Pith. sign in

REVIEW 3 major objections 7 minor 27 references

Explicit Sign-Magnitude Encoders Enable Power-Efficient Multipliers

T0 review · 3 major / 7 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 Explicitly decomposing a 4-bit two's-complement multiplier into a sign-magnitude encoder plus a sign-magnitude multiplier cuts switching activity by 12.9% for zero-centered inputs, while keeping the circuit logic-equivalent.

desk verdict A clean, honest synthesis-level study; the 12.9% switching-activity drop is plausible, but the 'power-efficient' title outruns a zero-delay toggle proxy. read the letter →

arxiv 2507.18179 v1 pith:SE7C5HPF submitted 2025-07-24 cs.NE cs.ARcs.PF

Felix Arnold , Maxence Bouvier , Ryan Amaudruz , Renzo Andri , Lukas Cavigelli This is my paper · ORCID
classification cs.NEcs.ARcs.PF
keywords SwitchingactivitySign-magnitudeencodingTwo'scomplementMultiplieroptimizationFixed-pointarithmeticLogicsynthesisLow-powerdigitaldesignspaceexploration
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

This paper tries to establish that how a signed integer is encoded inside a multiplier matters for power, and that standard synthesis tools do not find the better encoding on their own. By explicitly wrapping a 4-bit two's-complement multiplier with sign-magnitude encoders on both input operands, and synthesizing each block separately, the overall circuit stays functionally identical but toggles less for the zero-centered value distributions typical of AI workloads. The headline result is a 12.9% reduction in switching activity at a standard deviation of 3.0 for the logic-equivalent configuration; allowing a slightly reduced input range raises the gain to 33%, and an all-sign-magnitude variant reaches 68–75%. If the switching-activity proxy holds in physical design, this is a drop-in circuit transformation for low-precision multiplier arrays.

What carries the argument

The load-bearing mechanism is the explicit decomposition of a two's-complement multiplier into per-input encoders and a sign-magnitude multiplier, with each block synthesized and optimized in isolation so the electronic design automation tool cannot re-merge them. Sign-magnitude separates the sign bit from a 3-bit unsigned magnitude, so the multiplier core shrinks to a 3×3 unsigned multiply plus a sign XOR, and values near zero toggle few magnitude bits; the extended sign-magnitude variant (SME) preserves the −8 code. Switching activity is estimated as each wire's toggles weighted by the transistor count of its fan-out cells, summed over all wires, and the paper models a two's-complement-in/two's-complement-out design as $s_{\mathrm{tot}} = 2 s_{\mathrm{enc}} + s_{\mathrm{mult}}$.

What would settle it

Take the baseline and decomposed multipliers, run both through placement and routing with timing-aware cell delays, apply the same normally distributed zero-centered stimulus (σ=3), and measure physical dynamic power including glitches; the claim predicts the decomposed circuit's measured energy is lower, and by roughly the size of the reported switching-activity gap.

Watch

Extended reading notes

Core claim

The central claim is that a 4-bit two's-complement multiplier, which standard synthesis maps to a single preconceived multiplier macro, can be explicitly decomposed into input encoders that convert operands to sign-magnitude (with an extended variant preserving the −8 code) and a sign-magnitude multiplier that returns a two's-complement result, and that synthesizing each block separately preserves logic equivalence while lowering switching activity for zero-centered inputs. With inputs drawn from a normal distribution with σ=3, the numerically equivalent configuration reduces switching activity by 12.9% relative to the baseline (24.5% at σ=2); if the asymmetric −8 value is clipped away, the reduction reaches 33% at σ=3 and 42.2% at σ=2, and an end-to-end sign-magnitude design reaches 68–75%. The paper further claims that a random design-space exploration guided by switching activity finds circuits with 5–10% lower switching activity than exploration guided only by transistor count, and that this is a general lever for power-aware synthesis.

Load-bearing premise

The result rests on assuming that counting how often each wire changes value, weighted by the size of the logic cells it feeds, accurately predicts real chip power use; if glitches, wire capacitance, or leakage change the ordering, the reported savings vanish in silicon.

