Record every sample from 3,072 channels and you push 571 megabits per second out of a head that can barely warm up. Recording hardware is the craft of keeping the signal and sending less.
Learn what an electrode hears, why it hisses, how a chip turns volts into bits, and why every brain implant is a fight between channel count, data rate and heat.
Pick an implant, then choose what leaves the chip: every sample, only spike events, or only counts. Watch the data rate and the power bill change by orders of magnitude. Then we build, piece by piece, the electrode, the amplifier, the converter and the budget underneath.
You need a feel for voltage and resistance (a battery pushes current through a resistor) and what a bit is. We build the rest from zero.
Built from primary sources: Musk and Neuralink 2019, Jun et al. 2017, Oxley et al. 2016, Maynard et al. 1997. Company and press figures are marked as claims. Part of the series Brain-computer interfaces, from the spike to the foundation model.
Channel counts, sample rates and bit depths are the papers’ (Neuralink 2019: 3,072 channels, 18.6 kHz, 10 bits; Neuropixels 2017: 384 channels, AP 30 kHz plus LFP 2.5 kHz, 10 bits). The 96-channel row assumes 30 kS/s at 16 bits, a typical setting, not a paper value. Illustrative: spike events at 10 per second per channel and 24 bits each, counts at 4 bits per 20 ms bin, a radio at 20 nJ per bit, a 320 mW heat budget, and 5.2 µW per channel in the amplifier from the Neuralink chip.
Chapter 0
Follow one spike from the tissue to the radio and see why every implant is a fight between channels, bits and heat
A woman lies in bed and wants to move a cursor on a screen. Her arm does not move, but the part of her brain that plans the movement still works. Hundreds of thousands of nerve cells near the surface of that region are busy. Each one sends tiny electrical pulses.
A neuron is a nerve cell. A spike is the short pulse it fires, lasting about one millisecond. When a neuron fires, a small current flows through the fluid around it. That current makes a small change in voltage nearby, and a metal tip placed close enough can sense it.
That metal tip is an electrode. One electrode, with the wire and the electronics behind it, gives one channel: a single stream of voltage readings over time. A brain-computer interface (BCI) is mostly a machine that collects many channels and turns them into a command.
Before the software can decode anything, the hardware must solve four problems, one after another. Each problem is a chapter of this lesson.
First, touch the signal. The electrode must sit close to the neurons, and the body must tolerate it for years. Second, hear it above the hiss. Every electrode adds random noise, and the spike is only a few times bigger. Third, turn volts into numbers. A chip must amplify, filter and round the voltage into a stream of integers. Fourth, move the numbers out. The implant sits in a closed skull, and it may only warm the tissue by a small amount.
The fourth problem is what makes the rest hard. Think of a garden hose and a bucket. A hose with ten times the flow fills the bucket ten times faster, and an implant with ten times the channels makes ten times the bits. But the heat the tissue can take does not grow. Something must give.
Why does the hardware deserve a whole lesson? Because every later step inherits its limits. A decoder, the program that turns neural signals into a cursor or a word, cannot recover a spike the electrode never heard. It cannot undo noise the amplifier added. And it cannot use data that the radio had no energy to send. Good decoders start from good hardware, and a team that reads a datasheet well can judge any claim about a BCI.
The same chain also appears in a wearable cap with no surgery at all. The signals are weaker and blurrier, so the amplifiers must be quieter, but the four problems stay the same: touch, hear, digitise and move. We meet both worlds here, the needle and the cap, and use one set of tools for both.
Here is the whole chain in one picture. The numbers come from the Neuralink chip described in its 2019 white paper (the paper’s Table 1 and text). The size of the spike itself, 100 µV, is a typical round number for the lesson, not a measurement. Drag the slider, or press the button, and watch what happens to the spike at each stage.
The plot shows the same 1.5 ms of voltage at each stage, measured at the input of the chip. Step along the chain and watch the noise arrive, then the rounding. Press the button to ride the whole way.
From the 2019 Neuralink white paper: input-referred noise 5.9 µV rms (3 Hz to 10 kHz), gain 42.9 to 59.4 dB, 7.2 mV peak-to-peak input range, 10-bit converter at 19.3 kHz. Illustrative: the 100 µV spike and its shape, and the 50 dB gain we pick inside the paper’s range. The noise drawn is random, seeded so it repeats on every visit.
Notice what the chain does to the spike. It starts as a clean dip. The electrode adds hiss. The amplifier makes everything bigger, and the converter turns it into stair steps. At the end, the spike is nothing but a list of integers: 29 of them for one spike, at 19.3 kHz.
Now multiply. One channel at 19.3 kHz and 10 bits is 193 kilobits per second (kb/s). The paper’s largest system had 3,072 channels at 18.6 kHz. That is 3,072 × 18,600 × 10 bits, or 571 megabits per second (Mb/s). The paper’s answer was a USB-C cable. A wire needs no radio, but it also means no wireless implant. That tension is the plot of this lesson.
This lesson follows the signal from the neuron outward, then looks at the real devices people have built, then at hardware that never touches the brain at all.
Chapter 1
Learn why a spike fades fast with distance, and why one electrode hears only the neurons that sit close to its tip
Stand in a huge stadium full of people, all talking. Hold a microphone at arm’s length from one person. You hear that person clearly. You also hear a murmur from everyone else, which you cannot separate into voices.
An electrode in the brain is that microphone. The neurons near its tip are the people at arm’s length. The thousands farther away are the murmur. The first job of recording hardware is to decide which voices count, and the answer depends on distance.
Brain tissue is a salty fluid, so it conducts electricity. When a neuron fires, ions rush through its membrane. That makes a small current flow out into the fluid around it. A current through a conductor makes a voltage, a push that depends on where you measure.
