pith. sign in

arxiv: 2606.19913 · v1 · pith:ENDOT2YLnew · submitted 2026-06-18 · 💻 cs.AR

Design and Evaluation of Energy-Efficient Whisper Dot-Product Kernel Offloading on a CGLA Architecture

Takuto Ando , Yu Eto , Ayumu Takeuchi , Yasuhiko Nakashima This is my paper

Pith reviewed 2026-06-26 15:27 UTC · model grok-4.3

classification 💻 cs.AR
keywords Whisperdot-product kernelCGLAenergy efficiencykernel offloadingASRFPGA prototypeASIC projection
0
0 comments X

The pith

Offloading Whisper dot-product kernels to IMAX CGLA cuts PDP to 11.58J for the tiny model, 2.35 times below Jetson AGX Orin.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that dot-product operations dominate Whisper execution time, reaching 90.6 percent of FP16 time and 87.1 percent of Q8_0 time on an ARM Cortex-A72. By combining kernel mapping, 32KB local memory, burst scheduling, inline FP16-to-FP32 conversion, 2-way SIMD FMA, column-wise multithreading, and mixed host-accelerator execution, the design targets this bottleneck on the IMAX programmable CGLA architecture. An FPGA prototype and 28nm ASIC projection at 840MHz are used to evaluate the approach. The resulting PDP of 11.58J for Whisper-tiny.en Q8_0 is shown to be lower than that of Jetson AGX Orin and RTX 4090 under TDP-based comparison, with the advantage decreasing for larger models.

Core claim

The authors implement dot-product kernel offloading on IMAX using kernel mapping, local-memory sizing to 32KB, burst scheduling of length 16, inline FP16-to-FP32 conversion, 2-way SIMD FMA on a 64-bit datapath, column-wise multithreading, and mixed execution in which aligned vector segments run on IMAX while residual segments run on the host CPU. For Whisper-tiny.en, 32KB local memory and burst length 16 jointly minimize PDP and EDP. Under a TDP-based cross-platform comparison, the projected IMAX records a PDP of 11.58J for Whisper-tiny.en Q8_0, 2.35 times lower than Jetson AGX Orin at 27.16J and 10.48 times lower than RTX 4090 at 121.38J. The same design extends to Whisper-base.en and Whisp

What carries the argument

IMAX, a programmable Coarse-Grained Linear Arrays (CGLA) architecture, with burst scheduling, SIMD FMA, and mixed execution for dot-product kernel offloading.

If this is right

  • 32KB local memory and burst length 16 jointly minimize PDP and EDP for Whisper-tiny.en.
  • The PDP advantage decreases for Whisper-base.en and Whisper-small.en as local-memory coverage falls to about 66.5 percent.
  • The design supports both FP16 and Q8_0 quantization through the same offloading strategy with inline conversion.
  • Mixed execution allows residual vector segments to run concurrently on the host CPU.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The kernel mapping and scheduling techniques could be adapted to other transformer-based models where matrix multiplications dominate compute time.
  • Programmable CGLA designs may provide a flexible alternative to fixed-function accelerators when AI models continue to evolve.
  • Edge deployment would need to verify whether memory bandwidth and software overheads remain within the bounds assumed in the projection.

Load-bearing premise

The 28nm ASIC projection at 840MHz combined with the TDP-based cross-platform comparison accurately captures real performance without unaccounted overheads from memory bandwidth, software stack, or thermal limits.

What would settle it

Fabrication and measurement of the 28nm IMAX ASIC running the Whisper-tiny.en Q8_0 dot-product kernel to determine whether actual PDP reaches the projected 11.58J value under equivalent conditions.

