ASBIT event telemetry and the distributed MEA front end

What the work claims

Lee, Lee, Leung, Laiwalla, Lopez-Gordo, Larson and Nurmikko report ASBIT, an asynchronous, sparse, binary protocol for wireless sensor networks.¹ The idea is to treat each sensor node like an event-based pixel: it only talks when it detects a change, encodes the event with a unique Gold-code address, and backscatters the packet to a single external transceiver near 900 MHz. Because the data are sparse and asynchronous, the network reuses spectrum, code and timing rather than reserving fixed slots for every node.

The authors fabricated 78 prototype communication chips in TSMC's 65 nm mixed-signal/RF low-power CMOS process. Each die is 300 × 300 µm, carries its own power-harvesting coil, rectifier, regulator, oscillator, Gold-code generator, digital finite-state machine and BPSK backscatter modulator. They then measured backscattered signals from the real chips and, by replaying recorded background packets, extrapolated the system's behavior to networks of 200, 500, 1,000 and 2,500 nodes.

At an average event rate of 50 Hz and an average SNR of 1.7 dB, the free-running-oscillator version supports about 750 nodes while keeping the event error rate (EER) at 1.19 × 10^-3; with sparser background firing, the same EER holds out to roughly 2,500 nodes. A second design that derives its clock from a frequency divider on the incoming RF carrier supports up to 2,500 nodes at 50 Hz with an EER of 1.03 × 10^-3, a more than three-fold improvement over the free-running case. The aggregate network capacity is governed by the total event rate: about 100,000 events per second can be collected while keeping EER below 10^-3.

How it works

Conventional random-access or time-division protocols force every node to occupy a slot or bandwidth allocation even when it has nothing to say. That is wasteful for biological tissue, where most electrodes are silent most of the time. ASBIT flips the contract. A node transmits only when an on-chip threshold-crossing circuit reports a binary event, and the event is multiplied by a quasi-orthogonal Gold code that acts as the node's address. The receiver runs a bank of matched filters to separate the superimposed backscattered signals.

The prototype uses a free-running on-chip oscillator to clock the digital logic. That keeps the chip simple and low power, but the oscillator drifts by about ±1,000 ppm with supply and temperature. Drift smears the Gold-code waveform, so the receiver must use several matched-filter variants tuned to different clock offsets. The trade-off is latency: more filters mean more demodulation time per node. The RF-carrier-divided clock removes most of the drift and lets the receiver work with a narrower filter bank, but it consumes more power because the divider must run continuously.

The authors also show that the same physical layer can carry real neural data. They transmitted over 8,000 channels of previously recorded primate cortical spikes through the simulated ASBIT link and decoded hand kinematics with a spiking neural network. That is a communication-system demonstration, not an in-vivo recording, but it connects the protocol directly to the kind of sparse spike traffic an MEA produces.

Where a skeptic should push

The most load-bearing assumption is event sparsity. ASBIT's capacity scales only because most nodes are quiet most of the time. A dense, high-firing organoid, or an acquisition strategy that reports every threshold crossing on every channel, could push the aggregate event rate past the roughly 100,000 events per second knee where EER stays below 10^-3. The paper does not report what happens when events cluster in time, as they do during network bursts.

The thousand-node scaling is also partly simulated. The authors built 78 chips and then synthesized larger populations by replaying recorded background packets with replacement. The method is reasonable for studying interference statistics, but it does not prove that 1,000 independent radios would all lock onto power, stay within the same dynamic range, and maintain the same 1.7 dB SNR. The benchtop link used a three-coil near-field setup with a relay coil, 9 mm of air and 1 mm of glass; conductive tissue or culture medium would change both power transfer and backscatter amplitude.

Finally, the chips demonstrated here are communication-only. The paper assumes, but does not implement, the front-end amplifiers, ADCs or spike detectors that would sit between an electrode and the Gold-code encoder. Those blocks will dominate power, noise and area, and they will set the true event rate. ASBIT is a telemetry protocol first; calling it a complete MEA front end is premature.

What it means for distributed MEA telemetry

The non-obvious opportunity is architectural. Today's high-density MEAs route hundreds or thousands of channels off a single substrate through wire bonds, flip-chip bumps or capacitive links. That geometry is why planar arrays sit under a 3D organoid and why 3D integration is hard. ASBIT points to a different topology: a cloud of submillimetre wireless nodes distributed around or inside the tissue, each responsible for a small neighborhood of electrodes and reporting only events. Power, clock and data would all come through the RF field rather than through physical traces. For organoids, that could make it easier to record from the interior of a thick culture without drilling electrodes through it.

The second opportunity is data-rate engineering. If the front end converts each electrode's waveform into a sparse spike train at the node, the acquisition chain no longer has to ship raw samples off-chip. A 1,000-channel array sampling at 25 kHz and 16 bits produces 400 Mb/s before compression; the same array reporting asynchronous threshold crossings at a few hertz per channel produces kilobits per second. ASBIT's protocol is designed precisely for that regime, and the Gold-code addressing removes the need for a central arbiter to schedule transmissions.

The genuine threat is that the sparsity assumption may not survive contact with real tissue. Organoid networks can burst synchronously, producing correlated threshold crossings across many channels at once. In that limit the aggregate event rate spikes, packets collide, and the EER climbs. The paper's capacity rule is simple and useful - smaller networks tolerate higher firing rates, larger networks require sparser firing - but it is derived from replayed data, not from a living preparation. There is also the practical question of RF exposure: a 900 MHz carrier delivering enough power to run hundreds of wireless nodes inside a culture dish raises specific absorption rate and heating questions that the paper does not address for in-vitro use.

From an instrumentation standpoint, the most immediate lesson is about timing. ASBIT gives event time stamps, not synchronized sampled waveforms. For spike sorting, phase analysis or any method that requires sub-millisecond alignment across channels, the ±1,000 ppm free-running clock and the event-driven packet timing are limitations, not features. A frequency-divided carrier improves this but at higher power. Anyone building an organoid array around ASBIT would need to decide whether the application can live with event-time telemetry or whether a hybrid architecture - event packets plus a low-rate synchronized common clock - is necessary.

The bottom line

ASBIT is a strong communication proof-of-concept: a 300 µm backscattering chip, a working Gold-code protocol, and credible scaling analysis up to thousands of nodes. It is not yet a biopotential front end. For organoid MEAs, the value is a concrete design target for wireless, event-driven acquisition. The protocol looks most attractive for applications that can tolerate sparse, asynchronous event reports - for example, long-term culture monitoring or closed-loop stimulation triggered by population bursts. It is less attractive for experiments that need dense, phase-locked waveforms from every channel. What would confirm the opportunity is a follow-up that integrates ASBIT with a real neural amplifier and spike detector and records from a living network.