pith. sign in

arxiv: 2607.01607 · v1 · pith:LE42G73Anew · submitted 2026-07-02 · 💻 cs.AR

MxGLUT: A Reconfigurable LUT-Centric Broadcast Dataflow Accelerator for Mixed-Precision GEMM

Weiyu Zhou , Chen Ding , Mingyuan Liu , Liangyu Gan , Yukun Feng , Hao Jia , Haoming Chu , Lirong Zheng
show 2 more authors
Ning Ma Yuxiang Huan
This is my paper

Pith reviewed 2026-07-03 04:30 UTC · model grok-4.3

classification 💻 cs.AR
keywords LLM inferencemixed-precision GEMMLUT-based acceleratorreconfigurable dataflowFP8 quantizationweight-only quantizationprefill and decodearea and energy efficiency
0
0 comments X

The pith

MxGLUT unifies FP8-INT4 and FP8-FP8 GEMM computation inside one LUT-based accelerator for LLM inference.

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

The paper sets out to demonstrate that a reconfigurable LUT-centric broadcast dataflow can execute both integer and floating-point mixed-precision matrix multiplications under a single mechanism. This removes the need for separate floating-point datapaths while adapting the data movement pattern to the distinct demands of prefill and decode stages. A sympathetic reader would care because weight-only quantization leaves activations in FP8, yet existing accelerators either duplicate hardware or use static schedules that waste resources across phases. If the design works as described, it would cut multiplier area and power while preserving model accuracy on families such as Llama.

Core claim

MxGLUT organizes FP8-INT4 and FP8-FP8 GEMM under one LUT-based compute path by means of mixed-precision LUT-based processing elements and a reconfigurable broadcast dataflow that localizes partial-sum accumulation in prefill and reuses weights in decode; synthesis in 28 nm CMOS at 200 MHz shows multiplier area reduced by up to 56.92 percent and power by up to 78.35 percent, with accelerator-level area efficiency of 0.492 TFLOPS/mm² and energy efficiency of 11.58 TFLOPS/W, plus up to 2.16× prefill and 1.49× decode latency gains at no more than 1.70 percent perplexity increase.

What carries the argument

The mixed-precision LUT-based processing element (MxLPE) together with the reconfigurable LUT-centric broadcast (RLB) dataflow, which executes both precision modes without dedicated FP multipliers or extra datapaths.

If this is right

  • Multiplier area drops by up to 56.92 percent and power by up to 78.35 percent when FP8-FP8 support is added.
  • Adding native FP8-FP8 mode reduces accelerator area efficiency by only 2.57 percent and energy efficiency by only 3.34 percent relative to an FP8-INT4-only baseline.
  • Prefill latency improves up to 2.16× and decode latency up to 1.49× across the Llama family while normalized energy falls to 0.44× and 0.71× respectively.
  • Perplexity degradation stays within 1.70 percent when the accelerator runs the full inference workload.

Where Pith is reading between the lines

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

  • The same RLB schedule could be retargeted to other attention-heavy workloads such as long-context retrieval if the broadcast widths are scaled accordingly.
  • Because the design already collapses two precision modes into one datapath, extending the LUT tables to additional low-bit formats would require only modest additional configuration logic.
  • System-level energy savings would increase further if on-chip memory bandwidth is matched to the localized accumulation pattern rather than relying on off-chip DRAM for partial sums.

Load-bearing premise

Post-synthesis area, power, and latency numbers obtained at 200 MHz in 28 nm CMOS will translate directly to real silicon once process variation, interconnect parasitics, and memory-subsystem integration are taken into account.

What would settle it

Fabricate the MxGLUT design in the target 28 nm process and measure actual silicon area, power at 200 MHz, and end-to-end LLM inference energy/latency against the reported synthesis figures; a discrepancy larger than the claimed margins would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2607.01607 by Chen Ding, Hao Jia, Haoming Chu, Liangyu Gan, Lirong Zheng, Mingyuan Liu, Ning Ma, Weiyu Zhou, Yukun Feng, Yuxiang Huan.