Editorial extensions

If this is right

  • A logic-equivalent 4-bit multiplier can cut switching activity by 12.9% at σ=3 and 24.5% at σ=2 without changing the two's-complement interface, so existing multiply-accumulate data paths could adopt it as a drop-in swap.
  • Gains grow as inputs concentrate near zero: clipping or excluding the asymmetric −8 value yields up to 33% at σ=3 and 42.2% at σ=2 (configurations C and D).
  • An end-to-end sign-magnitude pipeline (configuration E) reduces switching activity by 68–75% at σ=2–3, provided the system can store and consume sign-magnitude values.
  • Synthesis optimization that selects circuits by switching activity rather than transistor count adds a further 5–10% reduction, so power-aware design-space exploration is a complementary lever.
  • The decomposed designs add 16.7–41.7% logic depth, so meeting a target clock may require pipelining at the encoder output rather than a direct swap.

Reading between the lines

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

  • If the switching-activity ordering survives place-and-route, the same decomposition should transfer to the larger multiplier widths the authors mention testing internally, making it a candidate for low-precision MAC arrays in AI accelerators.
  • The σ-dependence implies a co-design rule the paper does not spell out: quantizing weights and activations to a symmetric input range is what unlocks the largest gains, so the encoding choice and the quantization scheme should be set together.
  • A further implicit consequence is that transistor count is a poor proxy for dynamic power in these circuits; power-aware search over graph representations is the lever that closes the gap.
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

3 major / 7 minor

Summary. The paper proposes decomposing a 4-bit two's-complement multiplier into a two's-complement-to-sign-magnitude encoder followed by an SME-to-TC multiplier, synthesizing each block separately, and claims this reduces switching activity for zero-centered input distributions while preserving functional equivalence for the numerically equivalent configuration B. Using a cost-weighted toggle-count SwAct metric from zero-delay post-synthesis simulation, configuration B achieves 12.9% lower SwAct than the baseline at sigma = 3, and configurations with a reduced input range achieve larger reductions (up to 75.5% at sigma = 2 for the pure SM configuration). The paper also claims a guided random-search synthesis framework gives a further 5-10% improvement when selecting for switching activity. The authors verify all circuits over the full input space and list several limitations, including the absence of timing analysis and place-and-route.

Significance. If the switching-activity reductions survive physical design, the decomposition is an attractive, drop-in transformation for AI workloads with zero-centered data. The paper's strengths are the full input-space functional verification, the reproducible open-source synthesis flow, the explicit enumeration of limitations, and the large design-space exploration dataset. However, the evidence for the central 'power-efficient' claim is a simplified toggle proxy, and the headline sigma = 3 number is an in-sample selection result. The qualitative direction of the result is plausible and partly supported by out-of-sample sigma = 2 and sigma = 4 columns, but the quantitative claims outrun the evidence as presented.

