PaperlayerView exact revision on arXiv

Source-linked AI summary

Benchmarking spiking neural networks across sensing modalities on edge devices

Xin Du, Di Yu, Changze Lv, Yuqi Zhang, Zhuo Chen, Wentao Tong, Helin Zheng, Weisong Zhang, Xiaofan Zhao, Linshan Jiang, Shijie Ji, Hui Fang, Xiaoqing Zheng, Gang Pan, Shuiguang Deng

arXiv:2609.00026v1cs.NE

TL;DR

Systematic evidence remains limited on when SNNs provide practical advantages across sensing modalities, despite edge systems needing reliable performance under resource constraints. This paper benchmarks SNNs across five modalities and multiple devices, finding broadly comparable performance to ANNs across most workloads but a particularly favorable result for wireless sensing.

  • Problem

    Systematic evidence remains limited on when SNNs provide practical edge-sensing advantages, despite the need to maintain reliable performance across diverse modalities under resource constraints.

    Source: MAIN TEXT INTRODUCTION · p. 2

  • Method

    The paper benchmarks SNNs across five modalities, 24 datasets, seven edge devices, and four neuromorphic platforms, varying encodings, neuron models, and architectures under unified protocols.

    Source: MAIN TEXT INTRODUCTION · p. 2

  • Results

    SNNs were broadly comparable to ANNs across most modalities, while wireless sensing was consistently favorable; deployment benefits also varied across accuracy, latency, energy, and memory.

    Source: MAIN TEXT INTRODUCTION · p. 3

  • Takeaways & Limitations

    SNN deployment should be assessed as a trade-off across modality, model design, encoding, hardware, accuracy, latency, energy, and memory rather than energy alone.

    Source: MAIN TEXT INTRODUCTION · p. 3

    The benchmark does not cover the full diversity of edge-sensing tasks, and its mechanistic analysis focuses on wireless sensing.

    Source: DISCUSSION · p. 9

Abstract

from arXiv · show

Edge computing systems need to support diverse sensing workloads under tight energy and memory constraints, thereby motivating deployment-aware model selection. Spiking neural networks (SNNs) are a promising alternative to conventional artificial neural networks (ANNs), yet systematic evidence for when and why they provide practical advantages remains limited. Here, we present a benchmark of SNNs across five sensing modalities and multiple edge devices, systematically evaluating spike encoding, neuron models, and network topologies under consistent training and deployment protocols. We find that SNN advantages are strongly modality-dependent: while SNNs achieve performance broadly comparable to ANNs across most workloads, wireless sensing emerges as a particularly favorable domain. Frequency-domain and feature-space analyses further explain this result by showing that spiking dynamics naturally align with the spectral-temporal structure of wireless signals. Our deployment analysis further shows that SNN advantages are not one-dimensional, with energy gains often accompanied by modality-dependent system costs. Finally, we provide an open-source framework for reproducible benchmarking and deployment profiling, offering a practical foundation for algorithm-software-hardware co-design on emerging edge and neuromorphic computing platforms.

MAIN TEXT INTRODUCTION

The introduction frames SNN effectiveness as dependent on sensing modality and deployment constraints, motivating a unified benchmark rather than a universal claim of SNN superiority. It reports broadly comparable performance across most modalities, a favorable wireless domain, and the release of a reproducible benchmarking framework.

Motivation

  • Edge intelligence must support diverse sensing workloads under energy, latency, and memory constraints; despite event-driven processing and sparse activation, SNN benefits remain inconsistent across applications.This motivates evaluating practical performance across modalities rather than treating SNN advantages as an architectural property alone.Source: MAIN TEXT INTRODUCTION · p. 2

Evaluation challenge

  • Encoding, neuron dynamics, and network topology interact with input statistics and hardware constraints, making isolated benchmarks and aggregate spike or operation counts incomplete.The modalities also differ in sampling rate, sparsity, noise, temporal continuity, and frequency structure, while prior evidence used inconsistent protocols.Source: MAIN TEXT INTRODUCTION · p. 2