read the original abstract

In this paper, we implement and evaluate Whisper dot-product kernel offloading on IMAX, a programmable Coarse-Grained Linear Arrays (CGLAs) architecture. Whisper-tiny.en profiling on an ARM Cortex-A72 shows that dot-product operations account for 90.6% of FP16 execution time and 87.1% of Q8_0 execution time. To address this kernel bottleneck, we combine kernel mapping, local-memory sizing, and burst scheduling. The implementation uses inline FP16-to-FP32 conversion, 2-way SIMD FMA on a 64-bit datapath, column-wise multithreading, and mixed execution in which aligned vector segments run on IMAX and residual segments run concurrently on the host CPU. We evaluate the design with an FPGA prototype and a 28nm ASIC projection at 840MHz. For Whisper-tiny.en, 32KB local memory and burst length 16 jointly minimize PDP and EDP. Under a TDP-based cross-platform comparison, the projected IMAX records a PDP of 11.58J for Whisper-tiny.en Q8_0, 2.35x lower than Jetson AGX Orin (27.16J) and 10.48x lower than RTX 4090 (121.38J). The same design extends to Whisper-base.en and Whisper-small.en, where the PDP gap narrows as 32KB local-memory coverage drops from 93.8% for tiny to about 66.5% for base and small. These results position IMAX as a programmable architecture for lower-PDP local ASR in the tiny-model regime.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper implements and evaluates offloading of the dot-product kernel from Whisper ASR models (tiny.en, base.en, small.en) to the IMAX programmable CGLA architecture. It profiles the kernel on ARM Cortex-A72 (90.6% FP16 and 87.1% Q8_0 time), applies kernel mapping with inline FP16-to-FP32 conversion, 2-way SIMD FMA, column-wise multithreading, burst scheduling, and mixed host/IMAX execution. Evaluation uses an FPGA prototype plus a 28 nm ASIC projection at 840 MHz; with 32 KB local memory and burst length 16, the design reports a TDP-based PDP of 11.58 J for Whisper-tiny.en Q8_0 (2.35× lower than Jetson AGX Orin at 27.16 J and 10.48× lower than RTX 4090 at 121.38 J), with the advantage narrowing for larger models as local-memory coverage drops.

Significance. If the ASIC projection and TDP comparison are shown to be accurate, the work provides concrete evidence that a programmable CGLA can deliver lower PDP than contemporary GPUs and SoCs for the dominant kernel in tiny Whisper models, supporting the case for specialized but programmable accelerators in on-device ASR.

major comments (2)
  1. [Evaluation / ASIC projection] Evaluation / ASIC projection paragraph: the 840 MHz frequency, power, and effective throughput values used for the 28 nm projection are stated without derivation steps, synthesis reports, or explicit validation against the measured FPGA data; because the headline PDP numbers (11.58 J, 2.35×/10.48×) rest directly on this projection, the absence of these details prevents verification that memory-bandwidth, burst-scheduling, and mixed-execution overheads have been folded in.
  2. [Evaluation / cross-platform comparison] Evaluation / cross-platform comparison paragraph: the TDP-based method is used to claim superiority over Jetson AGX Orin and RTX 4090, yet no quantitative accounting is given for unmodeled factors (host/IMAX data movement, software-stack costs, thermal throttling, or effective memory bandwidth) that the skeptic note flags as potentially inflating the reported advantage.
minor comments (2)
  1. [Abstract / §3] Abstract and §3: the exact definition of “PDP” (energy per inference or per token) and the precise Whisper input lengths used for the reported figures should be stated explicitly to allow direct reproduction.
  2. [Kernel mapping section] The paper mentions “inline FP16-to-FP32 conversion” and “2-way SIMD FMA on a 64-bit datapath” but does not provide the corresponding instruction or datapath diagram; a small figure would clarify the mapping.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and will revise the manuscript to enhance the transparency of the ASIC projection and cross-platform evaluation.

read point-by-point responses

  1. Referee: [Evaluation / ASIC projection] Evaluation / ASIC projection paragraph: the 840 MHz frequency, power, and effective throughput values used for the 28 nm projection are stated without derivation steps, synthesis reports, or explicit validation against the measured FPGA data; because the headline PDP numbers (11.58 J, 2.35×/10.48×) rest directly on this projection, the absence of these details prevents verification that memory-bandwidth, burst-scheduling, and mixed-execution overheads have been folded in.