The electrode measures the voltage at its own tip, against a reference electrode far away. This is an extracellular recording, because the metal sits outside the cell. It is less invasive than poking a glass tube into one neuron, and it lets us watch many neurons for months.
A tiny current source in a uniform conductor makes a voltage that falls as one over the distance. Twice as far means half the voltage. Real neurons are more complex, and near the cell the fall can be steeper. The one-over-distance rule is a toy, but it shows the right shape: strong close up, weak far away.
Now add noise. In Chapter 2 we see why every electrode has some. The Neuropixels paper measured a mean noise of 5.1 µV rms (root mean square, a kind of average size) in its spike band, across 384 sites. A spike is only detected when it rises clearly above that noise. A common rule sets a threshold at some multiple of the noise, say 4 times.
Worked example, with toy neuron numbers. Say a neuron gives 150 µV at 20 µm. The noise is 5.1 µV, so a 4× threshold is 20.4 µV. Step one: set V(r) equal to the threshold, 150 × 20 / r = 20.4. Step two: solve, r = 3,000 / 20.4 = 147 µm. Past 147 µm, that neuron is lost in the hiss.
How many neurons live inside that distance? Cortex holds on the order of tens of thousands of neurons in each cubic millimetre. Take 50,000 as a round number. A sphere of radius 0.147 mm has volume 4/3 × π × 0.1473, or 0.0133 mm3. That gives about 665 neurons. Try the device. The dots are neurons in a toy slice of tissue.
Each dot is one neuron: its distance from the tip across, its spike size up. Dots above the dashed threshold are heard. Drag the noise and the threshold, and watch the heard crowd shrink or grow.
Noise default 5.1 µV rms is the Neuropixels paper’s mean for its spike band. Illustrative: the 150 µV at 20 µm, the one-over-distance fall, the spread of neuron sizes (0.4 to 1.6 times), 50,000 neurons per mm3, and the 4× rule. The 220 dots are a seeded random sample of a sphere of radius 300 µm.
Two lessons fall out of the device. First, halving the noise doubles the reach, so the detection radius grows as one over the threshold. A quiet electrode hears many more neurons. Second, the count grows with the cube of the radius. Doubling the reach means about eight times as many neurons.
The same electrode also hears slow voltage waves. The local field potential (LFP) is the summed, slow voltage of thousands of nearby neurons, below about 300 Hz (the Neuropixels LFP band reaches 1,000 Hz). The Neuropixels chip splits each channel into a spike band of 0.3 to 10 kHz and an LFP band of 0.5 to 1,000 Hz, and digitises them separately, so you can keep both.
Symptom: weeks after surgery, a channel that once heard clear spikes goes quiet. Measurement: the spike amplitude, plotted by day, sags, while the noise stays flat. The model above explains this. If scar tissue or small shifts move the neurons from 40 µm to 100 µm, the amplitude falls to 40% of what it was.
The fix is partly in the hardware (smaller, softer implants that move less with the brain, Chapters 6 and 7) and partly in software (decoders that adapt to a changing signal). The software half is for later lessons in this series.
Chapter 2
Compute why every electrode hisses, and see how a smaller or better coated tip changes how much of the spike you keep
You turn up a radio between stations. It hisses. You did not cause the hiss, and the radio is not broken. Heat makes the electrons inside every resistor jiggle at random, and the jiggle shows up as a voltage.
An electrode is a resistor too. A spike of 100 µV must stand out above its hiss. If the hiss is 5 µV, you see the spike clearly. If it is 50 µV, you do not. So the first number to know about any electrode is how much it hisses.
Resistance measures how hard it is to push current through something. Its unit is the ohm (Ω). A kilohm (kΩ) is a thousand ohms. Impedance is resistance that can change with the speed of the signal. A metal tip in salty fluid is not a plain resistor. Ions pile up at its surface and form a thin layer that acts like a tiny capacitor. So its impedance depends on frequency. By convention, papers quote it at 1 kHz, the middle of the spike band.
Three real numbers, all measured at 1 kHz. The Neuropixels paper reports a mean of 149 kΩ (standard deviation 6) for its titanium nitride sites, in saline. The Neuralink 2019 paper reports 36.97 kΩ (standard deviation 4.68, 257 electrodes) for gold sites coated with a polymer called PEDOT:PSS, and 56.46 kΩ (standard deviation 7.10, 588 electrodes) for iridium oxide. Both are benchtop values. The Neuralink sites are very small, so the coating does the work.
Why do coatings help? The impedance at the surface is roughly inversely proportional to the real surface area. A rough or porous coating has a huge area packed into a tiny footprint. So a small site stays small, but its impedance falls. Make a site four times smaller without a coating, and the impedance goes up about four times.
Physicists found that a resistor at temperature T makes a random voltage whose size depends on three things: its resistance, the temperature, and the bandwidth you listen to. It is called thermal noise, or Johnson noise.
Worked example, step by step, for the 149 kΩ Neuropixels site. Step one: 4 × 1.38×10−23 × 310 = 1.71×10−20. Step two: multiply by R = 149,000 to get 2.55×10−15. Step three: multiply by 9,700 Hz to get 2.47×10−11. Step four: take the square root, 4.97×10−6 V. That is 5.0 µV.
The paper measured 5.1 µV rms across its whole recording chain. The match is close, but treat it as a ballpark, not a proof. Only the resistive part of the impedance makes thermal noise, and the amplifier adds noise of its own. The two sources add in quadrature: total = √(thermal2 + amplifier2). The same sum for PEDOT:PSS at 37 kΩ gives 2.5 µV, and for iridium oxide at 56.5 kΩ gives 3.1 µV. Play with it.