Benchmark design

  • The benchmark spans five modalities, 24 datasets, seven edge devices, four neuromorphic simulators, and 4,662 evaluations under unified protocols.It assesses encoding strategies, neuron models, and architectures using accuracy, latency, energy, and memory measures.Source: MAIN TEXT INTRODUCTION · p. 2

Main findings

  • SNNs perform broadly comparably to ANN baselines across most modalities, while wireless sensing emerges as a consistently favorable domain.The introduction emphasizes that SNN benefits are selective rather than universal.Source: MAIN TEXT INTRODUCTION · p. 3

Wireless analysis

  • Wireless analyses associate this advantage with alignment between channel-dynamic signal structure and the low-pass filtering induced by spiking neuronal dynamics.The proposed mechanism is specific to wireless signal dynamics and does not imply a shared advantage across all modalities.Source: MAIN TEXT INTRODUCTION · p. 3

Deployment analysis

  • SNN deployment must be assessed across accuracy, latency, energy, and memory because energy gains do not guarantee lower latency, memory use, or system-level advantages.The trade-offs depend on modality, model design, encoding strategy, and hardware substrate.Source: MAIN TEXT INTRODUCTION · p. 3

Framework

  • Soul–NeuSim integrates multimodal processing, SNN construction, edge profiling, and neuromorphic simulation to support reproducible comparisons and future co-design.The framework is released as open source.Source: MAIN TEXT INTRODUCTION · p. 3

RESULTS · A modality-aware benchmark for SNN edge sensing · SNN design choices exhibit strong modality dependence

Across five sensing modalities and multiple edge devices, the benchmark finds that SNN advantages are modality-dependent rather than universal, with the clearest advantage in wireless sensing. It evaluates how encoding, neuron dynamics, and network topology shape SNN performance and deployment behavior.

A modality-aware benchmark for SNN edge sensing

  • The benchmark compares spike encoding, LIF neuron dynamics, and network topologies ranging from convolutional backbones to transformer-inspired spiking models.These design dimensions are evaluated for their effects on information preservation, performance, efficiency, connectivity, temporal modeling, and practical deployment.Source: A modality-aware benchmark for SNN… · p. 3
  • The framework evaluates model accuracy and deployment behavior across five sensing modalities under unified protocols, comparing SNNs and ANNs in relation to modality, model design, and hardware constraints.Its analyses cover SNN design choices, wireless signal structure in relation to spiking dynamics, and whether algorithmic benefits translate into gains across edge and neuromorphic devices.Source: A modality-aware benchmark for SNN… · p. 3

SNN design choices exhibit strong modality dependence

  • Across modalities, Direct encoding achieved the highest relative accuracy but was not spike-compatible; among practical encodings, wireless performance was sensitive to encoding choice.For wireless sensing, TTFS and phase coding substantially outperformed rate coding, while Burst and temporal-switch coding reduced performance; motion accuracy remained relatively high across most schemes.Source: SNN design choices exhibit strong… · p. 4
  • No neuron model dominated across modalities: TLIF improved over LIF by up to 4.79% on BullyDetect, while unsuitable designs degraded UrbanSound8K by up to 5.99%.Vision differences were relatively small, motion rankings were stable within 0.45%, and certain variants such as RPLIF remained in the upper tier on acoustic and neuromorphic tasks.Source: SNN design choices exhibit strong… · p. 4
  • SNNs and ANNs showed broadly comparable accuracy on vision, acoustic, motion, and neuromorphic tasks, while SNNs most consistently outperformed ANNs on wireless sensing with a matched LeNet backbone.The comparison shows that SNN advantages are selective rather than universal across edge sensing workloads.Source: SNN design choices exhibit strong… · p. 4
  • Transformer-style SNNs, including Spikformer and QKFormer, achieved the strongest overall performance and more consistent behavior across evaluated modalities than conventional spiking models.Architecture alone did not determine whether SNNs outperformed ANNs.Source: SNN design choices exhibit strong… · p. 4

Wireless sensing emerges as a favorable domain for SNNs · Spectral and feature-space analyses explain the wireless advantage

SNNs perform especially well on wireless sensing, where they match or outperform ANN baselines, often with compact models. Their advantage is associated with neuronal dynamics that align with wireless signals’ low-frequency discriminative structure, suppress noise, and yield more separable features.

