acceptodds
Under review as a conference paper at ICLR 2027

A New Perspective on ANN-to-SNN Conversion: Hardware-Aligned Multi-Threshold Neuron Dynamics for Neuromorphic Inference

Abstract

By accumulating spikes over multiple timesteps, rate coding approximates the real-valued activations of an ANN, a principle that underpins numerous ANN-to-SNN conversion methods. However, the converted SNNs typically incur loss in accuracy compared to the original ANN. Existing methods either increase a number of timesteps or introduce more complex neuronal dynamics to reduce this accuracy gap. Yet, both approaches ultimately lead to the same outcome: redundancy either in information or in computation, thereby suffering from efficiency degradation. To address this issue, we propose **PRIME** (**P**otential-**R**esidual **I**ntegration for **M**ulti-threshold **E**ncoding), a **training-free** multi-threshold concurrent spiking neuron framework redesigned from a hardware data representation perspective: a) We introduce the Residual Membrane Potential Amplification (RMPA) mechanism, which replaces rate coding with Integration Coding by actively amplifying the residual membrane potential. RMPA enables the binary representation of spikes to carry substantially more information, achieving near-lossless conversion with as few as two timesteps. b) Complementarily, we implement concurrent spike firing in multi-threshold neurons, where spikes are decoded directly from the binary representation of floating-point data, eliminating membrane potential–threshold comparisons and replacing multiplications with lightweight bit-shift and accumulation operations that align with hardware computation primitives. At and , we achieve competitive performance across both vision and language benchmarks with only 40.6% of the FP32 bit-transmission cost (81.3% of FP16) and 10.5% of the FP32 multiplication energy consumption (35.5% of FP16). The PRIME framework establishes a new design principle for efficient ANN-to-SNN conversion, bridging the gap between algorithmic efficiency and hardware constraints.

Primary area:
applications to neuroscience & cognitive science
Keywords:
ANN-to-SNN, multi-threshold neuron, neuron dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RzbyOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PRIME, a training-free ANN-to-SNN conversion framework that replaces conventional rate coding with residual-potential amplification and multi-bit spike encoding. PRIME further maps floating-point sign/exponent/mantissa bits directly to spike representations and uses shift-and-add operations to reduce multiplication cost. Experiments on CNNs, vision Transformers, and LLMs show near-lossless conversion with very low timesteps, typically T=2,D=3T=2, D=3, while the paper also reports analytical reductions in inter-layer bit transmission and computational energy.

Critical strengths
  1. The paper presents a technically interesting and coherent redesign of ANN-to-SNN conversion.
    The connection between residual membrane potential, binary representation, direct bit parsing, and shift-add inference is conceptually clean and gives the method a clear hardware-oriented motivation. 
  2. The empirical results are strong and broad.
    PRIME achieves near-lossless, training-free conversion with very low timesteps across CNNs, vision Transformers, and LLMs, which makes the method substantially more convincing than a narrow proof-of-concept.
Critical weaknesses
  1. The relationship to modern ANN quantization and digit-serial arithmetic is not sufficiently addressed.
    Numerically, PRIME keeps the exponent and decomposes the mantissa into low-bit digits/spikes, which makes it quite close to low-bit or bit-serial inference. Without stronger comparisons, it is difficult to judge what is fundamentally gained by using the SNN formulation rather than INT8/INT4 ANN inference. 
  2. The hardware-efficiency claims are still supported mainly by analytical estimates.
    The energy analysis accounts for several relevant primitives, but there is no accelerator-level validation. Also, the very low timestep count should be interpreted together with the actual bit budget, since T=2,D=3T=2,D=3 still uses a relatively rich multi-bit representation.
Questions for authors
  1. How does PRIME compare with strong low-bit ANN baselines under a matched accuracy and hardware budget?
    In particular, comparisons against INT8/INT4 activation quantization or digit-serial inference would help clarify whether PRIME provides an advantage beyond simply changing the numerical representation. 
  2. What is the specifically SNN-related benefit of the proposed representation?
    With multi-bit spikes and parallelizable timesteps, PRIME begins to resemble a hardware-oriented floating-point decomposition. I would like the authors to clarify what is gained from retaining the SNN interpretation rather than viewing PRIME as a general numerical encoding scheme (which can also be available for ANN quantization encoding). 
  3. How much of the reported energy advantage is expected to remain in a complete hardware implementation?
    A clearer end-to-end accounting including control, LOD, shifting, accumulation, storage, and the remaining floating-point operations would materially affect my assessment of the hardware significance.
2 comments

We thank the reviewer for raising this important question. We agree that, from the perspective of low-level numerical operations, PRIME shares certain similarities with low-bit ANN quantization and digit-serial arithmetic. However, we would like to further clarify that the core research objective of PRIME is not to propose a general-purpose numerical encoding format independent of SNNs, but rather to address the long-standing problems of encoding redundancy, conversion error, and hardware execution efficiency in ANN-to-SNN conversion.

