pith. machine review for the scientific record. sign in

arxiv: 2511.06838 · v4 · submitted 2025-11-10 · 💻 cs.AR · cs.LG

P3-LLM: An Integrated NPU-PIM Accelerator for Edge LLM Inference Using Hybrid Numerical Formats

Yuzong Chen , Chao Fang , Xilai Dai , Yuheng Wu , Thierry Tambe , Marian Verhelst , Mohamed S. Abdelfattah This is my paper

Pith reviewed 2026-05-18 00:12 UTC · model grok-4.3

classification 💻 cs.AR cs.LG
keywords edge LLM inferenceNPU-PIM acceleratorhybrid numerical formatsmixed-precision quantizationprocessing-in-memoryaccelerator co-designoperator fusionDRAM-based PIM
0
0 comments X

The pith

Hybrid numerical formats let low-precision PIM units accelerate edge LLM inference while preserving accuracy.

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 show that NPU-PIM systems for large language model inference can overcome the area and power costs of high-precision DRAM compute units by pairing a mixed-precision quantization scheme with hardware co-design. Different model operands receive hybrid numerical formats chosen for compression efficiency and low accuracy loss; this choice permits the PIM units themselves to run at reduced precision and still deliver higher throughput under fixed area budgets. Operator fusion then removes most runtime dequantization cost in the dataflow. A sympathetic reader would care because edge devices lack the memory bandwidth and power budget of data-center hardware, so any practical way to run capable LLMs locally improves latency, privacy, and accessibility.

Core claim

P3-LLM introduces a flexible mixed-precision quantization scheme that applies hybrid numerical formats to different LLM operands for high compression and minimal accuracy loss. An efficient PIM accelerator is then built with enhanced compute units that support these formats, allowing low-precision PIM operation under iso-area constraints and thereby raising computation throughput. Low-precision dataflow is further optimized by operator fusion to cut dequantization overhead, producing higher accuracy than prior KV-cache and weight-activation quantization methods together with average speedups of 4.9× over HBM-PIM, 2.0× over Ecco, and 3.4× over Pimba across diverse LLMs and tasks.

What carries the argument

Hybrid numerical formats applied to different LLM operands, which enable co-design of low-precision PIM compute units that raise throughput under iso-area constraints in DRAM technology.

If this is right

  • The accelerator achieves higher accuracy than state-of-the-art KV-cache quantization and weight-activation quantization algorithms.
  • Average speedups reach 4.9× over HBM-PIM, 2.0× over Ecco, and 3.4× over Pimba.
  • Operator fusion minimizes runtime dequantization overhead for low-precision dataflow across LLM modules.
  • The design supports diverse LLMs and tasks on edge hardware under iso-area constraints.

Where Pith is reading between the lines

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

  • The same operand-specific format selection might reduce energy in other memory-bound accelerators such as those for vision or speech models.
  • Hardware prototypes could expose additional fusion opportunities when the NPU and PIM share a common low-precision data path.
  • Extending the approach to dynamic format selection at runtime could further adapt to varying sequence lengths without retraining.

Load-bearing premise

The hybrid numerical formats chosen for different operands will maintain acceptable accuracy while allowing the PIM compute units to be built at low precision under realistic iso-area constraints in actual DRAM technology.

What would settle it

A cycle-accurate or post-layout simulation of the low-precision PIM units in real DRAM process parameters that shows either throughput gains below the claimed multiples or accuracy loss exceeding the reported levels on the same models and tasks.

Figures

Figures reproduced from arXiv: 2511.06838 by Chao Fang, Marian Verhelst, Mohamed S. Abdelfattah, Thierry Tambe, Xilai Dai, Yuheng Wu, Yuzong Chen.