Wireless sensing emerges as a favorable domain for SNNs

  • On Fi-HAR and UT-HAR, SNNs reach near-ceiling performance, with SEW-ResNet18 at 100.0% and SpikingResFormer at 99.73%, matching or exceeding the strongest ANN baselines.The cited ANN examples are BiLSTM at 99.69% and ResNet18 at 98.11%.Source: Wireless sensing emerges as a… · p. 4
  • Top-performing SNNs retain compact models, including Spikformer at 0.45M parameters versus ResNet50’s 23.55M, an approximate 52× reduction; results favor their accuracy-complexity frontier.QKFormer uses 1.62M parameters, while lightweight RNNs with 0.03M parameters show substantial performance degradation.Source: Wireless sensing emerges as a… · pp. 4–5
  • On Widar, SpikingResFormer scores 81.42±0.32% versus ResNet18’s 71.70±0.33%; on BullyDetect, QKFormer scores 85.58±0.39% versus ResNet50’s 73.48±1.72%.The reported advantages are 9.72% and 12.10 points, respectively.Source: Wireless sensing emerges as a… · p. 5
  • On AOPHand, Spikformer scores 95.12±0.42% versus the ANN best of 72.54±0.80%; on MMActivity, QKFormer scores 80.00±2.50% versus ViT’s 65.83±5.77%.The reported margins are 22.58 and 14.17 points, respectively.Source: Wireless sensing emerges as a… · p. 5

Spectral and feature-space analyses explain the wireless advantage

  • LIF neurons attenuate high-frequency components as a temporal low-pass filter, whereas ReLU preserves more; real-world spectra and GradCAM also show SNN emphasis on global low-frequency features.The analysis uses a synthesized multi-tone signal and compares real-world feature maps after LIF and ReLU processing.Source: Spectral and feature-space analyses… · p. 5
  • Wireless task information concentrates in low-frequency energy envelopes and Doppler shifts, while high-frequency bands are largely environmental scattering noise; FMS is consistently higher in low-frequency bins across β settings.Increasing β monotonically suppresses non-DC frequencies and narrows the effective passband, yet SNNs can exploit discriminative high-frequency features beyond a traditional −3dB cutoff.Source: Spectral and feature-space analyses… · p. 5
  • SNN features form more separable wireless-data clusters, with a silhouette score of 0.60 versus the ANN’s 0.39; on noisy samples, SNNs filter high-frequency noise that can contaminate ANN features.The ANN struggles when signals fluctuate heavily, while the SNN can classify using cleaner low-frequency motion data.Source: Spectral and feature-space analyses… · p. 6
  • SNN reconstruction suppresses noise and preserves samples with greater fidelity than ANN reconstruction; effectiveness depends on aligning spike encoding, neuronal dynamics, and architectural bias with signal spectral structure.ANN reconstructions contain few useful features and mostly noise, whereas SNN reconstructions act as low-pass temporal filtering.Source: Spectral and feature-space analyses… · p. 6

Deployment benefits are multi-dimensional and hardware dependent

SNN deployment benefits vary by modality, architecture, and hardware: energy reductions can come with higher latency or memory use. The results support evaluating deployment as a multi-objective trade-off across accuracy, latency, energy, and memory.