The trace is a 100 µV spike buried in the noise your choices make. Pick a real coating or drag the impedance. Watch the dashed line: the threshold at 4 times the noise.
Impedances are measured values from the Neuropixels paper (149 kΩ, titanium nitride, saline) and the Neuralink 2019 paper (iridium oxide and PEDOT:PSS, benchtop). The formula is exact for a pure resistor at 310 K. Illustrative: using the full impedance as the resistor, the amplifier noise of 1 µV, the 100 µV spike and its shape. The noise on the trace is random, seeded.
The impedance slider is log scale, so each step is a multiple. Moving from 149 to 37 kΩ cuts the noise by half, because 37 is a quarter of 149 and noise goes as the square root. Likewise, a wider bandwidth lets in more noise. That is why spike detection filters out everything above 10 kHz.
The amplifier also has an input impedance. The electrode and the amplifier input form a voltage divider: the fraction of the spike that survives is Zin / (Zin + Zelectrode). With a 10 MΩ input and a 149 kΩ electrode, you keep 98.5%. With a 1 MΩ input you keep 87%. This is why chips put the first amplifier stage right next to the electrode: short wires keep unwanted capacitance low and the divider kind.
Symptom: a channel shows a huge, slow, drifting signal, or none at all. Measurement: inject a small known current at 1 kHz and read the voltage, which gives the impedance. A broken wire reads as very high impedance. A cracked insulation lets current leak away and reads as very low. A healthy coated site stays near its factory value. The fix is to drop the channel from the decoder, or to re-reference around it.
Chapter 3
Work out how many bits a recording chip needs, and see why ten is enough when the noise is a few microvolts
You want to weigh a coin on a bathroom scale. The scale reads in whole kilograms. The coin weighs five grams. The needle never moves. Now suppose the scale reads in grams: you learn the weight exactly. A neural recording chip has the same problem, in volts.
The spike is a few hundred microvolts. The chip’s converter works in volts, in a range of a few millivolts. In between sit three parts: an amplifier, a filter, and the converter. Together they are called the analog front end. Analog means the signal is still a smooth voltage. Digital means it is a list of integers.
The amplifier multiplies the voltage. Its gain is the ratio of output to input. Engineers write gain in decibels (dB), where 20 dB is a factor of 10, 40 dB is 100, and 60 dB is 1,000. The Neuralink 2019 chip has a gain of 42.9 to 59.4 dB, which is a factor of about 140 to 940. The Neuropixels probe has a gain you can select from 50 to 2,500.
The amplifier is the place where noise gets decided. Any noise it adds is multiplied along with the signal, so a good chip makes its own noise small next to the electrode’s hiss. That is why the paper quotes input-referred noise: the noise of the whole chain divided by the gain, so it can be compared with the spike at the input. The Neuralink chip has 5.9 µV rms (3 Hz to 10 kHz).
Slow voltage drifts and the local field potential (Chapter 1) can be thousands of times bigger than a spike. If you amplify everything, these big slow waves use up the converter’s range. So the chip first applies a band-pass filter, a filter that keeps a band of frequencies and removes the rest. The Neuropixels probe splits each channel into an action potential band (0.3 to 10 kHz) and an LFP band (0.5 to 1,000 Hz) and treats them with different gain: 250 for the LFP band, selectable for the spike band.
The analog-to-digital converter (ADC) measures the voltage many times per second. Each measurement is a sample. Each sample is rounded to the nearest of 2n levels, where n is the number of bits. Ten bits give 1,024 levels. The gap between neighbouring levels is the least significant bit, or LSB.
Rounding adds an error. If the true value falls anywhere in a step, with equal chance, the error is spread evenly across the step. Its rms size is the LSB divided by the square root of 12. This is the quantization noise. (The lesson on quantization noise derives it from scratch.) Sampling fast enough is the other half. The sampling rate must exceed twice the highest frequency you keep, or the fast parts fold back as false slow ones. The lesson on the discrete Fourier transform shows this folding.
Worked example with the Neuralink numbers. Step one: LSB = 7.2 mV / 1,024 = 7.03 µV at the input. Step two: q = 7.03 / 3.464 = 2.03 µV. Step three: the analog noise is 5.9 µV, so the total, adding in quadrature, is √(5.92 + 2.032) = 6.24 µV. Rounding costs only 6% more noise. We use 5.9 µV here as the analog noise before rounding, which is an assumption, since the paper quotes one total figure.
Now try 8 bits. The LSB is 28.1 µV and q is 8.1 µV. The total is 10.0 µV, or 70% worse. Try 12 bits. The LSB is 1.76 µV and q is 0.51 µV. The total is 5.92 µV, barely better than ten bits. The device shows all three.
The smooth line is the analog spike with 5.9 µV of noise. The stair steps are what the converter keeps. Drag the bits down and watch the steps swallow the spike. Then shrink the range and watch them get finer.
From the Neuralink 2019 paper: 7.2 mV peak-to-peak input range, 10 bits, 5.9 µV rms input-referred noise. Illustrative: the 100 µV spike, treating 5.9 µV as the noise before rounding, and the seeded random noise. A smaller range here means more gain ahead of the converter, which is why a real chip needs the filter to remove big slow waves first.
Why not use 16 bits and be safe? Each extra bit costs chip area, power and data, and every extra bit adds a tenth to the data rate at ten bits. When quantization noise is already well below the analog noise, more bits buy nothing but heat. This is the reasoning behind the Neuropixels paper’s choice of 10-bit converters to save base area and power.
