Pith. sign in

REVIEW 4 major objections 6 minor 49 references

Bhasha-Rupantarika: Algorithm-Hardware Co-design approach for Multilingual Neural Machine Translation

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read FP4 quantization of a distilled 200-language translation model, paired with a custom FPGA accelerator, delivers a 4.1x smaller footprint, 4.2x faster inference, and 66 tokens/s throughput — without losing translation quality, according to t

desk verdict Concrete FPGA accelerator for 4-bit NLLB-200, but missing translation-quality metrics and uneven hardware comparisons undermine the central claims. read the letter →

arxiv 2510.10676 v1 pith:FKV3GYRM submitted 2025-10-12 cs.AR cs.CLcs.ROeess.AS

Mukul Lokhande , Tanushree Dewangan , Mohd Sharik Mansoori , Tejas Chaudhari , Akarsh J. , Damayanti Lokhande , Adam Teman , Santosh Kumar Vishvakarma This is my paper · ORCID
classification cs.ARcs.CLcs.ROeess.AS
keywords multilingualneuralmachinetranslationmodelquantizationFP4precisionFPGAacceleratoralgorithm-hardwareco-designNLLB-200Indiclanguagestransformerinference
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

Bhasha-Rupantarika is an algorithm-hardware co-design for running multilingual translation on resource-limited devices. The paper claims that a distilled 200-language translation model, quantized to FP4, drops to 0.56 GB (4.1x smaller than FP32), infers 4.2x faster, and reaches 66 tokens/s (a 4.8x improvement), all without compromising translation ability. To back that up, it introduces an FPGA accelerator (NLPE) with a mixed-precision SIMD multiply-accumulate array and a CORDIC-based activation unit; on a ZCU104 board the design uses 1.96x fewer LUTs and 1.65x fewer FFs than prior NLP accelerators, with 2.2x and 4.6x higher throughput. If these results hold, a single bidirectional model can translate between Indian and international languages on IoT-class hardware, a step toward real-time translation in rural areas.

What carries the argument

The load-bearing mechanism is the NLPE accelerator, built around two reusable units: a SIMD multiply-accumulate engine whose 4-bit RMMEC blocks can be configured as multipliers or exponent comparators, allowing INT4, FP4, FP8, and BF16 operations in the same datapath; and a CORDIC-based FASST vector array that delivers SoftMax, sigmoid, tanh, and related nonlinearities at FP8/BF16 precision. An output-stationary systolic MAC array keeps partial sums in place while weights and activations stream through, and all units are pipelined to overlap computation with memory traffic. On the algorithm side, the claim rests on quantizing a distilled encoder-decoder transformer (NLLB-200, 600M parameters

What would settle it

Run the proposed NLPE and the two prior accelerators from the paper's comparison on the same FPGA board with the same NLLB-200 INT4/FP4 workload and the same synthesis flow, and see whether the 1.96x LUT reduction and 2.2x/4.6x throughput gains persist. Separately, compute BLEU or COMET scores for FP4 versus FP32 on the same query set to test the 'without compromising translation capability' claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that sub-octet quantization is what makes a 200-language neural translation model deployable on edge hardware, and that a co-designed accelerator lets the savings appear in real silicon. Concretely, the distilled NLLB-200 model (a 200-language transformer), post-training quantized to 4-bit floating point with small trainable adapters, keeps translations usable while cutting the footprint to 0.56 GB and enabling 66 tokens/s on the proposed NLPE engine. The accelerator's SIMD MAC array uses reconfigurable 4-bit RMMEC blocks that switch between multiplier and exponent-comparator roles, and its CORDIC-based FASST unit computes SoftMax, sigmoid, tanh, and related nonl

Load-bearing premise

The load-bearing premise is that the hardware comparison is fair: the new design is measured on one board and model while the compared accelerators use different boards and models, so the claimed resource and throughput advantages could partly reflect an easier workload rather than a better design.

Editorial extensions

If this is right

  • A 200-language translation model can fit in 0.56 GB, small enough for edge and IoT storage budgets.
  • At 66 tokens/s, interactive translation on a low-power FPGA becomes realistic rather than batch-only.
  • A single bidirectional model replaces separate Indian-to-English, English-to-foreign, and reverse models, cutting total model count and memory overhead.
  • The same accelerator datapath can be reconfigured across precisions and activation types, so future transformer variants can deploy without redesigning the MAC or nonlinear units.
  • The reported resource savings leave headroom on the same FPGA class for larger or additional workloads.