Figure 1
Figure 1. Figure 1: Illustration of LLM architecture. rithm, while remaining within the PIM area constraints. 4) Through comprehensive evaluation, we demonstrate that P 3 -LLM achieves higher accuracy than SoTA LLM quantization algorithms Oaken [38], QuaRot [2], and QoQ [47], while offering an average of 4.9×, 2.0×, and 3.4× speedups over the SoTA LLM accelerators HBM￾PIM [43], Ecco [8], and Pimba [39], respectively. II. BACK… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of PIM architectures for LLM decoding acceleration. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of LLM operands: (a) Memory footprint of various LLMs at a 4K context length across different batch sizes. (b) Impact of quantization [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Roofline analysis of an NPU-PIM system and the proposed P [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: KV-cache distribution (in absolute value) of Wikitext-2 dataset from representative layers and heads of Llama-2-7B and Llama-3.1-8B. The context [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) The architecture of P3 -LLM. (b) Operator mapping on P3 -LLM during decoding. (c) The quantized dataflow of three GEMV operations: weights @ activations, query @ key, attention-score @ value. For clarity, we use ”@” and ”×” to denote GEMV and element-wise multiplication, respectively. C. Weight Quantization For weight quantization, we build upon an existing nu￾merical format, BitMoD [7], which adaptive… view at source ↗
Figure 7
Figure 7. Figure 7: Command timing of HBM-PIM and P3 -LLM. executed on NPU using high-precision arithmetic. For linear layers, the dequantization scaling is performed after matrix multiplication. For Q · KT , since the post-RoPE key cache contains per-channel smoothing factors (SSF), we fuse SSF into query via element-wise multiplication prior to FP8 quantiza￾tion. Similarly, for P · V , we fuse the per-value-head scaling fac… view at source ↗
Figure 8
Figure 8. Figure 8: Normalized layer-wise key-cache quantization error of Llama-2-7B [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Normalized speedup (↑) vs. batch sizes (BS) for different accelerator systems. The context length is 4K [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Normalized energy consumption (↓) vs. batch sizes (BS) and the breakdown for attention and linear layers. The context length is 4K. Llama-3 and Mistral offer inherent data reuse opportunities that HBM-PIM fail to exploit. Furthermore, both NPU and HBM-PIM deliver lower performance than Ecco that leverages quantization to reduce the demand of memory bandwidth. On the other hand, P3 -LLM offers substantial … view at source ↗
Figure 11
Figure 11. Figure 11: Normalized speedup (↑) of Pimba and P3 -LLM. The context length is 4K [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Normalized latency (↓) across a wide range of batch sizes from 2 to 64. The context length is 4K. tectural techniques proposed by P3 -LLM. Four designs are evaluated: (1) The baseline HBM-PIM system; (2) A PIM accelerator supporting W4A8KV4 models without attention￾score quantization (PIM-W4A8KV4); (3) A PIM accelera￾tor incorporating our throughput-enhanced PCU to accelerate W4A8KV4 models (PIM-W4A8KV4-T… view at source ↗
read the original abstract

The substantial memory bandwidth and computational demands of large language models (LLMs) present critical challenges for efficient inference. To tackle this, the literature has explored heterogeneous systems that combine neural processing units (NPUs) with DRAM-based processing-in-memory (PIM) for LLM acceleration. However, the high-precision PIM compute units incur significant area and power overhead in DRAM technology, limiting the effective computation throughput. In this paper, we introduce P3-LLM, a novel NPU-PIM integrated accelerator for edge LLM inference. Our approach is threefold: First, we propose a flexible mixed-precision quantization scheme, which leverages hybrid numerical formats to quantize different LLM operands with high compression efficiency and minimal accuracy loss. Second, we architect an efficient PIM accelerator for P3-LLM, featuring enhanced compute units to support hybrid numerical formats. Our careful choice of numerical formats allows to co-design low-precision PIM compute units that significantly boost the computation throughput under iso-area constraints. Third, we optimize the low-precision dataflow of different LLM modules by applying operator fusion to minimize the overhead of runtime dequantization. Evaluations on diverse LLMs and tasks demonstrate that P3-LLM achieves higher accuracy than state-of-the-art KV-cache quantization and weight-activation quantization algorithms. Combining the proposed quantization scheme with low-precision PIM architecture co-design, P3-LLM yields an average of $4.9\times$, $2.0\times$, and $3.4\times$ speedups over state-of-the-art LLM accelerators HBM-PIM, Ecco, and Pimba, respectively. Code is available at https://github.com/yc2367/P3-LLM.

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 manuscript proposes P3-LLM, an integrated NPU-PIM accelerator for edge LLM inference. It introduces a flexible mixed-precision quantization scheme using hybrid numerical formats to quantize different LLM operands, an efficient PIM architecture with enhanced low-precision compute units co-designed for these formats to increase throughput under iso-area constraints in DRAM, and operator fusion to reduce runtime dequantization overhead in the low-precision dataflow. Evaluations claim higher accuracy than state-of-the-art KV-cache and weight-activation quantization methods, along with average speedups of 4.9× over HBM-PIM, 2.0× over Ecco, and 3.4× over Pimba.

Significance. If the hybrid-format co-design successfully enables low-precision PIM units while respecting realistic DRAM area budgets and maintaining accuracy, the work could meaningfully advance efficient edge LLM deployment by improving computation throughput in heterogeneous NPU-PIM systems. The open availability of code at the cited GitHub repository is a positive contribution to reproducibility.

major comments (2)
  1. [Section 4] Section 4 and associated architecture diagrams: the description of enhanced PIM compute units for hybrid numerical formats treats format conversion, dequantization logic, and multi-precision datapath support as area-neutral or negligible under iso-area constraints. In actual DRAM processes, even modest additional logic for operand routing or sense-amp sharing can increase effective area per compute unit and erode the density gain that underpins the reported throughput numbers; a quantitative area breakdown including these overheads is needed to substantiate the central speedup claims.
  2. [Evaluation] Evaluation section: the reported average speedups (4.9×, 2.0×, 3.4×) are presented without visible error bars, detailed descriptions of baseline implementations, or full experimental setup parameters, which prevents full verification of the performance advantages over HBM-PIM, Ecco, and Pimba.
minor comments (2)
  1. [Abstract] The abstract states that P3-LLM 'achieves higher accuracy' than SOTA methods but does not specify the exact accuracy metrics, models, or tasks where the gains are observed; adding this detail would improve clarity.
  2. Figure captions and table labels for the hybrid numerical formats and PIM datapath should explicitly indicate bit-widths and conversion points to aid reader understanding.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and have revised the paper to strengthen the substantiation of our claims.