Deployment benefits are multi-dimensional and hardware dependent

  • 3.07× and 2.84× were SNN-to-ANN latency ratios for event-sensing and wireless workloads, versus 1.63× for motion sensing.The latency gap was smaller on GPU-accelerated than CPU-based devices, suggesting parallelism can partly offset sequential timestep overhead.Source: Deployment benefits are… · p. 6
  • Up to 29.58× energy reductions occurred for neuromorphic event sensing, while vision showed reductions as low as 1.27×.The passage relates this pattern to stronger temporal sparsity in neuromorphic data and dense, frame-based vision inputs.Source: Deployment benefits are… · p. 6
  • SNN deployment should be evaluated as a multi-objective trade-off across accuracy, latency, energy, and memory, because lower energy can accompany higher latency or memory demand.Neuromorphic hardware simulations also showed that latency varied by platform, energy was positively associated with latency, and memory demand depended more on workload than chip configuration.Source: Deployment benefits are… · pp. 6–7
  • 65.56% (paired t-test, p=0.10) and 41.91% (paired t-test, p=0.13) were the mean memory increases for neuromorphic and wireless workloads, respectively.SNNs had higher mean memory use across modalities because membrane states and intermediate temporal activations must be retained during inference.Source: Deployment benefits are… · p. 6
  • SpikingVGG and SEW-ResNet had more compact normalized latency distributions than Spikformer, while QKFormer consumed 64.79–91.83% less energy than larger transformer-style or deeper architectures.QKFormer's strongest energy advantage was observed in wireless tasks; average memory demand was highest for wireless sensing and lowest for motion sensing.Source: Deployment benefits are… · p. 7
  • In neuromorphic hardware simulations, latency varied across platforms, energy was positively associated with latency, and wireless and acoustic workloads imposed greater memory demand than motion sensing.The passage attributes platform latency variation to spike-routing architecture and neuron-core parallelism, while memory depended more on workload than chip configuration.Source: Deployment benefits are… · p. 7

Soul-NeuSim supports reproducible benchmarking and neuromorphic co-design · DISCUSSION

SNNs benefit edge sensing selectively, with wireless workloads showing the clearest advantage when spiking dynamics align with signal structure. Soul–NeuSim provides a reproducible workflow for evaluating these modality- and hardware-dependent trade-offs and supporting co-design.

Soul-NeuSim supports reproducible benchmarking and neuromorphic co-design

  • Soul–NeuSim combines Soul’s multimodal benchmarking front end with NeuSim’s hardware-constrained deployment and neuromorphic simulation back end in a unified reproducible workflow.Soul standardizes data processing, encoding, model construction, training, and deployment evaluation while logging configurations, seeds, hyperparameters, and profiling outputs; NeuSim supports cycle-level estimates and compilation, partitioning, and physical-core mapping. The workflow enables consistent comparison and study of interactions among modality, model design, and deployment platform.Source: Soul-NeuSim supports reproducible… · p. 7, DISCUSSION · p. 9

DISCUSSION

  • SNNs showed no uniform advantage over ANNs across sensing modalities; wireless sensing was the clearest favorable domain, while performance elsewhere was broadly comparable and configuration-dependent.Several spiking architectures substantially outperformed ANN counterparts on wireless workloads, which were more sensitive to spike encoding and neuronal dynamics.Source: DISCUSSION · pp. 7–8
  • Wireless SNN advantages aligned with LIF neurons’ temporal filtering of high frequencies and preservation of slower variations, matching discriminative spectral-temporal structure in wireless signals.The spectral, feature-space, and reconstruction analyses support this mechanism, but do not establish SNN superiority for all temporal signals.Source: DISCUSSION · p. 8
  • Encoding, neuronal dynamics, and topology jointly shaped modality-dependent SNN performance, so design choices made without considering target signal characteristics may mislead.Encoding controls preserved information, neuronal dynamics accumulate, filter, and reset temporal evidence, and topology represents spatial, temporal, and cross-channel dependencies.Source: DISCUSSION · p. 8
  • On conventional edge devices, multi-timestep inference often increased SNN latency and membrane-state memory, while temporally sparse workloads—especially neuromorphic event sensing—showed substantial energy reductions.The results establish a multi-objective trade-off across accuracy, latency, energy, and memory rather than an energy-only comparison.Source: DISCUSSION · p. 8
  • Deployment trade-offs depended on hardware: GPUs partly reduced SNN latency penalties, CPUs exposed sequential overhead, and spiking transformers were more device-sensitive than convolutional models.Neuromorphic simulations further showed that routing, neuron-core parallelism, and workload characteristics jointly affect latency, energy, and memory.Source: DISCUSSION · p. 8
  • The benchmark covers five representative modalities but leaves other sensing domains, broader neuron models, physical neuromorphic deployments, and continual or online learning for future study.Mechanistic analysis focused on wireless sensing and LIF-type dynamics; real-device measurements and simulations offer complementary but incomplete deployment views.Source: DISCUSSION · p. 9

MATERIALS AND METHODS Experimental design

Soul-NeuSim benchmarked SNNs against ANN baselines across sensing modalities, model designs, and deployment platforms under controlled protocols. It also expanded wireless experiments to examine the domain’s apparent SNN advantage and its relationship to signal structure.