Reading between the lines

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

  • Because the reported hardware gains compare a small translation model on one board against larger models on other boards, a controlled replay on identical workloads is needed before treating the resource ratios as general.
  • If independent translation-quality scores confirm the FP4 result, 4-bit floating point could become a default edge precision for multilingual sequence-to-sequence models, extending beyond Indian languages.
  • The paper's pipeline sketches speech-to-text and text-to-speech around the translator; an integrated voice-to-voice system on the same FPGA is the natural next experiment, with end-to-end latency as the success metric.
  • The authors' scaling hypothesis suggests testing the same RMMEC/FASST datapath at larger model sizes; a positive result would extend the co-design from edge translation to cloud-scale chatbots.
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

4 major / 6 minor

Summary. The paper presents Bhasha-Rupantarika, an algorithm-hardware co-design flow for multilingual NMT on edge FPGAs. The software side applies BitsAndBytes-based post-training quantization and QLoRA fine-tuning to a distilled NLLB-200 600M model, reporting a 0.56 GB FP4 footprint, 4.1x model-size reduction, 4.2x inference speedup, and 66 tokens/s throughput. The hardware side proposes the NLPE accelerator with an RMMEC-based SIMD MAC unit and a CORDIC-based FASST activation vector array, synthesized on a ZCU104, with claimed 1.96x LUT and 1.65x FF reductions and 2.2x/4.6x throughput gains over OPU/HPTA. The stated goal is to enable low-resource multilingual translation, particularly between Indian and international languages.

Significance. The problem is relevant and timely: quantized multilingual NMT with a custom accelerator could be useful for edge deployment in resource-limited regions. Strengths include the public code/dataset link, the RTL implementation in SystemVerilog, the use of an iso-functional emulator for co-verification, and the comparison against an FP32 baseline for the quantization study. However, the central claims are not currently established: translation quality is never measured, the FP4 footprint number is internally inconsistent, and the hardware efficiency claims rest on non-apples-to-apples comparisons. If these issues are corrected, the work could make a meaningful contribution; as submitted, the evidence does not support the conclusions.

major comments (4)
  1. [Section III / Abstract / Conclusion] No translation-quality metric is reported. Section III states only that PTQ was evaluated on 1000 queries per language and produced 'satisfactory results'; no BLEU, chrF, COMET, or human evaluation is provided, and Figure 10 plots only model size, translation time, and tokens/s. The Conclusion then asserts 'without compromising translation capability.' The paper itself concedes in Section III that 'a comprehensive error analysis would highlight the exact translation performance...', acknowledging the absence. Since the headline 4.1x size reduction and 4.2x speedup are only meaningful if output quality is preserved, this missing evidence is load-bearing. The authors should add quantitative quality metrics for the FP32 baseline and each quantized precision, ideally broken down by language pair.
  2. [Section III, Fig. 10] The reported FP4 footprint of 0.56 GB is arithmetically inconsistent with a 600M-parameter model at 4 bits, which should occupy roughly 0.3 GB; 0.56 GB implies about 7.5 bits per parameter. If the number includes block-wise scale/zero-point overhead, embedding tables, or non-weight tensors, that breakdown must be provided. As written, the 'FP4' label and the 4.1x model-size reduction claim are not self-consistent and require either correction or a clear accounting of the extra storage.
  3. [Section III, Table IV] The headline hardware improvements are derived from cross-platform, cross-model comparisons. NLPE is evaluated for NLLB on a ZCU104, while Table IV compares it with NPE (BERT on Zynq Z-7100), HPTA (Swin-T on ZCU102), Q-BERT (MNLI on ZCU102), TPDS'22 (NMT on VCU118), and others. These differ in board, model, precision, and synthesis flow, so a smaller LUT count or higher GOPS figure can reflect different workload sizes rather than design superiority. The claimed 1.96x LUT reduction, 1.65x FF reduction, and 2.2x/4.6x throughput gains are therefore not established. The authors should either compare on the same model/FPGA or provide a principled normalization that accounts for model size, precision, and hardware resources.
  4. [Section III / Section II-B] The paper does not demonstrate end-to-end execution of NLLB on the NLPE. Section III describes synthesis and verification of the MAC and NAF units and reports a board-level resource table, but Figure 10's tokens/s appears to come from a software/GPU inference run, and the 684.48 GOPS entry in Table IV is unexplained. No full-transformer mapping, memory-bandwidth analysis, or measured end-to-end latency for the NLLB model on the FPGA is provided. The claim that the accelerator 'deploys' the quantized NLLB model is therefore not substantiated. Please separate unit-level results from system-level projections and provide an actual end-to-end FPGA demonstration or a carefully validated performance model.
minor comments (6)
  1. [Section II-A] The text alternates between 'FP4' and 'INT4' for the distilled NLLB variant, and BitsAndBytes 4-bit quantization is typically NF4 rather than FP4. Clarify the exact format used and use consistent terminology throughout.
  2. [Fig. 7] The caption refers to 'NVU' while the text refers to 'NMV'; unify the naming for the nonlinear vector array.
  3. [Table IV] The column heading 'Throughput GOPS' is ambiguous; some rows have a dash for BRAM/throughput. Add units, define GOPS, and explain how the 684.48 GOPS figure is derived for the NLPE.
  4. [References] Citation numbering is inconsistent, e.g., OPU is cited as both [8] and [21] in the same paragraph. Check all references for duplicate or mismatched numbering.
  5. [Section III] The text says 'Python 3.0,' which is likely outdated; use 'Python 3.x.' Also, the role of Qkeras in a BitsAndBytes/QLoRA pipeline is unclear and should be explained or removed.
  6. [Fig. 10] The figure lacks a legend or color labels; it is difficult to tell which series corresponds to model size, translation time, and tokens per second. Add axis clarity and a legend.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; software and hardware claims are measured outcomes, and self-citations serve only as comparison baselines.

