Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Jack Unit: An Area- and Energy-Efficient Multiply-Accumulate (MAC) Unit Supporting Diverse Data Formats

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a single multiply-accumulate unit can serve integer, floating-point, and microscaling (MX) data formats, cutting unit area by up to 2.01x and improving accelerator-level energy efficiency by up to 5.41x.

desk verdict A genuinely new multi-format MAC datapath with credible area/power numbers, but the FP and MX accuracy claims currently rest on a single-layer C simulation and need end-to-end validation before those modes can be taken at face value. read the letter →

arxiv 2507.04772 v1 pith:2RXLKAEX submitted 2025-07-07 cs.AR

Seock-Hwan Noh , Sungju Kim , Seohyun Kim , Daehoon Kim , Jaeha Kung , Yeseong Kim This is my paper · ORCID
classification cs.AR
keywords multiply-accumulateunitdataformatflexibilityintegerarithmeticfloating-pointmicroscalingformatscarry-savemultipliersub-wordparallelismAIacceleratorenergyefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to prove that one multiply-accumulate (MAC) unit can cover integer (INT), floating-point (FP), and microscaling (MX) data formats with bit-level flexibility, instead of a chip carrying a separate multiplier for every format. The move is to push the expensive parts of FP arithmetic—exponent alignment and significand shifting—down into a reconfigurable carry-save multiplier, so that accumulation happens in a cheap integer adder tree. If the claim holds, AI accelerators can replace several specialized compute fabrics with one compact unit, and the numbers the paper reports are 1.17–2.01x smaller area and 1.05–1.84x lower power per MAC, and 1.32–5.41x better energy efficiency for a full accelerator across five AI benchmarks. The paper also claims the unit supports MX formats, including a 7.13x energy-efficiency gain over a baseline running bfloat16.

What carries the argument

The central object is the reconstructed carry-save multiplier (CSM), built from four precision-scalable CSMs, each made of four 4-bit × 4-bit sub-multipliers that can fuse their partial products for 8-bit significands or stay separate for four 4-bit lanes. Its job is to perform exponent-difference alignment by shifting sub-product significands inside the multiplier, so the following accumulation can be done by an integer adder tree instead of a floating-point adder tree. The design also exploits 2D sub-word parallelism: sub-multipliers across the four CSMs that receive identical shift amounts share barrel shifters, reducing shifter count by 75% and shortening the critical path. A normalizer and rounder at the end produce a single 16-bit output, matching the accumulation-width convention of the baseline it is compared with.

What would settle it

Run the Jack unit's C-level model on all five benchmarks in bfloat16, FP8, and MXFP8 modes and compare against a GPU reference and a standard FP MAC unit: if any benchmark's numerical deviation markedly exceeds the 0.2% seen on the ConvNeXt-T layer, or if error grows with accumulation depth or MX block size, the FP and MX claim collapses. Independently re-synthesizing the layout with a different 65 nm flow or a tighter timing target would check whether the 1.17–2.01x area and 1.05–1.84x power ratios are robust to implementation choices.

Watch

Extended reading notes

Core claim

The paper's central discovery is that FP and MX multiply-accumulate operations can be rewritten so that significand alignment happens inside a precision-scalable carry-save multiplier, before any floating-point addition. Four bit-scalable CSMs are clustered into one reconstructed CSM, and sub-multipliers that need the same left-shift share barrel shifters, cutting shifter count by 75% and narrowing the integer adder tree's input width. In this design, FP8, bfloat16, INT8, INT4, MXINT8, and MXFP8 are all modes of one datapath rather than different datapaths. Measured on a synthesized 65 nm layout, the Jack unit occupies 1.17–2.01x less area and consumes 1.05–1.84x less power than three baselines, and an accelerator built from it reaches 1.32–5.41x higher energy efficiency than a baseline accelerator with the same throughput across five AI benchmarks, at the cost of a 6.65% longer inference latency in the directly compared modes.