Symptom: during a movement, a channel suddenly flat-lines at the top of its range. Measurement: the raw samples sit at 1,023 for many milliseconds. The cause is a large artifact, a muscle twitch or a cable movement, that pushes the voltage past the range. The converter cannot count above its top step. The fix is a high-pass filter ahead of the converter, a wider range, or an automatic reset that pulls the signal back to the middle.
Chapter 4
Compute what an implant can send, and find how many channels a heat budget can pay for
You need to move a whole house through a keyhole. You can carry the furniture in pieces, one item at a time, but each trip costs effort. Or you can send a short list of what the house contains. A brain implant faces the same choice with data.
The chips of Chapter 3 make a flood of integers. Chapter 0 gave the formula in a quiz. Here it is as a rule, then we use it on real systems and ask what it costs to move the bits.
Take the Neuralink 2019 white paper’s two systems. System A has 1,536 channels sampled at 19.3 kHz, 10 bits each. That is 1,536 × 19,300 × 10, or 296 Mb/s. System B has 3,072 channels at 18.6 kHz: 571 Mb/s. The Neuropixels probe has 384 channels, each digitised as a 30 kHz spike stream and a 2.5 kHz LFP stream, 10 bits each: 384 × 32,500 × 10, or 125 Mb/s.
These are rates inside the device, before any choice about what to send. The paper’s answer for System B is a USB-C cable that carries full-bandwidth data out of the implant. That works in the lab animals of the paper. A person who wants to walk around needs a wireless link, and radios are expensive in energy.
Power is energy per second. A radio spends some energy for each bit it sends. If it spends E joules per bit, then sending R bits per second costs E × R watts. We use 20 nanojoules per bit as an illustrative figure for a low-power link. The real figure depends on the radio, the distance and the standard, so the lesson does not claim it as a measurement.
The amplifiers cost power too. The Neuralink chip report says each analog pixel, one channel’s amplifier, uses 5.2 µW, and the whole 256-channel ASIC uses about 6 mW, or 23 µW per channel including everything on the chip. Whole recording systems cost far more: System A drew 550 mW for 1,536 channels (358 µW per channel) and System B 750 mW for 3,072 (244 µW per channel). Those totals include the programmable logic, the sensors and the USB link, not only the amplifiers.
Everything the implant burns turns into heat, and that heat goes into the brain. Even-Chen and colleagues (Nature Biomedical Engineering 2020) argue that the power budget of the implant limits the number of channels in a wireless interface, and that relaxing circuit specifications could save about a factor of ten. We quote no safety limit here, because we did not check one against its original source. The device below uses a round 320 mW as an illustrative heat budget, not a measured or regulatory figure.
The cure is to reduce the data on the chip, before the radio. There are three common levels. Send every sample (raw). Send only spike events: when a channel crosses its threshold, send a time stamp and a channel number. Or send bin counts: for each channel, count the threshold crossings in every 20 ms window, and send a small number.
Worked example for 3,072 channels, with the illustrative settings of the hero. Raw: 571 Mb/s, and 11.4 W of radio. Events: 3,072 × 10 per second × 24 bits is 0.737 Mb/s, and 14.7 mW of radio. Counts: 3,072 × 50 × 4 bits is 0.614 Mb/s. Adding the amplifiers, 3,072 × 5.2 µW = 16 mW, events cost about 31 mW in all. That is under a tenth of the budget. Raw costs 36 times the budget.
The price is information. Events and counts throw away the shape of the waveform. You can no longer tell two neurons on the same channel apart afterwards. Many decoders do not need that, as later lessons in this series show. But a lab that wants to study single neurons keeps the raw stream. Build your own budget in the device.
Pick what leaves the chip, then drag the channels, rate, bits and radio cost. The bars are log scale. The dashed line is your heat budget.
From the papers: the amplifier at 5.2 µW per channel (Neuralink 2019 chip), the starting channel count, rate and bits (Neuralink System B). Illustrative: the radio energy, the heat budget, 10 events per second per channel at 24 bits, and 50 bins per second at 4 bits. Sample rate and bits apply to the raw choice only. The amplifier cost is held at 5.2 µW per channel for every system.
Drag the channels up with the raw choice on. Even a radio at 5 nJ per bit cannot carry thousands of raw channels inside a modest budget. Switch to events, and the same budget pays for tens of thousands. Notice that at that point the amplifiers cost as much as the radio. The next limit is the front end itself.
Symptom: the implant’s own temperature sensor reads a few degrees above body temperature after a long session. Measurement: log the sensor next to the data rate (the Neuralink system includes a temperature sensor on its board). The fix is to cut the rate: drop the sample rate, send events, or turn off channels that carry no spikes.
Chapter 5
Compare a bed of nails with a single long needle, and see what each one samples from the tissue around it
Imagine you want to learn what a crowd is saying. You can send a hundred reporters, each to a different spot on one street. Or you can send one reporter who walks slowly down the street from end to end. The first plan gives many places at one depth. The second gives one place at many depths. Two famous silicon devices follow these two plans.
The Utah array was described by Maynard, Nordhausen and Normann in 1997, under the title “The Utah intracortical electrode array: a recording structure for potential brain-computer interfaces.” It is cut from a block of silicon into a square grid of 10 × 10 needles. As commonly described by the University of Utah group, each needle is about 1.5 mm long, the needles stand 400 µm apart, and a base about 4 mm square holds them. Only the platinum-coated tip of each needle is bare. The rest is insulated.
Surgeons press the array into the surface of the motor cortex so that every tip lands at about the same depth. Clinical arrays have 96 working channels (Even-Chen and colleagues describe their data as coming from “96-channel Utah multielectrode arrays”). Wires leave the array through a connector that passes through the skin, the “pedestal.”