    Authors: We agree that the derivation of the 840 MHz frequency, power estimates, and throughput for the 28 nm ASIC projection requires additional detail to enable verification. These values are obtained by scaling post-synthesis timing and power reports from the FPGA prototype using 28 nm standard-cell library characteristics, with throughput adjusted for the observed burst length and mixed-execution behavior measured on the prototype. In the revised manuscript we will insert a dedicated subsection that presents the scaling methodology, key synthesis metrics, and explicit confirmation that memory-bandwidth, burst-scheduling, and mixed-execution overheads measured on FPGA are propagated into the projection. revision: yes

  2. Referee: [Evaluation / cross-platform comparison] Evaluation / cross-platform comparison paragraph: the TDP-based method is used to claim superiority over Jetson AGX Orin and RTX 4090, yet no quantitative accounting is given for unmodeled factors (host/IMAX data movement, software-stack costs, thermal throttling, or effective memory bandwidth) that the skeptic note flags as potentially inflating the reported advantage.

    Authors: The TDP-based comparison follows the common practice in accelerator papers for first-order energy-efficiency ranking. We nevertheless accept that a fuller discussion of unmodeled factors is needed. The revised evaluation section will add a paragraph that (a) quantifies the data-movement cost under the mixed-execution scheme already implemented, (b) notes the absence of full software-stack and thermal-throttling measurements on the prototype, and (c) provides conservative bounds on how these factors could affect the reported PDP advantage. Because complete system-level quantification would require additional integration work beyond the current kernel-level prototype, the revision will be partial. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical hardware implementation and measurement study

full rationale

The paper reports an FPGA prototype implementation plus standard 28nm ASIC synthesis projection for a CGLA offload design. All reported PDP/EDP numbers and cross-platform comparisons derive from measured prototype behavior, synthesis reports, and TDP-based scaling; no equations, fitted parameters, or self-citations are used to derive the central claims. No self-definitional, fitted-input-called-prediction, or load-bearing self-citation patterns appear. The work is self-contained against external benchmarks (FPGA measurements, synthesis tools, published TDP figures for Jetson/RTX).

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The paper relies on two optimization parameters chosen to minimize PDP/EDP and one domain assumption about the accuracy of the ASIC projection; no new entities are postulated.

free parameters (2)
  • local memory size = 32KB
    Selected jointly with burst length to minimize PDP and EDP for Whisper-tiny.en
  • burst length = 16
    Selected jointly with local memory size to minimize PDP and EDP for Whisper-tiny.en
axioms (1)
  • domain assumption FPGA prototype and 28nm ASIC projection at 840MHz provide accurate performance estimates
    Invoked for all reported PDP and comparison numbers in the abstract

pith-pipeline@v0.9.1-grok · 5844 in / 1411 out tokens · 35429 ms · 2026-06-26T15:27:19.696371+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    ExpeL: LLM agents are experiential learners

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong- Jin Liu„ and Gao Huang. 2024. “ExpeL: LLM agents are experiential learners.” InProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference onInnovativeApplicationsofArtificialIntelligenceandFourteenth Symposium on Educational Advances in Artificial Intel...

    2024

  2. [2]

    HealAI:AHealthcareLLMforEffectiveMedicalDocumentation

    Sagar Goyal, Eti Rastogi, Sree Prasanna Rajagopal, Dong Yuan, Fen Zhao, Jai Chintagunta, Gautam Naik„ and Jeff Ward. 2024. “HealAI:AHealthcareLLMforEffectiveMedicalDocumentation.” InProceedings of the 17th ACM International Conference on Web Search and Data Mining, WSDM ’24, 1167–1168. New York, NY, USA: Association for Computing Machinery

    2024

  3. [3]

    A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu„ and Yi Yang. “A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges.”Vicinagearth1 (2024): 9

    2024

  4. [4]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave„ and Guillaume Lample. 2023. “LLaMA: Open and Efficient Foundation Language Models.” , arXiv:2302.13971

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    Long Short- Term Memory Recurrent Neural Network for Automatic Speech Recognition

    Jane Oruh, Serestina Viriri„ and Adekanmi Adegun. “Long Short- Term Memory Recurrent Neural Network for Automatic Speech Recognition.”IEEE Access10 (2022): 30069–30079

    2022

  6. [6]

