New multi-physics AI architecture boosts computing speed, efficiency

Chinese researchers at Peking University have developed a novel multi-physics computing architecture that significantly enhances processing speed and efficiency, achieving nearly a fourfold increase in performance. By integrating two innovative devices optimized for frequency generation, modulation, and in-memory computing, the system effectively matches frequency conversion across multiple physical domains—such as electrical current, charge, and light. This versatile architecture excels at complex operations like the Fourier Transform, a fundamental technique for converting signals into frequency-domain representations widely used in science and engineering. The new system boosts Fourier Transform processing speeds from approximately 130 billion to 500 billion operations per second while maintaining accuracy and reducing power consumption. This advancement addresses the limitations of traditional digital computing architectures, which struggle to meet the increasing demands of AI workloads. The Peking University team’s approach aligns with a broader global trend toward specialized computing paradigms—including neuromorphic, photonic, and analog architectures—that optimize specific mathematical functions to improve speed and energy efficiency. By enabling computations to run in their most efficient physical