acceptodds
Under review as a conference paper at ICLR 2027

Latency and Accuracy Tradeoffs in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) are attractive for low-power speech command recognition, yet their latency has received far less attention than their energy efficiency, and their multi-timestep execution is widely assumed to make them slower than quantized neural networks (QNNs). This paper challenges the assumption that more local timesteps necessarily imply higher network latency. By overlapping computation across adjacent layers at the timestep level, SNNs may complete execution in less time than a comparable bit-serial QNNs. However, this overlap relies on spikes firing on incomplete inputs, and a spike once generated cannot be withdrawn, so its error persists and reduces accuracy. Waiting for more input before firing would seem to improve accuracy at the cost of reduced overlap. Yet we find and prove that this intuition fails at some layers, where even a small increase in waiting can change spike timing and downstream computation, making the network both slower and less accurate. We therefore propose a Pipeline Delay Search (PDS) method which selects each layer’s delay by balancing task-level accuracy gains against added network latency. We then adapt the selected configurations through spike-based quantization-aware training and bounded tuning of firing thresholds and initial membrane potentials. Together, these steps form Falcon, a framework for Fine-grained Analysis of Latency and Controlled firing which systematically analyzes and optimizes SNN latency under a spatial analog compute-in-memory mapping with shared digital engines. We evaluate Falcon on Google Speech Commands V2 (GSC) and Spiking Speech Commands (SSC), achieving competitive accuracies of 96.31% and 83.02% at modeled network-core latencies of 119.64 and 124.00, respectively. Together, our analysis and results show that SNNs can compute more yet finish faster, and wait longer yet predict worse, highlighting why Falcon matters for both latency and accuracy.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking neural networks, Latency analysis, Energy efficiency, Speech recognition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_E8svOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Focusing on inter-layer pipelining in SNNs, this work analyzes how pipeline delays shape network accuracy and latency. The authors first introduce a delay-controlled firing mechanism and demonstrate that simply cutting local spike-count errors fails to guarantee higher task accuracy, as downstream processing relies heavily on spike timing. Based on this insight, they develop Falcon, a framework that searches for layer-wise delays before fine-tuning both network and neuron parameters. Tested on two speech recognition datasets using a modeled analog compute-in-memory architecture, Falcon delivers compelling trade-offs between latency and accuracy.

Critical strengths
  1. The paper clearly explains why waiting for additional input slots can reduce task accuracy, even as local spike-count fidelity improves. Together with the ablations in Figure 2, this analysis shows why uniform lookahead is an unreliable heuristic and supports the use of layer-wise delay search.
  2. The experimental controls are well designed and directly test the paper’s central claims. In addition to comparisons with matched bit-serial QNNs and full-lookahead baselines, the authors compare PDS with random schedules matched for both delay distribution and modeled core latency.
Critical weaknesses
  1. Translating the single-vector analog cost in Section 2 into full-layer latency across tokens and convolution windows depends on several assumptions that the paper does not specify. Without details such as the total number of tiles, exact analog timing values, and how contention at shared digital engines is handled, it is difficult to assess the relatively modest speedups over quantized baselines. Table 2 also compares different hardware paradigms: Falcon’s spatial analog mapping is evaluated against digital reconstructions of prior methods, making the method’s purely algorithmic benefit difficult to isolate.
  2. The experiments focus exclusively on speech command recognition, leaving open how the method performs on longer sequences, where full-window attention dependencies could bottleneck the pipeline. This limitation does not undermine the speech results, but additional benchmarks, such as CIFAR10-DVS for vision and sequential CIFAR-10 or ps-MNIST for sequence length, would provide a more complete empirical evaluation.
Questions for authors
  1. How do you scale per-vector analog costs to full layers across tokens and convolution steps, and what are the exact timing parameters, tile counts, and digital engine arbitration rules behind your latency numbers? Also, under equal hardware resources, can the matched QNN baseline overlap downstream execution at the token or output-group level once all bits for a group arrive, without waiting on the full upstream layer?
  2. If rate-coding slots and hardware budgets are held constant, how does the speedup over the matched QNN scale with sequence length? Specifically, do full-window attention or digital engine contention eventually bottleneck longer inputs?
  3. Can you provide accuracy and modeled latency results (against the matched QNN under the same hardware setup) on a non-speech benchmark with longer sequences, to show that the benefits extend beyond short audio clips?
4 comments
AuthorsOpenReview ↗

Response to Reviewer E8sv Part 1

1. Response to W1 / Q1 (latency accounting)

Thank you for this careful comment. We agree that these details should be stated explicitly, and we add an appendix with the full latency model to the revised paper.