    Robust Speech Recognition via Large-Scale Weak Supervision

    AlecRadford,JongWookKim,TaoXu,GregBrockman,Christine McLeavey„ and Ilya Sutskever. 2022. “Robust Speech Recognition via Large-Scale Weak Supervision.” , arXiv:2212.04356

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  7. [7]

    Design and implementation of smart voice assistant and recognizing academic words

    Ahmed Abougarair. “Design and implementation of smart voice assistant and recognizing academic words.”International Journal of Robotics and Automation8 (2022): 27–32

    2022

  8. [8]

    SLM:Bridge the Thin Gap Between Speech and Text Foundation Models

    Mingqiu Wang, Wei Han, Izhak Shafran, Zelin Wu, Chung-Cheng Chiu, Yuan Cao, Nanxin Chen, Yu Zhang, Hagen Soltau, Paul K. Rubenstein, Lukas Zilka, Dian Yu, Golan Pundak, Nikhil Sid- dhartha,JohanSchalkwyk„andYonghuiWu.2023. “SLM:Bridge the Thin Gap Between Speech and Text Foundation Models.” In2023 IEEE Automatic Speech Recognition and Understanding Worksh...

    2023

  9. [9]

    Assess- ingtheEffectivenessofAutomaticSpeechRecognitionTechnology in Emergency Medicine Settings: a Comparative Study of Four AI- Powered Engines

    XiaoLuo,LeZhou,KathleenAdelgais„andZhanZhang. “Assess- ingtheEffectivenessofAutomaticSpeechRecognitionTechnology in Emergency Medicine Settings: a Comparative Study of Four AI- Powered Engines.”Journal of Healthcare Informatics Research9, no. 3 (2025): 494–512

    2025

  10. [10]

    Energy and AI

    International Energy Agency 2025. “Energy and AI.” https: //www.iea.org/reports/energy-and-ai, Licence: CC BY 4.0, https:// creativecommons.org/licenses/by/4.0/, Accessed on June 14, 2025

    2025

  11. [11]

    2024 United States Data Center Energy Usage Report

    ArmanShehabi,SarahSmith,AlanHubbard,AdamNewkirk,Ning Lei, Md Abu Bakar Siddik, et al. 2024. “2024 United States Data Center Energy Usage Report.” Technical Report LBNL-2001637, Lawrence Berkeley National Laboratory. Accessed on May 30, 2025

    2024

  12. [12]

    IMAX: A Power-Efficient Multilevel Pipelined CGLA and Appli- cations

    Tomoya Akabe, Vu Trung Duong LE„ and Yasuhiko Nakashima. “IMAX: A Power-Efficient Multilevel Pipelined CGLA and Appli- cations.”IEEE Access13 (2025): 31899–31911

    2025

  13. [13]

    TinyMachineLearning (TinyML): Research trends and future application opportunities

    HuiHan,SilvanaTrimi„andSangM.Lee. “TinyMachineLearning (TinyML): Research trends and future application opportunities.” Array29 (2026): 100674

    2026

  14. [14]

    A Low-Latency Streaming On-Device Au- tomatic Speech Recognition System Using a CNN Acoustic Model on FPGA and a Language Model on Smartphone

    Jaehyun Park, Hyeonkyu Noh, Hyunjoon Nam, Won-Cheol Lee„ and Hong-June Park. “A Low-Latency Streaming On-Device Au- tomatic Speech Recognition System Using a CNN Acoustic Model on FPGA and a Language Model on Smartphone.”Electronics11, no. 12

  15. [15]

    Hardware Accelerator for Transformer basedEnd-to-EndAutomaticSpeechRecognitionSystem

    Shaarada D. Yamini, Ganesh S. Mirishkar, Anil Kumar Vuppala„ and Suresh Purini. 2023. “Hardware Accelerator for Transformer basedEnd-to-EndAutomaticSpeechRecognitionSystem.”In2023 IEEEInternationalParallelandDistributedProcessingSymposium Workshops (IPDPSW), 93–100

    2023

  16. [16]

    A Spiking LSTM Accelerator for Automatic Speech Recognition Application Based on FPGA