Figure 1
Figure 1. Figure 1: Overview of LLM inference heterogeneity and MxGLUT design motivations: (a) transformer computational characteristics with cross-layer arithmetic [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Instantiation of the unified LUT-based execution framework for FP8- [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Instantiation of the unified LUT-based execution framework for the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Microarchitecture of the mixed-precision LUT-based processing element (MxLPE), (a) LUT generator, (b) FP8-FP8 LUT generator, (c) shared FP8- [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overall architecture of MxGLUT. selected value and its negation. This leverages LUT-entry sign symmetry so that only eight entries need to be stored. The selected value is then directly forwarded to the FP32 adder to update the running partial sum. The per-entry FP shifters in the LUT block, controlled by a local state machine, pre￾shift the LUT entries associated with different bit-planes prior to accumul… view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the proposed reconfigurable LUT-centric broadcast (RLB) dataflow operating in (a) output-stationary (OS) mode and (b) weight [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Architectural evaluation results. (a) Area and power comparison between the proposed shared FP8-INT4 LUT generator and the conventional design [10]. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Layer-wise comparison of latency speedup and normalized energy across different dataflows for FP8-INT4 and FP8-FP8 GEMMs during the prefill [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: End-to-end inference latency speedup and normalized energy across different dataflow configurations for the Llama family during the prefill and [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Design space exploration of the multiplexer count ( [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
read the original abstract

Large language model (LLM) inference suffers from growing inefficiency across the prefill and decode phases, especially under weight-only quantization, where activations remain in FP8 while weights are compressed to low-bit integers. Existing LUT-based accelerators mainly target FP8-INT4 computation and still rely on separate floating-point (FP) datapaths for attention GEMM operations, leading to redundant hardware and non-unified mixed-precision execution. Moreover, their static dataflows are poorly matched to the distinct prefill and decode phases. To address these challenges, we propose MxGLUT, a reconfigurable LUT-centric broadcast (RLB) dataflow accelerator built on mixed-precision LUT-based processing elements (MxLPEs). Guided by a unified LUT-based execution framework, MxGLUT organizes both FP8-INT4 and FP8-FP8 GEMMs under a single LUT-based compute mechanism without dedicated FP multipliers or additional FP datapaths, and further adopts the RLB dataflow that localizes heavy partial-sum accumulation during the prefill phase and exploits weight reuse in the decode phase. Synthesized in UMC $28\,\mathrm{nm}$ CMOS at $200~\mathrm{MHz}$, MxGLUT reduces multiplier area by up to $56.92\%$ and power by up to $77.07\%$ and $78.35\%$ in FP8-INT4 and FP8-FP8 modes, respectively. At the accelerator level, MxGLUT achieves an area efficiency of $0.492~\mathrm{TFLOPS/mm^2}$ and an energy efficiency of $11.58~\mathrm{TFLOPS/W}$, while adding native FP8-FP8 support incurs only $2.57\%$ and $3.34\%$ reductions in area and energy efficiency, respectively, relative to the FP8-INT4-only FIGLUT baseline. Across the Llama family, MxGLUT achieves up to $2.16\times$ and $1.49\times$ latency speedup, and reduces normalized energy to $0.44\times$ and $0.71\times$ in prefill and decode, respectively, with at most $1.70\%$ perplexity increase.

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 / 0 minor

Summary. The manuscript proposes MxGLUT, a reconfigurable LUT-centric broadcast (RLB) dataflow accelerator for mixed-precision GEMM operations in large language model inference. It introduces mixed-precision LUT-based processing elements (MxLPEs) that unify FP8-INT4 and FP8-FP8 computations under a single LUT-based mechanism without dedicated floating-point multipliers or separate datapaths. The RLB dataflow is tailored to localize partial-sum accumulation in the prefill phase and exploit weight reuse in the decode phase. Based on synthesis in UMC 28nm CMOS at 200 MHz, the paper reports multiplier area reductions of up to 56.92%, power reductions of up to 77.07% and 78.35% in the respective modes, accelerator-level efficiencies of 0.492 TFLOPS/mm² and 11.58 TFLOPS/W, and up to 2.16× and 1.49× latency speedups in prefill and decode across Llama models with at most 1.70% perplexity increase compared to the FIGLUT baseline.