PRIME is complementary to, rather than a replacement for, quantization techniques such as INT8/INT4. Quantization mainly reduces computational overhead by lowering numerical precision, whereas PRIME improves the information-carrying capacity of each spike through the Residual Membrane Potential Amplification (RMPA) mechanism and Integration Coding, improving the upper bound of conversion error from linear decay with the number of timesteps in traditional rate coding to exponential decay, thereby reducing the information redundancy caused by repeated spikes in traditional SNNs. More importantly, quantized models can still be further optimized using the PRIME framework. Experiments on models quantized to formats such as FP8 and INT4 are presented in Appendix L, "SUPPORT FOR MORE FORMATS," where energy consumption can be further reduced under our TD setting. As shown Table A and Table B, these results demonstrate that PRIME does not merely change the numerical representation, but can be combined with existing ANN quantization techniques to further reduce computational overhead on top of low-bit models.

Table A: Estimated computational energy consumption of FP8 and INT4 quantized ANN baselines and their PRIME-converted counterparts under different (T, D) configurations.

FormatConfigurationEnergy (pJ)vs. Quantized ANN
FP8Baseline0.280100%
FP8PRIME T2D20.0937533.48%
INT4Baseline0.050100%
INT4PRIME T1D20.01687533.75%

Table B: Accuracy comparison before and after PRIME conversion under different (T, D) configurations: MMLU accuracy (%) on FP8-quantized Qwen3-0.6B and ARC-Challenge accuracy (%) on INT4-quantized LLaMA2-7B (FlatQuant W4A4).

ModelConfigurationMetricPerformance
Qwen3-0.6B FP8Quantized ANNMMLU Acc.26.78%
Qwen3-0.6B FP8PRIME T2D2MMLU Acc.26.66%
LLaMA2-7B INT4FlatQuant W4A4ARC-C44.37%
LLaMA2-7B INT4PRIME T1D2ARC-C40.44%

The core value of retaining the SNN framework lies in the fact that PRIME can not only be efficiently executed on existing digital hardware, but also preserve the potential energy-efficiency advantages of spike-driven computation on neuromorphic chips. Through its unique Residual Membrane Potential Amplification mechanism and Integration Coding mechanism, PRIME substantially increases the actual amount of information that can be represented by the same number of spikes, improving the error decay in ANN-to-SNN conversion from linear to exponential and significantly enhancing efficiency, while avoiding the inefficient operation of repeatedly firing spikes to represent the same information in traditional rate coding. We agree that PRIME's binary decomposition and shift-and-accumulate operations can also be applied to general ANN numerical encoding, but this is merely its implementation approach for existing digital hardware and does not change its fundamental SNN nature based on spike firing and membrane potential dynamics. Therefore, PRIME's advantage lies in simultaneously considering efficient spike encoding in SNNs, the potential for future deployment on neuromorphic hardware, and computational efficiency on existing GPU/NPU architectures, rather than merely proposing a new floating-point decomposition method.

We thank the reviewer for the valuable comments. We understand the reviewer's concerns regarding hardware energy-efficiency estimation and the actual bit budget. First, we would like to clarify that our energy consumption analysis considers not only multiplication operations but also major operations such as sign-bit processing, exponent computation, LOD, shift-and-accumulate operations, and cross-timestep accumulation.

Regarding the multi-bit representation issue mentioned by the reviewer, we agree that a low number of timesteps does not necessarily imply low communication overhead. Therefore, in the comparative experiments presented in our paper, we report both the number of timesteps and the actual number of transmitted bits. We also provide energy consumption and data transmission analyses under different (T,D) configurations in Appendix C, allowing trade-offs based on actual hardware conditions and accuracy requirements. As shown in Table 7 of Appendix C.2, under the representative configuration of T=2 and D=3, PRIME requires 13 bits for inter-layer transmission, compared with 16 bits for FP16 and 32 bits for FP32, corresponding to reductions of 18.75% and 59.38%, respectively.

