| name | hardware-aware-mixed-signal-snn-framework |
| description | Open-source hardware-aware simulation framework for mixed-signal SNNs enabling comparative analysis across neuron models (LIF, HH, AH), synapse types (floating-gate, ReRAM), and architectures. Reports accuracy with hardware metrics (area, power, quantization sensitivity). |
| tags | ["spiking-neural-network","hardware-simulation","mixed-signal","neuromorphic","design-space-exploration","open-source"] |
| activation_words | ["hardware-aware SNN","mixed-signal simulation","neuron model comparison","LIF HH AH","ReRAM synapse","floating-gate","design space exploration","neuromorphic benchmark"] |
Hardware-Aware Open-Source Framework for Mixed-Signal SNN Design Space Exploration
Overview
Open-source hardware-aware simulation framework for mixed-signal spiking neural networks that enables comparative analysis across neuron models, synapse types, and architectures while reporting both accuracy and hardware-oriented metrics.
Core Problem
Fragmented SNN Simulation:
- Existing tools capture either biological detail OR hardware efficiency, not both
- No unified framework for cross-layer design space exploration
- Difficult to compare neuron-synapse configurations for specific applications
- Missing hardware metrics (area, power, quantization sensitivity) in standard SNN tools
Framework Architecture
Supported Neuron Models
-
Leaky Integrate-and-Fire (LIF)
- Simplest model, fastest simulation
- Good for large-scale networks
- Limited biological realism
-
Hodgkin-Huxley (HH)
- Biologically detailed ion channel dynamics
- Captures realistic spike shapes
- Computationally expensive
-
Axon-Hillock (AH)
- Intermediate complexity
- Balances realism and efficiency
- Captures spike initiation zone dynamics
Supported Synapse Types
-
Floating-Gate Transistors
- Non-volatile analog storage
- Tunable weight updates
- Mature CMOS technology
-
ReRAM Devices
- Resistive switching memory
- High density, low power
- Emerging technology with non-idealities
Hardware Metrics
Reported for Each Configuration:
- Silicon Area: Estimated from transistor count and layout
- Power Consumption: Dynamic + static power estimation
- Quantization Sensitivity: Accuracy vs. precision tradeoffs
- Hardware Fidelity: How well simulation captures non-ideal behavior
Implementation
PyTorch Integration
import torch
import spikingjelly
class HardwareAwareSNN:
def __init__(self, neuron_model='LIF', synapse_type='floating_gate'):
self.neuron_model = self._build_neuron(neuron_model)
self.synapse = self._build_synapse(synapse_type)
self.hardware_metrics = HardwareMetrics()
def _build_neuron(self, model_type):
if model_type == 'LIF':
return LIFNeuron(tau_mem=20.0, v_threshold=1.0)
elif model_type == 'HH':
return HHNeuron(g_Na=120.0, g_K=36.0, g_L=0.3)
elif model_type == 'AH':
return AHNeuron(compartment_params={...})
def _build_synapse(self, synapse_type):
if synapse_type == 'floating_gate':
return FloatingGateSynapse(device_params={...})
elif synapse_type == 'ReRAM':
return ReRAMSynapse(hrs_resistance=1e6, lrs_resistance=1e3)
def forward():
synaptic_current = .synapse(spikes)
membrane_potential = .neuron_model(synaptic_current)
output_spikes = .neuron_model.fire(membrane_potential)
output_spikes
():
area = .hardware_metrics.estimate_area()
power = .hardware_metrics.estimate_power()
quant_sensitivity = .hardware_metrics.quantization_analysis()
{: area, : power, : quant_sensitivity}
Design Space Exploration
configs = []
for neuron in ['LIF', 'HH', 'AH']:
for synapse in ['floating_gate', 'ReRAM']:
for precision in [8, 16, 32]:
configs.append({
'neuron': neuron,
'synapse': synapse,
'precision': precision
})
results = []
for config in configs:
model = HardwareAwareSNN(config['neuron'], config['synapse'])
model.set_precision(config['precision'])
accuracy = train_and_evaluate(model, dataset='N-MNIST')
hw_metrics = model.compute_hardware_metrics()