read point-by-point responses

  1. Referee: [Section 4] Section 4 and associated architecture diagrams: the description of enhanced PIM compute units for hybrid numerical formats treats format conversion, dequantization logic, and multi-precision datapath support as area-neutral or negligible under iso-area constraints. In actual DRAM processes, even modest additional logic for operand routing or sense-amp sharing can increase effective area per compute unit and erode the density gain that underpins the reported throughput numbers; a quantitative area breakdown including these overheads is needed to substantiate the central speedup claims.

    Authors: We appreciate this point. Our hybrid format selection was deliberately chosen to maximize resource sharing across precisions and thereby limit extra logic. Nevertheless, we agree that an explicit breakdown is required for credibility. In the revised Section 4 we now include a quantitative area breakdown (Table IV and accompanying text) derived from synthesized layouts in the target DRAM process. The breakdown shows that format conversion, dequantization, and multi-precision routing together add less than 8 % to the per-unit area; the reported iso-area throughput gains remain intact after this overhead is accounted for. revision: yes

  2. Referee: [Evaluation] Evaluation section: the reported average speedups (4.9×, 2.0×, 3.4×) are presented without visible error bars, detailed descriptions of baseline implementations, or full experimental setup parameters, which prevents full verification of the performance advantages over HBM-PIM, Ecco, and Pimba.

    Authors: We concur that greater transparency aids verification. The revised Evaluation section now reports error bars (standard deviation over five independent runs), provides expanded descriptions of how each baseline (HBM-PIM, Ecco, Pimba) was implemented and configured, and includes a new table (Table VII) listing all key experimental parameters: hardware configurations, DRAM process assumptions, workload batch sizes, and simulation settings. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on external evaluations

full rationale

The paper proposes a hybrid-format quantization scheme and co-designed low-precision PIM units, then reports empirical speedups (4.9×/2.0×/3.4×) measured against external baselines HBM-PIM, Ecco, and Pimba on diverse LLMs. No derivation step reduces a claimed prediction or throughput result to a fitted parameter or self-citation by construction. Iso-area assumptions are design choices whose validity is tested via reported evaluations rather than defined into the result. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The design rests on standard assumptions about DRAM area/power trade-offs for PIM units and on the empirical claim that chosen hybrid formats preserve accuracy; no new physical entities are postulated.

axioms (1)
  • domain assumption High-precision PIM compute units incur significant area and power overhead in DRAM technology.
    Invoked in the abstract to motivate the low-precision co-design.

pith-pipeline@v0.9.0 · 5629 in / 1297 out tokens · 53800 ms · 2026-05-18T00:12:55.407587+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

79 extracted references · 79 canonical work pages · 9 internal anchors

  1. [1]

    AMD INSTINCT™ MI350X GPU

    AMD, “AMD INSTINCT™ MI350X GPU.” [Online]. Avail- able: https://www.amd.com/content/dam/amd/en/documents/instinct- tech-docs/product-briefs/amd-instinct-mi350x-gpu-brochure.pdf

    work page

  2. [2]

    QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs,

    S. Ashkboos, A. Mohtashami, M. L. Croci, B. Li, M. Jaggi, D. Alistarh, T. Hoefler, and J. Hensman, “QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs,”Advances in neural information processing systems (NeurIPS), 2024

    work page 2024

  3. [3]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond,”arXiv preprint arXiv:2308.12966, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding,

    Y . Bai, X. Lv, J. Zhang, H. Lyu, J. Tang, Z. Huang, Z. Du, X. Liu, A. Zeng, L. Hou, Y . Dong, J. Tang, and J. Li, “LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding,”Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 2024

    work page 2024

  5. [5]

    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, June 2017

    work page 2017

  6. [6]

    Language Models are Few-Shot Learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert- V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amode...

    work page 2020

  7. [7]

    BitMoD: Bit-serial Mixture-of- Datatype LLM Acceleration,

    Y . Chen, A. F. AbouElhamayed, X. Dai, Y . Wang, M. Andronic, G. A. Constantinides, and M. S. Abdelfattah, “BitMoD: Bit-serial Mixture-of- Datatype LLM Acceleration,”IEEE International Symposium on High- Performance Computer Architecture (HPCA), 2025

    work page 2025

  8. [8]

    Ecco: Improving Memory Band- width and Capacity for LLMs via Entropy-Aware Cache Compression,

    F. Cheng, C. Guo, C. Wei, J. Zhang, C. Zhou, E. Hanson, J. Zhang, X. Liu, H. H. Li, and Y . Chen, “Ecco: Improving Memory Band- width and Capacity for LLMs via Entropy-Aware Cache Compression,” ACM/IEEE 52nd Annual International Symposium on Computer Archi- tecture (ISCA), 2025

    work page 2025

  9. [9]

    Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

    P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord, “Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge,”arXiv preprint arXiv:1803.05457, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    DeepSeek R1

    DeepSeek AI, “DeepSeek R1.” [Online]. Available: https://github.com/ deepseek-ai/DeepSeek-R1

    work page

  11. [11]

    LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale

    T. Dettmers, M. Lewis, Y . Belkada, and L. Zettlemoyer, “LLM.int8(): 8-bit matrix multiplication for transformers at scale,”arXiv preprint arXiv:2208.07339, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  12. [12]

    The true Processing In Memory accelerator,

    F. Devaux, “The true Processing In Memory accelerator,”IEEE Hot Chips 31 Symposium (HCS), 2019

    work page 2019

  13. [13]

    Documenting large webtext corpora: A case study on the colossal clean crawled corpus,

    J. Dodge, A. Marasovic, G. Ilharco, D. Groeneveld, M. Mitchell, and M. Gardner, “Documenting large webtext corpora: A case study on the colossal clean crawled corpus,” inConference on Empirical Methods in Natural Language Processing (EMNLP), 2021

    work page 2021

  14. [14]

    Learning from Students: Applying t-Distributions to Explore Accurate and Efficient Formats for LLMs,

    J. Dotzel, Y . Chen, B. Kotb, S. Prasad, G. Wu, S. Li, M. S. Abdelfat- tah, and Z. Zhang, “Learning from Students: Applying t-Distributions to Explore Accurate and Efficient Formats for LLMs,”International Conference on Machine Learning (ICML), 2024

    work page 2024

  15. [15]

    Anda: Unlocking Efficient LLM Inference with a Variable-Length Grouped Activation Data Format,

    C. Fang, M. Shi, R. Geens, A. Symons, Z. Wang, and M. Verhelst, “Anda: Unlocking Efficient LLM Inference with a Variable-Length Grouped Activation Data Format,”IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2025

    work page 2025

  16. [16]

    GPTQ: Accurate Post-training Compression for Generative Pretrained Transformers,

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “GPTQ: Accurate Post-training Compression for Generative Pretrained Transformers,” International Conference on Learning Representations (ICLR), 2023

    work page 2023

  17. [17]

    The Pile: An 800GB Dataset of Diverse Text for Language Modeling

    L. Gao, S. Biderman, S. Black, L. Golding, T. Hoppe, C. Foster, J. Phang, H. He, A. Thite, N. Nabeshima, S. Presser, and C. Leahy, “The Pile: An 800GB Dataset of Diverse Text for Language Modeling,” arXiv preprint arXiv:2101.00027, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  18. [18]

    Albert Gu and Tri Dao

    L. Gao, J. Tow, B. Abbasi, S. Biderman, S. Black, A. DiPofi, C. Foster, L. Golding, J. Hsu, A. Le Noac’h, H. Li, K. McDonell, N. Muennighoff, C. Ociepa, J. Phang, L. Reynolds, H. Schoelkopf, A. Skowron, L. Sutawika, E. Tang, A. Thite, B. Wang, K. Wang, and A. Zou, “The Language Model Evaluation Harness,” 2024. [Online]. Available: https://zenodo.org/recor...

    work page arXiv 2024

  19. [19]

    Energy Cost Modelling for Optimizing Large Language Model Inference on Hardware Accelerators,

    R. Geens, M. Shi, A. Symons, C. Fang, and M. Verhelst, “Energy Cost Modelling for Optimizing Large Language Model Inference on Hardware Accelerators,” inIEEE 37th International System-on-Chip Conference (SOCC), 2024

    work page 2024

  20. [20]

    OliVe: Accelerating Large Language Models via Hardware-friendly Outlier-Victim Pair Quantization,

    C. Guo, J. Tang, W. Hu, J. Leng, C. Zhang, F. Yang, Y .-B. Liu, M. Guo, and Y . Zhu, “OliVe: Accelerating Large Language Models via Hardware-friendly Outlier-Victim Pair Quantization,”ACM/IEEE 50th Annual International Symposium on Computer Architecture (ISCA), 2023

    work page 2023

  21. [21]

    ANT: Exploiting Adaptive Numerical Data Type for Low-bit Deep Neural Network Quantization,

    C. Guo, C. Zhang, J. Leng, Z. Liu, F. Yang, Y .-B. Liu, M. Guo, and Y . Zhu, “ANT: Exploiting Adaptive Numerical Data Type for Low-bit Deep Neural Network Quantization,”IEEE/ACM 55th Annual International Symposium on Microarchitecture (MICRO), 2022

    work page 2022

  22. [22]

    Newton: A DRAM-maker’s Accelerator-in- Memory (AiM) Architecture for Machine Learning,

    M. He, C. Song, I. Kim, C. Jeong, S. Kim, I. Park, M. Thottethodi, and T. N. Vijaykumar, “Newton: A DRAM-maker’s Accelerator-in- Memory (AiM) Architecture for Machine Learning,”IEEE/ACM 53rd International Symposium on Microarchitecture (MICRO), 2020

    work page 2020

  23. [23]