major comments (3)
  1. [Section IV-C, Table II] The final circuits used for the results in Table II were selected as the best-SwAct circuits at sigma = 3 across 200 random-synthesis runs (Section IV-C states that 'the best SwAct (for sigma = 3) determines the final circuit selected across all runs'). Consequently, the headline 12.9% reduction for configuration B at sigma = 3 is an in-sample optimum and is expected to overstate the true gain. The fact that the sigma = 2 (-24.5%) and sigma = 4 (-4.0%) columns are out-of-sample and still favor B is reassuring, but the quantitative headline should be re-estimated with a selection rule that is independent of the evaluation metric (e.g., select on transistor count or area, or use a hold-out input stream), and the in-sample nature of the sigma = 3 column should be explicitly stated.
  2. [Section IV-B, Section VII-3/4] The abstract and title claim 'power-efficient multipliers' and 'power savings', but the only evidence is the SwAct proxy: a zero-delay, post-synthesis simulation in which each wire's toggles are weighted by the transistor count of its fan-out cells, output wires receive zero cost, and glitches, wire capacitance, cell-internal power, and leakage are not modeled. Limitations 3 and 4 explicitly acknowledge this. A 12.9% reduction in this proxy does not establish a 12.9% reduction in physical power, especially because configuration B increases cell-level depth by 41.7% and is synthesized as separate blocks with no timing closure across the encoder/multiplier boundary (Table III). The authors should either add timing-aware or post-layout power measurements for at least configurations A and B, or rephrase the claims as switching-activity reductions only.
  3. [Section VI, Fig. 7] The second headline claim, that switching-activity-driven DSE yields a further 5-10% improvement, is derived by comparing the min-SwAct circuit with the min-transistor circuit within the same set of generated circuits (arrows in Fig. 7). Since SwAct is exactly the metric used to select the final point, the minimum-SwAct point is expected to be below the minimum-transistor point by construction, and this gap does not measure the value of power-aware guidance. To support the claim, evaluate the selected circuits on a hold-out metric (e.g., a different input distribution or a physical power estimator) or compare the distribution of SwAct for circuits generated with different guidance rules.
minor comments (7)
  1. [Table II] The totals for rows B and D are inconsistent with the stated formula stot = 2*senc + smult (e.g., B at sigma = 2: 2*35 + 173 = 243, but the table reports 244; D at sigma = 3: 2*35 + 153 = 223, but the table reports 224). Recompute or note rounding.
  2. [Abstract/Conclusion vs Table II] The 'up to' values for the SM-domain configuration are inconsistent: Table II reports 75.5% at sigma = 2 for configuration E, while the Conclusion says 'up to 68.8%'; the abstract also mentions 33% for the reduced-range TC case, which should be tied to a specific sigma. Harmonize these numbers.
  3. [Section IV-A] The metric called 'number of transistors' is Yosys's default transistor-count model, not a physical transistor count; use 'estimated transistor count' consistently throughout the text and figures.
  4. [Section IV-A] The definition of depth should be tightened: the current wording says depth is 'directly related but not necessarily proportional' to delay, but the text immediately states cell delays are ignored; make explicit that depth is a cell-count depth with no timing model.
  5. [Section IV-B] The statement that output wires do not contribute to SwAct because they have no attached cells should be justified, since in a real datapath the multiplier outputs drive subsequent logic; at minimum, note this as an assumption that may affect comparisons across configurations.
  6. [Section VII-1] The claim that 'internal evaluation of configurations A and B with 8-bit input operands tend to indicate that the results are reproducible' is not supported by any reported data; either provide the data or remove the claim.
  7. [References] Reference [16] appears to be a non-peer-reviewed blog post; consider citing a primary or peer-reviewed source for the claim that LLM weights approximate a normal distribution.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SwAct comparisons are empirical post-synthesis measurements under a stated proxy model, not derivations from the model's own assumptions.

full rationale

The paper's claimed derivation chain is self-contained rather than circular. The central comparison is between measured post-synthesis switching activities for different RTL decompositions; SwAct is explicitly defined in Sections IV-A and IV-B as a cost-weighted toggle count under a zero-delay model, and the reported 12.9% figure is a simulation output, not a quantity forced by the definition. No fitted parameter is renamed as a prediction, and no equation reduces to its own inputs. The only self-citation, [3], appears in Sections I and IV-C as an inspiration for the random design-space exploration, but the optimization procedure is fully specified in the text and the framework is attributed to EPFL's [14]; it is therefore not load-bearing. Section IV-C's statement that 'the best SwAct (for σ = 3) determines the final circuit selected across all runs' makes the σ=3 headline in-sample, which is a selection-bias caveat; however, the σ=2 and σ=4 columns in Table II and the higher-effort runs in Section VI provide out-of-sample support, so the result is not definitionally forced. The paper's own limitations, especially Section VII items 3 and 4, concede that the SwAct proxy omits glitches, wire capacitance, and place-and-route effects, and Section V-B admits that combined synthesis was not run due to time constraints; these are validity threats to the physical power claim, but they are not circularity. The unsupported 8-bit extrapolation in Section VII.1 is an outlook assertion, not a load-bearing derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central result rests on the SwAct-as-power proxy and the assumed zero-centered input distribution; both are chosen rather than measured, so they count against the ledger. No new physical entities are introduced.