We understand the reviewer's concerns regarding the basis of our energy consumption estimates and the distinction between these estimates and actual hardware inference energy consumption. However, actual power consumption can only be measured after chip design and fabrication. Therefore, our power consumption analysis is based on the statistical results reported in [1]. Furthermore, to provide sufficient margin for practical hardware inference deployment, we adopt more conservative estimates in our theoretical evaluation. For example, for low-power bit-level operations such as Sign-bit XOR and LOD, we uniformly overestimate their energy consumption using the energy cost of an 8-bit addition. Unlike conventional ANN-to-SNN energy estimation methods, which completely ignore the energy consumption associated with spike firing, reset operations, and other processes (a single addition consumes 0.9 pJ, while a single multiplication consumes 3.7 pJ, meaning that approximately four additions or subtractions consume nearly the same energy as one multiplication; ignoring multiple such additions and subtractions may render their energy consumption analyses entirely invalid), these methods also fail to consider the data transmission overhead introduced by a large number of timesteps. We analyze our data transmission volume to determine its impact on data movement and memory access. Furthermore, we carefully account for the energy consumption of all computational processes, including LOD and other operations. Therefore, we expect that the energy-efficiency advantages of our approach will become even more significant once the chip is implemented.

[1] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pp. 10–14, 2014. doi: 10.1109/ISSCC.2014.6757323.

Reviewer_LwS5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRIME. It generates spikes by directly reading the binary representation of FP16 values and uses shift-and-add instead of floating-point multiplication to reduce the computational cost. But to me, this paper is closer to activation quantization and approximate multiplication than to a real SNN.

Critical strengths
  1. The paper tries to redesign ANN-to-SNN conversion from a hardware computing perspective, and the research question itself has some value.

  2. The method directly reads the binary representation of FP16 values and computes smaller partial products through shift-and-add operations. It is straightforward and easy to understand. Using a Leading-One Detector (LOD) to skip consecutive zero bits also has some engineering value.

Critical weaknesses
  1. I feel that this work is closer to quantization than SNN. This is also my main concern.

The authors describe their method as a new form of spiking neuron dynamics, but the actual computation in Section 3.3 does not need actual neuron dynamics. Section 3.3 directly reads the sign, exponent, and mantissa of FP16 values, encodes some bits into spikes, and then uses shift-and-add to perform multiplication. The whole process does not need actual membrane potential integration, threshold comparisons, or neuron firing dynamics.

The whole computation can be simplified to: FP16 Activation → Adaptive Bit Extraction → Low-bit Partial Products → Shift-and-Add → Output

To me, this looks more like adaptive approximate encoding of FP16 activations, followed by using this encoding to optimize multiplication. If we completely remove terms like Spike, Timestep, and Membrane Potential from this paper and directly use the same activation encoding and approximate multiplication kernel in a conventional ANN, the algorithm and computation results do not seem to change at all. So I would like the authors to clearly answer: What fundamentally distinguishes PRIME from an ANN with adaptive activation quantization and approximate multiplication?

  1. According to Section 3.3, the next layer needs to read the FP16 representation of the previous layer’s complete output to determine its own spike pattern.

So the first spike of the next layer may depend on the results from all timesteps of the previous layer. This means the authors actually know the complete activation in advance before splitting it into multiple spikes. Here, timesteps look more like data groups than an actual time dimension.

  1. Why not just use a low-precision ANN? What exactly is PRIME’s advantage over a QNN or an approximate multiplier with the same accuracy, weight precision, and hardware budget? If a conventional ANN can use exactly the same encoding and multiplication modules, why is the SNN representation in this paper needed?

Other technical concerns:

  1. The theoretical error bound in Proposition 1 seems to be wrong. The authors claim in Proposition 1 that:
εPRIME<θ2TD+1.\varepsilon_{\mathrm{PRIME}}<\frac{\theta}{2^{TD+1}}.