    LP-Spec: Leveraging LPDDR PIM for Efficient LLM Mobile Speculative Inference with Architecture- Dataflow Co-Optimization,

    S. He, Z. Zhu, Y . He, and T. Jia, “LP-Spec: Leveraging LPDDR PIM for Efficient LLM Mobile Speculative Inference with Architecture- Dataflow Co-Optimization,” inIEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2025

    work page 2025

  24. [24]

    Measuring Massive Multitask Language Understanding,

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt, “Measuring Massive Multitask Language Understanding,” inInternational Conference on Learning Representations (ICLR), 2021

    work page 2021

  25. [25]

    NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing,

    G. Heo, S. Lee, J. Cho, H. Choi, S. Lee, H. Ham, G. T. Kim, D. Mahajan, and J. Park, “NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing,”Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2024

    work page 2024

  26. [26]

    KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization,

    C. Hooper, S. Kim, H. Mohammadzadeh, M. W. Mahoney, Y . S. Shao, K. Keutzer, and A. Gholami, “KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization,”Advances in neural information processing systems (NeurIPS), 2024

    work page 2024

  27. [27]

    M-ANT: Efficient Low-bit Group Quantization 13 for LLMs via Mathematically Adaptive Numerical Type,

    W. Hu, H. Zhang, C. Guo, Y . Feng, R. Guan, Z. Hua, Z. Liu, Y . Guan, M. Guo, and J. Leng, “M-ANT: Efficient Low-bit Group Quantization 13 for LLMs via Mathematically Adaptive Numerical Type,”IEEE Interna- tional Symposium on High-Performance Computer Architecture (HPCA), 2025

    work page 2025

  28. [28]

    PLAIN: Leveraging High Internal Bandwidth in PIM for Accelerating Large Language Model Inference via Mixed-Precision Quantization,

    Y . Hu, F. Liu, Z. Wang, Y . Zhao, T. Yang, L. Jiang, and H. Guan, “PLAIN: Leveraging High Internal Bandwidth in PIM for Accelerating Large Language Model Inference via Mixed-Precision Quantization,” inIEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2025

    work page 2025

  29. [29]

    FIGNA: Integer unit-based accelerator design for fp-int gemm preserving numerical accuracy,

    J. Jang, Y . Kim, J. Lee, and J.-J. Kim, “FIGNA: Integer unit-based accelerator design for fp-int gemm preserving numerical accuracy,”IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2024

    work page 2024

  30. [30]

    BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference,

    W. Jang and T. Tambe, “BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference,”International Conference on Machine Learning (ICML), 2025

    work page 2025

  31. [31]

    High Bandwidth Memory DRAM,

    JEDEC, “High Bandwidth Memory DRAM,” 2021. [Online]. Available: https://www.jedec.org/standards-documents/docs/jesd235a

    work page 2021

  32. [32]

    High Bandwidth Memory (HBM3) DRAM,

    JEDEC, “High Bandwidth Memory (HBM3) DRAM,” 2025. [Online]. Available: https://www.jedec.org/standards-documents/docs/jesd238b01

    work page 2025

  33. [33]

    High Bandwidth Memory (HBM4) DRAM,

    JEDEC, “High Bandwidth Memory (HBM4) DRAM,” 2025. [Online]. Available: https://www.jedec.org/standards-documents/docs/jesd270-4

    work page 2025

  34. [34]

    Mistral 7B

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de Las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, and W. E. Sayed, “Mistral 7B,”arXiv preprint arXiv:2310.06825, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  35. [35]

    Ten Lessons From Three Generations Shaped Google’s TPUv4i: Industrial Product,

    N. P. Jouppi, D. H. Yoon, M. Ashcraft, M. Gottscho, T. B. Jablin, G. Kurian, J. Laudon, S. Li, P. C. Ma, X. Ma, T. Norrie, N. Patil, S. Prasad, C. Young, Z. Zhou, and D. A. Patterson, “Ten Lessons From Three Generations Shaped Google’s TPUv4i: Industrial Product,” ACM/IEEE 48th Annual International Symposium on Computer Archi- tecture (ISCA), 2021

    work page 2021

  36. [36]

    SK Hynix AI-Specific Computing Memory Solution: From AiM Device to Heterogeneous AiMX-xPU System for Comprehensive LLM Inference,

    G. Kim, J. Kim, N. Y . Kim, W. Shin, J.-H. Won, H. Joo, H. Choi, B. An, G. Shin, D. Yun, J. Kim, C. Kim, I.-H. Kim, J. Park, Y . Song, B. Yang, H. Lee, S. Park, W. Lee, S. Kim, Y . Park, Y . Jung, G.-H. Park, and E. Lim, “SK Hynix AI-Specific Computing Memory Solution: From AiM Device to Heterogeneous AiMX-xPU System for Comprehensive LLM Inference,”IEEE ...

    work page 2024

  37. [37]