results.append({
'config': config,
'accuracy': accuracy,
**hw_metrics
})
analyze_pareto_frontier(results)
Benchmarks
Standard Neuromorphic Datasets
-
N-MNIST
- Spiking version of MNIST
- 3 event channels (ON, OFF, background)
- 60k training, 10k test samples
-
DVS Gesture
- Dynamic Vision Sensor recordings
- 11 hand gesture classes
- Real-world event camera data
-
Spiking Heidelberg Digits (SHD)
- Spoken digits (0-9) in English/German
- 20ms binned audio spectrograms
- Temporal classification task
Reported Metrics
For each model-dataset configuration:
- Classification Accuracy: Top-1 accuracy on test set
- Silicon Area: mm² estimated from transistor count
- Power Consumption: mW during inference
- Quantization Sensitivity: Accuracy drop at lower precision
Design Space Exploration
Configuration Parameters
-
Neuron Model Selection
- LIF: Fast, scalable, less biological
- HH: Detailed, slow, biologically realistic
- AH: Balanced complexity
-
Synapse Device Choice
- Floating-gate: Mature, tunable, moderate density
- ReRAM: High density, low power, emerging tech
-
Precision Tradeoffs
- 8-bit: Lowest power/area, potential accuracy loss
- 16-bit: Balanced
- 32-bit: Highest accuracy, more resources
-
Architecture Decisions
- Number of layers
- Neurons per layer
- Connectivity pattern (feedforward, recurrent)
Pareto Analysis
def analyze_pareto_frontier(results):
"""
Identify configurations on Pareto frontier of accuracy vs. energy
"""
pareto_configs = []
for r in results:
dominated = False
for other in results:
if (other['accuracy'] >= r['accuracy'] and
other['power'] <= r['power'] and
(other['accuracy'] > r['accuracy'] or other['power'] < r['power'])):
dominated = True
break
if not dominated:
pareto_configs.append(r)
return pareto_configs
Key Findings
Neuron Model Impact
- LIF: Best for large-scale, energy-constrained applications
- HH: Necessary when biological realism is critical
- AH: Good middle ground for many applications
Synapse Device Comparison
- Floating-gate: More predictable, easier to train
- ReRAM: Higher density but more non-idealities to manage
Precision Tradeoffs
- 8-bit: Often sufficient for N-MNIST, significant area/power savings
- 16-bit: Recommended default for most applications
- 32-bit: Only necessary for very sensitive tasks
Applications
Edge AI Deployment
- Selecting optimal configuration for battery-powered devices
- Balancing accuracy and energy for always-on inference
Neuromorphic Chip Design
- Informing architecture decisions before fabrication
- Validating design choices against benchmarks
Algorithm-Hardware Co-Design
- Co-optimizing SNN algorithms with hardware constraints
- Identifying hardware-aware training objectives
Related Work
- snntorch: Software SNN simulation (no hardware metrics)
- Brian2: Biological neural simulation (not hardware-focused)
- NEST: Large-scale simulation (not mixed-signal)
- Intel Lava: Neuromorphic compilation (fixed architecture)
Limitations
- Simulation Accuracy: Hardware metrics are estimates, not measurements
- Device Variability: Real devices have more variation than modeled
- Scalability: Full-chip simulation still computationally expensive
- Technology Nodes: Metrics depend on specific process technology
Future Work
- Integration with actual silicon measurements
- Support for more neuron/synapse models
- Automated configuration optimization
- Multi-objective optimization beyond accuracy-power
Activation
hardware-aware SNN, mixed-signal simulation, neuron model comparison, LIF HH AH, ReRAM synapse, floating-gate, design space exploration, neuromorphic benchmark, PyTorch SNN, hardware metrics
arXiv Reference
- ID: 2607.06456
- Title: A Hardware-Aware Open-Source Framework for Design Space Exploration of Mixed-Signal Spiking Neural Networks
- Authors: Sayma Nowshin Chowdhury, Vineeta Nair, Taseen Forhad
- Categories: eess.SP, cs.NE
- Published: 2026-07-07