Load-bearing premise

The load-bearing premise is that accumulating floating-point products in an integer adder, without intermediate normalization or rounding, stays numerically close enough to standard FP MAC arithmetic for real models; the paper tests this on one layer of one model and assumes it carries over to the other benchmarks and MX modes.

Editorial extensions

If this is right

  • A single Jack unit can replace several dedicated-format MAC units, so an accelerator can expose bfloat16, FP8, INT8, INT4, and MX modes as different configurations of the same hardware rather than separate compute arrays.
  • Because FP accumulation flows through an integer adder tree, the floating-point and MX modes inherit the area and power profile of integer hardware, with numerical error measured at about 0.2% on the one tested layer.
  • At the accelerator level, the MAC-array area shrinks by about 1.93x and compute density rises about 1.80x at the same 400 MHz clock, making the unit useful wherever area per operation matters.
  • The MX modes are usable at low cost: MXINT8 gives about 7.13x better energy efficiency than the baseline in bfloat16, and MXFP8 gives about 4.98x over the baseline in FP8, because shared exponents cut the per-element multiply work.
  • The design carries a latency overhead of about 6.65% in the compared bfloat16 and INT4 modes, caused by the pipelined datapath and on-chip buffer access, which a system designer would trade against the energy and area gains.

Reading between the lines

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

  • A natural extension is to stress-test the integer-adder accumulation across all five benchmarks and across longer reduction chains, since the paper verifies numerical accuracy on only one layer of one model; if error grows with depth or block size, FP and MX modes would need occasional re-normalization.
  • The same significand-inside-multiplier trick could generalize to other formats, including larger FP widths or block-floating-point blocks of different sizes, because the shared-exponent handling is already parameterized across Jack units.
  • The 2D sub-word parallelism idea—grouping sub-multipliers that share shift amounts—could transfer to other multiplier-array styles, including bit-serial or transposed designs, wherever sub-word products align by identical shifts.
  • The measured latency penalty suggests the energy benefits could shrink on memory-bound workloads or larger arrays, where the pipelined datapath and buffer-access overhead dominate; an end-to-end system study would reveal where the crossover lies.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes the Jack unit, a multiply-accumulate (MAC) datapath intended to support INT, FP, and MX data formats in a single unit. The design combines a precision-scalable carry-save multiplier, in-multiplier exponent alignment that allows accumulation through an INT adder tree, and 2D sub-word parallelism to share shifters across sub-multipliers. The authors report RTL synthesis and P&R results in a 65 nm library for the Jack unit and three ablation baselines, and an accelerator-level comparison against a RaPiD-like systolic array on five AI benchmarks using SCALE-sim and CACTI. The headline quantitative claims are 1.17-2.01x smaller area and 1.05-1.84x lower power for the MAC unit, and 1.32-5.41x energy-efficiency improvement at the accelerator level.

Significance. The hardware-efficiency contribution is credible if the numbers are correct: the paper uses a standard Synopsys synthesis and P&R flow, compares against a sequence of ablations that isolate the three proposed design techniques, and includes a full accelerator-level evaluation with cycle-level simulation and memory models. The 2D sub-word parallelism idea is clearly explained and the ablation sequence is well chosen. However, the numerical correctness of the FP/MX accumulation path is only demonstrated on a single layer of one CNN, and no end-to-end accuracy is reported for any of the five benchmarks or any supported format. Because "supports FP and MX data formats" is a central claim of the paper, the current evidence supports the hardware-efficiency story but not the usability of the FP/MX modes. The paper is potentially a solid systems contribution after the accuracy gap is filled and the reported numbers are reconciled.