    Tingting Yin, Feihong Dong, Chao Chen, Chenghao Ouyang, Zheng Wang„ and Yongkui Yang. “A Spiking LSTM Accelerator for Automatic Speech Recognition Application Based on FPGA.” Electronics13, no. 5

  17. [17]

    Design and Implementation of Intelligent Speech Recog- nition System Based on FPGA

    HuaixiangHu,JiatongLi,ChunchunWu,XueyangLi„andYuping Chen. “Design and Implementation of Intelligent Speech Recog- nition System Based on FPGA.”Journal of Physics: Conference Series2171, no. 1 (2022): 012010

    2022

  18. [18]

    FPGA based Adaptive Hardware Ac- celeration for Multiple Deep Learning Tasks

    YufanLu,XiaojunZhai,SangeetSaha,ShoaibEhsan„andKlausD. McDonald-Maier. 2021. “FPGA based Adaptive Hardware Ac- celeration for Multiple Deep Learning Tasks.” In2021 IEEE 14th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC), 204–209

    2021

  19. [19]

    Optimizing FPGA-based Accelerator De- sign for Deep Convolutional Neural Networks

    Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao„ and Jason Cong. 2015. “Optimizing FPGA-based Accelerator De- sign for Deep Convolutional Neural Networks.” InProceedings of the 2015 ACM/SIGDA International Symposium on Field- Programmable Gate Arrays, FPGA ’15, 161–170. New York, NY, USA: Association for Computing Machinery

    2015

  20. [20]

    Bingbing Li, Santosh Pandey, Haowen Fang, Yanjun Lyu, Ji Li, JieyangChen,MimiXie,LipengWan,HangLiu„andCaiwenDing

  21. [21]

    FTRANS: Energy-Efficient Acceleration of Transformers using FPGA

    “FTRANS: Energy-Efficient Acceleration of Transformers using FPGA.” , arXiv:2007.08563

    work page arXiv 2007

  22. [22]

    Hardware Accelerator for Multi-Head Atten- tion and Position-Wise Feed-Forward in the Transformer

    Siyuan Lu, Meiqi Wang, Shuang Liang, Jun Lin„ and Zhongfeng Wang. 2020. “Hardware Accelerator for Multi-Head Atten- tion and Position-Wise Feed-Forward in the Transformer.” , arXiv:2009.08605

    work page arXiv 2020

  23. [23]

    EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi- TaskNLPInference

    Thierry Tambe, Coleman Hooper, Lillian Pentecost, Tianyu Jia, En-YuYang,MarcoDonato,VictorSanh,PaulWhatmough,Alexan- derM.Rush,DavidBrooks„andGu-YeonWei.2021. “EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi- TaskNLPInference.”InMICRO-54:54thAnnualIEEE/ACMInter- national Symposium on Microarchitecture, MICRO ’21, 830–844. New York, N...

    2021

  24. [24]

    Understand- ing the Potential of FPGA-based Spatial Acceleration for Large Language Model Inference

    HongzhengChen,JiahaoZhang,YixiaoDu,ShaojieXiang,Zichao Yue,NiansongZhang,YaohuiCai„andZhiruZhang. “Understand- ing the Potential of FPGA-based Spatial Acceleration for Large Language Model Inference.”ACM Trans. Reconfigurable Technol. Syst.18, no. 1

  25. [25]

    FlightLLM:EfficientLargeLan- guageModelInferencewithaCompleteMappingFlowonFPGAs

    Shulin Zeng, Jun Liu, Guohao Dai, Xinhao Yang, Tianyu Fu, HongyiWang,WenhengMa,HanboSun,ShiyaoLi,ZixiaoHuang, Yadong Dai, Jintao Li, Zehao Wang, Ruoyu Zhang, Kairui Wen, XuefeiNing„andYuWang.2024.“FlightLLM:EfficientLargeLan- guageModelInferencewithaCompleteMappingFlowonFPGAs.” InProceedings of the 2024 ACM/SIGDA International Symposium on Field Programma...