The Utah array is the workhorse of human intracortical BCI. The BrainGate safety report (Rubin et al., 2023) covers 14 adults who received one or two arrays between 2004 and 2021. The average implant lasted 872 days, which adds up to 12,203 days of experience. There were 68 device-related adverse events, including 6 serious ones. The most common was skin irritation around the pedestal. No event required removing the device.
Jun and colleagues (Nature, 2017) took the other plan. Their Neuropixels probe is one thin shank of silicon, 10 mm long, with a cross-section of 70 × 20 µm. It carries 960 recording sites made of titanium nitride, each 12 × 12 µm, in a four-column checkerboard. Only 384 can record at once, so a switch picks which 384 sites connect to the 384 channels. That is 2.5 sites per channel.
The big idea is that the amplifiers, filters and converters sit on the same silicon chip as the shank. The 6 × 9 mm base digitises the signal, so a thin digital cable leaves the probe and not a thousand fragile analog wires. The finished probe weighs about 250 mg. The paper reports well-isolated spiking from hundreds of neurons per probe, and more than 700 neurons at once from two probes across five brain structures in an awake mouse.
Efficiency check, using the paper’s Table 1. The base uses 17.5 mW for 384 channels, which is 17.5 / 384 = 45.6 µW per channel, for two bands per channel (30 kHz plus 2.5 kHz). The Neuralink chip of Chapter 6 reports about 23 µW per channel for one 19.3 kHz band. The designs differ, so the numbers show the scale, not a contest.
One more design choice deserves a name. Because Neuropixels has 960 sites but only 384 channels, it is a switchable probe. The experimenter picks which block of sites to read out, like choosing which part of a long street to listen to. The Utah array has no such switch: each needle is wired to its own channel for life. A switch buys flexibility, since you can move the listening window along the shank after the probe is in place. It costs a little extra circuit area on the chip.
Both devices sample the tissue sparsely. Chapter 1 gave a toy hearing range for a spike. Draw a circle of that radius around each site, and most of the tissue lies outside every circle. Try it. The picture is a side view of a 4 mm × 4 mm slab of cortex.
Choose a device. Each circle is the hearing range around a site. Drag the range to see how much of the slab each one reaches.
Geometry from the papers: Utah needles 1.5 mm long on a 400 µm pitch (as commonly described, 10 across, one row shown); Neuropixels 960 sites over a 10 mm shank, so 384 channels cover about 3.8 mm (approximate, 384 of 960 sites in a row of the shank). The hearing range is the toy of Chapter 1, not a measured value. The coverage percentage counts a flat slice, not a volume.
Even at 250 µm, neither covers most of the slab. This is the main reason the design question is about where to put sites, not only how many. The Utah plan reaches many columns of cortex. The Neuropixels plan reaches many layers of one column, from the surface down through deep structures.
Symptom: a neuron that fired clearly on Monday looks weaker on Friday, and a new one appears on a neighbouring channel. Measurement: the average spike shape on each channel, tracked day by day, drifts across channels. A cause is that brain tissue is very soft and silicon is orders of magnitude stiffer. The brain shifts with heartbeat and breathing, and rigid needles do not shift with it. Two fixes exist. One is to hold the device very still with a rigid mount, which the Neuropixels paper does for its chronic rat recordings. The other is software that tracks the drift, covered in later lessons.
Chapter 6
Build the 2019 Neuralink array on paper: threads, a sewing robot, a 256-channel chip, and the yield you actually get
A stiff needle in a soft brain is like a steel pole in a bowl of jelly. Every time the jelly wobbles, the pole pushes against it. A thread of thin plastic wobbles with the jelly instead. That is the idea behind flexible electrodes. The price is that a floppy thread cannot be pushed in by hand.
In 2019 a company white paper, listed under the name of Elon Musk and Neuralink in the Journal of Medical Internet Research, set out one answer. It is a company paper, so read it as an engineering report, not as a clinical trial. Its numbers are the source of this chapter, and each is the paper’s own.
A thread is a strip of polymer with metal wires inside. The paper describes threads from 5 to 50 µm wide, nominally 4 to 6 µm thick. A human hair is around 70 µm across. The paper builds each thread from a plastic called polyimide that wraps a thin gold trace, with up to two layers of conductor and up to three layers of insulation. Small gold pads at the surface are the electrodes. Each thread in the big array carries 32 electrodes. Chapter 2 showed that these tiny pads would hiss too much if left bare, so the paper coats them with PEDOT:PSS or iridium oxide.
A neurosurgical robot inserts the threads one at a time. A needle made of tungsten-rhenium wire, etched down to 24 µm, hooks a small loop at the head of a thread and carries it into the brain. The robot can place a thread anywhere on the surface with micron precision, so it can steer around the small blood vessels that cover the brain. The paper states a rate of six threads, 192 electrodes, per minute. A full array of 96 threads therefore takes 96 / 6 = 16 minutes of insertion.
Thousands of electrodes cannot each have a wire to the outside. The paper’s answer is a custom chip, an application-specific integrated circuit (ASIC), placed next to the threads. Each ASIC has 256 amplifiers it calls “analog pixels,” one per channel, with a 10-bit converter on the chip and circuits to serialise the output. Its numbers: 42.9 to 59.4 dB gain, 5.9 µV rms input-referred noise, 7.2 mV peak-to-peak input range, 19.3 kHz sampling, 5.2 µW per analog pixel, about 6 mW for the whole chip. These are the numbers of Chapter 3, now in context. Twelve chips make 3,072 channels, since 12 × 256 = 3,072.
