acceptodds
Under review as a conference paper at ICLR 2027

IMC-CLINIC: Coupled Loss-Informed Newton Iterations for Clipping in Analog In-Memory Computing

Abstract

Analog in-memory computing (IMC) offers a promising path toward energy-efficient large language model (LLM) inference by executing matrix multiplications (MatMul) directly within memory arrays in the analog domain. Its efficiency, however, comes with an additional source of error: limited-precision analog-to-digital converters (ADCs) quantize the accumulated analog partial sums and introduce output-side error, which is a different problem from conventional activation and weight quantization at the MatMul inputs. Clipping can mitigate both operand and ADC quantization errors by reducing their dynamic ranges, but the optimal clipping factors must jointly balance activation rounding and clipping, weight rounding and clipping, and ADC quantization. Existing clipping methods, designed for digital quantization, do not explicitly optimize this coupled IMC error. Common approaches rely on costly search-based calibration, leading to suboptimal accuracy and long calibration time. We introduce **IMC-CLINIC** (**C**oupled **L**oss-**I**nformed **N**ewton **I**terations for **C**lipping), a clipping calibration framework built around an analytical surrogate for IMC MatMul output error. The loss surrogate models operand quantization, accumulated clipping-induced bias, and ADC quantization jointly, allowing its gradient and approximate curvature to be evaluated directly from only a small calibration set. **IMC-CLINIC** then jointly optimizes activation and weight clipping factors using a safeguarded Newton-type method. We show that **IMC-CLINIC** reduces analog-IMC MatMul output error by effectively balancing operand and ADC quantization errors. Across multiple models and datasets, it improves average zero-shot accuracy by 6.5–11.5 percentage points over the grid search baseline while reducing calibration time by 10.0×–12.1×. Moreover, its analytical surrogate closely tracks empirical IMC output error, while its optimizer is fast and certified within 1% of the global optimum under the loss objective across all projections on two representative models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
In-memory Computing, LLM Quantization, Clipping, ADC Quantization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iah6OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IMC-CLINIC, a post-training clipping-calibration method for analog in-memory computing (IMC) inference. The key idea is to optimize activation and weight clipping jointly using an analytical surrogate for MatMul output error that includes operand rounding/clipping error, accumulated clipping bias, and ADC quantization error. The method uses analytical gradients and an approximate Hessian in a safeguarded Newton-style solver. Experiments on four LLMs under an 8-bit weight/activation and 9-bit ADC IMC simulation show substantially better perplexity and zero-shot accuracy than a sequential weight/activation grid-search baseline, while reducing calibration time.

Critical strengths

Well-motivated formulation. Unlike conventional digital clipping calibration, activation and weight clipping in analog IMC jointly control both operand quantization error and output-referred ADC error. The proposed objective explicitly models this coupling, including the potentially coherent bias induced by clipping.

Critical weaknesses
  1. Limited comparsion with direct empirical IMC-output-MSE optimization. The paper includes an ADC-aware alternating empirical-MSE search, but this comparison appears limited to one model and uses a relatively coarse coordinate-search procedure. A stronger matched comparison against direct empirical IMC-output-MSE optimization across models and under comparable evaluation budgets is necessary to substantiate the claim that the surrogate is both accurate and practically preferable.

  2. Missing of calibrated circuit-level model covering important nonidealities in analog IMC. Inhernet hardware nonidealities, such as device variation, nonlinearity, ADC transfer-function mismatch, IR drop may alter both the output-error decomposition and the optimal clipping policy. The injected Gaussian-noise study is useful but does not fully establish robustness to realistic analog IMC behavior.

Questions for authors
  1. Can the authors provide matched results against a stronger direct empirical optimizer of ADC-enabled IMC MatMul output MSE across all four evaluated models?

  2. How sensitive are the selected clipping factors and the relative advantage over the baselines to nonidealities not represented by the current simulator, such as nonuniform ADC transfer curves, correlated or signal-dependent ADC noise, analog accumulation nonlinearity, and device mismatch?

