SPAD arrays push imaging bandwidth from 100 Gbps toward 1 Tbps

By Grace Kim

A 1-megapixel single-photon avalanche diode (SPAD) array counting photons at 100 kHz produces 100 Gbps of data; an 8-megapixel (4K) array approaches 1 Tbps. These are illustrative scale calculations, not measured camera throughput: the technical account names no product model, test condition or power budget.

Against that constraint, Tristan Swedish, cofounder and CTO of Ubicept in Boston, says: “At Ubicept, we’ve developed adaptive closed-loop algorithms that can effectively compress photon count information in the most optimal way, given bandwidth and power constraints.” The account gives no compression ratio, latency, image-quality metric or hardware test conditions for that claim.

What does a SPAD pixel detect?

A SPAD pixel is a reverse-biased diode operated above its breakdown voltage. One incoming photoelectron can trigger an avalanche; surrounding electronics charge the pixel, detect and quench the avalanche, then recharge it. The recharge interval, on the order of nanoseconds, prevents the pixel from counting photons that arrive simultaneously.

The timestamped avalanche can feed a time-to-digital converter in a time-correlated single-photon-counting system, principally used for time-of-flight LiDAR. Many solid-state LiDAR designs pair vertical-cavity surface-emitting laser arrays with SPAD arrays; optics project the two-dimensional laser pattern onto a scene, and the SPAD records the reflected pattern for depth measurement, effectively forming a time-of-flight camera.

The design consequence is direct: the recharge time sets the first limit to compare against the expected photon-arrival rate.

Why does immediate digitization matter?

Unlike a CMOS image sensor, which digitizes accumulated photoelectrons through an analog-to-digital converter at readout, a SPAD registers each avalanche immediately. ADC read noise introduces count uncertainty in CMOS; the account describes SPAD output as zero read noise, although it supplies no uncertainty budget or laboratory comparison. Both sensor families use silicon and have similar spectral response and thermal-noise properties; CMOS calls the thermal component dark current, while SPAD arrays call it dark counts.

Each pixel can hold an N-bit counter, reset when an exposure begins and incremented by each detected photon. The avalanche can also increment digital logic beside the pixel. The counter can be read repeatedly during exposure without the stated read-noise penalty, and differential counting can alternate between increments and decrements.

Many 2D arrays now use 3D-stacked wafers, placing detection and recharge circuitry below a maximized photosensitive area. The added area permits larger counters and simple arithmetic beside each pixel. Placing the same imaging optics used with CMOS in front also turns the array into a camera that reads photon counts over time, including ambient visible light for general-purpose perception.

Immediate digitization makes near-pixel compression and other advanced signal processing practical without changing the optical architecture. The buying distinction is therefore where the arithmetic occurs, not which lens the camera uses.

Does soft saturation widen dynamic range?

A SPAD detects only one photon at a time. Its flux response is nonlinear, an effect the source calls “soft saturation,” and the nonlinearity effectively increases dynamic range. CMOS signal-to-noise ratio rises in proportion to flux, while the article says SPAD SNR rises in proportion to photons counted.

Digitized counts let the system adjust the response dynamically and allocate SNR across an extremely large range. SPAD cameras can also read out at very high rates and produce a large volume of scene information. The source presents these capabilities as solutions for low-light and high-dynamic-range scenes, where shot noise and motion blur constrain CMOS imaging.

This remains a conceptual performance claim: no measured SNR curve, illumination level, exposure time, dark-count rate or competing-camera result appears. The practical test is to measure saturation and SNR at the deployment’s light level rather than accept a dynamic-range label.

Can resolution scale without drowning in data?

SPAD array resolution and pixel pitch have advanced rapidly over the past few years. Manufacturers have announced 2-megapixel arrays with 5-µm pixels, and academic papers have shown 3-µm and even 1-µm pitches. Those dimensions make drop-in replacement of CMOS plausible in some applications, but the technical account provides no yield, power, timing or reliability data for any particular array.

The proposed 3D-stacked architecture can drastically reduce bandwidth by processing incoming counts, then counting and transmitting only the most important information. Ubicept presents this as access to operating domains and engineering tradeoffs unavailable with conventional CMOS. Resolution is therefore the visible specification; usable scale still depends on a measured compression and readout system.

Can photon counting solve low-light motion blur?

Most computer-vision systems face a low-light tradeoff: extending exposure collects more light but increases motion blur. Mobile-robot perception loops may run at 10 Hz, restricting exposures to 10 ms or less. At that setting, the source says each frame discards 90% of the available light.

A SPAD camera can integrate for the full available interval while compensating motion dynamically. Real-time optical-flow tracking and streaming photon-count updates can reduce or eliminate blur. The vision algorithm still receives an image frame, but it can use the cleaner input with the same downstream models. Optical-flow information may also travel in the frame data to improve downstream efficiency.

As machine-facing imaging systems increase, algorithms can preserve the information most relevant to a given task. Swedish writes: “This SPAD based perception paradigm is poised to provide a ‘digital retina’ for a wide range of computer vision and perception applications during the next few years.” That is a roadmap forecast, not measured performance.

Ubicept CEO and cofounder Sebastian Bauer discussed the technology in a podcast interview released May 26, 2026.

The adoption test is concrete: can a buyer obtain documented camera performance for low dark counts, bounded recharge time, manageable bandwidth and verified motion compensation under the deployment’s own lighting and timing constraints?