Pith. sign in

REVIEW 4 major objections 5 minor 110 references

By protecting only the sink and recent KV entries on a high-precision digital path, GoS reduces analog-noise perplexity from 33.91 to 11.95, near the clean 11.06.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:03 UTC pith:23YFAPAA

load-bearing objection A solid analog-CIM KV-protection proposal with a genuinely useful ownership protocol and a credible vulnerability analysis, but all quantitative claims rest on a single fitted noise model with no error bars — the result is conditional, not established. the 4 major comments →

arxiv 2607.29076 v1 pith:23YFAPAA submitted 2026-07-31 cs.AR

Selective KV Cache Protection for Noise-Resilient LLM Inference on Analog Compute-In-Memory Systems

Yuannuo Feng , Wenyong Zhou , Yuang Ma , Yizhe Chen , Wenshuai Yao , Yuxin Xie , Ngai Wong , Wang Kang This is my paper
classification cs.AR
keywords analog compute-in-memoryKV cacheLLM inferencehardware noiseattention sinkstoken-level vulnerabilityhybrid analog-digital architectureperplexity recovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Analog compute-in-memory (CIM) arrays promise energy-efficient LLM inference, but using them for the dynamic KV cache—the stored key/value vectors that attention reads—exposes inference to hardware noise. The paper claims that not every KV row needs equal precision: initial 'sink' tokens and a sliding window of recent tokens are disproportionately vulnerable, while middle tokens tolerate noise. It proposes GoS, a policy that keeps only those sensitive entries on a higher-precision digital path and stores the bulk on analog CIM. In simulations across nine LLMs, GoS lowers average perplexity under analog KV noise from 33.91 to 11.95, near the clean baseline of 11.06, and raises dynamic-KV programming-row utilization from 23.1% to 91.2%. The significance is that analog CIM attention becomes practical without storing the whole KV cache digitally.

Core claim

GoS (Guard-of-Sink) is a selective-protection scheme built on a measured asymmetry: corrupting the KV cache with analog noise hurts initial sink tokens and recent tokens far more than middle tokens. The paper argues the hardware need not make every KV row equally precise; it should preserve the small position-defined subset whose corruption propagates most broadly, and use analog density for the rest. Concretely, it pins m=8 sink tokens and a sliding recent window (total budget k=128, about 5% of the 2.56K working set) to a high-precision digital path, with a coalesced migration protocol that guarantees no token is ever absent from the global softmax. Under a chip-calibrated noise model with

What carries the argument

The central object is the token-level sensitivity profile of the KV cache under noise, grouped into three tiers: sink tokens (global impact), recent tokens (local-context impact), and middle bulk tokens (low individual attention, hence noise-tolerant). GoS turns this profile into an ownership protocol with three states—protected, pending-bulk, and active-bulk—and a threshold-coalesced programming batch (θ) that decouples logical token migration from physical tile activation. This lets the scheduler keep a bounded digital buffer of at most k+θ−1 entries, maintain exactly one serving path per token, and raise analog programming-row utilization.

Load-bearing premise

The load-bearing premise is that the chip-calibrated noise kernel used in simulation faithfully represents real dynamic KV programming and read noise for all nine models at all context lengths; if real device noise is correlated, state-dependent, or tile-size-dependent rather than the fitted additive kernel, the 5% protection budget may stop being sufficient.

What would settle it

Run one of the nine evaluated LLMs end to end on the actual prototype with a real dynamic KV cache at the calibrated noise level (σ=0.10); if measured noisy perplexity under sink-and-recent protection does not land close to the simulated 11.95–12.0 range, or if the per-position degradation pattern does not show the predicted sink/recent spike, the central claim fails. Even a single 8K-token sequence would test the prediction sharply.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If GoS holds, only about 5% of the KV working set needs high-precision storage; the rest can run on noisy analog arrays without losing much quality.
  • Average perplexity under analog KV noise drops from 33.91 to 11.95 across nine LLMs, approaching the clean 11.06, so the analog attention path becomes viable for models that currently collapse under noise.
  • Dynamic-KV programming-row utilization rises from 23.1% to 91.2%, meaning the analog arrays are no longer mostly idle during attention.
  • The bounded on-chip digital footprint (k+θ−1 entries) means context length can scale to 512K without growing protected storage; long contexts only add cold-tile HBM traffic.
  • Because GoS preserves full-attention semantics rather than truncating keys, it composes with sparsification and eviction policies rather than replacing them.