full rationale

The paper's derivation chain is self-contained. The quantization results (model size, translation time, tokens/s) are direct measurements of NLLB-600M after PTQ/QLoRA at each precision; no parameter is fitted to a target quantity and then reported as a prediction. The accelerator resource and throughput figures come from the authors' own Vivado and Cadence synthesis of their RTL, so the '1.96x fewer LUTs' and '684.48 GOPS' numbers are empirical outputs of their design flow, not consequences of a self-citation. The comparisons to Flex-PE [41], LPRE [31], XR-NPE [36], and QForce-RL [46] are benchmark tables; although those baselines are prior work by overlapping authors, the present design's numbers are independently synthesized in this flow, so the self-citations are not load-bearing in the sense of supplying the derivation. No uniqueness theorem or ansatz is imported from the authors' prior work. The '4.1x smaller footprint' is a direct arithmetic consequence of bit-width quantization rather than a circular step; any inconsistency with the ideal 4-bit parameter count (0.56 GB vs ~0.3 GB for 600M parameters) is an accuracy/soundness concern, not a circularity. Likewise, the absence of BLEU/chrF/COMET, and the paper's own concession that 'A comprehensive error analysis would highlight the exact translation performance...' (end of Section III), are missing-evidence problems for the claim of preserved translation quality, not circular derivations. The hardware comparison is vulnerable to benchmark-mismatch (different boards/models), but that is a validity-of-comparison issue outside the circularity taxonomy. Therefore no circular step is exhibited, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 2 invented entities

The paper introduces no fitted numeric parameters; its central claims rest instead on implicit premises: that 4-bit PTQ preserves NLLB quality (never measured), that cross-FPGA/cross-model published numbers are comparable baselines (Table IV), that the NAF-60%-of-ops workload characterization is accurate (Section III), and that the 1000-query-per-language set is representative (Section III). The hardware units (NLPE, RMMEC, FASST) are novel artifacts with only internal synthesis evidence, so they carry no independent falsifiable handle outside the paper.