Experimental design

  • The study evaluated spike encoding, neuronal dynamics, and network topology, compared SNNs with ANNs across modalities, and profiled accuracy, latency, energy, and memory on edge and simulated neuromorphic platforms.Additional wireless experiments and frequency-domain analyses examined how wireless signal structure relates to spiking dynamics; Soul-NeuSim recorded experiment configurations, seeds, hyperparameters, device settings, and profiling outputs.Source: MATERIALS AND METHODS Experimental… · pp. 9–10
  • The main benchmark covered five modalities with four representative datasets each, while the expanded wireless analysis added four datasets spanning WiFi channel-state information and millimeter-wave sensing.Matched comparisons used the same data split, preprocessing, training protocol, and deployment configuration; results are means ± standard deviations across three independent runs with fixed random seeds.Source: MATERIALS AND METHODS Experimental… · p. 10

Datasets and preprocessing

The benchmark covered five sensing modalities using modality-specific datasets and preprocessing. Inputs were temporally structured for SNN processing, with distinct default simulation lengths for neuromorphic and other workloads.

Datasets and preprocessing

  • The benchmark represented vision, acoustic, motion, wireless, and neuromorphic sensing with datasets spanning spatial perception, time-frequency signals, inertial series, channel-state or radar measurements, and sparse event streams.Neuromorphic data included event streams from sensors such as dynamic vision sensors and artificial cochleae.Source: Datasets and preprocessing · p. 10
  • Preprocessing preserved official train/test splits and applied modality-specific transformations, including 128×128 mel spectrograms, 128-length motion windows with stride 64, and amplitude extraction from wireless channel-state data when applicable.Vision used original resolution unless resizing was conventional, with training augmentation and normalization; neuromorphic events were converted to frames at discrete time steps.Source: Datasets and preprocessing · p. 10
  • Non-neuromorphic inputs used a default of four simulation timesteps, whereas neuromorphic workloads used 10 timesteps to preserve native event-driven temporal structure.Other modalities were converted into temporally structured inputs before SNN processing.Source: Datasets and preprocessing · p. 10

Spike-encoding strategies · Neuron models and surrogate-gradient training

The benchmark evaluates deployable spike-based input encodings alongside direct coding as an upper-bound reference, and varies neuron dynamics under surrogate-gradient training to isolate their effects across sensing modalities.

Spike-encoding strategies

  • Five spike-compatible strategies—rate, time-to-first-spike, burst, phase, and temporal-switch coding—transform continuous sensor inputs into binary event streams over a temporal window.These are the evaluated spike-encoding strategies.Source: Spike-encoding strategies · p. 10
  • Direct coding replicates normalized inputs across timesteps as constant analog drive, preserving precision but relying on real-valued synaptic input rather than discrete spikes, so it is not spike-compatible with spike-only hardware (24,51).It was evaluated as an upper-bound reference.Source: Spike-encoding strategies · p. 10
  • Timestep lengths and encoding parameters were selected by dataset, modality, and input format, with each configuration held constant across models within a benchmark setting for comparability.Detailed definitions and settings are in Supplementary Materials Section 8.Source: Spike-encoding strategies · pp. 10–11

Neuron models and surrogate-gradient training

  • Soul-NeuSim’s modular neuron factory supports 13 spiking neuron models and 13 surrogate-gradient formulations, enabling neuron-model evaluation independently of network topology.The model set includes representative LIF-family and related variants.Source: Neuron models and surrogate-gradient… · p. 11
  • Evaluated neuron variants differ in temporal integration, leakage, thresholding, reset behavior, and adaptive dynamics, allowing their effects on accuracy and deployment behavior to be assessed across sensing modalities.Model descriptions are provided in Supplementary Materials Sections 9 and 14.Source: Neuron models and surrogate-gradient… · p. 11
  • SNNs use surrogate-gradient backpropagation, with the arctangent surrogate as the default unless otherwise specified.The benchmark framework supports multiple surrogate-gradient formulations.Source: Neuron models and surrogate-gradient… · p. 11
  • Holding network topology and training settings fixed while changing neuron models isolates the effect of neuronal dynamics from architecture-level confounders.Source: Neuron models and surrogate-gradient… · p. 11

Network architectures and ANN baselines · Training protocol · Evaluation metrics

The benchmark compares diverse SNN architectures with a controlled LeNet ANN baseline, trains models using consistent dataset-level settings, and evaluates classification accuracy alongside deployment costs.

