Researchers have demonstrated analogue fast Fourier transforms on SONOS memory arrays, processing large audio and image signals with far less data movement. The advance could help battery-powered sensors, radar systems, wearables, and connected devices analyse information locally, rather than sending every raw signal back to the cloud.
The fast Fourier transform gets a hardware rethink
The fast Fourier transform, or FFT, is one of those invisible technologies doing essential work everywhere. It turns a time-based signal, such as a microphone recording or radar echo, into a map of its frequencies. That makes it easier for a device to recognise speech, analyse wireless signals, or build an image from reflected radar pulses.
For decades, the FFT has mostly lived inside digital processors. The algorithm is dramatically more efficient than calculating every frequency directly, but large signals still demand plenty of multiplications, additions, memory transfers, and coordination. In a small gadget running on a battery, all that movement can be as troublesome as the arithmetic itself, according to the Nature Electronics study.
The Sandia National Laboratories-led team took a different route. Instead of treating memory as a storage cupboard beside the processor, the researchers used memory cells as part of the computing machinery. It sounds slightly counterintuitive, but that's the appeal: the data doesn't have to shuttle back and forth quite so much.
Why analogue memory matters at the edge
Analogue in-memory computing works by storing numerical values as physical conductance levels. Apply voltages to an array, and the resulting currents naturally combine according to the circuit's electrical behaviour. In effect, the hardware performs many pieces of matrix arithmetic simultaneously, rather than handling every operation one after another.
That parallelism is particularly attractive for edge devices. Sensors in vehicles, wearables, industrial equipment, and remote monitoring systems often need to respond quickly while operating within tight power limits. A low-energy signal-processing engine could help them interpret sound, movement, radio signals, or images locally.
The broader idea isn't entirely new. Analogue hardware accelerators have already attracted interest for neural-network inference and other workloads built around matrix multiplication. The awkward part has been adapting that approach to the FFT, whose familiar divide-and-conquer structure doesn't look like one conveniently sized matrix.
A clever way to scale beyond one giant array
The researchers' solution was to express the FFT as a series of smaller matrix operations. These stages, often called butterfly operations, can run through modest memory arrays, with intermediate results passed along and the hardware reused. That avoids the enormous device count a straightforward, full-size analogue transform would require.
This matters because a brute-force analogue implementation would grow roughly with the square of the input size. Such a design might be manageable for a tiny signal, but it would become unwieldy for the large transforms used in modern communications and imaging. Breaking the job into reusable stages keeps the memory requirement growing much more gently.
The demonstration used SONOS, a charge-trapping memory technology. Its cells can hold programmable conductance states and operate at low power, which makes them a natural fit for analogue calculations. The setup handled a 65,536-point audio spectrum and activated more than half a million memory devices in one array, a scale that gives the result real engineering weight rather than merely laboratory novelty.
From intelligible speech to radar-ready images
Numbers on a chart are useful, but the best test is whether the output remains useful. In the audio experiments, speech reconstructed from the complex frequency data stayed intelligible. That suggests the analogue errors were contained well enough for a listener to recognise the result, rather than hearing a wash of electronic mush.
The team also explored two-dimensional transforms for image signals. These calculations are central to synthetic aperture radar, where systems combine measurements to create detailed pictures of terrain, ships, infrastructure, or weather. The work compared the analogue approach with digital implementations, including hardware built around a Xilinx Versal adaptive computing platform.
Precision remains the obvious catch. Analogue devices can drift, vary from one another, and respond to temperature or ageing. The researchers therefore used hardware-aware modelling and accuracy analysis to examine whether those imperfections would undermine real applications. The important question isn't always whether every bit matches a digital result; it's whether a speech recogniser, detector, or imaging system still makes the right call.
What this could mean for battery-powered devices
FFT calculations appear in far more products than most people realise. Wireless systems use them to handle many carrier channels, radar uses them to interpret echoes, and audio systems use them to create frequency representations for recognition and enhancement. In each case, local processing can reduce latency and limit the amount of raw data that needs to travel elsewhere.
That could be valuable in settings where connectivity is unreliable or expensive. A wearable might analyse a sound or motion pattern without constantly uploading data. An autonomous machine could react at once, while a remote sensor might run longer between charges. The savings won't come from a magical elimination of computing, but from performing many operations in parallel and reducing data movement.
Still, the technology isn't ready to send digital signal processors into retirement. Analogue systems give up some precision and flexibility in exchange for efficiency, and converting signals between analogue and digital formats can itself consume energy. Industry experience with hardware accelerators suggests that the most practical products will combine specialised analogue blocks with programmable digital logic.
The next chapter is likely hybrid, not analogue-only
The result is best viewed as an expansion of the signal-processing toolbox. Digital FFTs remain reliable, adaptable, and easy to program, while analogue arrays could handle repetitive, high-volume transforms close to the sensor. That division of labour may prove especially useful when a device needs speed and frugality more than perfect numerical exactness.
Manufacturing consistency, calibration, long-term reliability, and system-level integration will decide whether the concept moves from research demonstrations into commercial chips. Researchers also need to show how these arrays behave across broader workloads, harsher environments, and real products with strict certification requirements.
Even so, there's something satisfying about the arc of the story. A mathematical shortcut first made practical by digital computing is now being recast as a problem for physics and circuits. If the approach continues to mature, the quietest part of a sensor's electronics could become one of its smartest.
It's a promising step toward faster, leaner signal processing where the data is created.
Disclaimer: This article may have been created with AI assistance and reviewed by our editorial team. It is provided for general informational purposes only. Readers should verify information independently before relying on this content.
