The power and heat budget
Add up where the milliwatts go and you get the implant's total draw: the per-channel analog front-end (set by the NEF from Guide 2), the ADC and digital logic, on-chip DSP, and the radio. Because the analog and ADC terms replicate per channel, total power grows roughly linearly with channel count.
P_{\mathrm{total}} \approx N_{\mathrm{ch}}\,(P_{\mathrm{amp}} + P_{\mathrm{ADC}}) + P_{\mathrm{DSP}} + P_{\mathrm{radio}}The implant power ledger. The per-channel term scales with N; whether P_radio or the per-channel term dominates is exactly the on-chip-compute tradeoff of Guide 3.
The total implant power is the per-channel cost (amplifier plus ADC) times the number of channels, plus fixed costs for on-chip processing and the radio. Whether the radio or the per-channel term dominates is exactly what decides if on-chip compute pays off — and it's bounded by tissue heating.
- P_{\mathrm{total}}
- The whole implant's power draw.
- N_{\mathrm{ch}}(P_{\mathrm{amp}} + P_{\mathrm{ADC}})
- Per-channel amplifier and ADC power, multiplied by channel count.
- P_{\mathrm{DSP}}
- Power spent on on-chip digital processing.
- P_{\mathrm{radio}}
- Power spent transmitting data out wirelessly.
With thousands of channels, the N_{\mathrm{ch}} term can swamp everything else in the budget.
The ceiling is not battery life — it is biology. Every one of those milliwatts becomes heat deposited in tissue that cannot easily carry it away, and neural tissue tolerates only about a degree of sustained warming before function and safety are threatened. This is the single hardest constraint in the whole field.
p_{\mathrm{diss}} = \dfrac{P_{\mathrm{total}}}{A} \;\lesssim\; p_{\mathrm{safe}} \;\sim\; \text{a few tens of mW/cm}^2 \;\;(\Delta T \lesssim 1\,^{\circ}\mathrm{C})The heat-flux guideline: power dissipation per unit implant surface must stay low enough to keep the sustained tissue temperature rise around a degree. Commonly cited limits are on the order of tens of mW/cm².
It's not the total power that harms tissue but the power spread per unit surface area — the heat flux. Keep it low enough (of order tens of mW/cm²) that the tissue warms by only about a degree, or you risk cooking the cells you're trying to record from.
- p_{\mathrm{diss}}
- Power dissipated per unit implant surface area — the heat flux.
- A
- The implant's surface area over which power is spread.
- p_{\mathrm{safe}}
- The safe flux limit — commonly a few tens of mW/cm².
- \Delta T
- The sustained tissue temperature rise, kept around a degree.
Spreading the same total power over a larger surface lowers the flux and the temperature rise.
Inductive wireless power transfer
A sealed implant needs power without wires through the skin, which would be an infection route. The dominant solution is inductive power transfer: an external coil drives an alternating field that couples into an implanted coil, like a loosely coupled transformer across tissue. The achievable efficiency is governed by one figure of merit, the product of the coupling coefficient squared and the two coils' quality factors.
\eta_{\max} = \dfrac{k^2 Q_1 Q_2}{\left(1 + \sqrt{1 + k^2 Q_1 Q_2}\right)^2}Maximum two-coil link efficiency. Everything hinges on the product k²Q₁Q₂: distance and misalignment shrink k, so implants fight to keep Q high and the coils close and aligned.
The best power you can transfer between two coils hinges almost entirely on the product k^2 Q_1 Q_2. Distance and misalignment shrink the coupling k, so implants fight to keep the coil quality Q high and the two coils close and aligned.
- \eta_{\max}
- Maximum achievable power-transfer efficiency of the two-coil link.
- k
- The coupling coefficient — how well the coils share flux.
- Q_1,\ Q_2
- The quality factors of the transmit and receive coils.
- k^2 Q_1 Q_2
- The single product that governs efficiency.