Where Pith is reading between the lines

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

  • Inference: The paper leaves open whether the same positional tiers generalize to other noise sources; a direct test would re-run the per-position perplexity analysis under read noise only, or under correlated device variation.
  • Inference: The residual gap to clean (0.89 PPL on average) is attributed to projection-array noise that GoS deliberately does not touch, which implies further gains require fixing the projections rather than the KV path—a direction the paper does not pursue.
  • Inference: The sensitivity tiers suggest a digital counterpart: KV-quantization bit-widths could be assigned by position rather than uniformly, with sink and recent tokens kept at higher precision—a policy the paper does not evaluate.
  • Inference: The 5% budget is fixed while context grows, which suggests an even smaller fraction may suffice at long context; sweeping k at 512K would test whether the middle bulk becomes more noise-tolerant as attention becomes sparser.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper presents GoS (Guard-of-Sink), a hardware-algorithm co-design for protecting KV cache entries in analog compute-in-memory LLM inference. The authors first measure per-position PPL degradation under a chip-calibrated additive noise model (Eq. (14)) and identify sink and recent tokens as most vulnerable. They then partition the logical KV cache into a pinned sink region, a sliding recent region, and an analog bulk region (Eqs. (4)-(7)), with a three-state ownership protocol and threshold-coalesced migration to batch analog writes. The proposed attention computation (Eqs. (8)-(11)) uses a single global softmax and preserves full attention semantics. Simulations on nine LLMs report average WikiText-2 PPL improving from 33.91 (analog KV) to 11.95 (Prefill-Decode) and 11.51 (Prefill-Only), close to clean 11.06, with 3% energy/4% latency overhead, and claim dynamic-KV programming-row utilization increases from 23.1% to 91.2%. The paper is explicit that all PPL results are simulations: LLMs are not run end to end on the prototype and energy/PPA are design estimates.

Significance. The core formal contribution is clean: Eqs. (8)-(11) define a hybrid attention path with exactly one serving path per token and a global softmax, so there is no approximation beyond the analog noise itself. The ownership invariants (protected/pending/active-bulk) and the bounded on-chip footprint k+theta-1 are well stated and appear sound. The policy is also notable for requiring no learned predictor or per-request retraining; it uses only position indices. If the simulation results transfer to hardware, this would be a useful step for dynamic analog KV caching. However, the empirical evidence is not yet sufficient to establish transfer: the entire quality evaluation rests on a single fitted additive noise model, and the main quantitative claims are single-point estimates with no uncertainty. The 91.2% utilization figure is asserted but not substantiated with an experimental protocol. With additional validation, this could be a solid contribution; in its current form the central quality claim is conditional.

