Recording bandwidth · Grey Matter

Recording bandwidth is the amount of data per second that a recording of the brain produces and has to move out of the head, and it grows with every electrode added, until moving the data costs more power and heat than measuring it.

Recording bandwidth. Recording bandwidth is the amount of data per second that a recording of the brain produces and has to move out of the head, and it grows with every electrode added, until moving the data costs more power and heat than measuring it.

The arithmetic is simple and unforgiving. One channel sampled at 30 kHz with 10 bits is 300 kilobits per second; a thousand channels are 300 megabits per second, far beyond what the low-power radio links of today's implants carry. At the scale of the whole brain the numbers stop being engineering: about 86 billion neurons, each read at 1 kHz with 10 bits, would make

86×109×103 s−1×10 bits≈8.6×1014 bits per second,86 \times 10^{9} \times 10^{3}\ \text{s}^{-1} \times 10\ \text{bits} \approx 8.6 \times 10^{14}\ \text{bits per second},86×109×103 s−1×10 bits≈8.6×1014 bits per second,

close to 101510^{15}1015, and the 101410^{14}1014 to 101510^{15}1015 synapses between them would add three more orders of magnitude.

Implants reduce data before sending it. Detecting spikes on the chip and transmitting only their times, or the power in a few frequency bands, shrinks the stream by orders of magnitude; Neuralink's design did exactly this from its first published version.

Radio has a power-bandwidth trade. Sending more bits per second over radio costs more power per bit as rates rise, which is why light and ultrasound, which can be focused and reused in space, are proposed for future links.

Decoders need less than recordings produce. Most brain-computer interfaces today work from a few hundred features (spike counts or band power per channel) updated tens of times a second, so for them the harder limits are how many sensors can be placed and how long they last.

Bandwidth turns a measurement problem into an energy problem.

Every bit that leaves the head costs power, and power in the head becomes heat.

Questions: How can a cell record its own history without sending any data out? It grows a molecule that keeps a timeline. In the protein recorders published in 2023, cells assemble long chains from subunits carrying fluorescent tags, and a gene switched on by the signal being recorded changes which tag is added, so the order of tags along a chain stores when that signal was on. After days or weeks the tissue is fixed and the chains are read under a microscope, which removes the bandwidth problem entirely at the price of hours-scale time resolution and a single readout at the end. Does a wider data stream out of the brain expose more of the mind? More channels and richer features let a decoder find more than the one thing it was built for. Speech implants in motor cortex, designed to decode attempted speech, also picked up some inner speech during tasks such as counting, and the more of the raw signal leaves the head the more such secondary readings become possible later. Compressing data on the implant to only what one decoder needs is an engineering choice that also limits exposure, and ethicists have proposed safeguards of this kind, processing neural data where it is collected and sharing only what a service needs. How does an implant send out the data of a thousand electrodes through the skin? It does not send the raw signal. Sampled at 30 kHz with 10 bits, a thousand electrodes would produce about 300 megabits per second, far more than a low-power radio link through the skin carries on the power an implant can spend. Chips like those in the N1 amplify, digitise and detect spikes on the implant, then send only the times of spikes or compressed features per channel, which shrinks the stream by orders of magnitude and still carries what decoders of movement use.