence for the headline claim, which rests on Table I's held-out accuracy and Table II's measured FPGA performance. The RBMM identities (Eqs. 7-8 and 10) are algebraic reformulations of binary dot products and are not circular. There is no load-bearing self-citation chain or imported uniqueness theorem: Refs. [3], [9], and [10] are background citations only. The derivation chain is therefore self-contained with respect to the hardware contributions, and the accuracy claim is an empirical, calibrated result rather than a circular prediction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central accuracy claim depends on the SPS thresholds being fitted on per-task calibration data, while the hardware throughput claims depend on the chosen Npe and the synthesis/measurement assumptions. No new physical entities are introduced.

free parameters (2)
  • SPS threshold lambda_{i,k} = Head-wise thresholds in [0,1], granularity 0.05; for BERT-base h=12 heads per layer, 12 layers = 144 thresholds per task
    Searched per task on a 10% calibration set to minimize attention-map distortion vs. BiT (Eq. 6). These thresholds directly control the binarized attention pattern and contribute to the accuracy result.
  • Number of processing elements Npe = 32 on ZCU102 (16 on KV260)
    Chosen by hand to balance resource usage and throughput; the headline 3,894 GOPS figure uses Npe=32 (Table II).
assumptions (3)
  • standard math Binary dot product can be computed as 2*popcount(XNOR(bits)) - N for {-1,1} values and with the DC correction Eq. (7) for mixed {0,1} pairs.
    Assumed in Section III-B1, Eq. (7); this is a correct algebraic identity, but the engine's correctness depends on it.
  • domain assumption The SPS thresholds found on a 10% calibration sample of each GLUE benchmark are representative of the full task distribution and remain effective after fine-tuning.
    Stated in Section III-A3; if the calibration sample is not representative, the compared accuracy would change. No cross-task validation is provided.
  • domain assumption The FPGA implementation synthesized and ran at 300 MHz with the reported power and resource figures, and the GOPS counting is consistent across compared systems.
    Section IV-A reports timing closure and on-board evaluation; no measurement scripts or power breakdown are provided, so this is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of COBRA: Algorithm-Architecture Co-optimized Binary Transformer Accelerator for Edge Inference." pith.science (2026). https://pith.science/paper/BL5YLXM7

@misc{pith2026250416269,
  author       = {Pith},
  title        = {Pith review of: COBRA: Algorithm-Architecture Co-optimized Binary Transformer Accelerator for Edge Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BL5YLXM7}},
  note         = {Machine review of arXiv:2504.16269}
}
read the original abstract

Transformer-based models have demonstrated superior performance in various fields, including natural language processing and computer vision. However, their enormous model size and high demands in computation, memory, and communication limit their deployment to edge platforms for local, secure inference. Binary transformers offer a compact, low-complexity solution for edge deployment with reduced bandwidth needs and acceptable accuracy. However, existing binary transformers perform inefficiently on current hardware due to the lack of binary specific optimizations. To address this, we introduce COBRA, an algorithm-architecture co-optimized binary Transformer accelerator for edge computing. COBRA features a real 1-bit binary multiplication unit, enabling matrix operations with -1, 0, and +1 values, surpassing ternary methods. With further hardware-friendly optimizations in the attention block, COBRA achieves up to 3,894.7 GOPS throughput and 448.7 GOPS/Watt energy efficiency on edge FPGAs, delivering a 311x energy efficiency improvement over GPUs and a 3.5x throughput improvement over the state-of-the-art binary accelerator, with only negligible inference accuracy degradation.

Figures

Figures reproduced from arXiv: 2504.16269 by the authors.

Figure 1
Figure 1. General structure of binarized BERT with our Shifted [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Shifted Polarized Softmax (SPS) Search the matrix multiplication of binarized attention. BETA [18] employs a Compressor Tree Loop to create an accelerator specifically for fully binary transformers, supporting both matrix multiplication types and achieving improved efficiency and throughput on edge platforms. BAT [19] offers a co-design approach, creating a custom binarized transformer model with a specialized hardw… view at source ↗
Figure 3
Figure 3. Similarity and Correlation Comparisons Between BiT [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: A 6-bit example of our RBVM. 𝑨 𝑫𝒂𝒕𝒂𝒑𝒂𝒄𝒌 𝑩 𝑫𝒂𝒕𝒂𝒑𝒂𝒄𝒌𝟏 𝑩 𝑫𝒂𝒕𝒂𝒑𝒂𝒄𝒌𝑵𝒑𝒆 … A B On Chip Memory RBMM Engine 𝑶𝒖𝒕𝒑𝒖𝒕𝟐𝑩 𝑶𝒖𝒕𝒑𝒖𝒕𝟐𝑨 Off A B M2 -Chip DDR RAM M3 DC Return Return Output LayerNorm Unit Local Buffer AXI BUS 𝑹𝑩𝑴𝑴 𝑷𝑬𝑵𝒑𝒆 𝑹𝑩𝑴𝑴 𝑷𝑬 Npe 𝟐 𝑹𝑩𝑴𝑴 𝑷𝑬𝟏 COBRA Control Internal Read/Wr…
Figure 5
Figure 5. Figure 5: COBRA Hardware Architecture Overview. performance improvements while introducing more parameters to the model. It also increases the search time by over 20×. Therefore, we adopt the head-wise threshold as the default configuration for the SPS function in our COBRA mode…
Figure 6
Figure 6. Figure 6: The Architecture of RBMM Processing Elements (M1– [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: COBRA Accelerator Implementation Layout (ZCU102) [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 12 canonical work pages