major comments (4)
  1. [Experiment Setup / Eq. (14)] The load-bearing assumption is that Eq. (14)'s per-entry additive noise with fixed variance, validated by a held-out KL divergence of 0.003, represents dynamic KV programming/read noise for all models and context lengths. The KL divergence is a marginal distributional check; it does not validate joint/correlated noise, tile-size dependence, or state-dependent write noise. Since the tiering conclusion (sink/recent vulnerable, bulk tolerant) is derived under this independence assumption, a correlated noise component (e.g., a per-row or per-tile common-mode shift) could invalidate the sink/recent tiering and the 5% protection budget. Please add experiments with correlated noise with varying strength, sweep tile size, and report PPL for GoS under these conditions. Without this, the simulation evidence for the protection policy is incomplete.
  2. [Figures 4 and 10] The paper identifies sink/recent vulnerability using its own per-position PPL analysis under the same chip-calibrated noise model, then reports that protecting exactly those classes removes most degradation under that same model. This is partly self-confirming: the target classes are chosen on the same noise distribution used for evaluation. Although the policy is not fitted (it is a fixed positional rule), the evidence would be stronger if the tiering were derived from clean attention statistics or known attention-sink priors and then tested under a held-out noise condition, or if the ablation were repeated under a second, independently calibrated noise kernel. Please report this analysis.
  3. [Table 1 / Experiment Results] All headline quality numbers are single-point estimates. Table 1 and Figures 8-11 show no error bars, no standard deviations, and no seed variation; the simulated noise kernel is stochastic, so runs should be repeated. The average improvement from 33.91 to 11.95 masks per-model gaps (e.g., Qwen3-0.6B: 14.92 vs clean 12.35 in Prefill-Decode, 13.58 in Prefill-Only). Reporting confidence intervals and per-model residuals is necessary to support the claim of 'approaching clean baseline.' Without this, the central quantitative claim is not statistically grounded.
  4. [Abstract / Conclusion / Experiments] The claim that GoS raises dynamic-KV programming-row utilization from 23.1% to 91.2% appears in the abstract and conclusion, but no experimental setup, simulation, or measurement supporting these numbers is reported. Figure 5 is illustrative; Eq. (12) defines a metric but the values of n_b, B, R_tile, theta, and the measurement window are not given. Please provide a dedicated experiment with these parameters, including a validation of the threshold-coalesced scheduler and its effect on latency/energy.
minor comments (5)
  1. [Figure 4] Per-position PPL degradation is not described in terms of context length, model, temperature, and noise severity; please add the setup and error bars (or explain determinism).
  2. [Eq. (12)] Clarify whether a 'programming service b' corresponds to a tile batch and how partial rows are counted; define R_tile explicitly.
  3. [Figure 9] The temperature-to-sigma mapping (25,50,75,100,125 C to 0.1,...,0.5) is stated but not justified; please cite measurements or mark it as a parameterization.
  4. [Section 3.2] The statement that GoS does not recompute protected tokens using a clean digital projection is important; please move or emphasize it earlier to avoid the impression that digital protection removes projection noise.
  5. [Table 2] The normalized cost at 100 C is reported without the underlying energy/latency model; please state whether it includes HBM traffic for cold tiles, since the text says the bounded on-chip footprint does not remove HBM traffic.

Circularity Check

0 steps flagged

No significant circularity: the GoS policy is an empirical response to measured position-dependent vulnerability, not a fit repackaged as a prediction.

full rationale

The paper's derivation chain is self-contained. The central premise is that sink and recent tokens are disproportionately vulnerable to analog KV-cache noise, and that middle-bulk tokens are relatively tolerant. This is established by direct per-position perplexity measurements (Figure 4) under a chip-calibrated noise kernel (Eq. 14), and then independently tested by evaluating the fixed positional policy defined by Eqs. (4)-(7). The policy parameters (m=8, k=128) are not fitted to minimize the reported perplexity table; the token-selection ablation (Figure 10) shows that a random-k baseline (12.61 PPL) is substantially worse than the sink/recent policy (10.67 PPL), so the improvement is not forced by construction. The noise kernel itself is calibrated from measured CIM chips with held-out KL divergence 0.003, providing external grounding outside the PPL results. Self-citations (Feng et al. 2026a/b, Wang et al. 2025) appear only in related-work and novelty statements, not as load-bearing premises for the GoS mechanism or its evaluation. The paper's own Scope and Limitations transparently notes that the LLMs are not executed end-to-end on the prototype and that energy/latency/PPA values are design estimates; these are simulation-scope limitations, not circular arguments. No equation reduces to another by construction, and no fitted parameter is renamed as a prediction. The reported PPL recovery is a legitimate, if simulation-only, evaluation of a well-specified hardware-algorithm policy.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central design is parameterized by protection budget (m, k), a migration threshold, and a chip-calibrated noise kernel. The paper's headline PPL values depend on all of these. The only external empirical anchors are the attention-sink and recency phenomena from prior literature; the noise-tolerance of middle tokens is established in the same simulation loop used to evaluate the proposal, which is the main circularity burden.

