重大突破！芯片，大消息！

Core Viewpoint - The article highlights a significant breakthrough in the field of optical computing chips, specifically the development of the LightGen chip by researchers at Shanghai Jiao Tong University, which supports large-scale semantic media generation models [2][3]. Breakthrough Achievements - Researchers have successfully developed the LightGen chip, which is capable of achieving two orders of magnitude improvement in computing power and energy efficiency compared to top digital chips, even when using less advanced input devices [4]. - The LightGen chip overcomes three critical bottlenecks: integration of millions of optical neurons on a single chip, all-optical dimensional transformation, and a novel optical training algorithm that does not rely on ground truth [4]. - LightGen can perform a complete closed-loop process of input, understanding, semantic manipulation, and generation, enabling high-resolution image generation (≥512×512), 3D generation (NeRF), and high-definition video generation [4]. Advantages of Optical Computing - Optical computing is recognized for its scalability, low power consumption, ultra-high speed, wide bandwidth, and high parallelism, making it a key technology for rapid computation of large-scale data in AI, scientific computing, and multimodal perception [5]. - The article notes that traditional chip architectures are struggling to keep pace with the increasing demands for computing power and energy efficiency due to the rapid evolution of deep neural networks and large-scale generative models [3][5]. Recent Developments in Optical Computing - A recent advancement by the Shanghai Institute of Optics and Fine Mechanics has led to the development of a high-parallel optical computing integrated chip, achieving a parallelism greater than 100, which addresses the challenge of high-density information processing on optical chips [6][7]. - The new optical chip system includes a self-developed integrated micro-cavity optical frequency comb and a high-performance parallel computing core, enabling a theoretical peak computing power of over 2560 TOPS with a power efficiency of over 3.2 TOPS/W [7].