    Samsung PIM/PNM for Transfmer Based AI : Energy Efficiency on PIM/PNM Cluster,

    J. H. Kim, Y . Ro, J. So, S. Lee, S. Kang, Y . Cho, H. Kim, B. Kim, K. Kim, S.-S. Park, J.-S. Kim, S. Cha, W.-J. Lee, J. Jung, J. Lee, J. Lee, J. Song, S. Lee, J. Cho, J. Yu, and K. Sohn, “Samsung PIM/PNM for Transfmer Based AI : Energy Efficiency on PIM/PNM Cluster,”IEEE Hot Chips 35 Symposium (HCS), 2023

    work page 2023

  38. [38]

    Oaken: Fast and Efficient LLM Serving with Online-Offline Hybrid KV Cache Quantization,

    M. Kim, S. Hong, R. Ko, S. Choi, H. Lee, J. Kim, J.-Y . Kim, and J. Park, “Oaken: Fast and Efficient LLM Serving with Online-Offline Hybrid KV Cache Quantization,”ACM/IEEE 52nd Annual International Symposium on Computer Architecture (ISCA), 2025

    work page 2025

  39. [39]

    Pimba: A Processing- in-Memory Acceleration for Post-Transformer Large Language Model Serving,

    W. Kim, Y . Lee, Y . Kim, J. Hwang, S. Oh, J. Jung, A. Huseynov, W. G. Park, C. H. Park, D. Mahajan, and J. Park, “Pimba: A Processing- in-Memory Acceleration for Post-Transformer Large Language Model Serving,”IEEE/ACM 58th International Symposium on Microarchitec- ture (MICRO), 2025

    work page 2025

  40. [40]

    Tender: Accelerating Large Language Mod- els via Tensor Decomposition and Runtime Requantization,

    J. Lee, W. Lee, and J. Sim, “Tender: Accelerating Large Language Mod- els via Tensor Decomposition and Runtime Requantization,”ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA), 2024

    work page 2024

  41. [41]

    MX+: Pushing the Limits of Microscaling Formats for Efficient Large Language Model Serving,

    J. Lee, J. Park, S. Cha, J. Cho, and J. Sim, “MX+: Pushing the Limits of Microscaling Formats for Efficient Large Language Model Serving,” IEEE/ACM 58th Annual International Symposium on Microarchitecture (MICRO), 2025

    work page 2025

  42. [42]

    A 1ynm 1.25V 8Gb 16Gb/s/Pin GDDR6- Based Accelerator-in-Memory Supporting 1TFLOPS MAC Operation and Various Activation Functions for Deep Learning Application,

    S. J. Lee, K. Kim, S. Oh, J. Park, G. Hong, D. Y . Ka, K.-D. Hwang, J.-J. Park, K. Kang, J. Kim, J. Jeon, N. Y . Kim, Y . Kwon, K. Vladimir, W. Shin, J.-H. Won, M. Lee, H. Joo, H. Choi, J. Lee, D.-Y . Ko, Y . Jun, K. yeong Cho, I. Kim, C. Song, C. Jeong, D.-H. Kwon, J. Jang, I. Park, J. H. Chun, and J. Cho, “A 1ynm 1.25V 8Gb 16Gb/s/Pin GDDR6- Based Accele...

    work page 2023

  43. [43]

    Hardware Architecture and Software Stack for PIM Based on Commercial DRAM Technology: Industrial Product,

    S. Lee, S. Kang, J. Lee, H. Kim, E. Lee, S. young Seo, H. Yoon, S. Lee, K. Lim, H. Shin, J. Kim, S. O, A. Iyer, D. Wang, K. Sohn, and N. S. Kim, “Hardware Architecture and Software Stack for PIM Based on Commercial DRAM Technology: Industrial Product,”ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA), 2021

    work page 2021

  44. [44]

    H2-LLM: Hardware-Dataflow Co-Exploration for Heterogeneous Hybrid-Bonding-based Low-Batch LLM Inference,

    C. Li, Y . Yin, X. Wu, J. Zhu, Z. Gao, D. Niu, Q. Wu, X. Si, Y . Xie, C. Zhang, and G. Sun, “H2-LLM: Hardware-Dataflow Co-Exploration for Heterogeneous Hybrid-Bonding-based Low-Batch LLM Inference,” ACM/IEEE 52nd Annual International Symposium on Computer Archi- tecture (ISCA), 2025

    work page 2025

  45. [45]

    ORCHES: Orchestrated Test-Time- Compute-based LLM Reasoning on Collaborative GPU-PIM HEteroge- neous System,

    S. Li, Y . Chen, C. Li, Y . Fu, Z. Wang, Z. Yu, H. You, Z. Ye, W. Zhou, Y . Zhang, and Y . C. Lin, “ORCHES: Orchestrated Test-Time- Compute-based LLM Reasoning on Collaborative GPU-PIM HEteroge- neous System,”IEEE/ACM 58th Annual International Symposium on Microarchitecture (MICRO), 2025

    work page 2025

  46. [46]

    AWQ: Activation-aware Weight Quan- tization for 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 Quan- tization for LLM Compression and Acceleration,” inProceedings of Machine Learning and Systems (MLSys), 2024

    work page 2024

  47. [47]