major comments (4)
  1. [Sec. III-A2, footnote 3] The only numerical validation for the FP/MX accumulation path is a C simulation of the 2nd layer of ConvNeXt-T reporting 0.2% deviation from GPU results. The proposed scheme replaces the FP adder tree with an INT adder tree and skips intermediate normalization and rounding; its error behavior depends on activation exponent ranges, alignment shift widths, and accumulation lengths. Transformers, NeRF, and super-resolution models can differ in all three respects. Please report end-to-end task accuracy (e.g., top-1 accuracy, perplexity, PSNR/SSIM) for each benchmark and format, or at least a representative subset covering all five, and compare against the corresponding standard FP MAC. Without this, the claim that the unit "supports FP and MX data formats" is not established.
  2. [Sec. IV-B and Fig. 8] The energy-efficiency comparisons in bfloat16, FP8, and MX modes are reported with no accompanying accuracy results. Figure 8 shows only TFLOPS/W (or TOPS/W); the red- and blue-star comparisons are between different formats (Jack MXINT8 vs. baseline bfloat16; Jack MXFP8 vs. baseline FP8), so the energy gain is partly a precision/format change rather than a same-work comparison. Please provide per-mode accuracy results and, for cross-format comparisons, state the accuracy-equivalence assumption explicitly.
  3. [Sec. IV-A and Fig. 5] The absolute area and power values in the text do not match the figure. The text states that MAC-1 occupies approximately 11,084 um^2 and consumes 1.67 mW, while Fig. 5 reports 23,609 um^2 and 3.34 mW for the corresponding bar. Since the abstract's 1.17-2.01x area reduction and 1.05-1.84x power reduction claims are derived from these measurements, please correct the inconsistency and report one consistent set of P&R numbers for all four units.
  4. [Sec. III-C, footnote 4, and Sec. IV-B] The format definitions are inconsistent across the paper. Section III lists MXFP8 as {s:1, e:4, m:3}, while Section IV-B writes MXFP8 as {s:1, e shared:8, e local:4, m:7}; the latter notation is not the OCP MX format cited in reference [28]. In addition, footnote 4 states that a block size of 32 is "currently under standardization," but the cited OCP MX specification version 1.0 already defines a block size of 32. Please clarify the exact MX format and block-size compliance, because the bit widths directly affect both the area claims and the numerical accuracy of the MX modes.
minor comments (5)
  1. [Sec. II-A] The sentence describing the representable range of FP numbers is garbled: "|2^{B-1}, ((2-2^{-M})) x 2^B|" is not a standard range expression. Please rewrite it.
  2. [Fig. 3] The equation blocks in Fig. 3 have typographical inconsistencies in the subscripts: "MP1 x MQ1" appears twice and "MP0 x MQ0" is omitted. Please fix the subscripts.
  3. [Sec. III-A1, footnote 2] Footnote 2 reports a 42% layout overhead for non-pipelined MAC arrays but does not specify how this overhead is measured or whether it is included in the MAC-1 through MAC-3 numbers. Please clarify.
  4. [Table I] The row "Number of Multipliers" lists values such as 128x128 and 512x512; this appears to mean the number of MAC lanes or parallel multiply operations, not distinct multiplier instances. Please clarify the terminology.
  5. [Fig. 8] The energy-efficiency plot shows no per-benchmark breakdown or variability information, even though the text reports ranges such as 1.32-5.41x. Please consider showing per-benchmark points or a table with the individual values.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: area/power/energy claims are measured from implemented RTL/P&R designs, with one non-load-bearing self-citation and a thin single-layer FP-accuracy check that is a missing-evidence issue, not a circular one.

full rationale

The paper's quantitative claims (2.01x area reduction, 1.84x power reduction, 1.32-5.41x energy-efficiency gains) are obtained by synthesizing and place-and-routing the Jack unit and baselines in a commercial 65nm flow, with power estimated from switching activity and system-level energy from CACTI/SCALE-sim/HBM models. These are measurements against independently specified baselines, not quantities derived from the quantities they are said to predict. No parameter is fitted to a subset and then renamed a prediction, and no result is defined in terms of its own output. The one self-citation ([25] Flexblock) appears in a list of prior precision-scaling works and in related-work motivation; it is not invoked as the proof of any central design claim. Footnote 3's validation of the INT-adder FP accumulation on a single ConvNeXt-T layer is thin and would be a correctness-evidence concern for the FP/MX support claims, but it is external (C simulation vs. GPU) and not circular: the numerical behavior is checked, not assumed as an input. Accordingly, no circular step can be quoted or exhibited.