  1. [1]

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

    J. D. M.-W. C. Kenton and L. K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of naacL-HLT, vol. 1. Minneapolis, Minnesota, 2019, p. 2

    work page 2019

  2. [2]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    D. Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv: 2010.11929 , 2020

    arXiv 2010

  3. [3]

    Support for Stock Trend Prediction Using Transformers and Sentiment Analysis

    H. Kaeley, Y . Qiao, and N. Bagherzadeh, “Support for stock trend prediction using transformers and sentiment analysis,” arXiv preprint arXiv:2305.14368, 2023

    work page Pith review arXiv 2023

  4. [4]

    Progprompt: Generating situated robot task plans using large language models,

    I. Singh, V . Blukis, A. Mousavian, A. Goyal, D. Xu, J. Tremblay, D. Fox, J. Thomason, and A. Garg, “Progprompt: Generating situated robot task plans using large language models,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 11 523–11 530

    2023

  5. [5]

    Q8bert: Quantized 8bit bert,

    O. Zafrir, G. Boudoukh, P. Izsak, and M. Wasserblat, “Q8bert: Quantized 8bit bert,” in 2019 Fifth Workshop on Energy Efficient Machine Learning and Cognitive Computing-NeurIPS Edition (EMC2-NIPS). IEEE, 2019, pp. 36–39

    work page 2019

  6. [6]

    A fast post-training pruning framework for transform- ers,

    W. Kwon, S. Kim, M. W. Mahoney, J. Hassoun, K. Keutzer, and A. Gholami, “A fast post-training pruning framework for transform- ers,” Advances in Neural Information Processing Systems , vol. 35, pp. 24 101–24 116, 2022

    work page 2022

  7. [7]

    Accelerating binarized convolutional neural networks with software-programmable fpgas,

    R. Zhao, W. Song, W. Zhang, T. Xing, J.-H. Lin, M. Srivastava, R. Gupta, and Z. Zhang, “Accelerating binarized convolutional neural networks with software-programmable fpgas,” in Proceedings of the 2017 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays , ser. FPGA ’17. New York, NY , USA: Association for Computing Machinery, 2017, p. 15–...

    arXiv 2017

  8. [8]

    Fracbnn: Accurate and fpga-efficient binary neural networks with fractional activations,

    Y . Zhang, J. Pan, X. Liu, H. Chen, D. Chen, and Z. Zhang, “Fracbnn: Accurate and fpga-efficient binary neural networks with fractional activations,” in The 2021 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays , ser. FPGA ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 171–182. [Online]. Available: https://doi.or...

    arXiv 2021

Show all 25 references
  1. [9]

    A two-stage efficient 3- d cnn framework for eeg based emotion recognition,

    Y . Qiao, M. Alnemari, and N. Bagherzadeh, “A two-stage efficient 3- d cnn framework for eeg based emotion recognition,” in 2022 IEEE International Conference on Industrial Technology (ICIT). IEEE, 2022, pp. 1–8

    2022

  2. [10]

    Bnn an ideal architecture for acceleration with resistive in memory computation,

    A. Ding, Y . Qiao, and N. Bagherzadeh, “Bnn an ideal architecture for acceleration with resistive in memory computation,” IEEE Transactions on Emerging Topics in Computing , vol. 11, no. 2, pp. 281–291, 2023

    2023

  3. [11]

    BinaryBERT: Pushing the limit of bert quantization,

    H. Bai, W. Zhang, L. Hou, L. Shang, J. Jin, X. Jiang, Q. Liu, M. Lyu, and I. King, “BinaryBERT: Pushing the limit of bert quantization,”arXiv preprint arXiv:2012.15701, 2020

    2012 arXiv

  4. [12]

