pith. sign in

arxiv: 2604.12420 · v1 · submitted 2026-04-14 · 💻 cs.AR

HARP: Hadamard-Domain Write-and-Verify for Noise-Robust RRAM Programming

Ilhuan Choi , Jiwon Yoo , Yoona Lee , Yewon Jeong , Jason Jaesung Lee , Woo-Seok Choi This is my paper

Pith reviewed 2026-05-10 14:40 UTC · model grok-4.3

classification 💻 cs.AR
keywords RRAMwrite-and-verifyHadamard transformread noiseparallel verifyenergy efficiencyneural network accelerators
0
0 comments X

The pith

Hadamard-domain write-and-verify reduces read noise variance by a factor of N in RRAM columns without extra analog hardware.

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

The paper proposes replacing conventional one-hot verify reads in RRAM write-and-verify programming with N orthogonal Hadamard patterns applied to an N-cell column. Inverse Hadamard decoding then averages uncorrelated read noise by a factor of N and cancels common-mode disturbances. This approach keeps neural network accuracy high on tasks like CIFAR-10 even under severe noise, while using less latency and energy than repeated conventional reads for the same noise reduction.

Core claim

HD-PV performs N parallel verify reads using orthogonal Hadamard patterns on each column instead of sequential one-hot reads, then applies inverse Hadamard decoding to cut uncorrelated noise variance by N and remove common-mode effects. HARP further replaces full SAR ADC conversions with lightweight compare-only operations because only ternary update decisions are required during write-and-verify. The result maintains the original memory footprint while delivering accuracy within 1% of ideal under noise levels that cause over 20% loss with standard methods.

What carries the argument

Hadamard-Encoded Parallel-Verify (HD-PV), which substitutes orthogonal Hadamard read patterns for one-hot reads so that inverse decoding suppresses noise without added analog blocks.

If this is right

  • Conventional write-and-verify loses over 20% accuracy on CIFAR-10 under severe read noise.
  • HD-PV limits the accuracy loss to 0.6% and HARP to 1% on CIFAR-10, CIFAR-100, and keyword spotting under the same memory footprint.
  • HD-PV achieves up to 6.1x lower latency and 6.2x better energy efficiency than multi-read averaging for noise reduction.
  • HARP achieves 3.5x lower latency and 9.5x better energy efficiency than multi-read averaging while preserving accuracy.

Where Pith is reading between the lines

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

  • The basis-change technique may extend to other crossbar-based analog memories that face column-wise read noise.
  • It shows that measurement-basis redesign can improve signal quality in existing hardware without front-end redesign.

Load-bearing premise

Read noise is uncorrelated across cells in a column and common-mode disturbances can be fully canceled by inverse Hadamard decoding without introducing new error sources or requiring changes to the analog front-end.

What would settle it

A measurement showing that post-decoding noise variance in an RRAM column is not reduced by a factor close to N, or an accuracy test on CIFAR-10 where HD-PV or HARP loses more than 1% under the paper's severe read-noise model.

read the original abstract

Write-and-verify (WV) is essential for programming multi-level RRAM weights, yet under scaled-voltage and low-SNR conditions the verify read increasingly limits mapping accuracy, convergence speed and energy. We propose a Hadamard-domain WV framework that improves verify reliability without adding analog hardware. % without introducing additional analog blocks % while leveraging the existing analog front-end \emph{HD-PV} (Hadamard-Encoded Parallel-Verify) replaces conventional one-hot verify reads with $N$ orthogonal Hadamard patterns for an $N$-cell column. Changing the read basis without increasing the column-level read count, inverse Hadamard decoding reduces uncorrelated read-noise variance by a factor of $N$ and cancels common-mode disturbances. \emph{HARP} (Hadamard-based ADC-Energy-Reduced Parallel-Verify) further exploits the fact that WV needs only ternary update decisions, not full digital codes, and replaces SAR conversions with lightweight compare-only operations. Across CIFAR-10, CIFAR-100, and keyword spotting under severe read noise, conventional WV loses over 20\,\% accuracy on CIFAR-10, while HD-PV and HARP limit the loss to 0.6\,\% and 1\,\% under the same memory footprint. Compared to conventional multi-read averaging for noise reduction, HD-PV and HARP achieve comparable accuracy with up to $6.1\times$ and $3.5\times$ lower latency and $6.2\times$ and $9.5\times$ better energy efficiency, respectively. To the best of our knowledge, this is the first application of Hadamard-encoded verification to RRAM WV.