4 comments
AuthorsOpenReview ↗

Response to Reviewer iah6

We thank the reviewer for recognizing the motivation and mathematical formulation of IMC-CLINIC, and we appreciate the suggestions related to the strength of the experimental validation.

The two concerns are (1) the realism of analog IMC nonidealities of our simulation and (2) the comparison against our own stronger ADC-aware optimization baseline.

In response, we strengthen our evaluation along both points. The additional results below show that IMC-CLINIC (1) further provides robustness benefits in the presence of extensive nonidealities, and (2) is both more accurate and faster compared to the strengthened baseline across models and workloads.

Extended Robustness Evaluation Under Analog IMC Nonidealities

We thank the reviewer for highlighting the importance of realistic analog IMC nonidealities. We agree that evaluating these effects is important for understanding the robustness and applicability of IMC-CLINIC.

Our original simulation follows prior studies showing that Gaussian noise can approximate certain nonidealities in SRAM-based analog IMC, with validation against silicon measurements [1]. Prior work also demonstrates that systematic gain and offset errors can be mitigated through hardware calibration [2].

To directly address the reviewer's concern, we have extended our simulator to separately model six additional nonidealities: residual MAC gain/offset errors, residual ADC gain/offset errors, ADC integral nonlinearity (INL), static device/column mismatch, correlated noise, and signal-dependent noise. These models capture a broader range of analog imperfections beyond the original additive Gaussian-noise assumption, including residual static hardware variation, affine MAC/ADC errors, ADC transfer-function nonlinearity, and random noise with spatial correlation or signal dependence.

We provide detailed modeling assumptions and experimental results in Appendix H.2. Across the evaluated conditions, IMC-CLINIC maintains its accuracy advantage over the baselines. In several high-severity cases, particularly large MAC/ADC offsets and ADC INL, the no-clipping and W/A grid-search baselines degrade substantially, whereas IMC-CLINIC exhibits smaller accuracy degradation. For example, under 0.1 LSB MAC offset variation, the PPL of IMC-CLINIC increases from 12.27 to 13.40, whereas that of W/A grid search increases from 19.12 to 31.14. These results provide additional evidence of robustness under the modeled nonidealities.

Finally, we clarify that IMC-CLINIC primarily targets high-SNR IMC architectures with sufficiently linear MAC operations (Appendix A.1). While the extended experiments cover several important nonidealities, they do not constitute a fully silicon-calibrated circuit model. Extending the analytical surrogate to architectures with substantial IR drop, device variation, or temporal drift may require technology-specific modeling and remains an important direction for future work.

Additional Experiments against the Strengthened Baseline

We thank the reviewer for emphasizing the importance of stronger ADC-aware baselines. We agree that these comparisons help distinguish the benefits of the IMC-aware objective from those of the optimization algorithm.

Our main-table comparisons focus on established methods to demonstrate improvements over prior work. To further isolate the optimization contribution, we extend the ADC-aware empirical-MSE alternating search to all four models, including a finer 21-point search (Appendix H.1). All methods use the same calibration data and IMC configuration, and are evaluated on WikiText-2 perplexity and seven commonsense reasoning tasks.

IMC-CLINIC achieves the best downstream accuracy on three of four models. On LLaMA-3.1-8B, the 21-point search performs only 0.23 percentage points better, within the observed seed variation. Meanwhile, IMC-CLINIC achieves geometric-mean calibration speedups of 4.90X and 8.50X over the 11-point and 21-point searches, respectively, excluding initialization.

Increasing the empirical search resolution from 11 to 21 points provides little additional downstream improvement, suggesting diminishing returns from finer search. Separately, we verify ε-optimality with respect to our analytical surrogate (Appendix G), supporting the effectiveness of the proposed optimizer.

Together, these results demonstrate that IMC-CLINIC achieves competitive or better downstream accuracy than direct ADC-aware empirical optimization, while substantially reducing calibration cost.