But consider a simple counterexample: x=1.140625=(1.001001)2x=1.140625=(1.001001)_2, with θ=1\theta=1, T=2T=2, and D=3D=3. Following the encoding rules in Section 3.3.1, the two spikes are 100 and 100. Since the extra rounding bit is 1, the second spike becomes 101. So the final reconstructed value is:

x^=1+18+132=1.15625.\hat{x}=1+\frac18+\frac1{32}=1.15625.

The actual absolute error is:

∣x−x^∣=164>1128.|x-\hat{x}|=\frac1{64}>\frac1{128}.

This violates the error bound claimed in Proposition 1. How do the authors explain this counterexample?

  1. The SpikeLLM results cited in Table 2 come from the W2A16 configuration, while PRIME’s main method keeps FP16 weights. The two methods use different weight precisions, so the accuracy difference cannot be directly attributed to the encoding method.
Questions for authors

Please see above, embedded in the reviews.

7 comments

Thanks for your insightful comments. At first, we would like to point out that PRIME is a genuine spiking neuron framework based on the residual membrane potential amplification (RMPA) mechanism, derived from neurodynamic optimization and mainly designed for Neuromorphic Inference, as described in Section 2. According to our ANN-to-SNN method, ANN models, across different precision formats, can be efficiently converted into SNN models for deployment on neuromorphic chips with nearly-lossless accuracy.

Secondly, we also provide another equivalent GPU/NPU implementation version under existing binary hardware constraints. To adapt to the architecture of GPU/NPU chips, we innovatively interpret SNNs from a floating-point perspective as shown in Figure2. For example, the shift operation can be seen as the membrane potential amplification; the sign and exponent parts can be mapped to the firing threshold. Why this GPU/NPU implementation version is strictly equivalent to the neuronal implementation (PRIME in Section 2) is detailed in Section 3. Though the arithmetic looks like approximation mulitiplification, it is merely a straightforward SNN implementation on GPU/NPU under current binary hardware constraints.

In a short word, the SNN implementation on Neuromorphic chips and GPU/NPU chips are both considered in our work. On next-generation Neuromorphic chips, the proposed scheme naturally supports the threshold-comparison mechanism to emit spikes; on current-mainstream GPU/NPU chips, it effectively replace the threshold-comparison operation using shift-and-add mechanism . These two implementation are strictly equivalent and highly efficient on their own chip type.

We thank for the insights and comments. This concern origents from the execution mode of LIF neurons.

Two execution modes. Following previous works (SpikingJelly[1]), we distinguish single-step and multi-step execution of the same converted network. In single-step mode, the signal of one timestep is propagated layer by layer, so each layer's weights must be re-loaded at each timestep whereas in multi-step mode, each layer consumes all signals in all timesteps (T=N) of the previous layer , each weight is loaded once and reused T times. Since data loading accounts for an increasingly large share of power consumption in existing chips, many ANN-to-SNN studies have adopted the multi-step mode, where each layer computes the signals of all timesteps and feeds them to the next layer at once, avoiding repeated weight loading and greatly improving efficiency. Since our framework is co-designed for hardwares, the lower energy consumption with the multi-step mode is our first choice.whereas the single-step mode is also supported (see Appendix E, "Single-Step vs. Multi-Step Execution Modes").

[1] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothée Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. SpikingJelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023.

Thank for the insightful obersation of this paper and comments. The original design of PRIME originates from ANN-to-SNN conversion. Therefore, our purpose is to improve the encoding methods, the energy advantage of SNNs is manifested on neuromorphic chips. In addition, we also propose a method that can achieve low energy consumption on the existing GPU/NPU architectures.

We argue that traditional rate coding often transmits redundant information and is therefore inefficient, which motivated us to propose integration coding. Other works have also proposed distinctive coding schemes—e.g., the differential coding in DCGS [1], phase coding [2], and radix coding [3]—all of which advance the algorithmic performance of the ANN-to-SNN roadmaps.

[1] Zihan Huang, Wei Fang, Tong Bu, Peng Xue, Zecheng Hao, Wenxuan Liu, Yuanhong Tang, Zhaofei Yu, and Tiejun Huang. Differential coding for training-free ANN-to-SNN conversion. arXiv:2503.00301, 2025.

[2] Jaehyun Kim, Heesu Kim, Subin Huh, Jinho Lee, and Kiyoung Choi. Deep neural networks with weighted spikes. Neurocomputing, 311:373-386, 2018.