free parameters (4)
  • Input standard deviation sigma = 2.0, 3.0, 4.0 (sigma=3 as realistic case)
    Inputs are sampled from zero-mean normal distributions and clipped. The reported savings depend heavily on this chosen workload model; it is not fitted to data.
  • SwAct wire cost weight = Number of transistors in fan-out cells, using Yosys default model
    This cost model defines the metric being optimized and reported. The claimed power savings rest on this model rather than on measured energy.
  • Simulation length = 10,000 cycles
    One deterministic evaluation run per circuit is reported, with no variance or seed analysis.
  • Random-search effort and selection rule = 200 runs, 10 iterations of 20 steps; final selection by lowest SwAct at sigma=3
    Compute budget and selection policy. Table II numbers are the best-found circuits, making the sigma=3 column an in-sample optimum.
assumptions (5)
  • domain assumption Dynamic power dominates total power under high utilization.
    Section IV-B: SwAct is said to be directly related to power only if dynamic power dominates; no leakage or static power is modeled.
  • domain assumption Yosys cell transistor-count estimates are adequate proxies for area and fan-out capacitance.
    Section IV-A: transistor count defines both the area metric and the SwAct weighting; it is technology-dependent and not validated against a PDK.
  • domain assumption Zero-delay cycle-level simulation captures the relevant switching behavior.
    Section IV-B and Limitations 4: glitches, races, and timing effects are not simulated, and no place-and-route is performed.
  • domain assumption AI weight distributions are approximately zero-centered normals with sigma around 2 to 4.
    Section III-C: motivates the test distribution; the headline savings are computed for this assumed workload.
  • domain assumption The synthesized star-operator multiplier is the right baseline and EDA tools cannot reach the decomposition on their own.
    Section IV-C: Yosys is observed to map '*' to a Booth macro; the paper assumes this is representative of EDA behavior and that manual decomposition is needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explicit Sign-Magnitude Encoders Enable Power-Efficient Multipliers." pith.science (2026). https://pith.science/paper/SE7C5HPF

@misc{pith2026250718179,
  author       = {Pith},
  title        = {Pith review of: Explicit Sign-Magnitude Encoders Enable Power-Efficient Multipliers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SE7C5HPF}},
  note         = {Machine review of arXiv:2507.18179}
}
read the original abstract

This work presents a method to maximize power-efficiency of fixed point multiplier units by decomposing them into sub-components. First, an encoder block converts the operands from a two's complement to a sign magnitude representation, followed by a multiplier module which performs the compute operation and outputs the resulting value in the original format. This allows to leverage the power-efficiency of the Sign Magnitude encoding for the multiplication. To ensure the computing format is not altered, those two components are synthesized and optimized separately. Our method leads to significant power savings for input values centered around zero, as commonly encountered in AI workloads. Under a realistic input stream with values normally distributed with a standard deviation of 3.0, post-synthesis simulations of the 4-bit multiplier design show up to 12.9% lower switching activity compared to synthesis without decomposition. Those gains are achieved while ensuring compliance into any production-ready system as the overall circuit stays logic-equivalent. With the compliance lifted and a slightly smaller input range of -7 to +7, switching activity reductions can reach up to 33%. Additionally, we demonstrate that synthesis optimization methods based on switching-activity-driven design space exploration can yield a further 5-10% improvement in power-efficiency compared to a power agnostic approach.

Figures

Figures reproduced from arXiv: 2507.18179 by the authors.

Figure 1
Figure 1. Decomposition and optimization process utilizing a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution of input and output values for a 4-bit two’s complement multiplication [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the proposed switching activity model [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: High-level flow diagram 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 Chain Index 400 450 500 550 600 650 700 750 800 Number of Transistors 0 1 2 3 4 5 6 7 8 9 Iteration (a) TC→TC multiplier circuit 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 0 10 Chain Index 400 …
Figure 5
Figure 5. Figure 5: Optimization traces for 200 runs overlaid [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Transistor count and switching activity for different [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Transistor count and switching activity ( [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages

  1. [21]

    Datawidth- aware energy-efficient multipliers: A case for going sign magnitude

    Luc Waeijen, Hailong Jiao, Henk Corporaal, and Yifan He. Datawidth- aware energy-efficient multipliers: A case for going sign magnitude. In 2018 21st Euromicro Conference on Digital System Design (DSD), pages 54–61, 2018

    work page 2018

  2. [1]