  • From per-vector costs to full layers. Tokens and convolution windows are replicated spatially and processed in parallel, so one spike slot or bit-plane of a layer takes one analog pass.

  • Timing parameters. Each tile has Rtile=160R_{\mathrm{tile}}=160 rows, of which Nactive=16N_{\mathrm{active}}=16 are activated at once, and the shared ADCs read the columns in FMUX=16F_{\mathrm{MUX}}=16 groups. One analog cycle, covering input setup, array settling, and one ADC conversion, takes tcycle=11.75t_{\mathrm{cycle}}=11.75 ns, as reported for the 1-Mb ReRAM CIM macro of Xue et al. [1]. For input dimension KlK_l, one analog pass takes Ncycle,l=⌈min⁡(Kl,Rtile)/Nactive⌉ FMUXN_{\mathrm{cycle},l}=\lceil \min(K_l,R_{\mathrm{tile}})/N_{\mathrm{active}}\rceil\,F_{\mathrm{MUX}} cycles. All static-weight layers after the input encoder have Kl≥160K_l\ge160, so one pass takes τl=160 tcycle=1.88\tau_l=160\,t_{\mathrm{cycle}}=1.88 µs.

  • Tile counts. The fully spatial mapping of Table 1 uses 10,780, 12,348, and 21,100 logical 160×160160\times160 tiles for Falcon-Medium on GSC, Falcon-Large on GSC, and Falcon-Medium on SSC.

  • Digital engine arbitration. Attention runs on five shared 32×3232\times32 arrays and five vector engines at 100 MHz. A task starts once its inputs are ready and an engine is free. Tasks are served in order of their data-ready time, each on the earliest available engine.

  • Other assumptions. Buffers are sufficient, weights stay in the arrays, data transfers and partial-sum reductions are hidden behind computation, and ConSmax uses a fused lookup-table implementation.

References

[1] Cheng-Xin Xue, Wei-Hao Chen, Je-Syu Liu, et al. Embedded 1-Mb ReRAM-based computing-in-memory macro with multibit input and weight for CNN-based AI edge processors. IEEE Journal of Solid-State Circuits, 55(1):203–215, 2020.

2. Response to W1 (hardware paradigms in Table 2)

Thank you for raising this point. We agree that Table 2 is not a comparison of algorithms alone, and we would like to clarify two points.

  • Table 2 is a system-level comparison. Each method runs on the type of hardware it targets. The prior methods target digital hardware and do not report hardware latency, so we reconstruct it on digital arrays following their designs (Appendix F). They cannot be mapped directly onto analog CIM. For example, SpikeSCR and SpikCommander use 32-bit floating-point weights, which would require quantization and retraining and would change the original methods. In the revised paper, we will label Table 2 as a system-level comparison, explain this choice, and state the hardware of each method.

  • Table 1 shows the algorithmic benefit. In Table 1, the matched QNN, full lookahead, and all delay schedules share the same source QNN, architecture, and hardware model, so their differences come from the algorithm rather than the hardware.

3. Response to Q1 (overlap in the QNN baseline)

Thank you for this question. Yes, the QNN can overlap downstream execution at the token level, and we add this baseline to the revised paper. In both designs, a layer starts processing a token as soon as its inputs for that token are ready, instead of waiting for the full upstream layer. Both designs use the same scheduler, the same dispatch rules, and the hardware resources of Table 1. Under this baseline, the Balanced SNN remains faster.