Network architectures and ANN baselines

  • The framework provides 12 SNN backbones and hybrid models, including convolutional, residual, transformer-based, and hybrid attention-based designs.Evaluated architectures included SpikingVGG, SEW-ResNet, MS-ResNet, Spikformer, Meta-Spikeformer, QKFormer, and SpikingResFormer.Source: Network architectures and ANN baselines · p. 11
  • LeNet served as the shared lightweight ANN-SNN baseline, while broader SNN topology comparisons adapted architectures to non-vision inputs and controlled model capacity.Capacity was controlled by matching parameter counts or, where exact matching was not possible, aligning computational complexity.Source: Network architectures and ANN baselines · p. 11

Training protocol

  • Models used Adam, cosine annealing, an initial learning rate of 1×10−3, batch size 16, and 150 training epochs on NVIDIA RTX 4090 GPUs.Optimization settings were consistent within each dataset unless otherwise specified.Source: Training protocol · p. 11

Evaluation metrics

  • All tasks were classification tasks, assessed by top-1 accuracy, while deployment behavior was measured by inference latency, energy consumption, and peak memory footprint.Source: Evaluation metrics · p. 11
  • Latency measured wall-clock time per sample through prediction after warm-up, with timing conducted under identical batch-size and deployment settings and P90 used to capture runtime variability.Source: Evaluation metrics · pp. 11–12
  • Peak memory included parameters, intermediate activations, and SNN temporal states; operation comparisons used ANN FLOPs and SNN synaptic operations accounting for timesteps and presynaptic firing.The first SNN layer was treated as dense under direct input encoding, while later layers were measured as spike-driven synaptic operations.Source: Evaluation metrics · p. 12

Physical edge-device deployment · Wireless spectral analysis

The study profiled ANN and SNN deployment across seven physical edge devices using standardized conversion and measurement protocols, and analyzed wireless signals’ frequency structure to examine its relationship to spiking dynamics.

Physical edge-device deployment

  • Seven physical edge devices spanning CPU-, GPU-, and NPU-based execution were evaluated after model conversion through a standardized NCNN-based deployment interface.The devices included Raspberry Pi 4B, Jetson Nano, Jetson NX, Jetson AGX, Redmi K80, Pixel 6, and Huawei Mate 40.Source: Physical edge-device deployment · p. 12
  • Latency, energy consumption, and peak memory were recorded under identical batch-size and numerical-precision settings within each device.The same trained model, input representation, and preprocessing pipeline were used wherever the deployment backend supported them; conversion, scheduling, and metric collection were separated to reduce device-specific optimization bias.Source: Physical edge-device deployment · p. 12
  • Physical-device profiling was designed to quantify how ANN and SNN deployment behavior varies with sensing modality, model architecture, and hardware substrate.The profiling used consistent deployment settings to compare these factors.Source: Physical edge-device deployment · p. 12

Wireless spectral analysis

  • Wireless spectral analysis examined frequency-domain signal structure alongside the temporal filtering behavior of spiking neuronal dynamics.For inputs with explicit temporal structure, non-temporal dimensions were mean-pooled into proxy sequences and transformed using discrete Fourier transforms.Source: Wireless spectral analysis · p. 12
  • Millimeter-wave analysis computed feature-time and feature-Doppler energy maps, plus micro-Doppler spectrograms from selected high-variance temporal channels.Feature-time maps used instantaneous channel power, feature-Doppler maps used a time-axis discrete Fourier transform, and spectrograms used short-time Fourier transforms.Source: Wireless spectral analysis · p. 12
  • A Fisher-criterion discriminative spectrum estimated where class-discriminative information is concentrated across frequency bins, independently of neural architecture.Between-class and within-class scatter ratios from training data were normalized over the one-sided frequency grid into a probability mass function.Source: Wireless spectral analysis · p. 12

Frequency-matching score · Feature-space visualization and reconstruction analysis · Neuromorphic simulation with NeuSim

The section defines a frequency-matching measure and representation-level analyses for interpreting the wireless SNN advantage, then describes NeuSim’s simulation approach for neuromorphic deployment analysis.