[3] Zhehui Wang, Xiaozhe Gu, Rick Siow Mong Goh, Joey Tianyi Zhou, and Tao Luo. Efficient spiking neural networks with radix encoding. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3689-3701, 2022.

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

The paper proposes PRIME, an ANN-to-SNN conversion framework based on Potential-Residual Integration with Multi-threshold Encoding (RMPA). The central idea is to inject an input activation into the neuron’s membrane potential once, then iteratively amplify the residual potential and generate spikes using an adaptive threshold. The authors argue that this approach reduces threshold-storage requirements and enables efficient hardware implementation. They further propose parsing the FP16 activation representation directly and replacing conventional floating-point multiplication with a conditional shift-add mechanism. The paper reports reduced estimated energy consumption compared with conventional floating-point operations.

Critical strengths
  • The paper is generally clear and easy to follow. The method, mathematical formulation, and proposed hardware components are presented in a reasonably understandable manner. The figures help explain the encoding and multiplication procedures. Although some details of the hardware mapping remain insufficiently specified, the main presentation is not a major weakness.
  • Hardware-software co-design. The paper attempts to co-design activation representation and arithmetic operations, rather than considering the neuron model in isolation.
  • Potential reduction in multiplier cost. The conditional shift-add formulation offers a plausible way to replace conventional floating-point multiplication with simpler arithmetic operations for the proposed activation representation.
Critical weaknesses

Unclear distinction from SNN literature\textbf{Unclear distinction from SNN literature}

The method injects the input activation only at the first timestep. Subsequent timesteps operate on the residual potential to represent the same static activation. In the hardware-oriented implementation, the spike-generation procedure is further expressed as direct parsing of the FP16 mantissa using leading-one detection and bit manipulation.

Thus, the temporal dimension appears to serve primarily as a coding or serialization mechanism rather than as a means of performing temporal computation. The method consequently seems closer to a specialized activation representation combined with multiplier-free arithmetic than to a conventional ANN-to-SNN conversion framework. The paper does not sufficiently establish what capability or advantage the SNN formulation provides beyond this representation.

Insufficient evidence for accelerator-level efficiency\textbf{Insufficient evidence for accelerator-level efficiency}

The proposed conditional shift-add module is described at the level of individual multiplication operations. However, the resulting costs when these operations are replicated across complete neural-network layers are not sufficiently characterized. In particular, the paper does not provide enough detail to assess how the modules are mapped onto processing elements, how much parallelism can be exploited, or how trace alignment and accumulation affect execution.

The operation-level energy estimates are therefore insufficient to establish the claimed practical advantages at the accelerator level. A reduction in the cost of an individual multiplication does not necessarily translate into higher throughput or lower end-to-end energy.

Insufficiently established practical significance of the reported energy savings\textbf{Insufficiently established practical significance of the reported energy savings}

The reported energy figures are based on assumed operation costs. They do not, on their own, account for all implementation-dependent overheads or establish the energy consumption of a complete accelerator under comparable workloads. The paper should distinguish more clearly between estimated arithmetic energy and measured end-to-end inference energy.

Questions for authors
  • What distinguishes PRIME from activation quantization or bit-serial ANN computation?

    • Since the input is injected only once and subsequent timesteps process the residual representation of the same static activation, what computational capability is provided by the spiking-neuron formulation that is not already captured by direct activation decomposition and bit-level arithmetic? How does the proposed method differ fundamentally from a multiplier-free ANN implementation?
  • How are throughput and parallelism determined for a complete network?

    • Could the authors specify the accelerator dataflow, including how the proposed spike-generation and conditional shift-add modules are mapped to processing elements, how partial products are accumulated, and which operations can execute concurrently? In particular, how do the claimed timestep-level parallelism and latency translate into measured or estimated end-to-end inference latency?
  • What overheads are included in the energy estimates?

    • Do the reported energy results include the costs of control and leading-one detection, trace alignment, accumulation, data movement, and memory access? If not, how might these costs affect the reported energy advantage when the method is applied to complete convolutional or linear layers?
  • Are the hardware comparisons based on equivalent implementation assumptions?