Significance. Should the synthesis-based claims prove accurate upon fabrication and integration, MxGLUT would offer a meaningful contribution to hardware acceleration for quantized LLM inference by providing a unified compute mechanism for mixed precisions and a phase-adaptive dataflow, potentially leading to more efficient designs that reduce hardware redundancy in attention and linear layers.

major comments (2)
  1. Abstract: The central efficiency claims, such as the 56.92% multiplier area reduction, 77.07% and 78.35% power reductions, 0.492 TFLOPS/mm² area efficiency, and 11.58 TFLOPS/W energy efficiency, along with the 2.57% and 3.34% overheads relative to FIGLUT, are presented without any error bars, details on the verification methodology, or discussion of post-synthesis versus post-layout differences; these metrics are load-bearing for the paper's assertions about the benefits of the MxLPE and RLB design.
  2. Performance claims section: The manuscript does not address how post-synthesis estimates at 200 MHz in 28nm CMOS account for interconnect parasitics, process variation, clock skew, or off-chip memory bandwidth and latency penalties, which the RLB dataflow is said to exploit; this omission directly affects the credibility of the reported system-level latency speedups (up to 2.16× prefill, 1.49× decode) and energy reductions (to 0.44× and 0.71×).

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments on the presentation of our synthesis results and the assumptions underlying our performance claims. We address each point below and will make targeted revisions to improve transparency without altering the core technical contributions.

read point-by-point responses

  1. Referee: Abstract: The central efficiency claims, such as the 56.92% multiplier area reduction, 77.07% and 78.35% power reductions, 0.492 TFLOPS/mm² area efficiency, and 11.58 TFLOPS/W energy efficiency, along with the 2.57% and 3.34% overheads relative to FIGLUT, are presented without any error bars, details on the verification methodology, or discussion of post-synthesis versus post-layout differences; these metrics are load-bearing for the paper's assertions about the benefits of the MxLPE and RLB design.

    Authors: The reported figures come from a single deterministic post-synthesis run in Synopsys Design Compiler using the UMC 28 nm library at a 200 MHz target clock period. Because the input netlist, constraints, and library are fixed, repeated synthesis runs produce identical results and statistical error bars are not applicable. We will add a dedicated paragraph in the evaluation section (and a brief note in the abstract if space allows) that (1) describes the exact synthesis flow, tool versions, power analysis settings, and area reporting methodology, and (2) explicitly states that all numbers are post-synthesis estimates that exclude place-and-route parasitics. We will also note that the RLB dataflow was deliberately designed to localize partial-sum traffic precisely to reduce the sensitivity to interconnect delay. revision: partial

  2. Referee: Performance claims section: The manuscript does not address how post-synthesis estimates at 200 MHz in 28nm CMOS account for interconnect parasitics, process variation, clock skew, or off-chip memory bandwidth and latency penalties, which the RLB dataflow is said to exploit; this omission directly affects the credibility of the reported system-level latency speedups (up to 2.16× prefill, 1.49× decode) and energy reductions (to 0.44× and 0.71×).

    Authors: We agree that post-synthesis results omit several second-order effects. The 200 MHz constraint is applied at the synthesis stage with ideal clock assumptions; process variation and clock skew are not modeled. Off-chip bandwidth is assumed to be non-limiting because the RLB dataflow keeps the majority of partial-sum traffic on-chip during prefill and reuses weights locally during decode. We will expand the evaluation discussion to list these modeling assumptions explicitly and to qualify that the reported speedups and energy ratios are relative to the baseline under identical synthesis conditions. Because the current manuscript contains only synthesis results, we cannot supply quantitative post-layout or silicon-corrected numbers. revision: partial

standing simulated objections not resolved
  • Quantitative post-layout or silicon-level corrections for interconnect parasitics, process variation, clock skew, and off-chip memory penalties on the reported latency and energy figures.

Circularity Check

0 steps flagged

No circularity; all metrics are direct synthesis outputs vs external baselines

full rationale

The paper reports area, power, efficiency, and latency numbers exclusively from post-synthesis results in UMC 28nm at 200 MHz against the FIGLUT baseline and Llama models. No equations, fitted parameters, or self-citations are used to derive the headline claims; the RLB dataflow and MxLPE design choices are presented as engineering decisions whose benefits are measured externally. No self-definitional, fitted-input, or self-citation-load-bearing steps exist in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on standard assumptions of CMOS synthesis tools producing accurate area/power estimates and on the premise that the described dataflow matches LLM phase characteristics without unmodeled overheads.

