← The energy program · Thermodynamic sampling
Machines that wake up.
Your brain runs on about twenty watts. It does that by keeping almost every neuron silent almost all of the time, and paying only for the ones that fire. A class of chip is built the same way: no clock driving a dense matrix through the machine, nothing happening at all until a signal arrives. This page runs that idea on your own device, and puts a number on the one thing that decides whether it wins.
How often is a neuron actually doing anything?
Predict first. Below is a chain of eight spiking cells, one driving the next, running for two hundred milliseconds. A conventional simulator updates every cell on every tick whether or not anything reached it. What fraction of that work do you expect to land on a cell with nothing to do — a tenth? Half? Then run it and read the count. The instrument checks itself before it answers: a cell's firing rate has had a closed form since Lapicque wrote it down in 1907, and the simulator must reproduce it, and must produce exactly nothing below threshold, before it reports a single number.
One row per cell, time left to right. Every mark is a spike — and every gap is the machine paying nothing.
What does one spike cost?
A spike arriving at a synapse does two things. It adds a number, which is arithmetic and is cheap. And before it can do that, the synapse's weight has to be fetched out of memory, which is data movement and is not. Predict which of those two dominates. It is the same question the memory wall answered for conventional computing, and it has the same answer here — the arithmetic is the small term. Which means the quantity that decides whether a spiking machine wins is not how fast it computes but how often it has to go and get a weight.
26 pJ per synaptic event on TrueNorth in 28 nm — measured on fabricated silicon, Merolla et al., Science 345 (2014). One of the few numbers in this field taken from a real chip rather than a simulation of one.
This review did not locate a published figure for it, for any part you can buy. So the instrument above counts the fetches exactly and says so, rather than quietly pricing them at zero — which is what a synaptic-operation count does.
One measurement is missing, and it is the interesting one.
The field already worked out where the line sits. Whether a spiking network beats its dense equivalent is decided by spikes per synapse per inference — cross it and you are re-reading weights more often than a dense pass would, and the dense pass wins. At least six papers put a number on it and every one is below two. Davidson and Furber derive about 1.72, and Steve Furber designed SpiNNaker, so this is the field sharpening its own tools. What is missing is one instrument reading: a meter on a memory rail while a spiking network runs. Not a new idea, not a new chip — a protocol and an afternoon. That is the most tractable kind of gap there is, and it is why the instrument above reports the count rather than guessing the joules.
Davidson & Furber, Frontiers in Neuroscience 15:651141 (2021).
Dampfhoffer et al., IEEE TETCI 7(3) (2023), the work that put memory accesses back into the accounting.
Yan et al., arXiv:2409.08290 — what the same comparison gives once the data movement is counted.
A number is only as good as the kind of number it is.
The most-quoted energy figure in neuromorphic computing is 23.6 picojoules per synaptic operation for Intel's Loihi. It comes from a table captioned pre-silicon — a circuit simulation, cited as a measurement for the better part of a decade. Sorting figures by the kind of evidence behind them is most of the work, and it is why every number on this page carries its label. The open library below does the same thing in code: it prices what is known, names what is not, and refuses to invent the difference.
The library Documentation Thermodynamic sampling The Energy Lab
Lapicque, J. Physiol. Pathol. Gen. 9 (1907) · Hodgkin and Huxley, J. Physiol. 117 (1952) · Mead, Proc. IEEE 78 (1990) · Mahowald, Caltech (1992) · Merolla et al., Science 345 (2014) · Davies et al., IEEE Micro 38(1) (2018) · Davidson and Furber, Front. Neurosci. 15 (2021) · Dampfhoffer et al., IEEE TETCI 7(3) (2023) · Yik et al., Nat. Commun. 16 (2025) · arXiv:2409.08290.