    • The paper compares the proposed arithmetic with conventional floating-point operations. Can the authors clarify whether the estimates assume comparable technology, precision, throughput, and parallelism? In particular, how would PRIME compare with an optimized INT8 or low-bit activation implementation, or with a multiplier-free shift-add network, under the same hardware assumptions?
4 comments

Thank you for your insightful comments. We agree that PRIME shares similarities with activation decomposition, bit-serial computation, and multiplier-free ANNs in terms of hardware implementation. However, we believe that optimizing for existing digital hardware does not imply a departure from the research paradigm of ANN-to-SNN conversion. Furthermore, directly parsing the FP16 mantissa and employing shift-and-accumulate operations represent an efficient implementation of PRIME's spike encoding mechanism on digital hardware, rather than the entirety of PRIME's contributions. PRIME is a genuine spiking neuron framework based on the residual membrane potential amplification (RMPA) mechanism, derived from neurodynamic optimization and mainly designed for Neuromorphic Inference, as described in Section 2. According to our ANN-to-SNN method, ANN models, across different precision formats, can be efficiently converted into SNN models for deployment on neuromorphic chips with nearly-lossless accuracy.

Secondly, we also provide another equivalent GPU/NPU implementation version under existing binary hardware constraints. To adapt to the architecture of GPU/NPU chips, we innovatively interpret SNNs from a floating-point perspective as shown in Figure2. For examples, the shift operation can be seen as the membrane potential amplification; the sign and exponent parts can be mapped to the firing threshold. Why this GPU/NPU implementation version is strictly equivalent to the neuronal implementation (PRIME in Section 2) is detailed in Section 3. Though the arithmetic looks like approximation mulitiplification, it is merely a straightforward SNN implementation on GPU/NPU under current binary hardware constraints. In a short word, the SNN implementation on Neuromorphic chips and GPU/NPU chips are both considered in our work. On next-generation Neuromorphic chips, the proposed scheme naturally supports the threshold-comparison mechanism to emit spikes; on current-mainstream GPU/NPU chips, it effectively replace the threshold-comparison operation using shift-and-add mechanism . These two implementation are strictly equivalent and highly efficient on their own chip type.

Below, we address the questions in Q1 one by one.

What distinguishes PRIME from activation quantization or bit-serial ANN computation?

PRIME shares similarities with activation quantization and bit-serial ANNs in terms of low-level operations, but their research motivations are different. Activation quantization mainly reduces computational overhead by lowering numerical precision, whereas PRIME originates from the optimization of SNN neuron dynamics. Through Residual Membrane Potential Amplification (RMPA) and Integration Coding, PRIME improves the information-carrying capacity of spikes, reducing ANN-to-SNN conversion error and the required number of timesteps. Binary parsing and shift-and-accumulate operations are efficient implementations of this spike encoding mechanism on existing digital hardware, rather than the core algorithm of PRIME itself.

Since the input is injected only once and subsequent timesteps process the residual representation of the same static activation, what computational capability is provided by the spiking-neuron formulation that is not already captured by direct activation decomposition and bit-level arithmetic?

The value of the spiking neuron formulation lies in enabling efficiently encoded activations to be directly integrated into the spike firing, membrane potential update, and synaptic integration mechanisms of SNNs, while providing a foundation for deployment on event-driven neuromorphic hardware. PRIME improves the information density of each spike through RMPA, achieving near-lossless conversion with very few timesteps, thereby reducing the temporal and information redundancy of traditional ANN-to-SNN methods.

How does the proposed method differ fundamentally from a multiplier-free ANN implementation?

Although both can replace conventional multiplication with shift-and-accumulate operations, they differ significantly in algorithm design and computational framework. Multiplier-free ANNs mainly optimize the implementation of conventional ANN numerical operations, whereas PRIME first constructs spiking neuron dynamics through RMPA and Integration Coding, and then maps them to hardware-friendly binary operations. PRIME's bit-level operations are not only an equivalent implementation of SNN dynamics on existing GPU/NPU, but also offer the potential for deployment on future neuromorphic chips.

We thank the reviewer for this insightful suggestion. We answer Q2 and weakness2 together in this reply. This work primarily focuses on optimizing the computational units of PRIME, rather than proposing an entirely new accelerator architecture. Specifically, we replace the conventional multiplication units with our proposed PRIME modules, while retaining the existing hardware architecture and implementation strategies for other components. Therefore, existing optimizations for large-scale multiplication, including processing-element mapping, parallel execution, and data reuse, can also be applied to PRIME. We consider it sufficient to evaluate only the multiplication computation and the weight movement. We do not claim the power consumption of the entire network, because it cannot be given accurately. However, once the ratio between data-transfer power and computation power is known (multiplication accounts for more than 99% of the inference process [1][2]), the ratio of the overall network power can be derived.