The 2019 chip does not decide which samples matter. It digitises everything and streams it out through a single USB-C cable. In the paper, spikes are detected in real time after the data leave the device, with an online algorithm tuned for BCI use. Company and press descriptions of later chips talk about spike detection on the chip itself. We treat those descriptions as claims and could not check them against a paper.
| System | Channels | Sampling | Power | Size and weight |
|---|---|---|---|---|
| A (Neuralink 2019) | 1,536 | 19.3 kHz | 550 mW | 24.5 × 20 × 1.65 mm, 11 g |
| B (Neuralink 2019) | 3,072 | 18.6 kHz | 750 mW | 23 × 18.5 × 2 mm, 15 g |
Not every electrode hears a neuron. The paper’s spiking yield is the share of channels that show spikes. It reports yields of up to 70% in chronically implanted electrodes, and 43.4% on one array of 1,280 channels, where many spikes appeared on several neighbouring channels. So: 3,072 × 70% = 2,150 channels with spikes, in the best case. And 1,280 × 43.4% = 556 on the single array. Build one in the device.
Each column is one thread of 32 electrodes. Drag the threads to place them and the yield to light the electrodes that hear a spike.
From the Neuralink 2019 paper: 32 electrodes per thread, 96 threads, 256 channels per chip, about 6 mW per chip, six threads per minute, yields of 43.4% (one 1,280-channel array) and up to 70%. Which electrodes light is a seeded random draw at the chosen yield. Chip power is the chip alone, not the whole system (System B drew 750 mW).
Look at what the yield does to the budget. If 43% of channels carry a spike, then 57% of the data you digitise and the energy you spend on it records nothing useful. That is a strong reason to detect spikes before sending, and to switch off silent channels.
Symptom: a decoder trained on 3,000 channels performs no better than one trained on 800. Measurement: look at the raster. Many spikes appear on several neighbouring channels at once, as the paper itself notes for high-density probes. The same neuron is being counted on three channels. The fix is to compare channels for shared spikes and merge or drop duplicates, a job for the signal-processing lessons later in this series.
Chapter 7
Trade the sharp view of a needle for a safer route, and see exactly what the extra distance blurs
Listen to a concert from the front row, from the car park, and from the next town. In the front row you hear each instrument. In the car park you hear a blend. In the next town you hear a low rumble. The orchestra did not change. The path between you and it did.
Every step from a needle to a scalp cap puts more material between the electrode and the neurons. The material spreads and weakens the electric field. Engineers accept that loss to avoid brain surgery, or to avoid surgery altogether. This chapter shows where each choice sits and what it costs.
A stent is a small mesh tube that doctors place inside a blood vessel to hold it open. Oxley and colleagues (Nature Biotechnology, 2016) put electrodes on a stent and called the result a stentrode. A catheter carries it up through the blood vessels and releases it in a vein that runs over the motor cortex. No one opens the skull. In sheep, the group recorded freely moving animals for up to 190 days. The paper reports that the spectral content and bandwidth were comparable to recordings from electrodes laid on the dura, the tough membrane over the brain, and that the vein stayed open for the whole time.
The first human results came from the SWITCH study (Mitchell et al., JAMA Neurology, 2023). Five people with severe paralysis were enrolled and four were analysed. The device went in through the jugular vein to a large vein over the sensorimotor cortex. All four finished 12 months with no serious adverse events, no vessel blockage and no device movement. The mean signal bandwidth was 233 Hz (standard deviation 16) and stayed steady. At least five attempted movement types were decoded offline, and every patient controlled a computer.
Read that bandwidth carefully. 233 Hz is far below the 10 kHz of a spike band. A stentrode in a vein sees slow field activity from many neurons, not single spikes. It is a coarse sensor with an excellent safety story.
Electrocorticography (ECoG) places electrodes on the surface of the brain, under the skull. Newer devices make the array from thin flexible film so it follows the folds of the brain. Precision Neuroscience describes such an array. It is a company product, and we could not find a peer-reviewed paper for it, so we make no numerical claim here. A surface array hears the summed activity of the neurons beneath each contact,.
Electroencephalography (EEG) puts electrodes on the scalp. The field passes through fluid, bone and skin on the way. Chapter 8 returns to this case, since a new wave of companies bets on better scalp sensors and better models.
The field spreads out as it travels. A spread like that is a spatial low-pass filter: fine detail is washed out and the peak falls. The standard picture of a blur is a Gaussian bump: a source of width σs seen through a blur of width σb becomes a bump of width √(σs2 + σb2), and its peak falls by the factor σs2 / (σs2 + σb2) in two dimensions. Worked example: a source 0.8 mm wide seen through a 10 mm blur has a peak of 0.64 / (0.64 + 100) = 0.6% of the original. That tiny fraction is why scalp signals are small and why EEG amplifiers need very low noise.
The device shows three sources on a 20 mm square of cortex. Two are 4.5 mm apart. Pick a recording site and see whether you can still tell them apart. The blur widths and sensor spacings are illustrative round numbers, one per recording site.
Choose where the sensor sits. The picture shows what that sensor could see: width and brightness. The dots are sensors at that site’s typical spacing.
Illustrative throughout: the three Gaussian sources (0.8 mm wide, two of them 4.5 mm apart), the blur widths (0.1, 1, 3 and 10 mm) and the sensor spacings (0.4, 1.5, 4 and 20 mm). The 233 Hz stentrode bandwidth, the 12 months and the four patients are from the SWITCH paper. Brightness is shown on the same scale for all four sites, so a dim picture means a weak signal.