free parameters (4)
  • sink budget m = 8
    Pinned sink tokens protected digitally; fixed default, not swept in main text; paper acknowledges optimal value may vary.
  • total protection budget k = 128 (recent window r = 120)
    Protection ratio 5% of the 2.56K dynamic-KV working set; no sensitivity analysis reported in the main text.
  • migration threshold theta = not stated in main text
    Determines coalescing batch size and bounded digital footprint k + theta - 1; directly affects the reported 91.2% utilization.
  • noise severity sigma = 0.10 default; mapped to 0.10-0.5 for 25-125C
    Configurable severity in the chip-calibrated noise kernel; a modeling choice not fitted to PPL targets, but still load-bearing for all quality results.
axioms (4)
  • domain assumption Sink tokens attract disproportionately large attention across positions and layers
    Invoked to justify protecting the first m tokens; based on Xiao et al. 2023, external prior empirical result, not re-derived here.
  • domain assumption Recent tokens concentrate local context and are heavily attended, while middle bulk tokens are noise-tolerant
    Motivates the sliding recent window and analog bulk assignment; supported by the paper's own Figure 4 and streaming-LLM literature, but treated as universal across models and contexts.
  • domain assumption Analog KV noise enters as additive program/read perturbations on individual K/V entries with a chip-calibrated distribution
    All PPL results use this noise model (Eq. 14); the paper validates KL divergence 0.003 on held-out chip data but assumes the calibrated kernel generalizes to arbitrary LLMs and long contexts.
  • domain assumption Q/K/V projection weight matrices are weight-stationary analog arrays programmed once at deployment
    Architecture premise in the Hardware Architecture section; if projections also require dynamic updates, the cost model and noise separation change.

pith-pipeline@v1.3.0-daily-deepseek · 11468 in / 15148 out tokens · 162299 ms · 2026-08-03T14:03:37.951271+00:00 · methodology

0 comments
read the original abstract

Analog compute-in-memory (CIM) arrays have emerged as a promising substrate for energy-efficient LLM inference, particularly for weight-stationary computations in linear layers. However, extending analog CIM to attention mechanisms introduces a fundamental challenge: KV cache operations demand repeated in-situ weight updates, and the resulting mismatch with the weight-stationary paradigm exposes dynamic computations to significant hardware noise, a critical problem that remains largely unexplored. In this paper, we present the first systematic study of dynamic attention computation on analog CIM arrays, revealing that initial and recent tokens exhibit disproportionate vulnerability to hardware noise. Motivated by this token-level insight, we propose a hierarchical token protection strategy that keeps sink tokens and a sliding recent-token window on a higher-precision digital path while processing the bulk KV cache on analog CIM. A co-designed scheduler combines analog programming, ownership transition, and bulk-MVM tile formation to bound digital overhead. Evaluations on nine LLMs show that our approach lowers average perplexity under analog noise from 33.91 to 11.95, close to the clean baseline of 11.06, while improving dynamic-KV programming-row utilization from 23.1\% to 91.2\%.

Figures

Figures reproduced from arXiv: 2607.29076 by Ngai Wong, Wang Kang, Wenshuai Yao, Wenyong Zhou, Yizhe Chen, Yuang Ma, Yuannuo Feng, Yuxin Xie.