We want to emphasize that our central contribution lies in the joint analytical formulation of operand and ADC errors and the development of an efficient optimization algorithm. With the additional evidence addressing the two experimental concerns, we believe now this will allow the reviewer to reassess our central contribution. We thank for the review again and welcome any further feedback.

AuthorsOpenReview ↗

Response to Reviewer iah6 (References)

[1] Zhang, W., Ando, S., Chen, Y. C., & Yoshioka, K. (2025). Asim: Modeling and analyzing inference accuracy of sram-based analog cim circuits. IEEE Transactions on Very Large Scale Integration (VLSI) Systems.

[2] Zhang, H., Bai, Y., Shen, J., Du, Y., & Du, L. (2024). Linearity performance of charge domain in-memory computing: Analysis and calibration. Integrated Circuits and Systems, 1(1), 43-52.

Reviewer_iah6OpenReview ↗

Response to Authors

Thank you for the detailed response and for extending the comparison with the ADC-aware empirical-MSE alternating-search baseline to all four evaluated models.

The strengthened comparison is helpful. In particular, using the same calibration data and IMC configuration, increasing the search resolution from 11 to 21 points, and reporting both downstream accuracy and calibration speed provide a more meaningful evaluation of the proposed analytical surrogate and optimizer. The reported results suggest that IMC-CLINIC is competitive with, and often preferable to, the stronger direct empirical-search baseline under the evaluated simulation setting.

However, my second concern is not yet fully addressed in the current response. The original experiments appear to use a particular IMC simulator/noise model, whereas practical analog IMC systems can exhibit additional effects, including nonuniform or nonlinear ADC transfer characteristics, correlated or signal-dependent ADC noise, analog accumulation nonlinearity, IR drop, and device mismatch. The response states that the method provides robustness benefits under “extensive nonidealities,” but it does not specify which additional nonidealities were evaluated, their parameter ranges, whether the clipping factors were re-calibrated under each setting, or the corresponding quantitative comparison against the baselines.

To substantiate the robustness claim, I encourage the authors to include a clearly documented sensitivity/robustness study covering at least several nonidealities not represented by the original simulator. This study should report the perturbation models and magnitudes, the calibration protocol, and the relative performance of IMC-CLINIC and relevant baselines. If such experiments are beyond the scope of the revision, the claims should be appropriately narrowed to the evaluated IMC noise and ADC model.

Overall, the expanded empirical-optimizer comparison substantially addresses my first concern. My remaining assessment depends primarily on clarification and evidence regarding robustness to broader analog IMC nonidealities.

AuthorsOpenReview ↗

Response to Reviewer iah6

We thank the reviewer for acknowledging that our additional ADC-aware empirical-MSE comparisons help clarify the optimization efficiency of IMC-CLINIC.

We also appreciate the request for more explicit details on analog IMC nonidealities. These experiments are included in Appendix H.2 of the revised manuscript, but our previous response did not sufficiently summarize their settings and quantitative results. We clarify them below.


Robustness to Analog IMC Nonidealities (Appendix H.2).

We extend our simulator to separately evaluate eight nonidealities, including MAC/ADC gain and offset variations, ADC integral nonlinearity (INL), static device mismatch, spatially correlated noise, and signal-dependent noise.

Importantly, all clipping factors are calibrated only once under nominal conditions and held fixed throughout the nonideality sweeps, without recalibration or modification of the analytical surrogate. This directly evaluates robustness to previously unmodeled hardware perturbations. The following results report WikiText-2 perplexity (PPL), where lower is better:

NonidealityPerturbationW/A Grid Search PPLIMC-CLINIC PPL
NominalNone19.1212.27
MAC gain variationσ = 2%18.2312.46
MAC offset variationσ = 0.1 LSB31.1413.40
ADC gain variationσ = 2%17.6412.61
ADC offset variationσ = 0.1 LSB32.1913.33
ADC integral nonlinearity (INL)Peak = 0.5 LSB332.1523.63
Static device mismatchσ = 4%19.9012.78
Spatially correlated analog noiseρ = 0–1~21.90~12.65
Signal-dependent analog noiseCoefficient = 0–221.90–22.13~12.65