Frequency-matching score

  • The frequency-matching score measures how much discriminative wireless spectral content is retained by the neuron response, modeled as a first-order low-pass filter controlled by membrane decay.Higher scores indicate stronger overlap between the frequencies emphasized by spiking dynamics and those carrying task-relevant information.Source: Frequency-matching score · p. 13
  • A maximum-deviation rule on the frequency-matching curve defines boundaries between under-filtering, stable filtering, and over-low-pass regimes.The score uses the DC-normalized power response on a one-sided frequency grid; its full derivation is in Supplementary Materials Section 4.Source: Frequency-matching score · p. 13

Feature-space visualization and reconstruction analysis

  • Trained ANN and SNN latent features are visualized with t-SNE and compared using silhouette scores to assess class-cluster separability on the same wireless task.The analysis tests whether SNN representations produce more separable class clusters than ANN representations.Source: Feature-space visualization and… · p. 13
  • Representative ANN and SNN reconstructions are qualitatively examined for retained noise-dominated components versus cleaner task-relevant structures, alongside spectral analyses to interpret the wireless SNN advantage.Additional visualization and reconstruction analyses are provided in Supplementary Materials Section 5.Source: Feature-space visualization and… · p. 13

Neuromorphic simulation with NeuSim

  • NeuSim models neuromorphic processors as neuromorphic cores connected by a two-dimensional mesh network-on-chip, partitioning SNN workloads across cores for parallel execution.Neuron capacity, synapse memory, buffer depth, chip dimensions, and routing policy are configurable; settings are summarized in Supplementary Materials Section 13.Source: Neuromorphic simulation with NeuSim · p. 13
  • NeuSim represents hardware with core and router abstractions: cores update neuron states, store synapse connectivity, and packetize spikes, while routers model interconnect traversal; XY routing is the default.Router modeling includes routing, virtual-channel allocation, switch arbitration, and crossbar traversal.Source: Neuromorphic simulation with NeuSim · p. 13
  • Cycle-level simulation estimates computation and communication delays and records inference latency, utilization, traffic, congestion, packet delay, and energy.Energy models use empirical power data from representative physical devices; NeuSim supports multithreaded execution with per-cycle synchronization and exposes its C++ engine to Python for integration with Soul.Source: Neuromorphic simulation with NeuSim · p. 13

Compilation, partitioning, and mapping · Supplementary Materials

NeuSim compiles trained SNNs through synapse extraction, capacity-constrained partitioning, and physical-core mapping. Its workflow supports alternative locality-aware partitioning and placement strategies, and the supplementary materials list Figs. S1–S10.

Compilation, partitioning, and mapping

  • NeuSim’s compilation pipeline extracts synapse connectivity, partitions neurons and synapses into logical cores, and maps those cores onto physical hardware.Partitioning respects single-core neuron and synapse capacities.Source: Compilation, partitioning, and mapping · p. 14
  • Linear layers become fully connected bipartite connectivity, while convolutional layers use structured local connectivity extracted through unfolding.Operator compositions are analytically fused to obtain final synaptic weights.Source: Compilation, partitioning, and mapping · p. 14
  • The network is represented as a directed acyclic graph of neuron-layer nodes and operator edges instead of a dense global adjacency matrix.Source: Compilation, partitioning, and mapping · p. 14
  • A locality-preserving partitioning option orders neurons with space-filling curves and uses binary segmentation to form capacity-valid clusters.Sequential partitioning is the default; the locality-preserving heuristic is user-defined.Source: Compilation, partitioning, and mapping · p. 14
  • Logical cores can be placed sequentially or with Hilbert space-filling curves (HSC) on the two-dimensional mesh (72).The compiler interfaces allow customized partitioning and placement strategies in Soul-NeuSim.Source: Compilation, partitioning, and mapping · p. 14

Supplementary Materials

Loading 2609.00026v1…