Regarding hardware mapping, throughput, and parallelism, our planned hardware implementation directly replaces conventional multiplication units with the proposed spike-generation and conditional shift-add modules. Therefore, existing PE mappings and parallel execution strategies for large-scale matrix multiplication can also be applied to our design. In addition, PRIME introduces timestep-level parallelism: when $T \leq2 $, spike generation at different timesteps is independent and can be executed concurrently. As illustrated in Figure 3, sign-bit processing, exponent computation, and mantissa computation can also be performed in parallel, while partial products are aligned according to their corresponding spike traces and accumulated to obtain the final output. Furthermore, PRIME adopts the Multi-Step execution mode described in Appendix E, completing all timestep computations within each layer before proceeding to the next layer, thereby reducing repeated weight loading. Under the representative T=2,D=3T=2,D=3 configuration, PRIME requires 13 bits per encoded value for inter-layer transmission, compared with 16 bits for FP16, further reducing communication requirements.

Regarding end-to-end inference latency, the independence between timesteps when T≤2T\leq2 allows parallel spike generation, eliminating the sequential dependency between these timesteps. Combined with reduced weight reloading in Multi-Step execution, this provides opportunities to reduce inference latency. However, the actual end-to-end latency and throughput depend on the available PE resources, memory bandwidth, partial-product accumulation, and hardware scheduling. Therefore, our current analysis demonstrates the potential for parallel execution and latency reduction, rather than a measured or quantitatively estimated end-to-end speedup, which requires a complete accelerator implementation.

[1] X. Hu, X. Geng, Z. Mao, J. Han and H. Jiang. A Low-Power Mixed-Precision Integrated Multiply-Accumulate Architecture for Quantized Deep Neural Networks. DATE 2025, pp. 1-7.

[2] M. Bouchakour, E. Bergeret, G. Sicard, K. Abdelouahab and F. Berry. A Low Power Fully Analog MAC Operation Using Standard CMOS Technology for Neural Network Inference. NEWCAS 2025, pp. 10-14.

We understand the reviewer's concerns regarding the basis of our energy consumption estimates and the distinction between these estimates and actual hardware inference energy consumption. However, actual power consumption can only be measured after chip design and fabrication. Therefore, our power consumption analysis is based on the statistical results reported in [1]. Furthermore, to provide sufficient margin for practical hardware inference deployment, we adopt more conservative estimates in our theoretical evaluation. For example, for low-power bit-level operations such as Sign-bit XOR and LOD, we uniformly overestimate their energy consumption using the energy cost of an 8-bit addition.

Unlike conventional ANN-to-SNN energy estimation methods, which completely ignore the energy consumption associated with spike firing, reset operations, and other processes [2] (a single addition consumes 0.9 pJ, while a single multiplication consumes 3.7 pJ, meaning that approximately four additions or subtractions consume nearly the same energy as one multiplication; ignoring multiple such additions and subtractions may render their energy consumption analyses entirely invalid), these methods also fail to consider the data transmission overhead introduced by a large number of timesteps [3,4]. In contrast, we analyze our data transmission volume to determine its impact on data movement and memory access. Furthermore, we carefully account for the energy consumption of all computational processes, including LOD and other operations. Therefore, we expect that the energy-efficiency advantages of our approach will become even more significant once the chip is implemented.

Regarding Q3, our energy estimation considers the major computational operations involved in PRIME, including sign-bit processing, exponent computation, leading-one detection (LOD), conditional shift-and-add operations, rounding, and cross-timestep accumulation, as detailed in Appendix C.1. For lightweight operations such as sign-bit XOR and LOD, we adopt conservative estimates by assigning them the energy cost of an 8-bit integer addition rather than neglecting their overhead. In addition to computational energy, we separately evaluate inter-layer data transmission costs in Appendix C.2. Since our proposed modules are intended to replace conventional multiplication units while retaining the existing execution framework, the same computational optimization can be applied to convolutional, linear, and Transformer layers. Therefore, the operation-level energy advantages are applicable to the multiplication computations in these layers, while the actual end-to-end energy savings depend on the relative contributions of computation, memory access, and other hardware overheads in the target accelerator.