    NVIDIA Blackwell Architecture Technical Overview

    work page

  3. [2]

    Low-power approximate multipliers using encoded partial products and approximate compressors

    Mohammad Saeed Ansari, Honglan Jiang, Bruce Cockburn, and Jie Han. Low-power approximate multipliers using encoded partial products and approximate compressors. IEEE Journal on Emerging and Selected Topics in Circuits and Systems , PP:1–1, 05 2018

    work page 2018

  4. [3]

    Late breaking results: The art of beating the odds with predictor-guided random design space exploration, 2025

    Felix Arnold, Maxence Bouvier, Ryan Amaudruz, Renzo Andri, and Lukas Cavigelli. Late breaking results: The art of beating the odds with predictor-guided random design space exploration, 2025

    work page 2025

  5. [4]

    Quarot: Outlier-free 4-bit inference in rotated llms

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms. Ad- vances in Neural Information Processing Systems , 37:100213–100240, 2024

    work page 2024

  6. [5]

    Andrew D. Booth. A signed binary multiplication technique. The Quarterly Journal of Mechanics and Applied Mathematics , 4(2):236– 240, 01 1951

    work page 1951

  7. [6]

    A low-power multiplier with the spurious power suppression technique

    Kuan-Hung Chen and Yuan-Sun Chu. A low-power multiplier with the spurious power suppression technique. IEEE Trans. Very Large Scale Integr. Syst., 15(7):846–850, July 2007

    work page 2007

  8. [7]

    Brain floating point (bfloat16), 2025

    Google Cloud. Brain floating point (bfloat16), 2025. Accessed: 2025- 03-20

    work page 2025

Show all 27 references
  1. [8]

    The case for 4-bit precision: k- bit inference scaling laws

    Tim Dettmers and Luke Zettlemoyer. The case for 4-bit precision: k- bit inference scaling laws. In International Conference on Machine Learning, pages 7750–7774. PMLR, 2023

    2023

  2. [9]

    Gomarl: Global optimization of multiplier using multi-agent reinforcement learning

    Yi Feng and Chao Wang. Gomarl: Global optimization of multiplier using multi-agent reinforcement learning. In 2024 2nd International Symposium of Electronics Design Automation (ISEDA) , pages 728–733. IEEE, 2024

    2024

  3. [10]

    Bias by design: Diversity quantification to mitigate structural bias effects in aig logic optimization

    Isabella Venancia Gardner, Marcel Walter, Yukio Miyasaka, Robert Wille, and Michael Cochez. Bias by design: Diversity quantification to mitigate structural bias effects in aig logic optimization. In 2025 Design, Automation & Test in Europe Conference & Exhibition (DATE). In press

    2025

  4. [11]

    Low power combinational multipliers using data-driven signal gating

    Nima Honarmand and Ali Afzali-Kusha. Low power combinational multipliers using data-driven signal gating. In APCCAS 2006 - 2006 IEEE Asia Pacific Conference on Circuits and Systems , pages 1430– 1433, 2006

    2006

  5. [12]

    Top-Down Digital VLSI Design, From Architectures to Gate-Level Circuits and FPGAs

    Hubert Kaeslin. Top-Down Digital VLSI Design, From Architectures to Gate-Level Circuits and FPGAs. Morgan Kaufmann Publishers, Boston, 12 2014

    2014

  6. [13]

    Switching in multipliers

    Jakub Jerzy Kalis. Switching in multipliers. Master’s thesis, Norwegian University of Science and Technology, Trondheim, Norway, June 2009. Available at https://ntnuopen.ntnu.no/ntnu-xmlui/bitstream/handle/ 11250/2369439/348846 FULLTEXT01.pdf?sequence=1&isAllowed=y

    2009

  7. [14]

    Late breaking results: Majority-inverter graph minimization by design space exploration

    Siang-Yun Lee, Alessandro Tempia Calvino, Heinz Riener, and Giovanni De Micheli. Late breaking results: Majority-inverter graph minimization by design space exploration. In Proceedings of the 61st ACM/IEEE Design Automation Conference , DAC ’24, New York, NY , USA, 2024. Assoc...