Figure 1
Figure 1. Figure 1: The rapid expansion of LLM context windows from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Perplexity degradation caused by analog hardware [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: Averaged attention map of Qwen3-8B under [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-position PPL degradation of Qwen3-8B from [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Dynamic-KV programming-row utilization under [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: GoS-enabled analog CIM accelerator architecture with hybrid analog-digital processing and 3D-integrated design. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Measured-chip calibration of the configurable CIM [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative attention case before and after GoS [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Token-selection ablation on WikiText-2 at 8K [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Temperature-dependent Qwen3-8B accuracy on [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗

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

110 extracted references · 10 linked inside Pith

  1. [1]

    Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer , booktitle=

  2. [2]

    International conference on machine learning , pages=

    Arjovsky, Martin and Chintala, Soumith and Bottou, L. International conference on machine learning , pages=. 2017 , organization=

    2017

  3. [3]

    2015 , organization=

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , booktitle=. 2015 , organization=

    2015

  4. [4]

    Tran, Linh and Pantic, Maja and Deisenroth, Marc Peter , journal=

  5. [5]

    Li, Xiuyu and Liu, Yijiang and Lian, Long and Yang, Huanrui and Dong, Zhen and Kang, Daniel and Zhang, Shanghang and Keutzer, Kurt , booktitle=

  6. [6]

    and Crafton, Brian and Raychowdhury, Arijit and Fang, Yan , booktitle=

    Lele, Ashwin Sanjay and Chang, Muya and Spetalnick, Samuel D. and Crafton, Brian and Raychowdhury, Arijit and Fang, Yan , booktitle=. 2023 , volume=

    2023

  7. [7]

    2018 , volume=

    Kooli, Maha and others , booktitle=. 2018 , volume=

    2018

  8. [8]

    2021 , volume=

    Reis, Dayane and others , booktitle=. 2021 , volume=

    2021

  9. [9]

    2017 , publisher=

    Gokmen, Tayfun and Onen, Murat and Haensch, Wilfried , journal=. 2017 , publisher=

    2017

  10. [10]

    2020 , volume=

    Gupta, Saransh and others , booktitle=. 2020 , volume=

    2020

  11. [11]

    2023 , volume=

    He, Wangxin and otherws , booktitle=. 2023 , volume=

    2023

  12. [12]

    2021 , volume=

    Jung, Giju and Fouda, Mohammed and Lee, Sugil and Lee, Jongeun and Eltawil, Ahmed and Kurdahi, Fadi , booktitle=. 2021 , volume=

    2021

  13. [13]

    2017 , volume=

    Chen, Lerong and others , booktitle=. 2017 , volume=

    2017

  14. [14]

    2021 , volume=

    Gao, Di and Huang, Qingrong and Zhang, Grace Li and Yin, Xunzhao and Li, Bing and Schlichtmann, Ulf and Zhuo, Cheng , booktitle=. 2021 , volume=

    2021

  15. [15]

    2015 , volume=

    Liu, Beiye and Hai Li and Yiran Chen and Xin Li and Qing Wu and Tingwen Huang , booktitle=. 2015 , volume=

    2015

  16. [16]

    2021 , volume=

    Lin, Jilan and others , journal=. 2021 , volume=

    2021

  17. [17]

    2021 , volume=

    Ye, Nanyang and Mei, Jingbiao and Fang, Zhicheng and Zhang, Yuwen and Zhang, Ziqing and Wu, Huaying and Liang, Xiaoyao , booktitle=. 2021 , volume=

    2021

  18. [18]

    Du, Xuefeng and others , journal=

  19. [19]

    Yuzhang Shang and Zhihang Yuan and Bin Xie and Bingzhe Wu and Yan Yan , booktitle=

  20. [20]

    2019 , volume=

    Liu, Tao and Wen, Wujie and Jiang, Lei and Wang, Yanzhi and Yang, Chengmo and Quan, Gang , booktitle=. 2019 , volume=

    2019

  21. [21]

    2023 , volume=

    Chen, Xuan-Jun and Kuan, Cynthia and Yang, Chia-Lin , booktitle=. 2023 , volume=

    2023

  22. [22]

    2023 , volume=

    Chen, Jia and Tu, Fengbin and Shao, Kunming and Tian, Fengshi and Huo, Xiao and Tsui, Chi-Ying and Cheng, Kwang-Ting , booktitle=. 2023 , volume=