QNN (µs)Balanced SNN (µs)
Falcon-Medium on GSC122.40111.26
Falcon-Large on GSC188.88170.22
Falcon-Medium on SSC, fully spatial122.82115.44

4. Response to W2 / Q3 (other modalities and longer sequences)

Thank you for this suggestion. We added CIFAR10-DVS for vision and ps-MNIST for longer sequences. Both use Falcon-Medium with a task-specific front-end, and the Balanced SNN and the matched QNN run on the same hardware. On CIFAR10-DVS, the SNN is 9.8% faster, and its accuracy is 1.35 percentage points lower. On ps-MNIST, full-window attention takes most of the latency of both designs, so their latencies differ by less than 2%, and the SNN is 0.50 percentage points less accurate. We add these results to the revised paper.

Accuracy (%), QNN / SNNLatency (µs), QNN / SNN
CIFAR10-DVS55.10 / 53.7596.18 / 86.78
ps-MNIST98.30 / 97.801356.12 / 1380.02
AuthorsOpenReview ↗

Response to Reviewer E8sv Part 2

5. Response to Q2 (scaling with sequence length)

Thank you for this question. We keep Falcon-Medium on GSC, T=7T=7, the Balanced delays, and the hardware fixed, and increase the sequence length from 98 to 392 and 784 tokens. The speedup shrinks as the sequence grows, from 9.1% at L=98L=98 to 5.0% at L=392L=392 and 3.5% at L=784L=784. Full-window attention becomes the bottleneck, and its share of the latency grows from about a quarter at L=98L=98 to more than 80% at L=784L=784. Digital engine contention is not a bottleneck. At every length, the latency grows because of the attention work itself, not because tasks wait for the shared engines. We add this experiment to the revised paper.

We hope these clarifications and new results address your concerns, and we would be grateful if you would take them into account in your assessment.

Thank you for the detailed response and the additional experiments regarding sequence length scaling capability and the experimental comparisons across CIFAR10-DVS and ps-MNIST. These address my concerns, so I have decided to raise my score to 3 weak acceptance.

AuthorsOpenReview ↗

Response to Reviewer E8sv

Thank you for your positive feedback and for raising your score. We will include all the new results and clarifications in the final version of the paper. Thank you again for your time and constructive comments.

Reviewer_zHHbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the latency-accuracy tradeoff in SNNs with pipelines, with an application in low-power speech command recognition. The tradeoff mainly manifests in waiting for more input spikes from the last layer before computing the outputs, which the authors argue could exhibit a non-monotonic increase in accuracy. The authors then propose a Pipeline Delay Search (PDS) method to search for an optimal tradeoff between pipeline delay due to waiting and accuracy.

Critical strengths
  • The studied low-power speech command recognition application is of interest and is a natural application where SNNs have the potential to outperform ANNs and QNNs.
Critical weaknesses
  • The identified "tradeoff" is essentially, as far as I understand, the same as the "unevenness error" and similar notions in the SNN literature showing differences in spike counts due to the temporal order of positive and negative spikes [1,2].
  • The proposed method only demonstrates marginal performance gains and compares only to a simplified QNN pipeline with Q/K/V parallel. The advantage of PDS further shrinks after fine-tuning compared to vanilla conversion.

[1] Hao et al. Reducing ANN-SNN Conversion Error through Residual Membrane Potential, AAAI 2023.

[2] Li et al. BSNN: Towards Faster and Better Conversion of Artificial Neural Networks to Spiking Neural Networks with Bistable Neurons, Frontiers in Neuroscience, 2022.

Questions for authors

What is the relation between the analysis in this work and the unevenness error?

1 comment
AuthorsOpenReview ↗

Response to Reviewer zHHb

Response to W1 / Q (relation to the unevenness error)

Thank you for the question and for pointing us to these two works, which we now cite and compare in the revised paper. We would like to clarify that the tradeoff we study is not the unevenness error itself. It is the tradeoff between latency and accuracy as the delay δ\delta changes.