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

The central hardware-efficiency claim rests on domain assumptions about numerical accuracy, throughput equivalence, and representative EDA results. There are no fitted model parameters in the mathematical sense; the listed free parameters are engineering design choices. No new physical entities are introduced.

free parameters (3)
  • MX block size = 32
    Used in evaluation; the MX specification is still under standardization (footnote 4). The choice affects the shared exponent overhead and energy comparisons.
  • Jack unit output width = 16-bit (FP16/INT16)
    Chosen to avoid accumulation errors, following RaPiD [41]; affects area/power of adder tree and normalizer.
  • Pipeline stages = 2
    The accelerators use 2-stage pipelined Jack units to reach 400 MHz; this is a design choice affecting latency and area.
assumptions (3)
  • domain assumption FP accumulation via INT adder without intermediate normalization or rounding maintains numerical accuracy comparable to standard FP MAC
    Stated in Sec. III-A2 footnote; validated only on 2nd layer of ConvNeXt-T with 0.2% error, assumed for all benchmarks.
  • domain assumption A 32x32 array of Jack units provides the same throughput as a 128x128 baseline array in bfloat16/INT8 and as a 512x512 baseline array in FP8/INT4 modes
    Assumed in Table I and Sec. IV-B; the exact multiplier count and sub-word mapping are not fully explained.
  • domain assumption Synopsys synthesis and P&R at typical PVT corner (1.1V, 25C) on a 65nm library yields representative area/power/delay
    Standard methodology, single corner only (Sec. IV-A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Jack Unit: An Area- and Energy-Efficient Multiply-Accumulate (MAC) Unit Supporting Diverse Data Formats." pith.science (2026). https://pith.science/paper/2RXLKAEX

@misc{pith2026250704772,
  author       = {Pith},
  title        = {Pith review of: Jack Unit: An Area- and Energy-Efficient Multiply-Accumulate (MAC) Unit Supporting Diverse Data Formats},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2RXLKAEX}},
  note         = {Machine review of arXiv:2507.04772}
}
read the original abstract

In this work, we introduce an area- and energy-efficient multiply-accumulate (MAC) unit, named Jack unit, that is a jack-of-all-trades, supporting various data formats such as integer (INT), floating point (FP), and microscaling data format (MX). It provides bit-level flexibility and enhances hardware efficiency by i) replacing the carry-save multiplier (CSM) in the FP multiplier with a precision-scalable CSM, ii) performing the adjustment of significands based on the exponent differences within the CSM, and iii) utilizing 2D sub-word parallelism. To assess effectiveness, we implemented the layout of the Jack unit and three baseline MAC units. Additionally, we designed an AI accelerator equipped with our Jack units to compare with a state-of-the-art AI accelerator supporting various data formats. The proposed MAC unit occupies 1.17~2.01x smaller area and consumes 1.05~1.84x lower power compared to the baseline MAC units. On five AI benchmarks, the accelerator designed with our Jack units improves energy efficiency by 1.32~5.41x over the baseline across various data formats.

Figures

Figures reproduced from arXiv: 2507.04772 by the authors.

Figure 1
Figure 1. (a) Hardware costs, i.e., area and power-delay product (PDP), [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustrations showing how the precision-scalable CSM operates on (a) 8-bit [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Computational flow of FP addition, (b) MAC unit that adds two multiplication results via a FP adder, (c) MAC unit that performs the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (a) Overall structure of the proposed Jack unit, (b) Structural diagram of an exponent extractor. (c-f) Activation of sub-modules in the Jack [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: (a) Area and (b) power breakdown of MAC units. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Area breakdowns of the accelerator designed with Jack units and [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: (a) Inference latency of the JACK unit-based accelerator in bfloat16 and INT4 modes. (b) Comparison of compute density (i.e., area efficiency) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Average energy efficiency of the JACK unit-based accelerator (J) [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [28]

    Ocp microscaling formats (mx) specification version 1.0

    Open Computer Project. Ocp microscaling formats (mx) specification version 1.0. 2023

    work page 2023

  2. [1]