free parameters (2)
  • Quantization precisions (FP8, INT8, INT4, FP4) = FP8/INT8/INT4/FP4
    Study variables chosen by the authors; the central claims (4.1x size, 4.2x speed, 66 tokens/s) are conditional on these specific formats and would differ under other formats.
  • Systolic array / MME dimensions = parameterized, not reported
    Section II.B: 'The size of the systolic array is kept parameterized depending on the dimensions of the matrices'; concrete array sizes and CORDIC iteration counts are not stated, so the throughput numbers cannot be recomputed externally.
assumptions (5)
  • ad hoc to paper BitsAndBytes 4-bit PTQ plus QLoRA low-rank adapters preserve NLLB-200 translation quality close to FP32.
    Section III: 'performance was evaluated on 1000 queries per language, yielding satisfactory results' — asserted without any reported accuracy metric; the entire 'without compromising translation capability' claim rests on this.
  • ad hoc to paper Resource/throughput numbers from OPU, HPTA, Q-BERT, TPDS'22 etc. (different models, FPGAs, flows) are directly comparable to the NLPE on ZCU104.
    Table IV compares NLLB-on-ZCU104 against BERT-on-Z-7100, Swin-T-on-ZCU102, MNLI-on-ZCU102, and NMT-on-VCU118; the 1.96x/2.2x/4.6x claims depend on this comparability.
  • domain assumption Workload characterization (NAF ≈ 60% of NLLB operations) is accurate.
    Section III: 'workload characterization indicates that the NAF accounts for up to 60% of the operations in the overall NLLB' — no profiling methodology or data is shown.
  • domain assumption NLLB-200 distilled INT4 600M model is a suitable high-quality backbone for Indic↔overseas translation.
    Section II.A relies on the pretrained model's capabilities; quality is assumed, not re-measured.
  • domain assumption The 1000-query-per-language evaluation set is representative.
    Section III: the set is not described (languages, domains, sources), yet it grounds every software claim.
invented entities (2)
  • NLPE (Neural Language Processing Engine) with RMMEC-based SIMD MAC and FASST CORDIC activation array
    purpose: Quantized multi-precision inference engine for the NLLB-200 transformer on FPGAs (basis of LUT/FF/throughput claims)
    New hardware modules introduced in this paper; evaluated only in the paper's own Questa-Sim/Vivado/Innovus flows — no fabricated silicon, no third-party re-implementation, no bitstream release with checksum.
  • RMMEC 4-bit reconfigurable multiplier/exponent-comparator block
    purpose: Enables 6xINT4/FP4, 3xFP8, 1xBF16 per-cycle MAC packing
    Core claimed hardware novelty; supported only by internal synthesis numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bhasha-Rupantarika: Algorithm-Hardware Co-design approach for Multilingual Neural Machine Translation." pith.science (2026). https://pith.science/paper/FKV3GYRM

@misc{pith2026251010676,
  author       = {Pith},
  title        = {Pith review of: Bhasha-Rupantarika: Algorithm-Hardware Co-design approach for Multilingual Neural Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKV3GYRM}},
  note         = {Machine review of arXiv:2510.10676}
}
read the original abstract

This paper introduces Bhasha-Rupantarika, a light and efficient multilingual translation system tailored through algorithm-hardware codesign for resource-limited settings. The method investigates model deployment at sub-octet precision levels (FP8, INT8, INT4, and FP4), with experimental results indicating a 4.1x reduction in model size (FP4) and a 4.2x speedup in inference speed, which correlates with an increased throughput of 66 tokens/s (improvement by 4.8x). This underscores the importance of ultra-low precision quantization for real-time deployment in IoT devices using FPGA accelerators, achieving performance on par with expectations. Our evaluation covers bidirectional translation between Indian and international languages, showcasing its adaptability in low-resource linguistic contexts. The FPGA deployment demonstrated a 1.96x reduction in LUTs and a 1.65x decrease in FFs, resulting in a 2.2x enhancement in throughput compared to OPU and a 4.6x enhancement compared to HPTA. Overall, the evaluation provides a viable solution based on quantisation-aware translation along with hardware efficiency suitable for deployable multilingual AI systems. The entire codes [https://github.com/mukullokhande99/Bhasha-Rupantarika/] and dataset for reproducibility are publicly available, facilitating rapid integration and further development by researchers.

Figures

Figures reproduced from arXiv: 2510.10676 by the authors.

Figure 2
Figure 2. High level algorithm pipeline description, comparison [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. (a) Conventional system-on-chip (SoC) architecture, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Illustration of Transformer encoder used, (a) Dense [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The detailed flow for NLPE, Memory control handles [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: The datapath for SIMD multiply-accumulate unit [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: The detailed Non-linear MIMD Vector Array datapath. [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]
Figure 8
Figure 8. Figure 8: The detailed CORDIC-based FASST (Floating-point [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: True outputs with Bhasha-Rupantarika (INT4) in different Indic and overseas languages, showing translation for the [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 11
Figure 11. Figure 11: ASIC performance metrics, comparison with SoTA [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 12
Figure 12. Figure 12: ASIC performance metrics, comparison with SoTA [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 2 linked inside Pith

  1. [1]

    Efficient Methods for Mapping Neural Machine Translator on FPGAs,

    Q. Li, X. Zhang, J. Xiong, W.-M. Hwu, and D. Chen, “Efficient Methods for Mapping Neural Machine Translator on FPGAs,”IEEE Transactions on Parallel and Distributed Systems, vol. 32, pp. 1866–1877, July 2021

