JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

The Bandwidth Imperative: Stevenson's Law and Why Channels Matter

Why the next leap in BCI is fundamentally a scaling problem — and roughly how many channels it might actually take.

From tens to millions

Every decoder you met in Volumes I and II was channel-starved. A clinical Utah array delivers on the order of a hundred intracortical channels; a research Neuropixels probe pushes into the hundreds; the whole human cortex holds on the order of 10^{10} neurons. The organizing question of this track is what changes as the channel count C climbs through 10^4, 10^5, and beyond toward whole-brain scale — and which walls we hit on the way.

Scaling stresses three resources at once: data (the bits streaming off a sealed, wet implant), power (which becomes heat the brain cannot dissipate), and information (which saturates as channels become redundant). We will treat each in turn, but they are coupled — a fact that makes whole-brain interfacing a genuine systems problem, not a matter of simply printing more electrodes.

Stevenson's law: the empirical trend

Stevenson and Kording noted that the number of neurons recorded simultaneously in a typical experiment has roughly doubled every seven years or so across decades — an exponential, but one far shallower than Moore's law for transistors. This is Stevenson's law: real, empirical, and sobering.

N(t) = N_0 \, 2^{(t - t_0)/\tau}, \qquad \tau \approx 7\ \text{years}

The empirical doubling trend. With a seven-year doubling time, going from ~10^2 to ~10^6 simultaneously recorded neurons is roughly fourteen doublings — a century at the historical rate, unless a technology discontinuity bends the curve.

Simultaneously recorded neurons over the decades, on a log axis — the near-straight line that defines Stevenson's law.

The hope of this decade is exactly such a discontinuity: high-density CMOS probe arrays and active silicon move thousands of sites onto one shank, bending the curve upward. Whether that bend is a one-time step or a new, steeper exponential is one of the open empirical questions of the field.

What more channels buy you

More channels observe more neurons, which can mean more decodable degrees of freedom (a hand's many joints, not just a cursor), finer kinematics, and access to more brain areas at once — the ingredients of richer, more natural neuroprostheses. The naive intuition is that decodable information grows in proportion to channel count.

I_{\text{naive}}(C) = C \cdot i_{\text{ch}}

The linear assumption: each channel contributes a fixed increment of information i_{\text{ch}}. It is a useful first guess — and, as Guide 4 shows, quantitatively wrong once neurons share signal and noise.

Drag the channel count and watch the achievable information rate. It rises steeply at first, then bends over — the gap between the linear dream and reality that this track is about.

That early steep rise is real and worth chasing — the difference between one and ten degrees of freedom is the difference between a cursor and a hand. The bend that follows is the honest part of the story, and the reason bandwidth is not simply a manufacturing target.

The three walls ahead

The rest of this track is a tour of three walls. The data wall (Guide 3): a million broadband channels produce hundreds of gigabits per second that no implant can wirelessly export. The power/thermal wall (Guide 2): every channel dissipates heat, and the brain tolerates barely a degree or two of warming. The information wall (Guide 4): shared variability and low-dimensional structure make added channels increasingly redundant.