    Al-Ashrafy et al

    M. Al-Ashrafy et al. An efficient implementation of floating point multiplier. In Proc. of SIECPC , 2011

    work page 2011

  3. [2]

    Ethos-N78

    ARM. Ethos-N78. https://developer.arm.com/Processors/Ethos-N78, 2023

    work page 2023

  4. [3]

    Baktash et al

    J.-A. Baktash et al. Gpt-4: A review on advancements and opportunities in natural language processing. In arXiv:2305.03195, 2023

    arXiv 2023

  5. [4]

    Berger et al

    G. Berger et al. Quicksrnet: Plain single-image super-resolution archi- tecture for faster inference on mobile platforms. In Proc. of CVPR , 2023

    work page 2023

  6. [5]

    Findings of the 2014 workshop on statistical machine translation

    Ond ˇrej Bojar et al. Findings of the 2014 workshop on statistical machine translation. In Proc. of WMT , 2014

    work page 2014

  7. [6]

    Microscaling data formats for deep learning

    Bita Darvish Rouhani et al. Microscaling data formats for deep learning. In Proc. of NeurIPS , 2023

    work page 2023

  8. [7]

    Imagenet: A large-scale hierarchical image database

    Jia Deng et al. Imagenet: A large-scale hierarchical image database. In Proc. of CVPR , 2009

    work page 2009

Show all 43 references
  1. [8]

    Dettmers et al

    Z. Dettmers et al. Gpt3.int8(): 8-bit matrix multiplication for transform- ers at scale. In Proc. of NeurIPS , 2022

    2022

  2. [9]

    Devlin et al

    J. Devlin et al. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. of NAACL-HTL, 2019

    2019

  3. [10]

    Fowers et al

    J. Fowers et al. A configurable cloud-scale dnn processor for real-time ai. In Proc. of ISCA , 2018

    2018

  4. [11]

    Hickmann et al

    B. Hickmann et al. Intel nervana neural network processor-t (nnp-t) fused floating point many-term dot product. In Proc. of ARITH , 2020

    2020

  5. [12]

    Pirm challenge on perceptual image enhancement on smartphones: report

    Andrey Ignatov et al. Pirm challenge on perceptual image enhancement on smartphones: report. In Proc. of ECCV Workshop , 2019

    2019

  6. [13]

    Standard High Bandwidth Memory (HBM) DRAM, 2015

    JEDEC. Standard High Bandwidth Memory (HBM) DRAM, 2015

    2015

  7. [14]

    Jouppi et al

    N. Jouppi et al. Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings. In Proc. of ISCA, 2023

    2023

  8. [15]

    Kalamkar et al

    D. Kalamkar et al. A study of bfloat16 for deep learning training. In arXiv:1905.12322, 2019

    1905 arXiv

  9. [16]

    Kuzmin et al

    A. Kuzmin et al. Fp8 quantization: The power of the exponent. In Proc. of NeurIPS, 2022

    2022

  10. [17]

    Lee et al

    J. Lee et al. Unpu: A 50.6tops/w unified deep neural network accelerator with 1b-to-16b fully-variable weight bit-precision. In Proc. of ISSCC , 2018

    2018

  11. [18]

    Liu et al

    Z. Liu et al. A convnet for the 2020s. In Proc. of CVPR , 2022

    2022

  12. [19]

    Lo et al

    Y .-C. Lo et al. Bucket getter: A bucket-based processing engine for low-bit block floating point (bfp) dnns. In Proc. of MICRO, 2023

    2023

  13. [20]

    Pointer sentinel mixture models

    Stephen Merity et al. Pointer sentinel mixture models. 2016

    2016

  14. [21]

    Mildenhall et al

    B. Mildenhall et al. Nerf: representing scenes as neural radiance fields for view synthesis. In Proc. of ECCV , 2020

    2020

  15. [22]

    Ben Mildenhall et al. Nerf. https://www.matthewtancik.com/nerf, 2020

    2020

  16. [23]

    Art-mac: Ap- proximate rounding and truncation based mac unit for fault-tolerant applications