Across all evaluated perturbations, IMC-CLINIC consistently outperforms conventional W/A grid-search clipping. Notably, under 0.5-LSB ADC INL, IMC-CLINIC achieves a PPL of 23.63 versus 332.15 for W/A grid search, despite neither method being recalibrated for this nonideality. These experiments demonstrate that the accuracy advantage of IMC-CLINIC extends beyond the nominal Gaussian-noise model to a broader range of modeled hardware imperfections.

We emphasize that these are behavioral sensitivity studies rather than a fully silicon-calibrated circuit model. Our framework targets high-SNR IMC architectures with sufficiently linear MAC operations; detailed modeling of IR drop and substantial accumulation nonlinearity remains outside the present scope.


We hope these quantitative results and the clarified calibration protocol directly address the reviewer's concerns, and we appreciate the continued feedback.

Reviewer_QYEHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel clipping method to mitigate ADC quantization errors in IMC. Since ADC faces an inherent trade-off between dynamic range and resolution, the proposed method applies clipping in the digital domain to truncate the long-tailed distributions of weights and activations. This reduces the dynamic range of the psum within each computing tile, thereby improving the ADC resolution. Beyond the conventional rounding and clipping errors considered in traditional quantization, the paper further incorporates ADC quantization error into its optimization objective. Finally, the paper employs second-order derivatives and Newton's method to jointly optimize the three parameters (\alpha), (\beta), and (\gamma), thereby minimizing the overall loss function (L(\theta)).

  1. The paper provides extensive experimental results to validate the approximations made in its theoretical derivations. 2.By combining a surrogate error metric with a safeguarded Newton optimization method, the paper achieves a 10× speedup in the search process and an accuracy improvement of 6.5–11.5 percentage points.

1.weak baseline. In the accuracy evaluation, the authors select W/A grid search as the baseline under the same IMC configuration. However, W/A grid search is designed to optimize weight and activation clipping for digital hardware and does not account for ADC quantization errors. Therefore, comparing the proposed method directly against this baseline makes it difficult for readers to assess how much accuracy improvement can actually be attributed to the proposed optimization method. 2.The paper demonstrates that the proposed optimization method nearly reaches the performance limit implied by the surrogate error metric. However, it remains unclear how large the gap is between this surrogate-based limit and the actual achievable accuracy. Given that W8A8 quantization in the digital domain can already achieve nearly lossless performance, it is worth noting that the 9-bit quantization results reported for the analog domain still show a noticeable accuracy degradation compared to FP16.

Critical strengths

See above

Critical weaknesses

See above

Questions for authors

See above

1 comment
AuthorsOpenReview ↗

Response to Reviewer QYEH

We thank the reviewer for recognizing the novelty of our ADC-aware clipping optimization for IMC and its improvements in both accuracy and efficiency. We also appreciate the reviewer’s questions regarding the strength of our baselines and the gap between the surrogate objective and achievable model accuracy. We address both concerns below.

Additional Experiments against the Strengthened Baseline

Please refer to our response to Reviewer iah6 in "Additional Experiments against the Strengthened Baseline" and Appendix H.1 in the revised manuscript.

Mismatch between Optimization Proxy and actual Achievable Accuracy

We acknowledge the fundamental mismatch between local surrogate objectives and end-to-end model accuracy. Directly optimizing the final downstream accuracy for every projection would require repeated full-model evaluations, leading to much longer calibration time. This is a fundamental trade-off in post-training quantization, and existing methods commonly adopt local objectives rather than directly optimizing task-level metrics ([1, 2, 3]).

The key question is how well the surrogate captures the sensitivity relevant to downstream accuracy, and our results suggest that it is already a well-aligned proxy. In our experiment in Appendix F.11, we additionally incorporate the empirical Fisher information matrix into our objective to account for downstream sensitivity. However, this extension does not provide meaningful accuracy improvements, suggesting that there are limited room for downstream sensitivity to improve over our surrogate.

