Stochastic & Neuromorphic Computing  /  Explore  /  Per-spike energy model

Per-spike energy model
what one inference costs

“This classification cost 4.2 nanojoules.” That is the sentence neuromorphic hardware is supposed to earn, and it is only honest if you can account for every picojoule. SC-NeuroCore tallies energy the way the chip spends it — per synaptic operation, per membrane update, per spike, per memory read, per routed event — against measured-class per-operation costs for real targets. The result exposes the core of the efficiency argument: synaptic energy scales with spikes, membrane updates scale with the clock.

The accounting — five costs, one total

Every inference decomposes into a handful of countable operations. A spike drives $N_{\text{in}}$ synaptic operations, each of which also reads a weight from memory; every neuron takes a membrane update on every timestep whether or not it fires; each spike costs a comparator-and-reset to generate and a hop to route. Multiply each count by its per-operation cost in picojoules and sum. The one asymmetry that matters is in the counts, not the costs.

$$E = \underbrace{n_{\text{syn}}(c_{\text{syn}}{+}c_{\text{mem}})}_{\text{scales with spikes}} + \underbrace{N_{\text{out}}\,T\,c_{\text{mem.upd}}}_{\text{scales with the clock}} + \underbrace{n_{\text{sp}}\,c_{\text{gen}}}_{}+\underbrace{n_{\text{sp}}\,c_{\text{route}}}_{}, \qquad n_{\text{syn}} = n_{\text{sp}}\,N_{\text{in}}$$
Live — energy of one layer, by operation
synaptic ops memory reads membrane updates spike gen + routing
total per spike clock floor (paid at zero activity) vs fully-dense

Turn activity down and the synaptic and memory bars shrink with it — a spiking network only pays for the neurons that actually fire, which is the whole event-driven argument. But the green membrane-update bar does not move: every neuron is still clocked every timestep, so it is a fixed floor paid even at zero activity. That floor is exactly why event-driven hardware, which skips the update for silent neurons, matters — and why a sparse workload on a densely-clocked device leaves efficiency on the table. The “vs fully-dense” figure is what the same layer would cost if every neuron fired every step.

The per-operation costs

The calculator uses the toolkit's hardware cost model — per-operation energies, in picojoules, for each target. These are literature-class figures for the named devices, not measurements taken on this page; they set the relative economics, which is what the model is for.

TargetSynOp (pJ)Membrane (pJ)Spike gen (pJ)Mem read (pJ)Routing (pJ)
Loihi 223.61.20.54.03.0
Akida10.00.80.33.01.0
Analog 28 nm1.60.20.10.50.3
FPGA Artix-730.03.01.08.00.3
FPGA iCE4050.05.02.010.00.5
Deeper: why per-spike accounting beats a single number
A single “pJ per SynOp” headline hides where the energy actually goes. The accountant keeps the breakdown per layer, so it can name the dominant cost — often the synaptic-plus-memory term for a wide layer, or the membrane floor for a deep, sparsely-active one — and report the true energy-per-spike after everything is counted, which is usually far above the bare SynOp cost once memory and routing are included. That is the difference between an accessible metric and a high-fidelity one, and the reason the module accounts rather than quotes.
Where it fits

This is the operational-energy model behind the numbers used elsewhere on the site: the carbon page turns these picojoules into grams of CO₂ over a device's life; the hardware and compiler pages choose targets whose costs this table encodes. It is deliberately the same accounting the toolkit runs, so a claim made here can be traced to a count of operations rather than a slogan.

Evidence boundary: the calculator runs the library's exact accounting arithmetic in your browser on the per-operation cost table above. Those per-operation energies are literature-class figures for the named targets, not measurements on this page; a real deployment's energy depends on the mapping, the data, and the device. No wall-clock or throughput figures are quoted.