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

1 major / 1 minor

Summary. The paper proposes HARP, a Hadamard-domain write-and-verify (WV) framework for multi-level RRAM programming that improves noise robustness without additional analog hardware. HD-PV replaces conventional one-hot verify reads with N orthogonal Hadamard patterns applied to an N-cell column, followed by inverse decoding that reduces uncorrelated read-noise variance by N and cancels common-mode terms. HARP further replaces full SAR ADC conversions with lightweight compare-only operations to obtain only the ternary update decisions required by WV. On CIFAR-10, CIFAR-100, and keyword-spotting benchmarks under severe read noise, the methods limit accuracy loss to 0.6% and 1% (versus >20% for conventional WV) at the same memory footprint, while achieving up to 6.1× lower latency and 9.5× better energy efficiency than multi-read averaging.

Significance. If the central claims hold under realistic device conditions, the work offers a mathematically grounded, hardware-compatible approach to mitigating read noise in RRAM-based analog accelerators for machine learning. By exploiting the orthogonality of Hadamard matrices to average noise without increasing read count or modifying the analog front-end, it could meaningfully improve mapping accuracy, convergence speed, and energy efficiency in scaled or low-SNR RRAM arrays. The reported quantitative gains on standard benchmarks are substantial and, if reproducible, would strengthen the case for RRAM in noisy edge-computing environments.

major comments (1)
  1. [Abstract / HD-PV description] The noise-variance reduction by exactly N and exact common-mode cancellation in HD-PV (Abstract and HD-PV framework description) rests on the assumption that read noise is spatially uncorrelated across cells in a column and that no column-shared disturbances (row-driver noise, IR drop, temperature gradients) survive the inverse transform. This assumption is load-bearing for the reported accuracy retention (0.6% loss on CIFAR-10) and for the ternary decisions in HARP, yet the manuscript provides no analysis, Monte-Carlo simulations, or measurements of non-ideal column-shared effects that would violate orthogonality or linearity in the existing sense amplifiers.
minor comments (1)
  1. [Abstract] The abstract states that the approach requires 'no analog-front-end changes,' but the inverse decoding step implicitly requires that the existing sense amplifiers maintain sufficient linearity and that pattern application preserves exact orthogonality; a short clarification of these interface assumptions would improve clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and the recommendation for major revision. The concern about unanalyzed non-ideal column-shared effects is well-taken, and we will strengthen the manuscript accordingly while defending the core mathematical claims under the stated assumptions.

read point-by-point responses

  1. Referee: [Abstract / HD-PV description] The noise-variance reduction by exactly N and exact common-mode cancellation in HD-PV (Abstract and HD-PV framework description) rests on the assumption that read noise is spatially uncorrelated across cells in a column and that no column-shared disturbances (row-driver noise, IR drop, temperature gradients) survive the inverse transform. This assumption is load-bearing for the reported accuracy retention (0.6% loss on CIFAR-10) and for the ternary decisions in HARP, yet the manuscript provides no analysis, Monte-Carlo simulations, or measurements of non-ideal column-shared effects that would violate orthogonality or linearity in the existing sense amplifiers.

    Authors: We agree that the exact factor-of-N variance reduction requires spatially uncorrelated read noise, which follows directly from the properties of the (normalized) Hadamard matrix: each decoded cell is a linear combination of N independent noise terms with coefficients of magnitude 1/sqrt(N). Common-mode cancellation is likewise exact for any additive term identical across the N cells, because such a term is projected entirely onto the DC basis vector of the Hadamard transform and does not appear in the differential decoded values used for HARP's ternary decisions. This holds as long as the disturbance remains additive and common during the N reads, independent of sense-amplifier linearity. We acknowledge, however, that real column-shared effects (IR drop, temperature gradients, row-driver noise) are rarely perfectly uniform and may introduce spatially varying or non-linear components that do not cancel completely. The current manuscript states the ideal-case assumptions but provides no Monte-Carlo analysis or robustness study of these non-idealities. In the revised version we will add a dedicated subsection and figure with Monte-Carlo simulations that inject realistic column-shared noise models (common-mode Gaussian, linear IR-drop gradients, and temperature variation) while keeping the same RRAM device parameters and ML benchmarks. These simulations will quantify any degradation in effective noise reduction and end-to-end accuracy, allowing readers to assess the practical margin of the reported 0.6% loss figure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; noise reduction follows from external Hadamard orthogonality

full rationale

The paper applies the standard mathematical property of Hadamard matrix orthogonality to achieve variance reduction by N and common-mode cancellation in the verify reads. This is not derived within the paper, fitted from data, or dependent on self-citations; it is a known linear-algebra fact. No equations reduce to self-definition, no fitted parameters are relabeled as predictions, and the accuracy results are simulation outcomes under explicit assumptions rather than tautological. The 'first application' statement is a novelty claim, not a load-bearing derivation step.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on the mathematical orthogonality of Hadamard matrices for noise averaging and on domain assumptions about RRAM read noise being uncorrelated and common-mode; no explicit free parameters are fitted in the abstract.