Two things happen as the blur grows. The two close sources first touch, then merge into one lump. And the picture dims. With the dimmer, wider picture, a coarse grid of sensors is enough, which is exactly why scalp caps use tens of sensors and needles use hundreds or thousands.
| Site | What it hears | Evidence in this chapter |
|---|---|---|
| In the tissue | Spikes of single neurons, to 10 kHz | Neuropixels 2017: spike band 0.3 to 10 kHz |
| On the cortex | Summed activity under each contact | Company description, no paper checked |
| In a vein | Slow field activity, 233 Hz mean bandwidth | SWITCH 2023, four patients, 12 months |
| On the scalp | Very blurred, weak field | Chapter 8 |
Symptom: a decoder for individual finger movements works at chance on a vein or scalp sensor. Measurement: the signal spectrum stops at a few hundred hertz, and neighbouring channels are nearly copies of each other. The fix is not more channels on the same blurred field. It is to ask coarser questions (a switch, a selection among a few choices) or to move the sensor closer.
Chapter 8
Work out what a sensor that never touches the skin must do to hear anything, and what many sensors can and cannot buy
A hospital EEG cap comes with a tube of cold gel. A technician parts your hair at every electrode and fills the gap with it. The gel makes a good electrical bridge to the skin. It also takes half an hour, dries out, and washes out of your hair. A product for daily use cannot ask for that.
So a new group of companies builds sensors that need no gel. Sabi, a company that describes a non-invasive “brain foundation model,” puts it in three steps on its own site: imaging with “novel neuroimaging sensors powered by custom ASICs,” collecting a large dataset, and mapping brain signals to thoughts with a foundation model. Everything the company says about its own performance in this chapter is a claim, marked as such. The physics below does not depend on any claim.
A wet electrode uses gel and gives the lowest skin impedance. A dry electrode presses bare metal or a conductive pad onto the skin, often through hair. Its impedance is much higher, typically hundreds of kilohms or more. A non-contact (capacitive) electrode never touches the skin. A thin insulating layer or an air gap sits between the metal and the skin, and the two form a small capacitor, a device that passes changing signals across a gap without a direct connection.
Chapter 2 showed that a high-impedance source needs an amplifier that draws almost no current. The same divider rule applies, now with a capacitor. At frequency f, a coupling capacitance C has an impedance of 1 / (2πfC). The sensor’s amplifier input has a resistance Rin. The fraction of the scalp voltage that reaches the amplifier is:
Worked example with illustrative parts. Take C = 1 pF and f = 10 Hz. Step one: 1 / (2π × 10 × 10−12) = 15.9 GΩ. Step two: if Rin = 10 GΩ, the ratio inside is 15.9 / 10 = 1.59. Step three: g = 1 / √(1 + 1.592) = 0.53. So half the signal is gone, and more is lost at lower frequencies. With Rin = 100 GΩ the ratio is 0.159 and g = 0.99. This is why non-contact sensors need amplifiers built for extremely high input resistance, and why designers put the first stage right at the sensor.
Now the claim that stands out: up to roughly 100,000 chip-scale non-contact sensors, each 1 to 5 mm across, on one cap (press accounts of the company; not verified here). Suppose each sensor adds its own random noise, independent of the others. If you average N of them, the independent noise falls by √N. With 100 sensors it falls tenfold.
But recall Chapter 7. The scalp sees the brain through a blur of about 10 mm (our illustrative figure). Sensors 1 to 5 mm apart then see nearly the same field. More of them do not add spatial detail. They add redundancy that cuts independent noise, and they can help tell brain signals from local artifacts. Whether that is enough to decode imagined words is an empirical question. Press accounts of the company report more than 100,000 hours of EEG for training and a target of about 30 words per minute of typing. These are claims, and we found no independent accuracy data.
Try the physics in the device below.
The curve shows what fraction of the scalp signal reaches the amplifier at each frequency. The shaded band is the usual EEG range. Move the capacitor and the input resistance. Then add sensors to average.
The formula is exact for an ideal series capacitor and an ideal input resistor. Illustrative: the capacitance range, the resistance range, a 20 µV scalp signal and 2 µV of independent noise per sensor. Averaging assumes the noise is independent and the signal identical across sensors; interference that all sensors share, such as mains hum, does not average away. The 100,000-sensor and 30 words per minute figures are company claims.
The sliders show the design pressure. To keep the fraction near 1, make C larger (a thinner gap, a better dielectric), make Rin larger (a better amplifier), or both. Each costs something: larger C means contact with hair or skin, and larger Rin means a more delicate input stage, which also picks up more interference.
A sensor of a few millimetres cannot carry a bulky amplifier. The input stage, the filter and the converter all go onto one chip, in the same spirit as the Neuralink ASIC. Press accounts report that the company claims its own EEG chip is more than 50 times as power-efficient as a standard part from Texas Instruments. That is a claim we could not check. Power matters for a head-worn product for the same reason as in Chapter 4: everything that runs on your head runs on a battery.
Symptom: a huge, steady wave drowns the brain signal. Measurement: look at the spectrum. A sharp peak at 50 or 60 Hz is mains hum. A wide, slow wobble that follows head movement is motion artifact. The fix for hum is a good shared reference and a differential input (it reads the difference between two points, so whatever both share cancels). The fix for motion is a stable contact and a way to measure the motion, so software can subtract it. Dry and non-contact sensors suffer both worse than wet ones.