    Vishesh Mishra, Divy Pandey, Saurabh Singh, Sagar Satapathy, Kaustav Goswami, Babita Jajodia, and Dip Sankar Banerjee. Art-mac: Ap- proximate rounding and truncation based mac unit for fault-tolerant applications. In Proc. of ISCAS , 2022

    2022

  17. [24]

    Muralimanohar

    N. Muralimanohar. CACTI 6.0: A Tool to Model Large Caches. www. hpl.hp.com/techreports/2009/HPL-2009-85.pdf, 2009

    2009

  18. [25]

    Noh et al

    S.-H. Noh et al. Flexblock: A flexible dnn training accelerator with multi-mode block floating point support. In journal of TC , 2023

    2023

  19. [26]

    NVIDIA H100 Tensor Core GPU Architecture

    NVIDIA. NVIDIA H100 Tensor Core GPU Architecture. https://www.advancedclustering.com/wp-content/uploads/2022/03/ gtc22-whitepaper-hopper.pdf, 2022

    2022

  20. [27]

    Park et al

    J.-S. Park et al. A multi-mode 8k-mac hw-utilization-aware neural processing unit with a unified multi-precision datapath in 4nm flagship mobile soc. In Proc. of ISSCC , 2022

    2022

  21. [29]

    Language models are unsupervised multitask learners

    Alec Radford et al. Language models are unsupervised multitask learners. 2019

    2019

  22. [30]

    With shared microexponents, a little shifting goes a long way

    Darvish Rouhani et al. With shared microexponents, a little shifting goes a long way. In Proc. of ISCA , 2023

    2023

  23. [31]

    Ryu et al

    S. Ryu et al. Bitblade: Area and energy-efficient precision-scalable neural network accelerator with bitwise summation. In Proc. of DAC , 2019

    2019

  24. [32]

    Samajdar et al

    A. Samajdar et al. A systematic methodology for characterizing scalability of dnn accelerators using scale-sim. In Proc. of ISPASS , 2020

    2020

  25. [33]

    Sharify et al

    S. Sharify et al. Loom: Exploiting weight and activation precisions to accelerate convolutional neural networks. In Proc. of DAC, 2018

    2018

  26. [34]

    Sharma et al

    H. Sharma et al. Bit fusion: Bit-level dynamically composable architec- ture for accelerating deep neural network. In Proc. of ISCA , 2018

    2018

  27. [35]

    Sun et al

    X. Sun et al. Hybrid 8-bit floating point (hfp8) training and inference for deep neural networks. In Proc. of NeurIPS , 2019

    2019

  28. [36]

    IC Compiler II Implementation User Guide: Version L- 2016.03, 2016

    Synopsys. IC Compiler II Implementation User Guide: Version L- 2016.03, 2016

    2016

  29. [37]

    Design Compiler User Guide: Version N-2017.09, 2017

    Synopsys. Design Compiler User Guide: Version N-2017.09, 2017

    2017

  30. [38]

    Synopsys primetime px power analysis solution achieves broad market adoption

    Synopsys. Synopsys primetime px power analysis solution achieves broad market adoption. https://news.synopsys.com/home?item=123041, 2024

    2024

  31. [39]

    Lynx: A smartnic-driven accelerator-centric archi- tecture for network servers

    Maroun Tork et al. Lynx: A smartnic-driven accelerator-centric archi- tecture for network servers. In Proc. of ASPLOS , 2020

    2020

  32. [40]

    Touvron et al

    H. Touvron et al. Llama: Open and efficient foundation language models. In arXiv:2302.13971, 2023

    2023 arXiv

  33. [41]

    Venkataramani et al

    S. Venkataramani et al. Rapid: Ai accelerator for ultra-low precision training and inference. In Proc. of ISCA , 2021

    2021

  34. [42]

    Wortsman et al

    M. Wortsman et al. Stable and low-precision training for large-scale vision-language models. In Proc. of NeurIPS , 2024

    2024

  35. [43]

    Wu et al

    X. Wu et al. Understanding int4 quantization for language models: latency speedup, composability, and failure cases. In Proc. of ICML , 2023

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.