axioms (2)
  • standard math Hadamard matrices are orthogonal, enabling inverse decoding that reduces uncorrelated noise variance by a factor of N
    Invoked to justify the HD-PV noise reduction claim
  • domain assumption RRAM column read noise is uncorrelated across cells and common-mode disturbances are present
    Required for the variance reduction and cancellation benefits to hold
invented entities (1)
  • HD-PV (Hadamard-Encoded Parallel-Verify) and HARP frameworks no independent evidence
    purpose: Replace conventional one-hot verify reads with orthogonal patterns and lightweight compare-only operations for noise-robust RRAM WV
    New proposed techniques introduced in the paper

pith-pipeline@v0.9.0 · 5631 in / 1562 out tokens · 35146 ms · 2026-05-10T14:40:28.778355+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    Tanner Andrulis, Joel S Emer, and Vivienne Sze. 2023. Raella: Reforming the arithmetic for efficient, low-resolution, and low-loss analog pim: No retraining required!. InProceedings of the 50th Annual International Symposium on Computer Architecture. 1–16

    work page 2023

  2. [2]

    Pai-Yu Chen, Xiaochen Peng, and Shimeng Yu. 2018. NeuroSim: A circuit-level macro model for benchmarking neuro-inspired architectures in online learning. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 37, 12 (2018), 3067–3080

    work page 2018

  3. [3]

    Justin M Correll, Lu Jie, Seungheun Song, Seungjong Lee, Junkang Zhu, Wei Tang, Luke Wormald, Jack Erhardt, Nicolas Breil, Roger Quon, et al. 2025. An 8-bit 20.7 TOPS/W Multilevel Cell ReRAM Macro With ADC-Assisted Bit-Serial Processing.IEEE Journal of Solid-State Circuits(2025)

    work page 2025

  4. [4]

    Sujan K Gonugondla, Ameya D Patil, and Naresh R Shanbhag. 2020. Swipe: En- hancing robustness of reram crossbars for in-memory computing. InProceedings of the 39th International Conference on Computer-Aided Design. 1–9

    work page 2020

  5. [5]

    2012.Hadamard transform optics

    Martin Harwit. 2012.Hadamard transform optics. Elsevier

    work page 2012

  6. [6]

    Wangxin He, Jian Meng, Sujan Kumar Gonugondla, Shimeng Yu, Naresh R Shanbhag, and Jae-sun Seo. 2023. Prive: efficient RRAM programming with chip verification for RRAM-based in-memory computing acceleration. In2023 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 1–6

    work page 2023

  7. [7]

    Zhezhi He, Jie Lin, Rickard Ewetz, Jiann-Shiun Yuan, and Deliang Fan. 2019. Noise injection adaption: End-to-end ReRAM crossbar non-ideal effect adaption for neural network mapping. InProceedings of the 56th Annual Design Automation Conference 2019. 1–6

    work page 2019

  8. [8]

    Mark Horowitz. 2014. 1.1 computing’s energy problem (and what we can do about it). In2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). IEEE, 10–14

    work page 2014

  9. [9]

    Harold Hotelling. 1944. Some improvements in weighing and other experimental techniques.The Annals of Mathematical Statistics15, 3 (1944), 297–306

    work page 1944

  10. [10]

    Yucong Huang, Jingyu He, Tim Kwang-Ting Cheng, Chi Ying Tsui, and Terry Tao Ye. 2024. RWriC: A Dynamic Writing Scheme for Variation Compensation for RRAM-based In-Memory Computing. InProceedings of the 61st ACM/IEEE Design Automation Conference. 1–6

    work page 2024

  11. [11]

    Je-Min Hung, Xueqing Li, Juejian Wu, and Meng-Fan Chang. 2020. Challenges and trends indeveloping nonvolatile memory-enabled computing chips for intel- ligent edge devices.IEEE Transactions on Electron Devices67, 4 (2020), 1444–1453

    work page 2020

  12. [12]

    Daniele Ielmini and Giacomo Pedretti. 2025. Resistive switching random-access memory (RRAM): Applications and requirements for memory and computing. Chemical Reviews(2025)

    work page 2025

  13. [13]

    Vinay Joshi, Manuel Le Gallo, Simon Haefeli, Irem Boybat, Sasidharan Rajalek- shmi Nandakumar, Christophe Piveteau, Martino Dazzi, Bipin Rajendran, Abu Sebastian, and Evangelos Eleftheriou. 2020. Accurate deep neural network infer- ence using computational phase-change memory.Nature communications11, 1 (2020), 2473

    work page 2020

  14. [14]

