BEIJING -- Chinese researchers have developed an innovative computational architecture that enhances computing power nearly fourfold, thereby opening new possibilities in fields like embodied intelligence, edge sensing, brain-inspired computing and communication systems.
The Fourier Transform serves as a frequency "translator," converting complex signals, including sound and images, into frequency-domain representations, which is a fundamental and widely used method across science and engineering.
Targeting this universal computational process, researchers from Peking University creatively integrated two novel devices suitable for frequency conversion into a multi-physics-domain architecture, resulting in a versatile hardware system capable of performing operations, including the Fourier Transform. Their breakthrough was published on Friday in the journal Nature Electronics.
"This architecture enables different computing paradigms to operate within their optimal physical domains, such as electrical current, charge or light, thereby improving computational efficiency," said Tao Yaoyu, a researcher at the Institute for Artificial Intelligence of this university.
Tao noted that the integrated system leverages the complementary advantages of the two devices in frequency generation, modulation and in-memory computing. While maintaining accuracy and reducing power consumption, it increases the speed of Fourier Transform computations from about 130 billion operations per second to approximately 500 billion, which is a several-fold improvement.
This computational architecture could help enable new hardware to operate efficiently and accelerate its application in areas including foundational AI models, embodied intelligence, autonomous driving, brain-computer interfaces and communication systems.
