Researchers at Sandia National Laboratories, Infineon Technologies and Arizona State University recently demonstrated a new approach for efficiently performing crucial FFT calculations with an analog system made up of a large array of memory components.
Their proposed method, introduced in a paper published in Nature Electronics, entails dividing a large Fourier transform into smaller calculations and then leveraging the electrical properties of memory cells to perform key operations in parallel on an analog chip, while completing other steps digitally.
Whatever the specific architecture used here it is, it is worth taking a step back and reminding yourself that the Fourier Transform is not inherently Discrete and Bandwith Limited, rather it is just that in practice computers always end up doing Discrete Fourier Transforms.
What is going on here then is likely that the Fourier Transform is, through the sleight-of-hand magic trickery of computers being kept non-discrete (or briefly made non-discrete during some optimal step of the processing chain) longer into the process of the calculation than traditional computers and their hardware do (which is start from the beginning in discrete terms and never deviate from them).
*I mean discrete in the sense of whole numbers/measurements made in units vs. the analog realm which is composed of measurements along continuums to some degree of precision.
Perfect timed with the RAM apocalypse if it can make things more efficient for non generalized devices. This is a good Veritasium video about analog computing which helps explain some of the benefits like this paper describes.
The most fun way to understand the benefits of this in my opinion is to check out how crazy powerful and surprising software digital synthesizers and effects can sound and then think about how offloading any kind of digital synthesis/processing to analog means is free computing cycles from the digital perspective.
If you tried to capture the surface of a swimming pool along with measurements of momentum and then use a digital algorithm to predict how the surface will fluctuate (assuming no further disturbances beyond boundary conditions) it would be very quickly a computationally overwhelming task… but a real swimming pool just keeps on doing it endlessly.