The unevenness error is what Eq. 7 measures, and it is the starting point of our analysis. Eq. 7 compares a neuron's spike count under delay δ\delta with its count under full lookahead. At δ=1\delta=1, it captures the error studied by Hao et al. [1] and Li et al. [2]. These works aim to reduce this error, whereas we study how it changes with the delay and what this means for accuracy and latency. Our analysis gives two results that these works do not study:

  • Theorem 1 relates this error to the overlap between layers. For a fixed input stream, a longer delay, i.e., less overlap with the next layer, never increases a layer's error, and each additional slot reduces the error by at most one spike. Waiting is therefore a direct way to trade latency for a smaller error.

  • Proposition 1 shows that this error does not determine accuracy. Even though the error of the changed layer never increases as its delay grows (Theorem 1), network accuracy can still drop. In Appendix E, we construct a network in which the error of the changed layer is zero at every delay, yet accuracy first drops and then recovers as the delay increases, because waiting shifts spike timing and changes the decisions of downstream layers. This is a counterexample to the intuition that waiting longer, which never increases the error, always improves accuracy.

To show this on real networks, we added an experiment that compares two ways of placing delays under the same latency budget. The first places delays where they reduce the unevenness error the most. The second is PDS, which places delays where they improve accuracy the most. If the tradeoff were only the unevenness error, the two should perform similarly. They do not. Without retraining, PDS is 0.83, 48.69, and 36.80 percentage points more accurate than placing delays by the error for Falcon-Medium on GSC, Falcon-Large on GSC, and Falcon-Medium on SSC. Reducing the unevenness error therefore does not tell us where to wait.

References

[1] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ANN-SNN conversion error through residual membrane potential. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):11–21, 2023.

[2] Yang Li, Dongcheng Zhao, and Yi Zeng. BSNN: Towards faster and better conversion of artificial neural networks to spiking neural networks with bistable neurons. Frontiers in Neuroscience, 16:991851, 2022.

Response to W2 (PDS gains and the QNN baseline)

Thank you for these comments. PDS decides where to wait, and its effect is largest in the zero-shot setting, where the delays are applied without retraining. In this setting, the Fastest schedule, which uses δ=1\delta=1 at all searchable IF stages, reaches only 11.64% validation accuracy for Falcon-Large on GSC. PDS raises it to 73.28% with only 3.82 µs more latency.

Fine-tuning lets the network adapt to its delays and compensate for part of a poor placement, so a smaller gap after fine-tuning is expected. PDS still helps in two cases. With a limited training budget of 10 epochs, PDS reaches the highest validation accuracy in all three settings (Appendix A.2), so a good placement reduces the training needed to recover accuracy. On the harder SSC task, after full training, PDS reaches 83.02% test accuracy for Falcon-Medium, above all five random placements (82.11±0.31%82.11\pm0.31\%) and 1.78 percentage points above Fastest, with only 7.58 µs more latency. The main contribution of this paper is the latency analysis, and PDS is the tool we use to place delays under a latency budget.

Regarding the QNN baseline, we chose the QNN pipeline with parallel Q/K/V as a case study because it is a common way to run a quantized Transformer. To show that our conclusion does not depend on this choice, we also evaluate a QNN that overlaps consecutive layers at the token level. A layer starts processing a token as soon as its inputs for that token are ready, instead of waiting for all tokens. We apply the same rule to the SNN and use the same hardware for both. Under this baseline, the Balanced SNN remains faster by 9.1% for Falcon-Medium on GSC, 9.9% for Falcon-Large on GSC, and 9.8% for Falcon-Medium on CIFAR10-DVS.

We hope these clarifications and new results address your concerns, and we would be grateful if you would take them into account in your assessment.

Reviewer_rV6rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies an interesting accuracy–latency tradeoff under a special full-lookahead IF neuron formulation, where firing decisions are made after accumulating inputs across all timesteps. The authors observe that downstream task accuracy can tolerate certain deviations in intermediate spike firing decisions relative to the full-lookahead reference. This observation creates opportunities for pipelining the execution of different timesteps across layers. The paper further proposes a simple greedy search algorithm to optimize the tradeoff between inference latency and accuracy.

