HomeInsightsAnalog AI Computing May Have Solved a Problem That Held Back In-Memory Processing
TechnologyNov 8, 2026, 9:35 PM

Analog AI Computing May Have Solved a Problem That Held Back In-Memory Processing

Kosmos Consult

Published by

Kosmos Consult2 min read

Researchers have demonstrated large Fourier transforms on analog in-memory computing hardware, addressing a longstanding obstacle to more energy-efficient signal processing.

Editorial image showing an analog computing chip and memory-array circuitry, with no invented labels.

Editorial image showing an analog computing chip and memory-array circuitry, with no invented labels.

On this page

AI hardware has a persistent problem: moving data around can consume as much energy as processing it. Traditional systems repeatedly shuttle information between memory and processors. Analog in-memory computing tries to reduce that movement by performing calculations where the data is stored.

But one important operation remained difficult to scale: the Fast Fourier Transform, or FFT. It is used to break complex signals into their frequency components, helping computers process audio, images, communications and other forms of data.

The Problem with Large Fourier Transforms

Analog in-memory systems are attractive because they can perform many calculations in parallel. Yet large FFT operations have traditionally been awkward to map onto these systems.

The research team addressed this by breaking a large transform into smaller calculations. Each stage is divided until the resulting operations fit within the physical memory array. The system then uses the electrical properties of memory cells to carry out important calculations in parallel.

The team demonstrated computations involving up to 65,536 points, using more than 524,000 charge-trapping memory devices in a single array. The work was published in Nature Electronics on October 19, 2026, and reported by Tech Xplore on November 8.

Why This Matters for AI Hardware

Fourier transforms appear in audio processing, image analysis, communications and scientific computing. Making them more efficient could help edge devices process signals without depending as heavily on power-hungry digital hardware.

That matters in places where power, heat and physical size are serious constraints: compact sensors, embedded systems, portable devices and specialised AI accelerators.

The research does not mean analog chips are about to replace GPUs. The demonstration is a specific advance in mapping FFT operations onto analog memory, and practical systems will still need to account for accuracy, noise, programmability and the cost of combining analog and digital processing.

The Next Question Is Commercial Usefulness

The exciting part is not simply that a large calculation ran on analog hardware. It is that a limitation of analog in-memory computing may be less restrictive than previously thought.

The next test is whether the approach can deliver meaningful energy and performance gains in real workloads, outside controlled demonstrations.

Research: https://www.nature.com/articles/s41928-026-01725-x Tech Xplore report: https://techxplore.com/news/2026-11-team-large-scale-fourier-analog.html

Similar insights