- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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?
- 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
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, , 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 to 5.0% at and 3.5% at . Full-window attention becomes the bottleneck, and its share of the latency grows from about a quarter at to more than 80% at . 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.
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.
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=160 rows, of which Nactive=16 are activated at once, and the shared ADCs read the columns in FMUX=16 groups. One analog cycle, covering input setup, array settling, and one ADC conversion, takes tcycle=11.75 ns, as reported for the 1-Mb ReRAM CIM macro of Xue et al. [1]. For input dimension Kl, one analog pass takes Ncycle,l=⌈min(Kl,Rtile)/Nactive⌉FMUX cycles. All static-weight layers after the input encoder have Kl≥160, so one pass takes τl=160tcycle=1.88 µs.
Tile counts. The fully spatial mapping of Table 1 uses 10,780, 12,348, and 21,100 logical 160×160 tiles for Falcon-Medium on GSC, Falcon-Large on GSC, and Falcon-Medium on SSC.
Digital engine arbitration. Attention runs on five shared 32×32 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.
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.