    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,” inProceedings of Machine Learning and Systems (MLSys), 2025

    work page 2025

  48. [48]

    SPARK: Scalable and Precision-Aware Acceleration of Neural Networks via Ef- ficient Encoding,

    F. Liu, N. Yang, H. Li, Z. Wang, Z. Song, S. Pei, and L. Jiang, “SPARK: Scalable and Precision-Aware Acceleration of Neural Networks via Ef- ficient Encoding,”IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2024

    work page 2024

  49. [49]

    KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache,

    Z. Liu, J. Yuan, H. Jin, S. Zhong, Z. Xu, V . Braverman, B. Chen, and X. Hu, “KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache,”International Conference on Machine Learning (ICML), 2024

    work page 2024

  50. [50]

    Ramulator 2.0: A Modern, Modular, and Extensible DRAM Simulator,

    H. Luo, Y . C. Tugrul, F. N. Bostanci, A. Olgun, A. G. Yaglikc ¸i, and O. Mutlu, “Ramulator 2.0: A Modern, Modular, and Extensible DRAM Simulator,”IEEE Computer Architecture Letters (CAL), vol. 23, pp. 112– 116, 2023

    work page 2023

  51. [51]

    Pointer sentinel mixture models,

    S. Merity, C. Xiong, J. Bradbury, and R. Socher, “Pointer sentinel mixture models,”International Conference on Learning Representations (ICLR), 2017

    work page 2017

  52. [52]

    Introducing Llama 3.1: Our most capable models to date

    Meta, “Introducing Llama 3.1: Our most capable models to date.” [Online]. Available: https://ai.meta.com/blog/meta-llama-3-1/

    work page

  53. [53]

    Llama-3.2-90B-Vision-Instruct

    Meta, “Llama-3.2-90B-Vision-Instruct.” [Online]. Available: https: //huggingface.co/meta-llama/Llama-3.2-90B-Vision-Instruct

    work page

  54. [54]

    Llama 3.2: Revolutionizing edge AI and vision with open, customizable models

    Meta, “Llama 3.2: Revolutionizing edge AI and vision with open, customizable models.” [Online]. Available: https://ai.meta.com/blog/ llama-3-2-connect-2024-vision-edge-mobile-devices/

    work page 2024

  55. [55]

    Meta Llama 2

    Meta, “Meta Llama 2.” [Online]. Available: https://github.com/meta- llama/llama

    work page

  56. [56]

    The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation

    Meta, “The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation.” [Online]. Available: https: //ai.meta.com/blog/llama-4-multimodal-intelligence/

    work page

  57. [57]

    FP8 Formats for Deep Learning

    P. Micikevicius, D. Stosic, N. Burgess, M. Cornea, P. K. Dubey, R. Grisenthwaite, S. Ha, A. Heinecke, P. Judd, J. Kamalu, N. Mellem- pudi, S. F. Oberman, M. Shoeybi, M. Siu, and H. Wu, “FP8 Formats for Deep Learning,”arXiv preprint arXiv:2209.05433, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  58. [58]

    Introducing NVFP4 for Efficient and Accurate Low-Precision Inference

    NVIDIA, “Introducing NVFP4 for Efficient and Accurate Low-Precision Inference.” [Online]. Available: https://developer.nvidia.com/blog/ introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/

    work page

  59. [59]

    NVIDIA Blackwell GPU Architecture

    NVIDIA, “NVIDIA Blackwell GPU Architecture.” [Online]. Available: https://resources.nvidia.com/en-us-blackwell-architecture/datasheet

    work page

  60. [60]

    Openai o3-mini

    Open AI, “Openai o3-mini.” [Online]. Available: https://openai.com/ index/openai-o3-mini/

    work page

  61. [61]

    Gsm8k dataset

    OpenAI, “Gsm8k dataset.” [Online]. Available: https://huggingface.co/ datasets/openai/gsm8k

    work page

  62. [62]

    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,”IEEE International Symposium on High- Performance Computer Architecture (HPCA), 2025

    work page 2025

  63. [63]

    AttAcc! Unleashing the Power of PIM for Batched Transformer- based Generative Model Inference,

    J. Park, J. Choi, K. Kyung, M. J. Kim, Y . Kwon, N. S. Kim, and J. H. Ahn, “AttAcc! Unleashing the Power of PIM for Batched Transformer- based Generative Model Inference,”Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2024

    work page 2024

  64. [64]

    MicroScopiQ: Ac- celerating Foundational Models through Outlier-Aware Microscaling Quantization,

    A. Ramachandran, S. Kundu, and T. Krishna, “MicroScopiQ: Ac- celerating Foundational Models through Outlier-Aware Microscaling Quantization,”ACM/IEEE 52nd Annual International Symposium on Computer Architecture (ISCA), 2025

    work page 2025

  65. [65]

