- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper presents a method called FCPQ for low precision post training quantization for SRAM-based ACIM. The method is built on a hardware constraint where the quantization parameters need to be shared across output features. This is a constraint that does not exist in GPU settings and is not addressed in most PTQ methods. The main contribution of the paper is the FCPQ method which has two key components:
- Feature Correlation Calibration (FCC): The method calculates normalized vectors per output channel. Then it calculates the dot product of each vector that is within the same group and share the quantization parameters. Then it creates this loss that tries to preserve the relations of the FP model to the quantized one.
- Correlation aware quantization optimization: Create a feature importance weight that scales the reconstruction objective giving more emphasis on features that are highly involved in the original relation structure and suffer larger relation changes after quantization. The paper does an evaluation on three different CNN architecture, three different quantization settings and on two benchmarks.
Critical strengths
- The setting is unexplored: PTQ method under hardware-imposed scale sharing across output channels in a problem not addressed by PTQ methods.
- The method is cheap to apply and works with just a calibration dataset. The direction of the results is consistent and the method seems to outperform the other baseline methods.
Critical weaknesses
Critical Weaknesses:
-
Section 4.5 is misleading. The text describes the macro as “fabrication-ready, post-layout,” while the caption of Table 5 reports “measured” accuracy on a “fabricated” chip. As a result these experiments are done in simulation something that is not mentioned anywhere. The simulator and noise-injection model are not described. The first row of Table 5 is the same as the first row of Table 1, so the authors imply that there in zero effect of the SRAM-based ACIM hardware on accuracy when operates on nominal conditions. Furthermore it is unclear how much is the reduced resolution in “Reduced ADC resolution”, what is the Accum. noise, σ = 2% FS, what does it mean 10% IR-drop and what is the 5% drift on SRAM-based ACIM?
-
The gains are not statistically supported and baseline comparison is unclear and incomplete. The gains that are presented in the tables of the paper are relatively small, as a results it is unclear if the method is consistently better or the improvement is within statistical error. The authors present results from two resent methods BOF4 and AnyBCQ that present results for LLMs. Especially the BOF4 method is not clear how it can be used for ACIM as it uses a codebook. There are methods that are missing from the comparison like PD-Quant and QDrop which have made for CNNs, reportedly beat BRECQ and the authors cite them in the related work.
Questions for authors
- Which framework did you use for the simulation, what is the noise model you used for evaluation and what are the details of the noise configurations on Table 5.
- Can you explain what does it mean 5% drift for an SRAM-based ACIM?
- What is the difference between the Adaptive rounding in section 3.2 and AdaRound? Adaround should be cited in section 3.2. In the AdaRound paper in equation (24) \beta is decreasing in this paper equation (7) \beta is increasing. Can you explain this difference?
- An equation with the total loss is missing in section 3.5. Do you use also the L_round that is specified in section 3.2?
- The method presented in FCC is a standard correlation metric, can you comment on the novelty of this contribution and if it has been used in other similar works?
- Can you elaborate if a codebook quantization method like BOF4 can be used in SRAM-based ACIM?
- Can you explain the process for obtaining the results for W3A3 and W2A4 results for BOF4, as the method is made for 4-bit quantization.
- The same hyperparameters have been used for the training of the all baseline methods. How are these hyperparameters have been selected? Are these the optimal hyperparameters for FCPQ. Can you demonstrate that these are the optimal hyperparameters for all the methods?
- On equation 18 you specify us, ue which define the activation interval. In the experiments this values are not specified. Can you explain how you specify these values and what values you use in the results.
4 comments
Q2: Baseline Scope and Statistical Support
Response
We sincerely thank the reviewer for this valuable comment. To assess the consistency of the gains and complete the baseline comparison, we report matched CNN experiments and paired uncertainty estimates.
(1) Statistical support. To test whether FCPQ performs consistently across architectures and bit widths, we compare four architectures and 12 model-bit settings under matched checkpoints, calibration data, group size, optimization budget, and evaluation split. FCPQ ranks first in 11 settings; the only exception is a near-tie on ResNet-50 W4A4. For ResNet-18 W4A4, we use paired bootstrap on the same 50,000 ImageNet validation images with 10,000 resamples.
| Paired comparison | Difference (pp) | 95% CI |
|---|---|---|
| FCPQ - AnyBCQ | +0.06 | [-0.03, 0.15] |
| FCPQ - PD-Quant | +0.11 | [0.02, 0.20] |
| FCPQ - QDrop | +0.23 | [0.12, 0.34] |
The intervals are above zero for PD-Quant and QDrop; the AnyBCQ comparison is close. The results across architectures and bit widths show that FCPQ's ranking is not limited to one setting.