In addition, we highlight that thanks to the local scope of our surrogate, we enables efficient optimization and empirical validation of ε-optimality across all projections, which would otherwise be considerably more challenging with more complex objectives.

References

[1] Frantar, E., Ashkboos, S., Hoefler, T., & Alistarh, D. (2022). Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323.

[2] Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., ... & Luo, P. (2024, May). Omniquant: Omnidirectionally calibrated quantization for large language models. In International Conference on Learning Representations (Vol. 2024, pp. 45472-45496).

[3] Li, Y., Gong, R., Tan, X., Yang, Y., Hu, P., Zhang, Q., ... & Gu, S. (2021). Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426.

Reviewer_mFZCOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

IMC-CLINIC addresses post-training clipping calibration for LLM inference on analog in-memory computing, where operand quantization and ADC quantization jointly affect MatMul error. The authors derive an output-MSE surrogate that combines operand rounding and clipping, a signed bias term, and an ADC contribution that depends on activation and weight scales. They optimize per-projection clipping factors using a safeguarded Newton-type method. Under the simulated 8-bit operand, 9-bit ADC setup, the method improves zero-shot accuracy and calibration time relative to a W/A grid search based on a digital objective. The paper also checks surrogate fidelity and provides a numerical near-optimality certificate for the surrogate in selected models.

Critical strengths
  • The paper models a real difference between analog IMC and ordinary digital quantization: clipping the operands also changes the effective ADC error after rescaling. The coupled objective is well motivated, and the ablation without the ADC term makes its importance clear. The authors also examine the approximations behind the surrogate rather than assuming they hold.
  • The optimization and calibration analysis is well done. The comparisons with gradient-based methods, calibration sensitivity studies, and branch-and-bound checks make a credible case that the proposed solver is efficient for the stated surrogate. This is very useful in an IMC deployment flow.
Critical weaknesses
  • The principal W/A grid-search baseline does not optimize the same analog-aware objective. That makes it hard to separate the value of the loss model from the value of Newton optimization. The authors already include an ADC-aware alternating search, which comes much closer on perplexity in the illustrated case. I would like its accuracy and calibration cost on all four models. This is the comparison that would directly establish the advantage of the optimizer. IMC-specific range-setting methods are also worth discussing, although they may not be directly interchangeable with operand clipping.
  • The system-level trade-off remains open. The default 9-bit simulation still has a meaningful gap to FP16 on some models, and the appendix reports a 10-bit no-clipping configuration with better perplexity than the default 9-bit calibrated one in a representative case. That does not by itself make the 9-bit result unattractive. For instance, an extra ADC bit may cost energy and area. But the paper does not quantify that trade-off. A quality-versus-ADC-energy or area curve, under a stated circuit model, would make the practical case much stronger. Also, the optimality certificate applies to the surrogate, not directly to task accuracy.
Questions for authors
  • Please report the ADC-aware alternating search, with its calibration time, for all four models in Table 1. How much of the gain over W/A grid search remains once the baseline objective accounts for the ADC?
  • How does IMC-CLINIC compare with (a) ADC input range clipping (OCC or CIMQ), (b) the configurable ADC range of Lee et al. (2024), and (c) a statistical setting of the ADC full-scale range (for example k sigma), at the same ADC bits? Can operand clipping and ADC range setting be combined?
  • Can you add an energy versus accuracy plot over ADC bits and methods, to show when 9-bit with IMC-CLINIC is better than a 10-bit ADC? Even an explicitly labeled first-order ADC energy model would clarify the trade-off.
  • The objective is per-projection output MSE, optimized layer by layer. Have you tried accounting for error propagation across layers, for example sequential calibration on quantized inputs or block-output weighting? The empirical-Fisher result suggests the per-layer objective is not fully aligned with task quality.
  • How much do the results depend on the Hadamard rotation? Does IMC-CLINIC still help without rotation, when activation outliers are larger?
  • The noise model covers additive noise at the ADC input only. For other IMC technologies (RRAM, PCM) with programming variation and drift, would the surrogate still be accurate, or is the method specific to high-SNR capacitor-based SRAM?
  • Appendix A notes that decode is often memory-bound and that weights do not fit on chip. Can you estimate the system-level energy benefit for an end-to-end LLM inference scenario, to motivate why accuracy at this ADC precision matters? Minor: please give the calibration time of the ADC-aware search for each model, and state whether the reported 10-12x speedup includes initialization.
2 comments
AuthorsOpenReview ↗