Chapter 9
Keep the numbers, the checklist and the core calculations in one place, and see where to go next
An electrode hears the neurons close to its tip, and the signal falls fast with distance. Its impedance sets a hiss that must stay below the spike. An amplifier, a filter and a converter turn volts into 10-bit integers. The rate of those integers, times the price of moving each bit, meets a heat limit. So the craft is to keep the signal and send less, and to place sensors as close as the risk allows.
| Item | Number or meaning | Source |
|---|---|---|
| Neuropixels probe | 384 channels, 960 sites, 10 mm shank (70 × 20 µm), 12 × 12 µm TiN sites | Jun et al. 2017 |
| Neuropixels noise and impedance | 5.1 ± 0.6 µV rms (spike band), 149 ± 6 kΩ at 1 kHz | Jun et al. 2017 |
| Neuropixels digitising | 10 bits; spike band 30 kHz, LFP band 2.5 kHz; base 17.5 mW | Jun et al. 2017 |
| Neuralink array (2019) | 96 threads × 32 electrodes = 3,072; six threads per minute; threads 5 to 50 µm wide, 4 to 6 µm thick | Musk and Neuralink 2019 |
| Neuralink coatings | PEDOT:PSS 37.0 kΩ (n=257), iridium oxide 56.5 kΩ (n=588), at 1 kHz | Musk and Neuralink 2019 |
| Neuralink ASIC (2019) | 256 channels, 42.9 to 59.4 dB, 5.9 µV rms, 7.2 mV range, 10 bits at 19.3 kHz, 5.2 µW per pixel, about 6 mW | Musk and Neuralink 2019 |
| Neuralink systems (2019) | A: 1,536 ch, 550 mW. B: 3,072 ch, 750 mW. Yield 43.4% (one array), up to 70% | Musk and Neuralink 2019 |
| Utah array | 10 × 10 needles, about 1.5 mm, 400 µm pitch; 96 working channels in clinical use | Maynard et al. 1997 (geometry as commonly described); Even-Chen et al. 2020 |
| BrainGate safety | 14 adults, 872 days on average, 68 device-related adverse events, 6 serious, no explants | Rubin et al. 2023 |
| Stentrode | Sheep up to 190 days. Humans: 4 analysed, 12 months, no serious events, mean bandwidth 233 Hz | Oxley et al. 2016; Mitchell et al. 2023 |
| Company and press claims | Later Neuralink implant of 1,024 electrodes; Sabi sensors up to about 100,000, chip 50 times more efficient, about 30 words per minute | claim not checked against a paper |
| Thermal noise | v = √(4kTRΔf); 149 kΩ over 9.7 kHz at 310 K gives 5.0 µV | Standard physics, worked here |
| Quantization | LSB = range / 2n; noise = LSB / √12; 7.2 mV at 10 bits gives 7.03 and 2.03 µV | Worked here |
| Data rate | channels × samples per second × bits. System B: 571 Mb/s | Worked here |
Five short functions cover every worked example in this lesson. The comments show the values we computed by hand, and we ran the code to confirm them.
python import math K_B = 1.380649e-23 # J/K def thermal_noise_uv(z_ohm, bw_hz, temp_k=310.0): """Johnson noise of a resistor, microvolts rms.""" return math.sqrt(4 * K_B * temp_k * z_ohm * bw_hz) * 1e6 def lsb_uv(range_mv, bits): return range_mv * 1000 / 2 ** bits def quant_uv(lsb): return lsb / math.sqrt(12) def rss(*parts): return math.sqrt(sum(p * p for p in parts)) # independent noise adds in quadrature def bitrate(channels, fs_hz, bits): return channels * fs_hz * bits # bits per second def radio_mw(bps, nj_per_bit): return bps * nj_per_bit * 1e-9 * 1e3 def max_channels(budget_mw, bps_per_channel, nj_per_bit, afe_uw=5.2): per_channel_mw = afe_uw * 1e-3 + radio_mw(bps_per_channel, nj_per_bit) return budget_mw / per_channel_mw thermal_noise_uv(149e3, 9700) # 5.0 uV Neuropixels site, spike band lsb = lsb_uv(7.2, 10) # 7.03 uV Neuralink chip step quant_uv(lsb) # 2.03 uV rss(5.9, quant_uv(lsb)) # 6.24 uV noise after rounding bitrate(3072, 18600, 10) / 1e6 # 571 Mb/s System B, raw bitrate(3072, 10, 24) / 1e6 # 0.737 Mb/s events only (illustrative rates) radio_mw(bitrate(3072, 18600, 10), 20) # 11,428 mW raw, at 20 nJ per bit (illustrative) max_channels(320, 10 * 24, 20) # 32,000 channels inside 320 mW, events only
Recording hardware sets the ceiling for everything after it. A decoder can only use the signal that survives the electrode, the chip and the link. The newest work moves smarts into the chip: spike detection, compression and even small decoders on the implant. Even-Chen and colleagues argue that this is how wireless implants can reach higher channel counts. Later lessons in this series cover what the decoders do with the data.
This lesson was built from primary sources opened for it: Musk and Neuralink, “An integrated brain-machine interface platform with thousands of channels,” J Med Internet Res 2019 (doi). Jun et al., “Fully integrated silicon probes for high-density recording of neural activity,” Nature 2017 (doi). Oxley et al., Nature Biotechnology 2016 (doi). Mitchell et al., JAMA Neurology 2023 (doi). Rubin et al., Neurology 2023 (doi). Even-Chen et al., Nature Biomedical Engineering 2020 (doi). Maynard, Nordhausen and Normann, Electroencephalography and Clinical Neurophysiology 1997 (doi), whose geometry we took from the University of Utah’s own descriptions, because the original is not open. Company statements come from sabi.com and press accounts, and are marked as claims. The toy numbers, the radio energy and the heat budget are ours and are labelled illustrative wherever they appear.
Now try it on your own: go back to the opening device and find the smallest data rate that still fits the budget for 3,072 channels. Then open Teach mode and explain the chain from spike to bit on the whiteboard.