(2) CNN baselines and matched tuning. To compare with the CNN-oriented methods cited in the paper, we evaluate QDrop and PD-Quant under the same checkpoint, W/A precision, mapping-group size, calibration data, evaluation split, and optimization budget. Each method retains its native objective and uses validation-selected parameters. QDrop selects its drop probability from {0.1, 0.2, 0.3, 0.4}; PD-Quant selects a multiplier from {0.25, 0.5, 1, 2} of its default prediction-difference coefficient.
On ImageNet-1K, FCPQ outperforms QDrop and PD-Quant in all 12 model and bit-width settings.
| Model | Bit | QDrop | PD-Quant | FCPQ |
|---|---|---|---|---|
| ResNet-18 | W4A4 | 69.18 | 69.32 | 69.41 |
| ResNet-18 | W3A3 | 65.53 | 66.11 | 66.42 |
| ResNet-18 | W2A4 | 56.52 | 56.53 | 58.49 |
| ResNet-50 | W4A4 | 74.82 | 75.08 | 75.20 |
| ResNet-50 | W3A3 | 71.27 | 71.84 | 72.80 |
| ResNet-50 | W2A4 | 66.76 | 66.94 | 68.34 |
| MobileNetV2 | W4A4 | 68.15 | 68.28 | 69.33 |
| MobileNetV2 | W3A3 | 62.86 | 63.14 | 64.31 |
| MobileNetV2 | W2A4 | 52.82 | 54.54 | 54.80 |
| RepVGG-A0 | W4A4 | 68.92 | 69.07 | 69.44 |
| RepVGG-A0 | W3A3 | 63.83 | 64.15 | 65.15 |
| RepVGG-A0 | W2A4 | 54.51 | 55.02 | 55.81 |
To examine the comparison on another dataset, we use the same matched protocol on CIFAR-10. FCPQ achieves the highest Top-1 accuracy among the three methods in 11 of the 12 settings.
| Model | Bit | QDrop | PD-Quant | FCPQ |
|---|---|---|---|---|
| ResNet-18 | W4A4 | 93.72 | 93.87 | 94.13 |
| ResNet-18 | W3A3 | 91.47 | 91.64 | 92.26 |
| ResNet-18 | W2A4 | 84.98 | 85.17 | 86.41 |
| ResNet-50 | W4A4 | 94.26 | 94.34 | 94.48 |
| ResNet-50 | W3A3 | 92.31 | 92.45 | 93.04 |
| ResNet-50 | W2A4 | 86.63 | 86.82 | 87.96 |
| MobileNetV2 | W4A4 | 92.74 | 92.82 | 93.15 |
| MobileNetV2 | W3A3 | 89.88 | 90.01 | 90.73 |
| MobileNetV2 | W2A4 | 82.32 | 82.50 | 83.02 |
| RepVGG-A0 | W4A4 | 92.93 | 93.06 | 92.97 |
| RepVGG-A0 | W3A3 | 90.35 | 90.44 | 91.21 |
| RepVGG-A0 | W2A4 | 82.30 | 82.51 | 83.85 |
(3) BOF4 as an algorithmic reference. We include BOF4 because it uses a low-bit PTQ parameterization different from conventional uniform quantization. Under the same CNN checkpoints, bit widths, and data, it serves as a useful algorithmic reference for assessing FCPQ on standard CNN quantization tasks and examining whether its advantage depends on a particular parameterization.
To support different weight precisions, we adapt the codebook size while retaining BOF4's block-wise normalization and codebook optimization procedure. Specifically, we independently optimize 16/8/4 codebook centroids for W4/W3/W2, respectively, and use the same specified A3/A4 activation quantizers as the other methods. W3 and W2 are thus lower-bit adaptations of the original BOF4 principle.
This comparison evaluates algorithmic accuracy only. We do not implement BOF4's codebook storage, decoding, or conductance mapping on ACIM. Its results should therefore not be interpreted as an ACIM hardware deployment or chip-level evaluation.
Q3: Relationship to AdaRound and Adaptive Rounding
Response
Thank you for asking us to clarify the relationship between FCPQ's adaptive rounding in Section 3.2 and AdaRound. FCPQ follows AdaRound's continuous relaxation, STE-based hard rounding, and rounding regularizer. The main difference is the β schedule: AdaRound decreases β, whereas FCPQ increases it linearly from 2 to 20 over 2,000 updates. To examine this difference, we conduct a controlled β-schedule ablation, keeping all other settings fixed.
| β schedule | Top-1 (%) ↑ | Reconstruction error ↓ |
|---|---|---|
| Fixed β=10 | 69.17 | 0.0678 |
| Increasing 2→20 | 69.41 | 0.0649 |
| Decreasing 20→2 | 69.22 | 0.0669 |
| No rounding regularizer | 68.96 | 0.0721 |
The increasing schedule gives the best accuracy and lowest reconstruction error in this comparison, supporting its use under our calibration setting.