    2023

  23. [23]

    2023 , volume=

    Lee, Mingyen and Tang, Wenjun and Chen, Yiming and Wu, Juejian and Zhong, Hongtao and Xu, Yixin and Liu, Yongpan and Yang, Huazhong and Narayanan, Vijaykrishnan and Li, Xueqing , booktitle=. 2023 , volume=

    2023

  24. [24]

    and Crafton, Brian and Konno, Shota and Wan, Zishen and Bhat, Ashwin and Khwa, Win-San and Chih, Yu-Der and Chang, Meng-Fan and Raychowdhury, Arijit , journal=

    Lele, Ashwin Sanjay and Chang, Muya and Spetalnick, Samuel D. and Crafton, Brian and Konno, Shota and Wan, Zishen and Bhat, Ashwin and Khwa, Win-San and Chih, Yu-Der and Chang, Meng-Fan and Raychowdhury, Arijit , journal=. 2024 , volume=

    2024

  25. [25]

    and Cady, Nathaniel C

    Krishnan, Gokul and Wang, Zhenyu and Yeo, Injune and Yang, Li and Meng, Jian and Liehr, Maximilian and Joshi, Rajiv V. and Cady, Nathaniel C. and Fan, Deliang and Seo, Jae-Sun and Cao, Yu , journal=. 2022 , volume=

    2022

  26. [26]

    2024 , volume=

    Liu, Shiwei and Mu, Chen and Jiang, Hao and Wang, Yunzhengmao and Zhang, Jinshan and Lin, Feng and Zhou, Keji and Liu, Qi and Chen, Chixiao , journal=. 2024 , volume=

    2024

  27. [27]

    Zhang Wenqiang, Gao Bin, Yao Peng, Tang Jianshi, Qian He and Wu Huaqiang , journal=

  28. [28]

    2019 , volume=

    Zhang, Wenqiang and Peng, Xiaochen and Wu, Huaqiang and Gao, Bin and He, Hu and Zhang, Youhui and Yu, Shimeng and Qian, He , booktitle=. 2019 , volume=

    2019

  29. [29]

    and Cady, Nathaniel C

    Charan, Gouranga and Hazra, Jubin and Beckmann, Karsten and Du, Xiaocong and Krishnan, Gokul and Joshi, Rajiv V. and Cady, Nathaniel C. and Cao, Yu , booktitle=. 2020 , volume=

    2020

  30. [30]

    Ji Lin and Chuang Gan and Song Han , booktitle=

  31. [31]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  32. [32]

    Denoising Diffusion Implicit Models , author=

  33. [33]

    Pseudo Numerical Methods for Diffusion Models on Manifolds , author=

  34. [34]

    Advances in neural information processing systems , volume=

    Generating diverse high-fidelity images with vq-vae-2 , author=. Advances in neural information processing systems , volume=

  35. [35]

    Advances in neural information processing systems , volume=

    Generative adversarial nets , author=. Advances in neural information processing systems , volume=

  36. [36]

    Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , pages=

    Diffusion probabilistic model made slim , author=. Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , pages=

  37. [37]

    Advances in neural information processing systems , volume=

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps , author=. Advances in neural information processing systems , volume=

  38. [38]

    2025 Design, Automation & Test in Europe Conference (DATE) , pages=

    HyIMC: Analog-Digital Hybrid In-Memory Computing SoC for High-Quality Low-Latency Speech Enhancement , author=. 2025 Design, Automation & Test in Europe Conference (DATE) , pages=. 2025 , organization=

    2025

  39. [39]

    2024 IEEE International Conference on Integrated Circuits, Technologies and Applications (ICTA) , pages=

    Device Characteristic-Aware Quantization for eFlash-Based In-Memory Computing SoC , author=. 2024 IEEE International Conference on Integrated Circuits, Technologies and Applications (ICTA) , pages=. 2024 , organization=

    2024

  40. [40]

    IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , volume=

    An End-to-End In-Memory Computing System Based on a 40-nm eFlash-Based IMC SoC: Circuits, Toolchains, and Systems Co-Design Framework , author=. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , volume=. 2024 , publisher=

    2024

  41. [41]

    2023 IEEE International Conference on Integrated Circuits, Technologies and Applications (ICTA) , pages=

    A 40nm 5-16Tops/W@ INT8 eFlash In-Memory Computing SoC Chip with Noise Suppression and Compensation Techniques to Improve the Accuracy , author=. 2023 IEEE International Conference on Integrated Circuits, Technologies and Applications (ICTA) , pages=. 2023 , organization=

    2023

  42. [42]

    IEEE Transactions on Circuits and Systems II: Express Briefs , volume=

    A mini tutorial of processing in memory: From principles, devices to prototypes , author=. IEEE Transactions on Circuits and Systems II: Express Briefs , volume=. 2022 , publisher=

    2022

  43. [43]

    Enhancing Robustness of Implicit Neural Representations Against Weight Perturbations , year=

    Zhou, Wenyong and others , booktitle=. Enhancing Robustness of Implicit Neural Representations Against Weight Perturbations , year=

  44. [44]

    A Hardware-Aware Neural Architecture Search Pareto Front Exploration for In-Memory Computing , year=

    Guan, Ziyi and others , booktitle=. A Hardware-Aware Neural Architecture Search Pareto Front Exploration for In-Memory Computing , year=

  45. [45]

    A Time- and Energy-Efficient CNN with Dense Connections on Memristor-Based Chips , year=

    Zhou, Wenyong and others , booktitle=. A Time- and Energy-Efficient CNN with Dense Connections on Memristor-Based Chips , year=

  46. [46]

    HPD: Hybrid Projection Decomposition for Robust State Space Models on Analog CIM Hardware , year=

    Feng, Yuannuo and Zhou, Wenyong and Lyu, Yuexi and Liu, Hanjie and Liu, Zhengwu and Wong, Ngai and Kang, Wang , booktitle=. HPD: Hybrid Projection Decomposition for Robust State Space Models on Analog CIM Hardware , year=

  47. [47]

    Extending Straight-Through Estimation for Robust Neural Networks on Analog CIM Hardware , year=

    Feng, Yuannuo and Zhou, Wenyong and Lyu, Yuexi and Zhang, Yixiang and Liu, Zhengwu and Wong, Ngai and Kang, Wang , booktitle=. Extending Straight-Through Estimation for Robust Neural Networks on Analog CIM Hardware , year=

  48. [48]

    Computing-In-Memory Neural Network Accelerators for Safety-Critical Systems: Can Small Device Variations Be Disastrous? , year=

    Yan, Zheyu and others , booktitle=. Computing-In-Memory Neural Network Accelerators for Safety-Critical Systems: Can Small Device Variations Be Disastrous? , year=

  49. [49]

    2025 , volume=

    Zhou, Wenyong and others , booktitle=. 2025 , volume=

    2025

  50. [50]

    2020 , keywords=

    Charan, Gouranga and others , booktitle=. 2020 , keywords=

    2020

  51. [51]

    2016 , volume=

    Chi, Ping and others , booktitle=. 2016 , volume=

    2016

  52. [52]

    Nature communications , volume=

    Optimised weight programming for analogue memory-based deep neural networks , author=. Nature communications , volume=. 2022 , publisher=

    2022

  53. [53]

    Frontiers in Neuroscience , volume=

    Mixed-precision deep learning based on computational memory , author=. Frontiers in Neuroscience , volume=. 2020 , publisher=

    2020

  54. [54]

    Nature Electronics , volume=

    In-memory computing with resistive switching devices , author=. Nature Electronics , volume=. 2018 , publisher=

    2018

  55. [55]

    IEEE Solid-State Circuits Magazine , volume=

    In-memory computing: Advances and prospects , author=. IEEE Solid-State Circuits Magazine , volume=. 2019 , publisher=

    2019

  56. [56]