    2021

  2. [2]

    Algorithm-hardware Co-design of Attention Mechanism on FPGA Devices,

    X. Zhang, Y . Wu, P. Zhou, X. Tang, and J. Hu, “Algorithm-hardware Co-design of Attention Mechanism on FPGA Devices,”ACM Trans. Embed. Comput. Syst., vol. 20, Sept. 2021

    2021

  3. [3]

    A Survey of Multilingual Neural Machine Translation,

    R. Dabre, C. Chu, and A. Kunchukuttan, “A Survey of Multilingual Neural Machine Translation,”ACM Comput. Surv., vol. 53, Sept. 2020

    2020

  4. [4]

    Natural Language Processing for Dialects of a Language: A Survey,

    A. Joshi, R. Dabre, D. Kanojia, Z. Li, H. Zhan, G. Haffari, and D. Dippold, “Natural Language Processing for Dialects of a Language: A Survey,”ACM Comput. Surv., vol. 57, Feb. 2025

    2025

  5. [5]

    SelfSeg: A Self-supervised Sub-word Segmentation Method for Neural Machine Translation,

    H. Song, R. Dabre, C. Chu, S. Kurohashi, and E. Sumita, “SelfSeg: A Self-supervised Sub-word Segmentation Method for Neural Machine Translation,”ACM Trans. Asian Low-Resour . Lang. Inf. Process., vol. 22, Aug. 2023

    2023

  6. [6]

    A Survey: Collaborative Hardware and Software Design in the Era of Large Language Models,

    C. Guo, F. Cheng, Z. Du,et al., “A Survey: Collaborative Hardware and Software Design in the Era of Large Language Models,”IEEE Circuits and Systems Magazine, vol. 25, pp. 35–57, Feb. 2025

    2025

  7. [7]

    Indic-ST: A Large-Scale Multilingual Corpus for Low-Resource Speech-to-Text Translation,

    N. Sethiya, S. Nair, P. Walia, and C. Maurya, “Indic-ST: A Large-Scale Multilingual Corpus for Low-Resource Speech-to-Text Translation,” ACM Trans. Asian Low-Resour . Lang. Inf. Process., vol. 24, June 2025

    2025

  8. [8]

    OPU: An FPGA- Based Overlay Processor for Convolutional Neural Networks,

    Y . Yu, C. Wu, T. Zhao, K. Wang, and L. He, “OPU: An FPGA- Based Overlay Processor for Convolutional Neural Networks,”IEEE Transactions on V ery Large Scale Integration (VLSI) Systems, vol. 28, pp. 35–47, Jan. 2020

    2020

Show all 49 references
  1. [9]

    An Energy-Efficient Reconfigurable LSTM Accelerator for Natural Language Processing,

    E. Azari and S. Vrudhula, “An Energy-Efficient Reconfigurable LSTM Accelerator for Natural Language Processing,” in2019 IEEE Interna- tional Conference on Big Data (Big Data), pp. 4450–4459, 2019

    2019

  2. [10]

    Bhasaanuvaad: A speech translation dataset for 13 indian languages,

    S. Jain, A. Sankar, D. Choudhary, D. Suman, N. Narasimhan, M. S. U. R. Khan, A. Kunchukuttan, M. M. Khapra, and R. Dabre, “Bhasaanuvaad: A speech translation dataset for 13 indian languages,”arXiv preprint arXiv:2411.04699, Nov. 2024

    2024 arXiv

  3. [11]

    Towards building text-to-speech systems for the next billion users,

    G. K. Kumar, S. Praveen, P. Kumar, M. M. Khapra, and K. Nandakumar, “Towards building text-to-speech systems for the next billion users,” in Icassp 2023-2023 ieee international conference on acoustics, speech and signal processing (icassp), pp. 1–5, IEEE, May 2023

    2023

  4. [12]

    Scaling neural machine translation to 200 languages,

    “Scaling neural machine translation to 200 languages,”Nature, vol. 630, pp. 841–846, June 2024

    2024

  5. [13]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” inInternational conference on machine learning, pp. 28492– 28518, PMLR, July 2023

    2023

  6. [14]

    ViTCoD: Vision Transformer Acceleration via Dedicated Algo- rithm 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 Algo- rithm and Accelerator Co-Design,” inIEEE International Symposium on High-Performance Computer Architecture (HPCA), pp. 273–286, 2023

    2023

  7. [15]

    ViA: A Novel Vision-Transformer Accelerator Based on FPGA,

    T. Wang, L. Gong, C. Wang, Y . Yang, Y . Gao, X. Zhou, and H. Chen, “ViA: A Novel Vision-Transformer Accelerator Based on FPGA,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, pp. 4088–4099, Nov. 2022

    2022

  8. [16]

    EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference,

    T. Tambe, C. Hooper, L. Pentecost,et al., “EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference,” inMICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture, MICRO ’21, (New York, NY , USA), p. 830–844, Association for Com...

