Neural data is already sparse
Look at a raw extracellular trace and you mostly see silence — baseline noise punctuated by the occasional action potential. Nyquist sampling forces you to encode all of that silence at tens of kilohertz. An event-driven front-end refuses to: it detects the events and transmits only their times and a few features. The organising idea, borrowed from silicon retinas and cochleas, is address-event representation — a spike is communicated as which channel, at what time, and nothing more.
This matters because the neuromorphic sensor front-end can then be built to stay quiet. Rather than a clock ticking through every sample, comparators wake the circuit only when a threshold is crossed. Sparsity in the brain becomes sparsity in the silicon, and sparsity in the silicon becomes energy saved.
On-chip spike detection and threshold crossing
The simplest and most deployed on-implant computation is threshold-crossing detection: keep only the times at which the filtered signal crosses a threshold, discard the waveform. For many motor BCIs this loses surprisingly little decodable information while compressing enormously — a key reason it appears in real fully-implantable systems. Its smarter cousins improve robustness to noise and baseline drift.
The nonlinear energy operator (Teager–Kaiser), a classic on-chip spike detector: it emphasises the instantaneous product of amplitude and frequency, sharpening spikes above background before a threshold is applied. It costs only a handful of arithmetic operations per sample.
These operators are cheap precisely because they are local in time and require no stored model — a good match to the energy budget of an implant. Full spike sorting (assigning each spike to a putative neuron) is more powerful but far costlier; a live design question is how much sorting to do on-chip versus how much to defer.
Energy scales with activity, not time
Here is the deep reason event-driven design wins. In a clocked digital system, dynamic energy is roughly proportional to the clock — you pay for every tick whether or not anything happened. In an event-driven system, dynamic energy is proportional to the number of events. A quiet brain is a cheap brain to listen to.
Dynamic energy over a window T is the per-event cost E_{ ext{ev}} times the total spike count summed over N channels with mean rates ar{r}_i. Because cortical firing is sparse (single-digit to tens of Hz), this can sit far below a clocked design that pays \propto f_{ ext{clk}} \cdot T regardless.
You cannot cheat the analog noise floor
Event-driven cleverness happens after the signal is already amplified, and amplification is where physics bites. Extracellular spikes are microvolts riding on thermal and biological noise, so the front-end needs a low-noise amplifier, often with chopper stabilisation to beat 1/f noise, all under a strict current budget. The fundamental trade between noise and power is captured by the noise-efficiency factor.
The noise-efficiency factor benchmarks an amplifier against an ideal single bipolar transistor: it ties input-referred noise V_{\mathrm{ni,rms}} to total supply current I_{\mathrm{tot}} and bandwidth. NEF = 1 is the physical ideal; real neural amplifiers sit a small multiple above it, and every channel pays this current toll.
So the event-driven front-end is a two-part story: an unavoidable, always-on analog cost set by thermal noise and the amplifier, plus a highly compressible, activity-proportional digital cost. Neuromorphic design shrinks the second aggressively; the first is a floor you negotiate with, not one you cross. After the front-end, the surviving question is what to compute with these sparse events — the subject of the next two guides.