    IEEE Journal of Solid-State Circuits , volume=

    C3SRAM: An in-memory-computing SRAM macro based on robust capacitive coupling computing mechanism , author=. IEEE Journal of Solid-State Circuits , volume=. 2020 , publisher=

    2020

  57. [57]

    ACM SIGARCH Computer Architecture News , volume=

    ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars , author=. ACM SIGARCH Computer Architecture News , volume=. 2016 , organization=

    2016

  58. [58]

    ACM SIGARCH Computer Architecture News , volume=

    Prime: A novel processing-in-memory architecture for neural network computation in reram-based main memory , author=. ACM SIGARCH Computer Architecture News , volume=. 2016 , organization=

    2016

  59. [59]

    Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems , pages=

    Puma: A programmable ultra-efficient memristor-based accelerator for machine learning inference , author=. Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems , pages=

  60. [60]

    Nature , volume=

    Equivalent-accuracy accelerated neural-network training using analogue memory , author=. Nature , volume=. 2018 , publisher=

    2018

  61. [61]

    Nature , volume=

    Fully hardware-implemented memristor convolutional neural network , author=. Nature , volume=. 2020 , publisher=

    2020

  62. [62]

    Proceedings of the 44th annual international symposium on computer architecture , pages=

    In-datacenter performance analysis of a tensor processing unit , author=. Proceedings of the 44th annual international symposium on computer architecture , pages=

  63. [63]

    Journal of Machine Learning Research , volume=

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity , author=. Journal of Machine Learning Research , volume=

  64. [64]

    arXiv preprint arXiv:1308.0850 , year=

    Generating sequences with recurrent neural networks , author=. arXiv preprint arXiv:1308.0850 , year=

    Pith/arXiv arXiv

  65. [65]

    Advances in Physics: X , volume=

    Neuromorphic computing using non-volatile memory , author=. Advances in Physics: X , volume=. 2017 , publisher=

    2017

  66. [66]

    2020 IEEE Custom Integrated Circuits Conference (CICC) , pages=

    Compute-in-memory with emerging nonvolatile-memories: Challenges and prospects , author=. 2020 IEEE Custom Integrated Circuits Conference (CICC) , pages=. 2020 , organization=

    2020

  67. [67]

    International conference on machine learning , pages=

    Batch normalization: Accelerating deep network training by reducing internal covariate shift , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  68. [68]

    Communications Engineering , volume=

    Improving the robustness of analog deep neural networks through a Bayes-optimized noise injection approach , author=. Communications Engineering , volume=. 2023 , publisher=

    2023

  69. [69]

    2024 Design, Automation & Test in Europe Conference & Exhibition (DATE) , pages=

    Towards Robust RRAM-Based Vision Transformer Models with Noise-Aware Knowledge Distillation , author=. 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE) , pages=. 2024 , organization=

    2024

  70. [70]

    Learning fast samplers for diffusion models by differentiating through sample quality , author=

  71. [71]

    arXiv preprint arXiv:2202.00512 , year=

    Progressive distillation for fast sampling of diffusion models , author=. arXiv preprint arXiv:2202.00512 , year=

    Pith/arXiv arXiv

  72. [72]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  73. [73]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  74. [74]

    Minds and machines , volume=

    GPT-3: Its nature, scope, limits, and consequences , author=. Minds and machines , volume=. 2020 , publisher=

    2020

  75. [75]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  76. [76]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    Pith/arXiv arXiv

  77. [77]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    OLMo: Accelerating the science of language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  78. [78]

    Proceedings of machine learning and systems , volume=

    Efficiently scaling transformer inference , author=. Proceedings of machine learning and systems , volume=

  79. [79]

    International Conference on Machine Learning , pages=

    Flexgen: High-throughput generative inference of large language models with a single gpu , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  80. [80]

    Proceedings of the 29th symposium on operating systems principles , pages=

    Efficient memory management for large language model serving with pagedattention , author=. Proceedings of the 29th symposium on operating systems principles , pages=

Showing first 80 references.