    2024

  8. [15]

    Ascend hifloat8 format for deep learning, 2024

    Yuanyong Luo, Zhongxing Zhang, Richard Wu, Hu Liu, Ying Jin, Kai Zheng, Minmin Wang, Zhanying He, Guipeng Hu, Luyao Chen, Tianchi Hu, Junsong Wang, Minqi Chen, Mikhaylov Dmitry, Korviakov Vladimir, Bobrin Maxim, Yuhao Hu, Guanfu Chen, and Zeyi Huang. Ascend hifloat8 format for...

    2024

  9. [16]

    Basic facts about language model internals, January 2023

    Beren Millidge and Eric Winsor. Basic facts about language model internals, January 2023. Accessed: 2025-04-04

    2023

  10. [17]

    Weighted-entropy- based quantization for deep neural networks

    Eunhyeok Park, Junwhan Ahn, and Sungjoo Yoo. Weighted-entropy- based quantization for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    2017

  11. [18]

    Deep-powerx: A deep learning-based framework for low-power approximate logic synthesis

    Ghasem Pasandi, Mackenzie Peterson, Moises Herrera, Shahin Nazarian, and Massoud Pedram. Deep-powerx: A deep learning-based framework for low-power approximate logic synthesis. In Proceedings of the ACM/IEEE International Symposium on Low Power Electronics and Design, pages 73...

    2020

  12. [19]

    The EPFL logic synthesis libraries, June 2022

    Mathias Soeken, Heinz Riener, Winston Haaswijk, Eleonora Testa, Bruno Schmitt, Giulia Meuli, Fereshte Mozafari, Siang-Yun Lee, Alessandro Tempia Calvino, and Giovanni Marakkalage, Dewmini Su- dara De Micheli. The EPFL logic synthesis libraries, June 2022. arXiv:1805.05121v3

    2022 arXiv

  13. [20]

    Tang and Joy Alinda Reyes

    Alvin Joseph J. Tang and Joy Alinda Reyes. Comparative analysis of low power multiplier architectures. In 2011 Fifth Asia Modelling Symposium, pages 270–274, 2011

    2011

  14. [22]

    C. S. Wallace. A suggestion for a fast multiplier. IEEE Transactions on Electronic Computers, EC-13(1):14–17, 1964

    1964

  15. [23]

    A hierarchical adaptive multi-task reinforcement learning framework for multiplier circuit design

    Zhihai Wang, Jie Wang, Dongsheng Zuo, Ji Yunjie, Xilin Xia, Yuzhe Ma, Jianye Hao, Mingxuan Yuan, Yongdong Zhang, and Feng Wu. A hierarchical adaptive multi-task reinforcement learning framework for multiplier circuit design. In Forty-first International Conference on Machine L...

    2024

  16. [24]

    Yosys open synthesis suite

    Claire Wolf. Yosys open synthesis suite. https://yosyshq.net/yosys/

  17. [25]

    Domac: Differentiable op- timization for high-speed multipliers and multiply-accumulators

    Chenhao Xue, Yi Ren, Jinwei Zhou, Kezhi Li, Chen Zhang, Yibo Lin, Lining Zhang, Qiang Xu, and Guangyu Sun. Domac: Differentiable op- timization for high-speed multipliers and multiply-accumulators. arXiv preprint arXiv:2503.23943, 2025

    2025 arXiv

  18. [26]

    High-speed booth encoded parallel multiplier design

    Wen-Chang Yeh and Chein-Wei Jen. High-speed booth encoded parallel multiplier design. IEEE Transactions on Computers , 49(7):692–701, 2000

    2000

  19. [27]

    Rl-mul: Multiplier design optimization with deep reinforcement learning

    Dongsheng Zuo, Yikang Ouyang, and Yuzhe Ma. Rl-mul: Multiplier design optimization with deep reinforcement learning. In 2023 60th ACM/IEEE Design Automation Conference (DAC) , pages 1–6. IEEE, 2023. 9

    2023

Pith tools

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