Response to Reviewer mFZC (Part 1)

We thank the reviewer for highlighting our contribution to identifying the critical quantization gap between IMC and conventional digital compute, and for recognizing our clipping optimization and analysis. We also appreciate the reviewer’s insightful feedback, particularly regarding how IMC-CLINIC can work with ADC input clipping and how rotation interacts with clipping. We address each point below.

Additional Experiments against the Strengthened Baseline

Please refer to our response to Reviewer iah6 in "Additional Experiments against the Strengthened Baseline" and Appendix H.1 in the revised manuscript

System-Level Tradeoff versus digital and higher-precision ADC

We thank the reviewer for pointing out that 10b ADC No-Clipping case performs better than 9b ADC IMC-CLINIC case. However, we clarify that IMC-CLINIC improves accuracy at its given ADC precision, not only for 9b case. The design choice of using 9b or 10b for accuracy-energy tradeoff remains, but IMC-CLINIC benefits both cases as shown in Appendix F.4.

We also acknowledge that a noticeable gap to FP16 remains. However, the incentive of analog IMC is that it offers substantial compute-energy advantages over digital MACs (approximately 24.9× under the comparison in Appendix A.3), which motivates our efforts to close this accuracy gap.

We agree that system-level accuracy–energy analysis is important. Such trade-offs and approaches to closing the accuracy gap with digital computation have been explored in [2], demonstrating that reconfigurable analog IMC systems can achieve 6–11× system-level energy-efficiency gains, depending on the accuracy requirements. IMC-CLINIC, on the other hand, focuses on improving accuracy at a given ADC precision, which is complementary to system-level ADC optimization.

In Figure 22 (Appendix H.5 in the revised manuscript), we further demonstrate that IMC-CLINIC shifts the accuracy–energy trade-off toward higher model fidelity. Following the system-level energy modeling framework of [2], we evaluate token-probability KL divergence and validation perplexity on WikiText-2 across ADC precisions from 8 to 12 bits. These results demonstrate that IMC-CLINIC improves model fidelity across ADC precisions, complementary to the fundamental accuracy-energy trade-off for analog IMC.

ADC-input clipping

We thank the reviewer for highlighting this. As discussed in Appendix B.2, directly clipping ADC inputs in analog domain can improve accuracy but requires additional hardware assumptions. We have added a subsection in Appendix H.4 of the revised manuscript to (1) explain the practical hardware constraints of analog-domain clipping and (2) provide experimental results under a more realistic ADC-input clipping scenario.

First, our IMC-CLINIC framework can optimize ADC-input clipping seamlessly. The surrogate loss (eq. 9) natively captures the ADC step size ΔADC\Delta_{\mathrm{ADC}} change in the LADC\mathcal{L}_{\mathrm{ADC}} term. The only addition to the optimization is the ADC-input clipping factor.

The reason why we do not incorporate this degree of freedom is due to how ADC-input clipping is typically done in hardware. It shrinks the ADC range by reducing the ADC reference voltage, which (1) incurs substantial reconfiguration overhead for projection-specific clipping factors and (2) increases the relative impact of analog noise and other circuit nonidealities.

We therefore provide results under a more hardware-practical scenario, where the ADC input-clipping factor is fixed across projections. The results show that ADC input clipping improves perplexity for all evaluated methods, while IMC-CLINIC consistently achieves the lowest perplexity.