A larger k^2 Q_1 Q_2 pushes efficiency toward 1, while a loose, distant coil (small k) kills it.
Two consequences follow. First, coupling k falls fast with separation and misalignment, so link efficiency is fragile to exactly the things surgery and daily life cannot fully control — implant depth, head movement, coil placement. Second, whatever power you fail to deliver efficiently is dissipated as heat in the tissue between the coils, feeding straight back into the thermal budget above. Power delivery and heating are not separate problems.
Wireless telemetry and the Shannon wall
Data must leave the implant wirelessly too, and here information theory sets the ceiling. A radio link of bandwidth B at a given signal-to-noise ratio can carry no more than the Shannon capacity, and every real link falls short of it.
C = B\,\log_2\!\left(1 + \mathrm{SNR}\right)Shannon capacity: the ceiling on bits per second. Raising rate by widening B or boosting transmit SNR both cost power — the fundamental link between throughput and the energy budget.
This is the hard ceiling on error-free bits per second over a link, set by bandwidth and signal-to-noise. You can go faster by widening the band or boosting SNR — but both cost power, which is the fundamental tie between throughput and the energy budget.
- C
- Channel capacity — the ceiling on reliable bits per second.
- B
- The bandwidth of the link.
- \mathrm{SNR}
- Signal-to-noise ratio — helps only through a logarithm.
Doubling bandwidth roughly doubles capacity, but you must quadruple SNR just to add a couple of bits.
For an implant the more telling metric is energy per bit: the radio power divided by the data rate. State-of-the-art implantable radios reach the picojoule-to-nanojoule-per-bit range, and this is the reason the raw firehose cannot simply be transmitted — at hundreds of Mbit/s, even a few nJ/bit would blow through the entire thermal budget on the radio alone.
E_b = \dfrac{P_{\mathrm{radio}}}{R_{\mathrm{data}}} \quad\Longrightarrow\quad P_{\mathrm{radio}} = E_b\,R_{\mathrm{data}}Energy per bit ties telemetry power directly to data rate; since R_data = N_ch·f_s·N_bits, the only way to cut radio power without dropping channels is to cut bits per channel — i.e. compress or detect on-chip.
Radio power equals the energy each bit costs times how many bits per second you send. Since the data rate is channels times sample rate times bits, the only way to cut radio power without dropping channels is to send fewer bits per channel — i.e. compress or detect right on the chip.
- E_b
- Energy spent per transmitted bit.
- P_{\mathrm{radio}}
- The radio's power draw.
- R_{\mathrm{data}}
- The data rate — channels times sample rate times bits.
At a fixed energy per bit, halving the bits per channel halves the radio power.
Compression as a power strategy
Now the loop closes. Guide 3 introduced on-chip spike detection and compression as ways to shrink R_{\mathrm{data}}; the equation above shows exactly why that matters — every bit you do not send is energy you do not spend on the radio. The decision rule is a simple inequality: include on-chip processing only when the telemetry energy it saves exceeds the compute energy it costs.
E_b\,\Delta R_{\mathrm{data}} \;>\; P_{\mathrm{compute}} \quad\Longrightarrow\quad \text{on-chip processing pays off}The compression break-even: on-chip work is worth it when the radio energy saved by the reduced data rate ΔR exceeds the extra compute power it demands.
On-chip crunching is worth it only when the radio energy you save by shrinking the data rate beats the extra power the computation itself burns. It's a plain cost-benefit line: save more watts on the radio than you spend on the processor.
- E_b\,\Delta R_{\mathrm{data}}
- Radio power saved by cutting the data rate by \Delta R_{\mathrm{data}}.
- P_{\mathrm{compute}}
- The extra power the on-chip compression or detection demands.
- \Delta R_{\mathrm{data}}
- The reduction in data rate that compression achieves.
If compressing on-chip saves more radio watts than the compressor draws, do it on the chip.