    2021

  9. [17]

    Hardware acceleration of fully quantized bert for efficient natural language processing,

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

    2021

  10. [18]

    AccelTran: A Sparsity-Aware Accelerator for Dy- namic Inference With Transformers,

    S. Tuli and N. K. Jha, “AccelTran: A Sparsity-Aware Accelerator for Dy- namic Inference With Transformers,”IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems, vol. 42, pp. 4038– 4051, Nov. 2023

    2023

  11. [19]

    A Fast and Flexible FPGA-based Accelerator for Natural Language Processing Neural Networks,

    S. Hur, S. Na, D. Kwon, J. Kim, A. Boutros, E. Nurvitadhi, and J. Kim, “A Fast and Flexible FPGA-based Accelerator for Natural Language Processing Neural Networks,”ACM Trans. Archit. Code Optim., vol. 20, Feb. 2023

    2023

  12. [20]

    A 3: Accelerating Attention Mechanisms in Neural Networks with Approximation,

    T. J. Ham, S. J. Jung,et al., “A 3: Accelerating Attention Mechanisms in Neural Networks with Approximation,” inIEEE International Sym- posium on High Performance Computer Architecture (HPCA), pp. 328– 341, 2020

    2020

  13. [21]

    Uni-OPU: An FPGA- Based Uniform Accelerator for Convolutional and Transposed Convo- lutional Networks,

    Y . Yu, T. Zhao, M. Wang, K. Wang, and L. He, “Uni-OPU: An FPGA- Based Uniform Accelerator for Convolutional and Transposed Convo- lutional Networks,”IEEE Transactions on V ery Large Scale Integration (VLSI) Systems, vol. 28, pp. 1545–1556, July 2020

    2020

  14. [22]

    HPTA: A High Performance Transformer Acceler- ator Based on FPGA,

    Y . Han and Q. Liu, “HPTA: A High Performance Transformer Acceler- ator Based on FPGA,” in2023 33rd International Conference on Field- Programmable Logic and Applications (FPL), pp. 27–33, 2023

    2023

  15. [23]

    ELSA: hardware-software co-design for efficient, lightweight self-attention mechanism in neural networks,

    T. J. Ham, Y . Lee,et al., “ELSA: hardware-software co-design for efficient, lightweight self-attention mechanism in neural networks,” in Proceedings of the 48th Annual International Symposium on Computer Architecture, ISCA ’21, p. 692–705, IEEE Press, 2021

    2021

  16. [24]

    EdgeLLM: A Highly Efficient CPU- FPGA Heterogeneous Edge Accelerator for Large Language Models,

    M. Huang, A. Shen, K. Li,et al., “EdgeLLM: A Highly Efficient CPU- FPGA Heterogeneous Edge Accelerator for Large Language Models,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 72, pp. 3352–3365, July 2025

    2025

  17. [25]

    A 16-nm SoC for Noise-Robust Speech and NLP Edge AI Inference With Bayesian Sound Source Separation and Attention-Based DNNs,

    T. Tambe, E.-Y . Yang, G. G. Ko, Y . Chai,et al., “A 16-nm SoC for Noise-Robust Speech and NLP Edge AI Inference With Bayesian Sound Source Separation and Attention-Based DNNs,”IEEE Journal of Solid- State Circuits, vol. 58, pp. 569–581, Feb. 2023

    2023

  18. [26]

    A configurable cloud-scale DNN pro- cessor for real-time AI,

    J. Fowers, K. Ovtcharov,et al., “A configurable cloud-scale DNN pro- cessor for real-time AI,” inProceedings of the 45th Annual International Symposium on Computer Architecture, ISCA ’18, p. 1–14, IEEE Press, 2018

    2018

  19. [27]

    Occamy: A 432-Core Dual-Chiplet Dual- HBM2E 768-DP-GFLOP/s RISC-V System for 8-to-64-bit Dense and Sparse Computing in 12-nm FinFET,