    2024

  26. [26]

    Towards Energy-Efficient Llama2 Architecture on Embedded FPGAs

    Han Xu, Xingyuan Wang„ and Shihao Ji. 2024. “Towards Energy-Efficient Llama2 Architecture on Embedded FPGAs.” In Proceedings of the 33rd ACM International Conference on Infor- mation and Knowledge Management, CIKM ’24, 5570–5571. New York, NY, USA: Association for Computing Machinery

    2024

  27. [27]

    LlamaF: An Efficient Llama2 Architecture Accelerator on Embedded FPGAs

    Han Xu, Yutong Li„ and Shihao Ji. 2024. “LlamaF: An Efficient Llama2 Architecture Accelerator on Embedded FPGAs.” In2024 IEEE 10th World Forum on Internet of Things (WF-IoT), 1–7

    2024

  28. [28]

    Nvidia Jetson AGX Orin series

    Leela S Karumbunathan 2022. “Nvidia Jetson AGX Orin series.” , A Giant Leap Forward for Robotics and Edge AI Applications. Technical Brief

    2022

  29. [29]

    ASurveyon Coarse-GrainedReconfigurableArchitecturesFromaPerformance Perspective

    ArturPodobas,KentaroSano„andSatoshiMatsuoka. “ASurveyon Coarse-GrainedReconfigurableArchitecturesFromaPerformance Perspective.”IEEE Access8 (2020): 146719–146743

    2020

  30. [30]

    Ultra-Elastic CGRAs for Irregular Loop Specialization

    Christopher Torng, Peitian Pan, Yanghui Ou, Cheng Tan„ and Christopher Batten. 2021. “Ultra-Elastic CGRAs for Irregular Loop Specialization.” In2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 412–425

    2021

  31. [31]

    Specializing CGRAs for Light- Weight Convolutional Neural Networks

    Jungi Lee, and Jongeun Lee. “Specializing CGRAs for Light- Weight Convolutional Neural Networks.”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems41, no. 10 (2022): 3387–3399

    2022

  32. [32]

    HiMap: Fast and Scalable High-Quality Map- ping on CGRA via Hierarchical Abstraction

    DhananjayaWijerathne,ZhaoyingLi,AnujPathania,TulikaMitra„ and Lothar Thiele. “HiMap: Fast and Scalable High-Quality Map- ping on CGRA via Hierarchical Abstraction.”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems41, no. 10 (2022): 3290–3303

    2022

  33. [33]

    Implementation and Performance Analysis of LLaMA on a CGLA

    Yu Eto, and Yasuhiko Nakashima. 2026. “Implementation and Performance Analysis of LLaMA on a CGLA.” InProceedings of the Fifth International Conference on Intelligent Systems and 13 of 14 Networks, 516–525. Singapore: Springer Nature Singapore

    2026

  34. [34]

    Energy Consumption Optimization of Multi- Dimensional U-Nets on CGLA

    Duong Thi Sang, Ren Imamura, Tomoya Akabe„ and Yasuhiko Nakashima. “Energy Consumption Optimization of Multi- Dimensional U-Nets on CGLA.”IEEE Access13 (2025): 29476– 29492

    2025

  35. [35]

    A CGRA-Based Approach for Accelerating Convolutional Neural Networks

    Masakazu Tanomoto, Shinya Takamaeda-Yamazaki, Jun Yao„ and Yasuhiko Nakashima. 2015. “A CGRA-Based Approach for Accelerating Convolutional Neural Networks.” In2015 IEEE 9th International Symposium on Embedded Multicore/Many-core Systems-on-Chip, 73–80

    2015

  36. [36]

    Energy-Efficient 3D Convolution Using Interposed Memory Accelerator eXtension 2 for Medical Image Processing

    Ren Imamura, Zhu Guangxian, Sang Duong Thi, Hoai Luan Pham, RenyuanZhang„andYasuhikoNakashima.2024.“Energy-Efficient 3D Convolution Using Interposed Memory Accelerator eXtension 2 for Medical Image Processing.” InProceedings of 2023 Inter- national Conference on Medical Imaging and Computer-Aided Diagnosis (MICAD 2023), edited by Ruidan Su, Yu-Dong Zhang,...