Q4: Complete Optimization Objective
Response
Thank you for asking us to clarify the complete optimization objective in Section 3.5. Yes. FCPQ includes the rounding regularizer defined in Section 3.2 at every calibration step. The complete scheduled loss is
Here, gates only the relation terms; remains always included during calibration. The component losses, , and are defined in Section 3.
Q5:Novelty of Feature Correlation Constraint (FCC)
Response
(1) To our knowledge, this is the first groupwise PTQ method designed to explicitly preserve feature correlations under hardware-imposed shared-scale constraints in ACIM. Previous works have explored preserving relationships among feature representations in knowledge distillation [1; 2]. Meanwhile, existing PTQ methods, such as AdaRound and BRECQ, mainly focus on minimizing reconstruction errors [3; 4]. (2) ACIM quantization requires multiple output features within the same mapping group to share quantization parameters, coupling their quantization errors. This makes preserving relationships among jointly quantized features essential for maintaining the original feature structure. Furthermore, existing reconstruction-based PTQ approaches do not explicitly constrain these groupwise feature relationships. To address these issues, our FCC specifically aligns the correlation matrices of floating-point and quantized features within each ACIM mapping group, preserving feature relationships under hardware-imposed shared-scale constraints. (3) The results shown in the tables below demonstrate that FCC improves ImageNet Top-1 accuracy from 68.75% to 69.18% and reduces correlation distortion from 0.1028 to 0.0724. Furthermore, comparisons with alternative relation objectives show that FCC achieves the lowest Pearson distortion and the highest CKA score with comparable reconstruction errors, supporting its effectiveness in preserving feature relationships during hardware-aware quantization.
| Method | FCC | CQO | Top-1 (%) | Gain |
|---|---|---|---|---|
| Adaptive reconstruction | ✗ | ✗ | 68.75 | -- |
| + FCC | ✓ | ✗ | 69.18 | +0.43 |
| + CQO | ✗ | ✓ | 69.03 | +0.28 |
| FCPQ | ✓ | ✓ | 69.41 | +0.66 |
| Relation objective | Reconstruction MSE ↓ | Pearson distortion ↓ | ImageNet Top-1 (%) ↑ | CKA ↑ |
|---|---|---|---|---|
| Covariance matching | 0.0688 | 0.0836 | 68.93 | 0.842 |
| Variance-normalized MSE | 0.0684 | 0.0657 | 69.20 | 0.875 |
| Pearson FCC | 0.0685 | 0.0498 | 69.41 | 0.916 |
[1] F. Tung and G. Mori, "Similarity-preserving knowledge distillation," in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1365–1374, 2019. [2] J. Yim, D. Joo, J. Bae, and J. Kim, "A gift from knowledge distillation: Fast optimization, network minimization and transfer learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7130–7138, 2017. [3] M. Nagel, R. A. Amjad, M. van Baalen, C. Louizos, and T. Blankevoort, "Up or down? Adaptive rounding for post-training quantization," in Proceedings of the International Conference on Machine Learning (ICML), pp. 7197–7206, 2020. [4] Y. Li, R. Gong, X. Tan, Y. Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu, "BRECQ: Pushing the limit of post-training quantization by block reconstruction," in Proceedings of the International Conference on Learning Representations (ICLR), 2021.
Q6: Hyperparameter Selection and Activation Interval Configuration
Response
Thank you for asking how the hyperparameters and activation interval are selected.
(1) Hyperparameter selection. For a fair comparison, we use the same checkpoints, W/A bit widths, mapping-group size, calibration data, and evaluation split. As stated in Section 4.1, all trainable methods use 2,000 Adam iterations, batch size 64, and learning rate in the reported comparison. Each method retains its native objective and selects its method-specific parameters on the same held-out calibration split. The baseline search ranges and selected settings are:
| Method | Search range | Selected setting |
|---|---|---|
| QDrop | Drop probability: {0.1, 0.2, 0.3, 0.4} | 0.2 |
| PD-Quant | Prediction-difference coefficient: {0.25, 0.5, 1, 2} × default | 1 × default |
| AdaRound / BRECQ | Learning rate: {, }; native rounding or block parameters | ; validation-selected native settings |
To examine whether conventional hyperparameter tuning explains FCPQ's gain, we compare default and tuned adaptive reconstruction with FCPQ under the same experimental conditions.
| Method | Optimization strategy | Top-1 (%) |
|---|---|---|
| Adaptive reconstruction | Default configuration | 68.75 |
| Adaptive reconstruction | Tuned configuration | 68.81 |
| FCPQ | FCC + CQO | 69.41 |
FCPQ remains ahead of the tuned reconstruction baseline, supporting the contribution of its relation-preserving design. The selected settings are the best within the tested ranges; they do not establish global optimality for FCPQ or every baseline.
