Neuropixels probes

The bottleneck in high-density recording was never how small you could make an electrode site. It was wiring. A passive probe needs one conductor per site running out of the brain to an external amplifier, so site count is limited by how many wires the shank can carry and how much tissue that displaces.

Neuropixels moves the electronics onto the probe. Shank and base are a single 130 nm CMOS chip: the 10 mm shank carries the 960 recording sites and the switches that choose among them, and the 6 × 9 mm base, which stays outside the brain, filters, amplifies, multiplexes and digitises — so what leaves the probe is a digital stream on a single thin cable rather than a thousand analog wires. Albert Lee and Timothy Harris at HHMI’s Janelia Research Campus conceived it with Barundeb Dutta of imec, in Leuven; James Jun and colleagues described it in 2017.

Why that is the whole story

Every hard constraint in the design is a consequence of putting active circuitry on a probe that goes into a brain.

Power, therefore heat. Tissue tolerates very little temperature rise, which sets a budget for the entire amplifier array. Per-channel power has to be small enough that hundreds of channels stay within it, and that budget — not achievable noise performance in isolation — is what determines the front-end design.

Where the heat is made matters as much as how much. Thermal simulations for NeuroSeeker, a parallel imec probe, allowed 4.5 mW in the implanted shank but 45 mW in the base for the same 1 °C rise in the brain. Neuropixels dissipates nothing in the shank and all of its power, 17.5 mW or about 46 µW a channel, in the base. A buffer amplifier under every site was tried, at 1.3 mW for the shank, and gave no advantage — only more noise and more sensitivity to light.

Noise within that budget. Amplifier input-referred noise trades against current consumption. The site count is only useful if each channel is still quiet enough to resolve a spike, so the design lives at a specific point on that trade rather than at the low-noise extreme.

Neuropixels lands at about 5 µV RMS of noise in the spike band, helped by porous titanium nitride sites of low impedance, about 150 kΩ at 1 kHz; spikes can be as small as tens of microvolts. NeuroSeeker put an amplifier under each of its 1,356 sites, on 3 µW apiece, and time-multiplexed them along the shank: 12.4 µV RMS with half the sites on, up to 2.5 times worse with all of them.

Site selection, then multiplexing. The 960 sites feed 384 recording channels, so which sites are live is a programmable choice made by an on-chip switch matrix rather than a fixed wiring pattern — you choose your depth range after the probe is already in the brain. Downstream of that, groups of channels share converters and are sampled in sequence, which leaves a small known skew between channels.

The wiring limit did not disappear; it moved onto the chip. The switches sit in the shank, and every live site still needs its own metal line up a shank 70 µm wide, which caps how many can be live at once; Neuropixels has 384, covering 3.84 mm at full density. The choice is set before recording rather than scanned, because amplifiers that pass signals down to 0.5 Hz while rejecting the electrode’s DC offset settle slowly. The converters are 32 ten-bit units, each taking 12 channels in turn within every 33 µs sample period. Ten bits saved base area and power but cannot span spikes and slow potentials in one range, so each channel is split into a spike band, 0.3–10 kHz sampled at 30 kHz, and a local-field band, 0.5 Hz–1 kHz at 2.5 kHz, amplified and digitised separately.

What changed because of it

Recording from many neurons across several brain areas at once stopped being a specialist achievement and became something a normal lab does. That shifted the questions: from what a neuron responds to, toward how activity is distributed across a population and how it evolves over time — the same move microelectrode arrays started, carried far enough to change what most systems neuroscience is about.

The many areas are what needed the electronics: one thin cable per probe lets several probes into one small head at once. In 2019 Nicholas Steinmetz and colleagues inserted two or three at a time into mice making visual decisions and, over 39 sessions, recorded some 30,000 neurons in 42 brain regions, about 750 per session. Activity before a movement turned up almost everywhere, whichever way the mouse turned; neurons that predicted the choice were rare, and confined to parts of the neocortex, basal ganglia and midbrain. And activity before the stimulus, read across the population recorded together, predicted whether the mouse would respond on that trial: engagement meant more activity in subcortical regions and less in neocortex.

Some questions need the areas in the same millisecond. With up to six probes per mouse, Joshua Siegle and colleagues at the Allen Institute (2021) recorded up to eight visual areas of cortex and thalamus at once and cross-correlated spike trains between them. Of about two million pairs of cortical units, some 16,000 had a sharp peak within 10 ms, and which area tended to lead ranked the areas much as their anatomical hierarchy does — during visual stimulation, not during spontaneous activity. A lag that short between two areas cannot be read from recordings made on different days.

By 2025 it could be standardised across laboratories. The International Brain Laboratory, twelve labs running one task, made 699 insertions in 139 mice, mostly two probes at a time, covering 279 brain areas and yielding 621,733 units.

Four shanks and a smaller base

Neuropixels 2.0 (Steinmetz and colleagues, 2021) was shrunk for chronic implants in small animals. An implant a mouse can carry without hindrance must weigh under about 3 g; two probes and their shared headstage weigh about 1.1 g. The four-shank version has 5,120 sites — 1,280 per shank, in two aligned columns at 15 µm vertical pitch — on shanks 70 × 24 µm in section and 250 µm apart, still read through 384 channels. The base shrank to 2.2 × 8.7 mm while its power rose to 36.5 mW, and the two-band compromise went: one band, 0.5 Hz–10 kHz, at 14 bits and 30 kHz, or 161 Mbit/s a probe. The shank warms by less than 1 °C; the recording channel alone is a little noisier, 7.2 µV RMS against 5.4.

The new layout serves an algorithm. When the brain moves along the shank, a neuron’s spikes slide from site to site; with sites in straight columns 15 µm apart, software can register the recording as imaging registers frames, and the same neurons were followed for more than two months — with finer timing than chronic two-photon imaging gives, and in deeper structures than it reaches.

What it still cannot do

It hears spikes, not cells. Each neuron is inferred from the voltage near the shank by spike sorting — not yet truly automatic, four of its developers wrote in a 2018 review — and its type can at best be guessed from the waveform. Of the International Brain Laboratory’s 621,733 units, 75,708, about one in eight, passed quality control as well-isolated neurons.

The shank is rigid silicon, built for rodents. The same review named the uses it serves poorly: chronic implants in primates and other large animals, where tissue moves against a rigid probe; implants meant to last a year or more; and electrical microstimulation, which it cannot deliver.

And the wiring limit is still there, one level down. The single-shank 2.0 probe can connect two sites to each channel to span twice the length, but the channel then averages them: spikes come out at half size and, with electrode noise dominant, signal-to-noise should fall to $1/\sqrt{2}$, about 71%, of single-bank recording. It fell to 63.5%, and fewer neurons could be sorted than from the two banks recorded separately.