That seemed like a major limitation,
T. Patrick Xiao told Tech Xplore regarding the historical inability of analog hardware to run the fast Fourier transform algorithm.
Researchers demonstrated a new method for performing fast Fourier transform calculations using an analog system built from a large array of memory components, according to Tech Xplore.
Fourier transforms identify frequencies that make up a signal to reveal patterns in sounds and images. Digital processors typically execute these calculations using the fast Fourier transform, which breaks down large transforms into smaller ones to achieve speedups.
Analog in-memory computing systems offer high energy efficiency, but adapting the fast Fourier transform algorithm to them had remained unproven until now.
Dividing Large Transforms
The research team addressed the mismatch between analog memory arrays and the fast Fourier transform algorithm by dividing each Fourier transform into smaller transforms, recursively as needed. They stopped the recursive breakdown when the subproblem could fit into the memory array, according to Tech Xplore.
The team built their system using SONOS flash memory cells. Each cell traps electrical charge to retain values without continuous power.
The flash memory cells stored discrete Fourier transform coefficients. The system utilized more than half a million flash devices simultaneously within the same array to compute matrix multiplications.
Scaling Analog Computing
Previous analog chips could process Fourier transforms of up to 128 points. The new system successfully processed 65,536 points, according to Tech Xplore.
An arxiv.org submission describing the research notes that the system performed analog computations on up to 524,000 charge-trapping memory devices simultaneously. The setup produced a discrete Fourier transform scale that is more than 500 times larger than any previous analog demonstration.
Edge Computing Applications
The new approach reduces the energy required to process signals within tight power budgets. This efficiency suits edge computing systems with limited energy budgets.
Future workloads will combine fast Fourier transform signal processing with artificial intelligence algorithms. The team plans to test the hardware on complex real-world workloads that require both signal processing and AI tasks on a single chip.
How SONOS Flash Memory Stores Data and Processes Transforms
How does SONOS flash memory store data during calculations?
Each SONOS memory cell traps electrical charge to retain values without power. These stored values are used in the discrete Fourier transform operations.
What is the size difference between previous analog chips and the new system?
Previous analog chips processed Fourier transforms of up to 128 points. The research team’s system processed 65,536 points, utilizing over half a million flash devices simultaneously.
Which organizations collaborated on this research?
The research findings were published in Nature Electronics.