Error Propagation

(Regarding the proxy objective choice and its mismatch to final task accuracy, please refer to our response to Reviewer QYEH in “Mismatch between Optimization Proxy and actual Achievable Accuracy”)

We thank the reviewer for highlighting the design choice of whether to account for error propagation across blocks. During development, we evaluated both approaches: (1) using quantized projection inputs that capture errors propagated from preceding blocks, and (2) using FP projection inputs. We found that accounting for upstream error propagation did not meaningfully improve accuracy. For example, approach (1) achieves a slightly worse 12.6612.66 WikiText-2 PPL on LLaMA-3.2-3B in the default setting, while approach (2) achieves 12.2712.27. We therefore adopted approach (2), which additionally allows clipping calibration to be performed independently and in parallel.

AuthorsOpenReview ↗

Response to Reviewer mFZC (Part 2)

Dependence on Hadamard Rotation

We thank the reviewer for pointing this out. We agree that Hadamard rotation may implicitly benefit IMC-CLINIC by reducing the clipping sensitivity of certain channels. For example, removing/clipping certain spike channels can have a catastrophic effect on model outputs [1], whereas rotation amortizes these spikes across channels. The interaction between rotation and clipping is therefore worth exploring.

However, our use of Hadamard rotation is for improving the quantization quality across all methods. ADC quantization is highly sensitive to extreme activation values: these spikes and outliers expand the activation dynamic range, requiring a proportionally wider ADC input range and thus substantially increasing quantization error.

We provide additional experiments in Appendix H.3 of the revised manuscript, showing that clipping alone does not address activation outliers/spikes, and that Hadamard rotation is essential for achieving reasonable accuracy with analog IMC.

Extended Robustness Evaluation Under Analog IMC Nonidealities

Please refer to our response to Reviewer iah6 in "Extended Robustness Evaluation Under Analog IMC Nonidealities" and Appendix H.2 in the revised manuscript

Clarification Questions

  • The reported speedup only considers “clipping calibration time”, which does not include times spent on initialization (e.g., loading model weights, rotating weights offline).
  • The clipping calibration times across models for the ADC-aware search are provided in Appendix H.1 in Figure 14 and 15, with the speedup summarized in Table 15.

References

[1] Sun, M., Chen, X., Kolter, J. Z., & Liu, Z. (2024). Massive activations in large language models. arXiv preprint arXiv:2402.17762.

[2] Chen, Y. C., Stouraiti, O., & Verma, N. (2026). SMART-AIMC: Sensitivity-Aware Mixed ADC Resolution with Fine-Grained Tunability for Analog In-Memory Computing. arXiv preprint arXiv:2611.01707.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response

We thank the reviewers for recognizing our contributions to the mathematical formation and clipping methodology, and for providing thoughtful feedback. We also thank the Area Chair for organizing the review and discussion. We appreciate the time and effort devoted to the review process.

We provide additional experiment results and clarifications in Appendix H in the revised manuscript. To address the reviewers' concerns, we include extensive comparisons to our strengthed ADC-aware baseline (Appendix H.1) and robustness studies covering a broad range of analog IMC nonidealities (Appendix H.2). We also provide a rotation ablation (Appendix H.3) and additional ADC input-range clipping experiments (Appendix H.4) in response to Reviewer mFZC's thoughtful suggestions. Furthermore, we conduct additional experiments to address Reviewers mFZC and QYEH's questions regarding IMC-CLINIC's impact on the accuracy–energy trade-off across different ADC precisions (Appendix H.5).

AuthorsOpenReview ↗

Follow-Up on Reviewer Responses

We thank the reviewers again for their constructive feedback. As the discussion period is about to end, we would greatly appreciate your feedback on whether our responses have adequately addressed your concerns. We would also welcome the opportunity to provide any further clarifications while the discussion period remains open.

We sincerely thank the Area Chairs and reviewers for their time and effort throughout the review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.