Ultra-low-power inference (implant power budget)
The defining constraint of on-implant decoding is thermal: dissipated power heats surrounding tissue, and widely-cited safety guidance keeps chronic temperature rise to roughly a degree Celsius and power density to a few tens of milliwatts per square centimetre. Add the limits of wireless or harvested power and battery life, and the compute budget for an always-on decoder falls toward the sub-milliwatt range. Neuromorphic techniques — event-driven operation, in-memory compute, aggressive sparsity — are pursued precisely to fit decoding into this envelope.
Meeting a power budget is a whole-system problem: amplification, digitisation, the decoder, memory, and telemetry all draw power, and the radio to send results off-body is often the largest term. This is why decoding on-device (sending compact commands, not raw data) is the leverage point, and why reported decoder-only energy figures can be misleading.
Exact thermal limits depend on tissue, perfusion, and the standard applied; the qualitative point — a very tight, safety-bounded budget — is what drives the architecture.