    Amr Khashaba, Ahmed Elkholy, Karim M Megawer, Mostafa Gamal Ahmed, and Pavan Kumar Hanumolu. 2019. A low-noise frequency synthesizer using multiphase generation and combining techniques.IEEE Journal of Solid-State Circuits55, 3 (2019), 592–601

    work page 2019

  15. [15]

    Wantong Li, Xiaoyu Sun, Shanshi Huang, Hongwu Jiang, and Shimeng Yu. 2022. A 40-nm MLC-RRAM compute-in-memory macro with sparsity control, on-chip write-verify, and temperature-independent ADC references.IEEE Journal of Solid-State Circuits57, 9 (2022), 2868–2877

    work page 2022

  16. [16]

    Ziqi Meng, Yanan Sun, and Weikang Qian. 2022. Write or not: Programming scheme optimization for RRAM-based neuromorphic computing. InProceedings of the 59th ACM/IEEE Design Automation Conference. 985–990

    work page 2022

  17. [17]

    Xiaochen Peng, Shanshi Huang, Hongwu Jiang, Anni Lu, and Shimeng Yu. 2020. DNN+ NeuroSim V2. 0: An end-to-end benchmarking framework for compute-in- memory accelerators for on-chip training.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems40, 11 (2020), 2306–2319

    work page 2020

  18. [18]

    Hyeri Roh and Woo-Seok Choi. 2023. A context-aware readout system for sparse touch sensing array using ultra-low-power always-on event detection.IEEE Transactions on Circuits and Systems II: Express Briefs70, 9 (2023), 3719–3723

    work page 2023

  19. [19]

    Wonbo Shim, Jae-sun Seo, and Shimeng Yu. 2020. Two-step write–verify scheme and impact of the read noise in multilevel RRAM-based inference engine.Semi- conductor Science and Technology35, 11 (2020), 115026

    work page 2020

  20. [20]

    Joonghyun Song, Jiwon Shin, Hanseok Kim, and Woo-Seok Choi. 2022. Energy- Efficient High-Accuracy Spiking Neural Network Inference Using Time-Domain Neurons. In2022 IEEE 4th International Conference on Artificial Intelligence Cir- cuits and Systems (AICAS). 5–8. doi:10.1109/AICAS54282.2022.9870009

    work page doi:10.1109/aicas54282.2022.9870009 2022

  21. [21]

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. 2017. Efficient processing of deep neural networks: A tutorial and survey.Proc. IEEE105, 12 (2017), 2295–2329

    work page 2017

  22. [22]

    Xiyuan Tang, Jiaxin Liu, Yi Shen, Shaolan Li, Linxiao Shen, Arindam Sanyal, Kareem Ragab, and Nan Sun. 2022. Low-power SAR ADC design: Overview and survey of state-of-the-art techniques.IEEE Transactions on Circuits and Systems I: Regular Papers69, 6 (2022), 2249–2262

    work page 2022

  23. [23]

    Weier Wan, Rajkumar Kubendran, Clemens Schaefer, Sukru Burc Eryilmaz, Wenqiang Zhang, Dabin Wu, Stephen Deiss, Priyanka Raina, He Qian, Bin Gao, et al. 2022. A compute-in-memory chip based on resistive random-access memory. Nature608, 7923 (2022), 504–512

    work page 2022

  24. [24]

    Tian Xie, Shimeng Yu, and Shaolan Li. 2023. A high-parallelism RRAM-based compute-in-memory macro with intrinsic impedance boosting and in-ADC computing.IEEE Journal on Exploratory Solid-State Computational Devices and Circuits9, 1 (2023), 38–46

    work page 2023

  25. [25]

    Zheyu Yan, Xiaobo Sharon Hu, and Yiyu Shi. 2022. Swim: Selective write- verify for computing-in-memory neural accelerators. InProceedings of the 59th ACM/IEEE Design Automation Conference. 277–282

    work page 2022

  26. [26]

    Peng Yao, Huaqiang Wu, Bin Gao, Jianshi Tang, Qingtian Zhang, Wenqiang Zhang, J Joshua Yang, and He Qian. 2020. Fully hardware-implemented memristor convolutional neural network.Nature577, 7792 (2020), 641–646

    work page 2020

  27. [27]

    Shimeng Yu, Wonbo Shim, Xiaochen Peng, and Yandong Luo. 2021. RRAM for compute-in-memory: From inference to training.IEEE Transactions on Circuits and Systems I: Regular Papers68, 7 (2021), 2753–2765

    work page 2021

  28. [28]

    Grace Li Zhang, Bing Li, Xing Huang, Chen Shen, Shuhang Zhang, Florin Burcea, Helmut Graeb, Tsung-Yi Ho, Hai Li, and Ulf Schlichtmann. 2021. An efficient programming framework for memristor-based neuromorphic computing. In 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 1068–1073

    work page 2021