[1]Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pp. 10–14, 2014. doi: 10.1109/ISSCC.2014.6757323.

[2]Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ANN-SNN conversion. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), pp. 2501–2508, 2022. doi: 10.24963/ijcai.2022/347.

[3]Sangwoo Hwang, Seunghyun Lee, Dahoon Park, Donghun Lee, and Jaeha Kung. SpikedAttention: Training-free and fully spike-driven transformer-to-SNN conversion with winner-oriented spike shift for softmax operation. Advances in Neural Information Processing Systems, 37:67422–67445, 2024.

[4] Fabrizio Ottati, Chang Gao, Qinyu Chen, Giovanni Brignone, Mario R. Casu, Jason K. Eshraghian, and Luciano Lavagno. To spike or not to spike: A digital hardware perspective on deep learning acceleration. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 13(4):1015–1025, 2023. doi: 10.1109/JETCAS.2023.3330432.

We thank the reviewer for the insightful questions. The energy consumption comparisons in this paper are based on a unified set of hardware operation energy assumptions, with the relevant data mainly derived from reference [1], as detailed in Appendix C. We use the same energy parameters for basic operations and compare conventional floating-point multiplication with PRIME's shift-and-accumulate implementation to evaluate the theoretical energy consumption differences between different computation methods.

To maintain consistency in process technology and hardware assumptions as much as possible, our estimates use basic operation energy data from the same source. For large-scale deployment, PRIME's conditional shift-and-add module can replace the conventional multiplication unit in a processing element (PE), potentially retaining the existing parallel computing structure.

PRIME is complementary to low-bit quantization methods such as INT8/INT4, and quantized ANNs can still be further optimized using the PRIME framework. Experiments on FP8 and INT4 quantized models in Appendix L demonstrate that PRIME can further reduce estimated computational energy consumption under different (T, D) configurations. For multiplier-free shift-and-add networks, PRIME's potential advantage comes from the constraints on spike numerical patterns. Under the typical (T=2, D=3) configuration, each spike–mantissa multiplication requires at most one addition or subtraction per timestep, as shown in Appendix D, whereas direct bit-wise shift-and-accumulate operations on a conventional FP16 mantissa may require more additions. This provides PRIME with the potential to reduce computational complexity.

[1] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pp. 10–14, 2014. doi: 10.1109/ISSCC.2014.6757323.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A Brief Response to Common Concerns

Dear All Reviewers,

Thank you again for your detailed feedback. Below are our brief responses to several key points.

Contribution. First, We would like to reiterate that our main contribution is a high-accuracy and efficient ANN-to-SNN framework, applicable to both neuromorphic chips and GPUs/NPUs. By leveraging RMPA and Integration Coding, PRIME significantly tightens the general error bound for ANN-to-SNN conversion. Thanks to the inherent nature of SNNs, PRIME is naturally suited for efficient deployment on neuromorphic chips. However, on modern GPU/NPU chips, basic SNN operations such as threshold comparison , membrane potential reset, and repeated input injection are not efficiently supported, resulting in inefficient execution of SNN algorithms. Therefore, we co-design a functionally equivalent GPU/NPU implementation under existing binary hardware constraints. Section 3.3 details why the two are strictly equivalent.

Relation to quantization. For more efficient SNN execution on GPUs/NPUs, we redesign SNN computational operations under GPUs/NPUs' hardware constraints to achieve a functionally equivalent hardware mapping. With the same GPU/NPU architectures, it is unsurprising that the primitive operations of our method may resembles those of ANN quantization. But, in essence, our PRIME is a conversion method, which can be used together with general ANNs or quantized ANNs. It is orthogonal to quantization and can be applied to quantized models with different precisions (e.g., FP8/INT4), enabling lower energy consumption while preserving accuracy (Appendix L "SUPPORT FOR MORE FORMATS").

Single input injection. This is one of our innovations and an adaptation to current mainstream GPU/NPU architectures, which avoids repeated weight loading. At the same time, our framework also supports "repeated input injections" , as described in Appendix E ("Single-Step vs. Multi-Step Execution Modes").

Finally, we would like to emphasize that, even setting aside the hardware contributions, we believe our ANN-to-SNN method is also a strong contribution in its own right.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.