Critical strengths

The hardware efficiency evaluation is more comprehensive than many prior SNN studies, incorporating data-dependent energy costs rather than relying solely on operation-count-based estimates. This provides a more realistic assessment of hardware efficiency.

Under the assumption of a full-lookahead IF baseline, the problem is well-motivated. The proposed solution is straightforward yet effective in exploring the accuracy–latency tradeoff, with promising experimental results.

Critical weaknesses

My primary concern is the scope and generality of the studied problem. Early firing with incomplete input information is an interesting question with potentially broad system-level implications. However, the paper focuses on a specific full-lookahead IF formulation, rather than the conventional IF/LIF neurons commonly used in prior SNN research.

Specifically, the full-lookahead IF baseline accumulates inputs across all timesteps before generating output spikes (delta=T). In contrast, conventional IF neurons integrate inputs and make firing decisions at each timestep (delta=1). Consequently, the inter-layer pipelining illustrated in Figure 1(a) can be implemented without changing the outputs of conventional IF neurons. Prior SNN acceleration approaches have already exploited the temporal independence of GEMM operations and inter-layer pipelining to reduce latency without introducing approximation errors.

More importantly, I find the problem formulation and its claimed generality somewhat misleading. Both the title and introduction present the accuracy–latency tradeoff as a general challenge in SNN pipelining, suggesting that inter-layer pipelining inherently introduces errors due to incomplete spike inputs. However, as discussed above, this tradeoff primarily arises from the specific full-lookahead IF formulation adopted as the accuracy reference. I believe this distinction should be explicitly clarified and justified.

Furthermore, the full-lookahead formulation is conceptually related to the Few-Spikes (FS) neuron proposed by Stöckl and Maass (2021), which already supports pipelined execution across layers. Koopman et al. (2024) have also investigated relaxing spike synchronization to enable earlier firing, including its impact on network dynamics and accuracy. In addition, numerous prior works have explored lossless inter-layer pipelining and timestep-parallel execution for conventional IF/LIF neurons. While these approaches differ from Falcon's specific partial-lookahead mechanism, I believe the paper should discuss these closely related works and more clearly distinguish its novel contributions from existing SNN pipelining and acceleration techniques.

For these reasons, while the technical execution appears sound, I find the scope of its contribution rather narrow and I am not sufficiently convinced of the novelty and broader impact of the proposed methodology on the SNN research community.

Questions for authors

Most of my concerns are above in the weakness section.

Further questions: The SNN with (\delta=1) suffers significant accuracy degradation immediately after conversion. How would a conventional IF-based SNN trained from scratch perform under the same architecture and latency constraints? Additionally, how would other rate-based ANN-to-SNN conversion methods that directly use conventional IF neurons ((\delta=1)) compare with the proposed approach?

Could the authors clarify the exact QNN configuration considered in this work, particularly regarding weight and activation precision? The current formulation appears to focus on activation quantization, which heavily motivates the use of full-lookahead IF ((\delta=T)) to preserve activation quantization equivalence. However, on other non-CIM hardware, an alternative QNN that bit-serializes lower-precision weights while retaining higher-precision activations could achieve comparable execution latency under a matched bit-serial dataflow, while also reducing weight storage requirements (likely better accuracy for QNN baselines). Such a configuration would also change the motivation for the (\delta=T) conversion formulation. Does this mean that the proposed method and its reported advantages are primarily applicable to activation-quantized QNNs and their converted SNNs under the CIM architecture considered in this paper?

The hardware evaluation is currently limited to CIM. Assuming the use of full-lookahead IF, the proposed framework appears to be portable to other hardware platforms, such as GPUs. Could similar latency improvements be achieved on GPUs?

4 comments
AuthorsOpenReview ↗

Response to Reviewer rV6r Part1

1. Response to W1 (full lookahead versus conventional IF)

Thank you for this observation. Conventional IF is in fact part of our design space. It is the δ=1\delta=1 endpoint of our kernel, and full lookahead is the δ=T\delta=T endpoint. The scope of this paper is the intermediate states between them, with a separate δ\delta per layer.

