ENERGY
Energy scales with spikes
not parameters
Energy scales with spikes not parameters
A dense ANN computes every weight every timestep, whether the input changed or not. A spiking network computes only when neurons fire. On battery-bound edge systems — robots, drones, satellites, unattended sensors — that difference is architectural, not incremental.
pip install thrindex
The model, precisely
E = Σ_layer (pre_spikes × fan_out) × coefficient
Spike events from the pre-synaptic layer, all timesteps
pre_spikes
Post-synaptic connections per neuron
fan_out
Energy per syn-op (default: 0.5 pJ)
coefficient
Spike events from the pre-synaptic layer, all timesteps
Reduction in manual workflow coordination across teams
Energy per syn-op (default: 0.5 pJ)
coefficient
Post-synaptic connections per neuron
fan_out
A real network
A real network
100 timesteps, 10% mean spike rate:
100 timesteps, 10% mean spike rate:
Architecture
Dense(700→512) → LIF → Dense(512→20) → LIF
1
Architecture
Dense(700→512) → LIF → Dense(512→20) → LIF
Layer 1
7,000 spikes × 512 fan-out = 3,584,000 syn-ops
2
Layer 1
7,000 spikes × 512 fan-out = 3,584,000 syn-ops
Layer 2
5,120 spikes × 20 fan-out = 102,400 syn-ops
3
Layer 2
5,120 spikes × 20 fan-out = 102,400 syn-ops
Total
3,686,400 syn-ops × 0.5 pJ ≈ 1.84 µJ per inference
4
Total
3,686,400 syn-ops × 0.5 pJ ≈ 1.84 µJ per inference
Where the math matters
Where the math matters
Sparse inputs, always-on duty, tight power budgets, thermal limits.
Sparse inputs, always-on duty, tight power budgets, thermal limits.
Sparse inputs
Event cameras, audio onset, vibration — activity is bursty, not continuous.
Always-on duty
Network silent when nothing changes. Power collapses toward µW.
Tight power budget
Battery mass is mission mass on drones and orbital payloads.
Thermal limits
Less computation → less heat → smaller cooling on edge nodes.
What this is and isn't
What this is and isn't
A modeled estimate for architecture comparison — not a hardware power measurement.
A modeled estimate for architecture comparison — not a hardware power measurement.
This is
A modeled estimate from spike counts. Useful for comparing architectures before silicon is on the bench.
This is
A modeled estimate from spike counts. Useful for comparing architectures before silicon is on the bench.
This isn't
A hardware power measurement. Target-specific calibration requires device datasheets and HIL validation.
This isn't
A hardware power measurement. Target-specific calibration requires device datasheets and HIL validation.
This is
A modeled estimate from spike counts. Useful for comparing architectures before silicon is on the bench.
This isn't
A hardware power measurement. Target-specific calibration requires device datasheets and HIL validation.
Install. Run a template. Inspect the raster. No hardware. No account.
pip install thrindex
Install. Run a template. Inspect the raster. No hardware. No account.
pip install thrindex
Install. Run a template. Inspect the raster. No hardware. No account.
pip install thrindex