    BiBERT: Accurate fully binarized BERT,

    H. Qin, Y . Ding, M. Zhang, Q. Yan, A. Liu, Q. Dang, Z. Liu, and X. Liu, “BiBERT: Accurate fully binarized BERT,” arXiv preprint arXiv:2203.06390, 2022

    2022 arXiv

  5. [13]

    BiT: Robustly binarized multi-distilled transformer,

    Z. Liu, B. Oguz, A. Pappu, L. Xiao, S. Yih, M. Li, R. Krishnamoorthi, and Y . Mehdad, “BiT: Robustly binarized multi-distilled transformer,” Advances in neural information processing systems, vol. 35, pp. 14 303– 14 316, 2022

    2022

  6. [14]

    TernaryBERT: Distillation-aware ultra-low bit BERT,

    W. Zhang, L. Hou, Y . Yin, L. Shang, X. Chen, X. Jiang, and Q. Liu, “TernaryBERT: Distillation-aware ultra-low bit BERT,” arXiv preprint arXiv:2009.12812, 2020

    2009 arXiv

  7. [15]

    Yodann: An architecture for ultralow power binary-weight cnn acceleration,

    R. Andri, L. Cavigelli, D. Rossi, and L. Benini, “Yodann: An architecture for ultralow power binary-weight cnn acceleration,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 37, no. 1, pp. 48–60, 2017

    2017

  8. [16]

    Xnorbin: A 95 top/s/w hardware accelerator for binary convolutional neural networks,

    A. Al Bahou, G. Karunaratne, R. Andri, L. Cavigelli, and L. Benini, “Xnorbin: A 95 top/s/w hardware accelerator for binary convolutional neural networks,” in 2018 IEEE Symposium in Low-Power and High- Speed Chips (COOL CHIPS) . IEEE, 2018, pp. 1–3

    2018

  9. [17]

    Vaqf: Fully automatic software-hardware co-design frame- work for low-bit vision transformer,

    M. Sun, H. Ma, G. Kang, Y . Jiang, T. Chen, X. Ma, Z. Wang, and Y . Wang, “Vaqf: Fully automatic software-hardware co-design frame- work for low-bit vision transformer,” arXiv preprint arXiv:2201.06618 , 2022

    2022 arXiv

  10. [18]

    BETA: Binarized Energy-Efficient Transformer Accelerator at the Edge,

    Y . Ji, C. Fang, and Z. Wang, “BETA: Binarized Energy-Efficient Transformer Accelerator at the Edge,” arXiv preprint arXiv:2401.11851, 2024

    2024 arXiv

  11. [19]

    Co-designing binarized transformer and hardware accelerator for efficient end-to-end edge deployment,

    Y . Ji, C. Fang, S. Ma, H. Shao, and Z. Wang, “Co-designing binarized transformer and hardware accelerator for efficient end-to-end edge deployment,” arXiv preprint arXiv:2407.12070 , 2024

    2024 arXiv

  12. [20]

    Flashattention: fast and memory-efficient exact attention with io-awareness,

    T. Dao, D. Y . Fu, S. Ermon, A. Rudra, and C. R ´e, “Flashattention: fast and memory-efficient exact attention with io-awareness,” in Proceedings of the 36th International Conference on Neural Information Processing Systems, ser. NIPS ’22. Red Hook, NY , USA: Curran Associates...

    2022

  13. [21]

    GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding,

    A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. R. Bowman, “GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding,” 2019. [Online]. Available: https://arxiv.org/abs/1804.07461

    2019 arXiv

  14. [22]

    Hardware acceleration of fully quan- tized bert for efficient natural language processing,

    Z. Liu, G. Li, and J. Cheng, “Hardware acceleration of fully quan- tized bert for efficient natural language processing,” in 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE) , 2021, pp. 513–516

    2021

  15. [23]

    Cosa:co-operative systolic arrays for multi-head attention mechanism in neural network using hybrid data reuse and fusion methodologies,

    Z. Wang, G. Wang, H. Jiang, N. Xu, and G. He, “Cosa:co-operative systolic arrays for multi-head attention mechanism in neural network using hybrid data reuse and fusion methodologies,” in 2023 60th ACM/IEEE Design Automation Conference (DAC) , 2023, pp. 1–6

    2023

  16. [24]

    Transfru: Efficient deployment of transformers on fpga with full resource utilization,

    H. Wang, Y . Bai, J. Yu, and K. Wang, “Transfru: Efficient deployment of transformers on fpga with full resource utilization,” in 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), 2024, pp. 521–526

    2024

  17. [25]

    Hamming weight — wikipedia, the free encyclopedia

    Wikipedia, “Hamming weight — wikipedia, the free encyclopedia.” [Online]. Available: https://en.wikipedia.org/wiki/Hamming weight

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.