A neuron that decides after δ\delta input slots lets the next layer start δ\delta slots later, so a smaller δ\delta gives more overlap. However, it also makes the spike count depend more on the order in which inputs arrive, which costs accuracy. Our response to W2/Q1 shows this loss for conventional IF, both after conversion and when trained from scratch. PDS therefore chooses δ\delta for each layer to balance latency and accuracy. We will make this scope clearer in the revised paper.

2. Response to W2 / Q1 (conventional IF and the accuracy–latency tradeoff)

Thank you for raising this point. Below we clarify and justify this distinction, and we will revise the introduction and Section 2 to make it explicit in the paper.

You are right that conventional IF can be pipelined without changing its own outputs. However, conventional IF is still part of the tradeoff we study. A δ=1\delta=1 network is less accurate than the same network with δ=T\delta=T, whether it is obtained by conversion or trained directly. Following your suggestion, we trained Falcon-Medium on GSC from scratch with conventional IF (δ=1\delta=1) at every IF layer. We also converted the matched QNN to an SNN with conventional IF (δ=1\delta=1) at every IF layer, following the standard rate-based conversion [1]:

Falcon-Medium on GSCTest acc. (%)Latency (µs)
δ=1\delta=1, conversion94.28115.82
δ=1\delta=1, trained from scratch95.67115.82
δ=T\delta=T (full lookahead, = QNN accuracy)96.68252.44

We add these results to Appendix L of the revised paper. Both δ=1\delta=1 networks stay below δ=T\delta=T. This is because a δ=1\delta=1 neuron decides on partial input, so its spike count depends on the order in which inputs arrive. The lossless pipelining of conventional IF therefore comes with an accuracy cost, which places conventional IF at one end of the tradeoff rather than outside it.

References

[1] Zhanglu Yan, Jun Zhou, and Weng-Fai Wong. CQ$^{+}$ training: Minimizing accuracy loss in conversion from convolutional neural networks to spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11600–11611, 2023.

3. Response to W3 (comparison with related work)

Thank you for pointing us to these works. In the revised paper, we have added a discussion to the Preliminaries section (Section 2) that cites them and explains how our work differs from FS neurons [2], Koopman et al. [3], and SNN accelerators with timestep parallelism and inter-layer pipelining.

References

[2] Christoph Stöckl and Wolfgang Maass. Optimized spiking neurons can classify images with high accuracy through temporal coding with two spikes. Nature Machine Intelligence, 3(3):230–238, 2021.

[3] Roel Koopman, Amirreza Yousefzadeh, Mahyar Shahsavari, Guangzhi Tang, and Manolis Sifalakis. Exploring the limitations of layer synchronization in spiking neural networks. Transactions on Machine Learning Research, 2025.

4. Response to Q2 (QNN precision and weight-bit-serial QNNs)

Thank you for the question. Our QNN uses 4-bit convolutional and linear weights with per-output-channel scales, 3-bit tensor-wise intermediate activations (including Q/K/V), and a 4-bit attention map, so T=23−1=7T=2^3-1=7. We will state this configuration explicitly in the revised paper.

Our latency analysis targets analog CIM, a platform well suited to SNNs. In CIM, multi-bit weights stay in the array and inputs are applied bit-serially [4]. Each SNN timestep therefore reuses the stored weights without reading them again, which reduces data-movement energy, and a spike slot costs the same analog pass as an activation bit-plane. This is the condition under which timestep pipelining can reduce latency.

Thank you also for suggesting the weight-bit-serial QNN. We agree that it is a valuable design point on non-CIM hardware, and it may well offer competitive latency and accuracy there. It follows a different dataflow, with its own tradeoffs, such as fetching weights from memory and moving multi-bit activations between layers. In the revised paper, we will state clearly that our latency conclusions apply to activation-serial dataflows such as analog CIM, and that on other hardware, other designs such as this weight-bit-serial QNN may be preferable.