    P. Scheffler, T. Benz,et al., “Occamy: A 432-Core Dual-Chiplet Dual- HBM2E 768-DP-GFLOP/s RISC-V System for 8-to-64-bit Dense and Sparse Computing in 12-nm FinFET,”IEEE Journal of Solid-State Circuits, vol. 60, Apr. 2025

    2025

  20. [28]

    Maestro: A 302 GFLOPS/W and 19.8 GFLOPS RISC-V Vector-Tensor Architecture for Wearable Ultrasound Edge Computing,

    M. Sinigagliaet al., “Maestro: A 302 GFLOPS/W and 19.8 GFLOPS RISC-V Vector-Tensor Architecture for Wearable Ultrasound Edge Computing,”IEEE Trans. on Circuits and Syst.- I, pp. 1–15, 2025

    2025

  21. [29]

    Parallel Accurate Minifloat MACCs for NN Inference on Versal FPGAs,

    H. J. Damsgaard, K. J. HoBfeld, and J. Nurmi, “Parallel Accurate Minifloat MACCs for NN Inference on Versal FPGAs,”IEEE Trans. Comp.-Aided Des. Integ. Cir . Syst., vol. 44, pp. 2181–2194, June 2025

    2025

  22. [30]

    NVIDIA Blackwell Platform: Advancing Generative AI and Accelerated Computing,

    A. Tirumala and R. Wong, “NVIDIA Blackwell Platform: Advancing Generative AI and Accelerated Computing,” inIEEE Hot Chips Sym- posium (HCS), vol. 36, pp. 1–33, 2024

    2024

  23. [31]

    LPRE: Logarithmic Posit-enabled Reconfigurable edge-AI Engine,

    O. Kokane, M. Lokhande, G. Raut, A. Teman, and S. K. Vishvakarma, “LPRE: Logarithmic Posit-enabled Reconfigurable edge-AI Engine,” in 2025 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1–5, May 2025

    2025

  24. [32]

    AMD XDNA NPU in Ryzen AI Processors,

    A. Rico, S. Pareek,et al., “AMD XDNA NPU in Ryzen AI Processors,” IEEE Micro, vol. 44, pp. 73–82, Nov. 2024

    2024

  25. [33]

    A Multi-Mode 8k-MAC HW-Utilization- Aware Neural Processing Unit With a Unified Multi-Precision Datapath in 4-nm Flagship Mobile SoC,

    J.-S. Park, C. Park,et al., “A Multi-Mode 8k-MAC HW-Utilization- Aware Neural Processing Unit With a Unified Multi-Precision Datapath in 4-nm Flagship Mobile SoC,”IEEE Journal of Solid-State Circuits, vol. 58, pp. 189–202, Jan. 2023

    2023

  26. [34]

    A Low-Cost Floating-Point FMA Unit Supporting Package Operations for HPC-AI Applications,

    H. Tan, J. Zhang, X. He, L. Huang, Y . Wang, and L. Xiao, “A Low-Cost Floating-Point FMA Unit Supporting Package Operations for HPC-AI Applications,”IEEE Trans. on Circuits and Systems II: Express Briefs, vol. 71, pp. 3488–3492, July 2024

    2024

  27. [35]

    RaPiD: AI Accelerator for Ultra-low Precision Training and Inference,

    S. Venkataramani, V . Srinivasan,et al., “RaPiD: AI Accelerator for Ultra-low Precision Training and Inference,”ACM/IEEE 48th Annual International Symposium on Computer Architecture, pp. 153–166, 2021

    2021

  28. [36]

    XR-NPE: High-Throughput Mixed-precision SIMD Neural Processing Engine for Extended Reality Perception Workloads,

    T. Chaudhari, T. Dewangan, M. Lokhande, S. K. Vishvakarma,et al., “XR-NPE: High-Throughput Mixed-precision SIMD Neural Processing Engine for Extended Reality Perception Workloads,”arXiv preprint arXiv:2508.13049, 2025

    2025 arXiv

  29. [37]

    Exploring Hardware Ac- tivation Function Design: CORDIC Architecture in Diverse Floating Formats,

    M. Basavaraju, V . Rayapati, and M. Rao, “Exploring Hardware Ac- tivation Function Design: CORDIC Architecture in Diverse Floating Formats,” in2024 25th International Symposium on Quality Electronic Design (ISQED), pp. 1–8, 2024

    2024

  30. [38]

    Retrospective: A CORDIC Based Configurable Activation Function for NN Applications,