free parameters (1)
  • Operating frequency = 200 MHz
    200 MHz chosen for UMC 28nm synthesis; directly affects reported TFLOPS/W and area efficiency numbers.
axioms (2)
  • domain assumption Post-synthesis estimates in commercial 28nm process accurately predict fabricated chip metrics without significant deviation from parasitics or variation.
    Invoked when reporting all area, power, and efficiency numbers from synthesis.
  • domain assumption The RLB dataflow incurs no hidden control or routing overheads that would alter the claimed prefill/decode speedups.
    Required for the latency and energy reduction claims across Llama models.

pith-pipeline@v0.9.1-grok · 5974 in / 1756 out tokens · 39736 ms · 2026-07-03T04:30:49.512901+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

34 extracted references · 10 canonical work pages · 3 internal anchors

  1. [1]

    GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “Gptq: Accurate post-training quantization for generative pre-trained transformers,”arXiv preprint arXiv:2210.17323, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  2. [2]

    Awq: Activation-aware weight quanti- zation for on-device llm compression and acceleration,

    J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, and S. Han, “Awq: Activation-aware weight quanti- zation for on-device llm compression and acceleration,”Proceedings of machine learning and systems, vol. 6, pp. 87–100, 2024. 13

    2024

  3. [3]

    Mitigating quantization errors due to activation spikes in glu-based llms,

    J. Yang, H. Kim, and Y . Kim, “Mitigating quantization errors due to activation spikes in glu-based llms,”arXiv preprint arXiv:2405.14428, 2024

    work page arXiv 2024

  4. [4]

    Qqq: Quality quattuor-bit quantization for large language models,

    Y . Zhang, P. Zhang, M. Huang, J. Xiang, Y . Wang, C. Wang, Y . Zhang, L. Yu, C. Liu, and W. Lin, “Qqq: Quality quattuor-bit quantization for large language models,”arXiv preprint arXiv:2406.09904, 2024

    work page arXiv 2024

  5. [5]

    Qserve: W4a8kv4 quantization and system co-design for efficient llm serving,

    Y . Lin, H. Tang, S. Yang, Z. Zhang, G. Xiao, C. Gan, and S. Han, “Qserve: W4a8kv4 quantization and system co-design for efficient llm serving,”arXiv preprint arXiv:2405.04532, 2024

    work page arXiv 2024

  6. [6]

    Post-training quantization of openpangu models for efficient deployment on atlas a2,

    Y . Luo, H. Zheng, H. Meng, W. Liu, and P. Zhang, “Post-training quantization of openpangu models for efficient deployment on atlas a2,” 2026. [Online]. Available: https://arxiv.org/abs/2512.23367

    work page arXiv 2026

  7. [7]

    Zeroquant-fp: A leap forward in llms post- training w4a8 quantization using floating-point formats,

    X. Wu, Z. Yao, and Y . He, “Zeroquant-fp: A leap forward in llms post- training w4a8 quantization using floating-point formats,”arXiv preprint arXiv:2307.09782, 2023

    work page arXiv 2023

  8. [8]

    A survey of low-bit large language models: Basics, systems, and algorithms,

    R. Gong, Y . Ding, Z. Wang, C. Lv, X. Zheng, J. Du, H. Qin, J. Guo, M. Magno, and X. Liu, “A survey of low-bit large language models: Basics, systems, and algorithms,”arXiv preprint arXiv:2409.16694, 2024

    work page arXiv 2024

  9. [9]

    Lut-gemm: Quantized matrix multiplication based on luts for efficient inference in large-scale generative language models,

    G. Park, B. Park, M. Kim, S. Lee, J. Kim, B. Kwon, S. J. Kwon, B. Kim, Y . Lee, and D. Lee, “Lut-gemm: Quantized matrix multiplication based on luts for efficient inference in large-scale generative language models,” arXiv preprint arXiv:2206.09557, 2022

    work page arXiv 2022

  10. [10]

    Figlut: An energy-efficient accelerator design for fp-int gemm using look-up tables,

    G. Park, H. Kwon, J. Kim, J. Bae, B. Park, D. Lee, and Y . Lee, “Figlut: An energy-efficient accelerator design for fp-int gemm using look-up tables,” in2025 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 2025, pp. 1098–1111

    2025

  11. [11]

    Alternating Multi-bit Quantization for Recurrent Neural Networks

    C. Xu, J. Yao, Z. Lin, W. Ou, Y . Cao, Z. Wang, and H. Zha, “Alternating multi-bit quantization for recurrent neural networks,”arXiv preprint arXiv:1802.00150, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  12. [12]