References

[4] Cheng-Xin Xue, Wei-Hao Chen, Je-Syu Liu, et al. Embedded 1-Mb ReRAM-based computing-in-memory macro with multibit input and weight for CNN-based AI edge processors. IEEE Journal of Solid-State Circuits, 55(1):203–215, 2020.

AuthorsOpenReview ↗

Response to Reviewer rV6r Part2

5. Response to Q3 (portability to GPUs)

Thank you for the question. Yes, the framework is portable. Our current analysis is carried out for the dataflow of analog CIM, but it is not limited to CIM. For hardware that executes the network in the same order, it can be applied by replacing the per-operation cycle counts. On GPUs, full-lookahead IF is in fact a natural fit. GPUs usually execute the network layer by layer and fold the timesteps into the batch dimension, so each layer already receives all TT timesteps of its input before the next layer starts. Smaller delays pay off mainly on spatial architectures such as analog CIM, which are better suited to SNNs. There, consecutive layers can overlap, and the weights stay in the array across timesteps. We will make this clearer in the revised paper.

We hope these clarifications and new results address your concerns, and we would be grateful if you would take them into account in your assessment.

Thank you for your responses and the new results. I appreciate the additional efforts. Unfortunately, my main concern regarding the scope of this work remains.

I fully understand that the scope of this paper is focused on SNNs with δ>1\delta>1 running on CIM. While I agree that SNN+CIM is a popular research direction, I believe the current problem setting is overly restrictive and does not provide sufficiently generalizable insights to the broader community.

Given the substantial body of SNN research based on standard LIF neurons with δ=1\delta=1, together with their natural support for pipelining and GEMM-parallel execution (e.g., the lower latency demonstrated in your newly added table), I do not see a compelling reason to switch to the δ>1\delta>1 paradigm, which is more favorable to the proposed method. Furthermore, while CIM, especially analog CIM, is an active research topic in the hardware community, I am not convinced that a CIM-specific speedup strategy tailored to SNNs with δ>1\delta>1 has sufficiently broad relevance to the ICLR community. Even at the system level, as discussed in your newly added appendix section, the requirement of at least four consecutive layers to achieve speedup over QNNs further restricts the applicability of the proposed approach.

Again, I appreciate your efforts during the rebuttal, and some of my concerns have been addressed. However, my primary concern regarding the practical impact and broader relevance of this work remains unresolved, and I do not see how it can be adequately addressed within the current scope of the manuscript.

Best,

rV6r

AuthorsOpenReview ↗

Response to Reviewer rV6r

Thank you for the follow-up and for reading our response so carefully. Thank you also for raising these tradeoffs. They helped us think more clearly about the relation between different delays. We fully understand and respect your judgment. Whatever your final assessment is, we sincerely appreciate the time and care you put into our paper and your detailed and meaningful comments.

We would like to share three short thoughts on the points you raised.

1. Why use δ>1\delta>1. We agree that δ=1\delta=1 is the most natural choice for an SNN. However, δ=1\delta=1 also has a cost. When the network is more complex, it needs a large time window TT to keep its accuracy. Instead of increasing TT, we raise δ\delta to 2 or 3 at a few layers. This keeps the time window small and recovers much of the accuracy, with minimal added latency. For example, without retraining, the δ=1\delta=1 schedule reaches only 11.34% test accuracy for Falcon-Large on GSC with T=7T=7. Waiting longer at three selected layers raises it to 71.29% with only 2.1% more latency. We will upload a new version of the paper that adds this to the discussion.

2. Relevance beyond analog CIM. We understand this concern. The latency numbers are indeed specific to analog CIM. The analysis of how the delay changes the spike count and the accuracy (Theorem 1 and Proposition 1) does not depend on the hardware. The latency model can also be applied to other hardware that executes the network in the same order, by replacing the per-operation cycle counts.

3. The four-layer condition. This condition is not a requirement of our method. It is a result of the analysis. Our goal is not to argue that SNNs are always faster than QNNs. It is to state when they are, so that it can guide the design of SNNs at the network level.

Thank you again. We hope these points make our work clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.