    O. Kokane, G. Raut, S. Ullah, M. Lokhande, A. Teman, A. Kumar, and S. K. Vishvakarma, “Retrospective: A CORDIC Based Configurable Activation Function for NN Applications,” in2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), vol. 1, pp. 1–6, 2025

    2025

  31. [39]

    ASTRA: Reconfigurable Training Architecture Design for Nonlinear Softmax and Activation Functions in Transform- ers,

    H. Shao and Z. Wang, “ASTRA: Reconfigurable Training Architecture Design for Nonlinear Softmax and Activation Functions in Transform- ers,”IEEE Transactions on V ery Large Scale Integration (VLSI) Systems, vol. 33, pp. 2054–2058, July 2025

  32. [40]

    A Flexible Template for Edge Generative AI With High-Accuracy Accelerated Softmax and GELU,

    A. Belano, Y . Tortorella, A. Garofalo,et al., “A Flexible Template for Edge Generative AI With High-Accuracy Accelerated Softmax and GELU,”IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 15, pp. 200–216, June 2025

    2025

  33. [41]

    Flex-PE: Flexible and SIMD Multiprecision Processing Element for AI Workloads,

    M. Lokhande, G. Raut, and S. K. Vishvakarma, “Flex-PE: Flexible and SIMD Multiprecision Processing Element for AI Workloads,”IEEE Trans. VLSI Syst., vol. 33, pp. 1610–1623, June 2025

    2025

  34. [42]

    A Reconfigurable Processing Element for Multiple- Precision Floating/Fixed-Point HPC,

    B. Li, K. Li,et al., “A Reconfigurable Processing Element for Multiple- Precision Floating/Fixed-Point HPC,”IEEE Trans. Circuits Syst. II, vol. 71, pp. 1401–1405, Mar. 2024

    2024

  35. [43]

    QuantMAC: Enhancing Hardware Performance in DNNs With Quantize Enabled Multiply-Accumulate Unit,

    N. Ashar, G. Raut, V . Trivedi, S. K. Vishvakarma, and A. Ku- mar, “QuantMAC: Enhancing Hardware Performance in DNNs With Quantize Enabled Multiply-Accumulate Unit,”IEEE Access, vol. 12, pp. 43600–43614, 2024

    2024

  36. [44]

    ReNPU: A Resource-efficient Multi-Mode Neural Processing Unit with Unified Multi-Precision Datapath for Mobile AI Workloads,

    M. Lokhande, S. J. Chand, A. Jain, S. Kumar, and S. K. Vishvakarma, “ReNPU: A Resource-efficient Multi-Mode Neural Processing Unit with Unified Multi-Precision Datapath for Mobile AI Workloads,”Authorea Preprints, 2025

    2025

  37. [45]

    ReAFM: A Reconfigurable Nonlinear Activation Function Module for Neural Networks,

    X. Wu, S. Liang, M. Wang, and Z. Wang, “ReAFM: A Reconfigurable Nonlinear Activation Function Module for Neural Networks,”IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 70, pp. 2660–2664, July 2023

    2023

  38. [46]

    QForce- RL: Quantized FPGA-Optimized Reinforcement Learning Compute En- gine,

    A. Jha, T. Dewangan, M. Lokhande, and S. K. Vishvakarma, “QForce- RL: Quantized FPGA-Optimized Reinforcement Learning Compute En- gine,”29th International Symposium on VLSI Design and Test, July 2025

    2025

  39. [47]

    Flex-SFU: Activation Func- tion Acceleration with Non-Uniform Piecewise Approximation,

    R. Andri, E. Reggiani, and L. Cavigelli, “Flex-SFU: Activation Func- tion Acceleration with Non-Uniform Piecewise Approximation,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, pp. 1–1, 2025

    2025

  40. [48]

    TEA-S: A Tiny and Efficient Architecture for PLAC-Based Softmax in Transformers,

    Z. Mei, H. Dong, Y . Wang, and H. Pan, “TEA-S: A Tiny and Efficient Architecture for PLAC-Based Softmax in Transformers,”IEEE Transac- tions on Circuits and Systems II: Express Briefs, vol. 70, pp. 3594–3598, Sept. 2023

    2023

  41. [49]

    ShortcutFusion: From Tensorflow to FPGA-Based Accelerator With a Reuse-Aware Memory Allocation for Shortcut Data,

    D. T. Nguyen, H. Je, T. N. Nguyen,et al., “ShortcutFusion: From Tensorflow to FPGA-Based Accelerator With a Reuse-Aware Memory Allocation for Shortcut Data,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 69, pp. 2477–2489, June 2022

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.