    Lut tensor core: A software-hardware co-design for lut-based low-bit llm inference,

    Z. Mo, L. Wang, J. Wei, Z. Zeng, S. Cao, L. Ma, N. Jing, T. Cao, J. Xue, F. Yanget al., “Lut tensor core: A software-hardware co-design for lut-based low-bit llm inference,” inProceedings of the 52nd Annual International Symposium on Computer Architecture, 2025, pp. 514–528

    2025

  13. [13]

    M. V . Maceiras,Extending Vector Processing Units for Enhanced Linear Algebra Performance. Chalmers Tekniska Hogskola (Sweden), 2024

    2024

  14. [14]

    Gemmini: Enabling systematic deep- learning architecture evaluation via full-stack integration,

    H. Genc, S. Kim, A. Amid, A. Haj-Ali, V . Iyer, P. Prakash, J. Zhao, D. Grubb, H. Liew, H. Maoet al., “Gemmini: Enabling systematic deep- learning architecture evaluation via full-stack integration,” in2021 58th ACM/IEEE Design Automation Conference (DAC). IEEE, 2021, pp. 769–774

    2021

  15. [15]

    The Llama 3 Herd of Models

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  17. [17]

    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, C. Hooper, M. Donato, P. N. Whatmough, A. M. Rush, D. Brooks, and G.-Y . Wei, “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, no. 2, pp. 569–581, 2022

    2022

  18. [18]

    S. Kim, S. Kim, W. Jo, S. Kim, S. Hong, and H.-J. Yoo, “20.5 c- transformer: A 2.6-18.1µj/token homogeneous dnn-transformer/spiking- transformer processor with big-little network and implicit weight gener- ation for large language models,” in2024 IEEE International Solid-State Circuits Conference (ISSCC), vol. 67. IEEE, 2024, pp. 368–370

    2024

  19. [19]

    A 22nm 54.94tflops/w transformer fine-tuning pro- cessor with exponent-stationary re-computing, aggressive linear fitting, and logarithmic domain multiplicating,

    Y . Wang, X. Yang, Y . Qin, Z. Zhao, R. Guo, Z. Yue, H. Han, S. Wei, Y . Hu, and S. Yin, “A 22nm 54.94tflops/w transformer fine-tuning pro- cessor with exponent-stationary re-computing, aggressive linear fitting, and logarithmic domain multiplicating,” in2024 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), 2024, pp. 1–2

    2024

  20. [20]

    Dadiannao: A machine-learning supercomputer,

    Y . Chen, T. Luo, S. Liu, S. Zhang, L. He, J. Wang, L. Li, T. Chen, Z. Xu, N. Sunet al., “Dadiannao: A machine-learning supercomputer,” in2014 47th Annual IEEE/ACM international symposium on microarchitecture. IEEE, 2014, pp. 609–622

    2014

  21. [21]

    In-datacenter performance analysis of a tensor processing unit,

    N. P. Jouppi, C. Young, N. Patil, D. Patterson, G. Agrawal, R. Bajwa, S. Bates, S. Bhatia, N. Boden, A. Borcherset al., “In-datacenter performance analysis of a tensor processing unit,” inProceedings of the 44th annual international symposium on computer architecture, 2017, pp. 1–12

    2017

  22. [22]

    Inca: Input-stationary dataflow at outside- the-box thinking about deep learning accelerators,

    B. Kim, S. Li, and H. Li, “Inca: Input-stationary dataflow at outside- the-box thinking about deep learning accelerators,” in2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 2023, pp. 29–41

    2023

  23. [23]

    Dpima: A dram-based processing-in-memory accelerator for privacy-preserving machine learning,

    B. Kim, “Dpima: A dram-based processing-in-memory accelerator for privacy-preserving machine learning,” in2025 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). IEEE, 2025, pp. 1–7

    2025

  24. [24]

    Dip: A scalable, energy-efficient systolic array for matrix multiplication acceleration,

    A. J. Abdelmaksoud, S. Agwa, and T. Prodromakis, “Dip: A scalable, energy-efficient systolic array for matrix multiplication acceleration,” IEEE Transactions on Circuits and Systems I: Regular Papers, 2025

    2025

  25. [25]