(2) Activation interval. With normalized optimization progress , and define the relation-guidance window, not activation clipping values. Equation (18) sets to zero outside the interval and follows a half-sine within it. This lets reconstruction first establish numerical responses, then introduces relation guidance that returns to zero at the end of the window. We select FCPQ's parameters on the held-out calibration split, searching and over the ranges below with . The reported results use and , so relation guidance peaks at .
| FCPQ parameter | Search range | Selected value |
|---|---|---|
| {0.01, 0.03, 0.10, 0.30} | 0.10 | |
| {0.25, 0.50, 0.75} | 0.50 | |
| {0.05, 0.10, 0.20} | 0.10 | |
| {0.05, 0.10, 0.20} | 0.10 | |
| {0.70, 0.80, 0.90}, with | 0.90 |
To examine how the activation interval affects performance, we compare the selected window with alternative windows, including a full-duration window, under the same calibration setting.
| Relation window | Top-1 (%) ↑ | Reconstruction error ↓ | Correlation distortion ↓ |
|---|---|---|---|
| 69.31 | 0.0651 | 0.0537 | |
| 69.41 | 0.0649 | 0.0498 | |
| 69.33 | 0.0654 | 0.0521 | |
| 69.28 | 0.0657 | 0.0542 |
The selected window achieves the best accuracy and lowest errors in this comparison, supporting the chosen interval.
Dear Reviewer w3FE,
We sincerely thank you for recognizing the unexplored setting of PTQ under hardware-imposed scale sharing across output channels, as well as the practical value of a method that can be applied using only a calibration dataset. We also appreciate your positive assessment that the results are directionally consistent and that FCPQ outperforms the baseline methods. Thank you for your constructive comments and suggestions. Below, we provide detailed responses to address your concerns.
Q1: Clarification of Hardware Simulation and Noise Modeling
Response
We sincerely thank the reviewer for this valuable comment. To clarify the hardware evaluation, we separate post-layout circuit characterization, network-level error injection, and physical chip measurement.
We have revised Section 4.5 and the Table 5 caption to identify these results as a network-level evaluation based on post-layout macro characteristics, rather than as accuracy measured on a fabricated 22-nm chip.
(1) Evaluation scope and framework. To clarify which framework is used for the accuracy evaluation, we use PyTorch 2.4.1 with CUDA 12.4 to run ResNet-18/ImageNet-1K inference and inject the hardware effects before the corresponding conversion or accumulation step. The injected parameters come from the post-layout characterization of the 22-nm SRAM-based ACIM macro. The macro uses a nominal 0.8 V supply, 100 MHz clock, and a 4-bit SAR ADC. The nominal row is the unperturbed software reference with the nominal quantization setting.
To keep the comparison fixed, we use W4A4 ResNet-18 with mapping-group size 32, the same pretrained checkpoint, a 1,024-image calibration subset, an independent 4,096-image diagnostic set, and the complete ImageNet-1K validation set for final evaluation. Trainable quantization methods use 2,000 Adam iterations with batch size 64 and learning rate 10−2; calibration indices, diagnostic indices, loader order, and evaluation split are shared across methods.
To correct potentially misleading experimental wording in the original paper, we have revised the first two rows of Table 5 as follows and made the same change in the paper text and caption.
The nominal simulation shows an accuracy drop for all three methods, clarifying the effect of ACIM hardware under nominal conditions.
To check whether FCPQ's advantage also holds in physical chip evaluation, we evaluate W4A4 ResNet-18 on ImageNet-1K using the ACIM platform used for chip-level evaluation, including an ACIM-based near-sensor neural network computing chip (NS-Engine). FCPQ achieves the highest chip-measured Top-1 among the five methods.
(2) Noise and hardware configurations. To specify the conditions in Table 5, we evaluate each nonideality separately while keeping the model, quantization setting, data, and optimization protocol fixed. The low-supply condition uses 0.75 V. For reduced ADC resolution, we change the nominal 4-bit ADC to 3-bit and keep the same full-scale clipping. For accumulation noise, we add independent zero-mean Gaussian noise, n∼N(0,(0.02FS)2), to the accumulated analog output immediately before ADC quantization. For IR drop, we apply a row-dependent linear attenuation that increases from 0 to 10% across 144 MAC rows. For effective conductance variation, we use the static cell-wise model specified below.
(3) Meaning of 5% variation. For the SRAM-based ACIM macro, 5% effective conductance variation represents a static cell-wise variation in the effective conductance relative to its programmed value G0: G=G0(1+δ), where δ∼N(0,0.052) is sampled once per effective cell, clipped to [−0.15,0.15], and held fixed during inference.