    With Shared Microexponents, A Little Shifting Goes a Long Way,

    B. D. Rouhani, R. Zhao, V . Elango, R. Shafipour, M. Hall, M. Mes- makhosroshahi, A. More, L. Melnick, M. Golub, G. Varatkar, L. Shao, G. Kolhe, D. Melts, J. Klar, R. L’Heureux, M. Perry, D. Burger, E. S. Chung, Z. Deng, S. Naghshineh, J. Park, and M. Naumov, “With Shared Microexponents, A Little Shifting Goes a Long Way,”ACM/IEEE 50th 14 Annual Internati...

    work page 2023

  66. [66]

    IANUS: Integrated Accelerator based on NPU-PIM Unified Memory System,

    M. Seo, X. T. Nguyen, S. J. Hwang, Y . Kwon, G. Kim, C. Y . Park, I.-H. Kim, J. Park, J. Kim, W. Shin, J.-H. Won, H. Choi, K. Kim, D. Kwon, C. Jeong, S. Lee, Y . Choi, W. Byun, S. Baek, H.-J. Lee, and J. Kim, “IANUS: Integrated Accelerator based on NPU-PIM Unified Memory System,”Proceedings of the 29th ACM International Conference on Ar- chitectural Suppo...

    work page 2024

  67. [67]

    OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models,

    W. Shao, M. Chen, Z. Zhang, P. Xu, L. Zhao, Z. Li, K. Zhang, P. Gao, Y . J. Qiao, and P. Luo, “OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models,”International Conference on Learning Representations (ICLR), 2024

    work page 2024

  68. [68]

    RoFormer: Enhanced Transformer with Rotary Position Embedding,

    J. Su, Y . Lu, S. Pan, B. Wen, and Y . Liu, “RoFormer: Enhanced Transformer with Rotary Position Embedding,”Neurocomputing, 2024

    work page 2024

  69. [69]

    LLaMA: Open and Efficient Foundation Language Models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “LLaMA: Open and Efficient Foundation Language Models,”arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  70. [70]

    FP8 versus INT8 for efficient deep learning inference,

    M. van Baalen, A. Kuzmin, S. S. Nair, Y . Ren, E. Mahurin, C. Patel, S. Subramanian, S. Lee, M. Nagel, J. B. Soriaga, and T. Blankevoort, “FP8 versus INT8 for efficient deep learning inference,”arXiv preprint arXiv:2303.17951, 2023

    work page arXiv 2023

  71. [71]

    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

  72. [72]

    SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models,

    G. Xiao, J. Lin, M. Seznec, J. Demouth, and S. Han, “SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models,”International Conference on Machine Learning (ICML), 2023

    work page 2023

  73. [73]

    Amove: Accelerating LLMs through Mitigating Outliers and Salient Points via Fine-Grained Grouped Vectorized Data Type,

    X. Xie, L. Wang, L. Xiao, M. Han, L. Liu, X. Xu, J. Wang, Z. Song, and X. Liao, “Amove: Accelerating LLMs through Mitigating Outliers and Salient Points via Fine-Grained Grouped Vectorized Data Type,” IEEE/ACM 58th International Symposium on Microarchitecture (MI- CRO), 2025

    work page 2025

  74. [74]

    Qwen2.5 Technical Report

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y . Fan, Y . Su, Y . Zhang, Y . Wan, Y . Liu, Z. Cui, Z. Zhang, and Z. Qi...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  75. [75]

    FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving,

    Z. Ye, L. Chen, R. Lai, W. Lin, Y . Zhang, S. Wang, T. Chen, B. Kasikci, V . Grover, A. Krishnamurthy, and L. Ceze, “FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving,”Proceed- ings of Machine Learning and Systems (MLSys), 2025

    work page 2025

  76. [76]

    Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batch- ing,

    S. Yun, K. Kyung, J. Cho, J. Choi, J. Kim, B. Kim, S. Lee, K. Sohn, and J. H. Ahn, “Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batch- ing,”IEEE/ACM 57th International Symposium on Microarchitecture (MICRO), 2024

    work page 2024

  77. [77]

    SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration,

    J. Zhang, J. Wei, P. Zhang, J. Zhu, and J. Chen, “SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration,”International Conference on Learning Representations (ICLR), 2025

    work page 2025

  78. [78]

    DistServe: Disaggregating Prefill and Decoding for Goodput- optimized Large Language Model Serving,

    Y . Zhong, S. Liu, J. Chen, J. Hu, Y . Zhu, X. Liu, X. Jin, and H. Zhang, “DistServe: Disaggregating Prefill and Decoding for Goodput- optimized Large Language Model Serving,” in18th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2024

    work page 2024

  79. [79]

    A Survey on Efficient Inference for Large Language Models

    Z. Zhou, X. Ning, K. Hong, T. Fu, J. Xu, S. Li, Y . Lou, L. Wang, Z. Yuan, X. Li, S. Yan, G. Dai, X. Zhang, Y . Dong, and Y . Wang, “A Survey on Efficient Inference for Large Language Models,”arXiv preprint arXiv:2404.14294, 2024. 15

    work page internal anchor Pith review Pith/arXiv arXiv 2024