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

    R. Balasubramonian, A. B. Kahng, N. Muralimanohar, A. Shafiee, and V . Srinivas, “Cacti 7: New tools for interconnect exploration in innovative off-chip memories,”ACM Transactions on Architecture and Code Optimization (TACO), vol. 14, no. 2, pp. 1–25, 2017

    2017

  26. [26]

    Timeloop: A systematic approach to dnn accelerator evaluation,

    A. Parashar, P. Raina, Y . S. Shao, Y .-H. Chen, V . A. Ying, A. Mukkara, R. Venkatesan, B. Khailany, S. W. Keckler, and J. Emer, “Timeloop: A systematic approach to dnn accelerator evaluation,” in2019 IEEE inter- national symposium on performance analysis of systems and software (ISPASS). IEEE, 2019, pp. 304–315

    2019

  27. [27]

    Winning both the accuracy of floating point activation and the simplicity of integer arithmetic,

    Y . Kim, J. Jang, J. Lee, J. Park, J. Kim, B. Kim, S. J. Kwon, D. Lee et al., “Winning both the accuracy of floating point activation and the simplicity of integer arithmetic,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  28. [28]

    Desa: Dataflow efficient systolic array for acceleration of transformers,

    Z. Wang, H. Fan, and G. He, “Desa: Dataflow efficient systolic array for acceleration of transformers,”IEEE Transactions on Computers, 2025

    2025

  29. [29]

    Pointer sentinel mixture models,

    S. Merity, C. Xiong, J. Bradbury, and R. Socher, “Pointer sentinel mixture models,” 2016

    2016

  30. [30]

    Shiftaddllm: Accelerating pretrained llms via post-training multiplication-less reparameterization,

    H. You, Y . Guo, Y . Fu, W. Zhou, H. Shi, X. Zhang, S. Kundu, A. Yazdanbakhsh, and Y . C. Lin, “Shiftaddllm: Accelerating pretrained llms via post-training multiplication-less reparameterization,”Advances in Neural Information Processing Systems, vol. 37, pp. 24 822–24 848, 2024

    2024

  31. [31]

    22.9 a 12nm 18.1tflops/w sparse transformer processor with entropy-based early exit, mixed-precision predication and fine-grained power management,

    T. Tambe, J. Zhang, C. Hooper, T. Jia, P. N. Whatmough, J. Zuckerman, M. C. D. Santos, E. J. Loscalzo, D. Giri, K. Shepard, L. Carloni, A. Rush, D. Brooks, and G.-Y . Wei, “22.9 a 12nm 18.1tflops/w sparse transformer processor with entropy-based early exit, mixed-precision predication and fine-grained power management,” in2023 IEEE Inter- national Solid-S...

    2023

  32. [32]

    A 28nm 27.5 tops/w approximate-computing- based transformer processor with asymptotic sparsity speculating and out-of-order computing,

    Y . Wang, Y . Qin, D. Deng, J. Wei, Y . Zhou, Y . Fan, T. Chen, H. Sun, L. Liu, S. Weiet al., “A 28nm 27.5 tops/w approximate-computing- based transformer processor with asymptotic sparsity speculating and out-of-order computing,” in2022 IEEE international solid-state circuits conference (ISSCC), vol. 65. IEEE, 2022, pp. 1–3

    2022

  33. [33]

    T-rex: A 68-to-567µs/token 0.41-to-3.95µj/token transformer accelerator with reduced external memory access and enhanced hardware utilization in 16nm finfet,

    S. Moon, M. Li, G. K. Chen, P. C. Knag, R. K. Krishnamurthy, and M. Seok, “T-rex: A 68-to-567µs/token 0.41-to-3.95µj/token transformer accelerator with reduced external memory access and enhanced hardware utilization in 16nm finfet,” in2025 IEEE International Solid-State Circuits Conference (ISSCC), vol. 68. IEEE, 2025, pp. 406–408

    2025

  34. [34]

    A 22nm 25.08 tops/w multi- task transformer accelerator with mixed precision structured sparsity and two-stage task-adaptive power management,

    Z. Fan, Q. Zhang, P. Abillama, S. Shoori, J. Lee, C.-W. Tseng, W. Meng, H.-S. Kim, D. Blaauw, and D. Sylvester, “A 22nm 25.08 tops/w multi- task transformer accelerator with mixed precision structured sparsity and two-stage task-adaptive power management,” in2025 Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). IEEE, 2025, pp. 1–3

    2025