The binding constraint is heat, not FLOPS
It is easy to be seduced by throughput numbers, but a fully implantable system is not limited by how many operations it can do — it is limited by how much heat it may shed into living tissue. Chronic warming must stay within roughly a degree, which caps total dissipation for a cortical implant at only a few to tens of milliwatts. Amplifier, compute and telemetry all draw from that single shared purse.
The system-level budget: analog front-end plus compute plus transmit power must fit under a heat limit set by the maximum tolerable temperature rise \Delta T_{\max} and the thermal resistance R_{ heta} to the tissue. Neuromorphic design earns its keep by shrinking P_{\mathrm{comp}} and, via compression, P_{\mathrm{tx}}.
This reframes the whole track. On-implant compute is not attractive because it is fast; it is attractive because computing locally can shrink the telemetry bill, and telemetry is often the largest single power sink. Sometimes the cheapest joule is the bit you never transmit. See power budget and tissue heating and the thermal and power-budget limits.
Latency for the closed loop
The other budget is time. A closed-loop application — cursor control, a prosthetic limb, responsive stimulation — needs end-to-end latency below the thresholds at which the user's control or perception degrades. Moving compute on-device removes the wireless round-trip, but the on-chip inference itself then has its own latency budget to honour. See on-device closed-loop latency and latency and stability limits.
Event-driven, always-on neuromorphic hardware is a natural fit here: because it reacts to spikes as they arrive rather than waiting for a sampling frame or a batch, it can in principle deliver a decision with very short and, importantly, very predictable latency — and predictability matters as much as the mean for stable closed-loop control.
The honest ledger: demonstrated, emerging, aspirational
Intellectual honesty demands separating three columns. Demonstrated: on-chip spike detection and compression running in real implantable systems, cutting the telemetry rate by large factors; spiking decoders matching simple conventional decoders on external neuromorphic boards at markedly lower energy; memristor crossbars computing decode-scale matrix products in the lab.
Emerging: on-chip learning that adapts against drift; larger-scale analog in-memory decoding with error tolerance. Aspirational: a fully implanted, chronically running, in-memory analog spiking decoder inside a human, learning on-device against representational drift. That last system does not yet exist, and conflating its components' demos with the finished article is the field's most common form of overclaiming.
The missing benchmarks
A quieter open problem is measurement itself. Energy-per-inference numbers are rarely comparable across papers: different tasks, different signal quality, and different accounting of what is inside the boundary. The field lacks standardised, closed-loop, in-the-loop benchmarks — a neuromorphic analogue of the few-shot decoding benchmarks emerging elsewhere — and progress will stay hard to compare until standards bodies converge on them.
Where it is going
An honest reading points to a staged trajectory rather than a single leap. Near-term: compression and spike-detection ASICs become standard equipment in high-channel-count implants — the least glamorous pillar, but the one that actually unblocks scaling. Medium-term: hybrid systems pair a neuromorphic front-end with a conventional or edge decoder, splitting the work across the skin. Longer-term: co-adaptive on-chip learning against drift, and in-memory analog decode if device reliability matures enough to tame the non-idealities.
None of this dissolves the deeper ceilings the final track of this volume will confront — the amplifier's noise floor, the thermal wall, the irreducible variability of the brain itself. Neuromorphic engineering does not repeal those limits; it lets a BCI operate closer to them, inside the unforgiving budget of the skull. For the wider horizon, see an honest reading of the next twenty years.