    2024

  37. [37]

    Implementation andEvaluationofLLMonaCGLA

    Hitoaki Uetani, and Yasuhiko Nakashima. 2024. “ Implementation andEvaluationofLLMonaCGLA.”In2024TwelfthInternational Symposium on Computing and Networking (CANDAR), 252–258. IEEE Computer Society

    2024

  38. [38]

    Optimizing Matrix- VectorOperationsWithCGLAforHigh-PerformanceApproximate k-NN Search

    Dohyun Kim, and Yasuhiko Nakashima. “Optimizing Matrix- VectorOperationsWithCGLAforHigh-PerformanceApproximate k-NN Search.”IEEE Access13 (2025): 111087–111097

    2025

  39. [39]

    Universal 32/64-bit CGRA for Lightweight Cryptography in Securing IoT Data Transmis- sion

    SangDuongThi,HoaiLuanPham,VuTrungDuongLe,ThiDiem Tran, Ren Imamura, Quoc Duy Nam Nguyen, Thi Hong Tran„ and Yasuhiko Nakashima. 2023. “Universal 32/64-bit CGRA for Lightweight Cryptography in Securing IoT Data Transmis- sion.” In2023 IEEE 16th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC), 419–425

    2023

  40. [40]

    whisper.cpp: Port of OpenAI’s Whis- per model in C/C++

    Georgi Gerganov 2023. “whisper.cpp: Port of OpenAI’s Whis- per model in C/C++.” https://github.com/ggerganov/whisper.cpp, Accessed: 2025-07-24

    2023

  41. [41]

    TimDettmers,MikeLewis,YounesBelkada„andLukeZettlemoyer

  42. [42]

    LLM.int8():8-bitMatrixMultiplicationforTransformersat Scale

    “LLM.int8():8-bitMatrixMultiplicationforTransformersat Scale.” ,

  43. [43]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler„ and Dan Alistarh

  44. [44]

    OPTQ: Accurate Quantization for Generative Pre-trained Transformers

    “OPTQ: Accurate Quantization for Generative Pre-trained Transformers.”InTheEleventhInternationalConferenceonLearn- ing Representations,

  45. [45]

    VersalPowerDemo

    XilinxInc.2025. “VersalPowerDemo.”https://github.com/Xilinx/ pm_demo/tree/master?tab=readme-ov-file, Accessed on May 30, 2025

    2025

  46. [46]

    NVIDIAADAGPUARCHITECTURE: Designed to deliver outstanding gaming and creating, professional graphics, AI, and compute performance

    NvidiaCorporation2023.“NVIDIAADAGPUARCHITECTURE: Designed to deliver outstanding gaming and creating, professional graphics, AI, and compute performance.” , Whitepaper

  47. [47]

    LibriSpeech: An ASR corpus based on public domain audio books

    Vassil Panayotov, Guoguo Chen, Daniel Povey„ and Sanjeev Khu- danpur. 2015. “LibriSpeech: An ASR corpus based on public domain audio books.” In2015 IEEE International Conference on Acoustics, Speech and Signal Processing, 5206–5210

    2015

  48. [48]

    ARM Cortex-A72 Ar- chitecture Deep Dive

    Mike Humrick 2016, January). “ARM Cortex-A72 Ar- chitecture Deep Dive.” https://www.tomshardware.com/reviews/ arm-cortex-a72-architecture,4424.html, Accessed on May 30, 2025

    2016

  49. [49]

    DC Ultra: Concurrent Timing, Area, Power and Test Optimization

    Inc. Synopsys 2024. “DC Ultra: Concurrent Timing, Area, Power and Test Optimization.” https://www.synopsys.com/ implementation-and-signoff/rtl-synthesis-test/dc-ultra.html, Accessed on May 25, 2025

    2024

  50. [50]

    Intel®Xeon®w5-2455X Processor (30M Cache, 3.20 GHz) Specifications

    Intel Corporation 2025. “Intel®Xeon®w5-2455X Processor (30M Cache, 3.20 GHz) Specifications.” https://www.intel.co.jp/content/www/jp/ja/products/sku/ 233420/intel-xeon-w52455x-processor-30m-cache-3-20-ghz/ specifications.html, Accessed on May 30, 2025. 14 of 14 Concurrency and Computation: Practice and Experience, 2026

    2025