光计算芯片，新突破

Core Viewpoint - Shanghai Jiao Tong University has achieved a breakthrough in the field of next-generation optical computing chips, successfully realizing an all-optical computing chip that supports large-scale semantic media generation models, published in the journal "Science" on December 19 [1] Group 1: Optical Computing Breakthrough - The rapid evolution of deep neural networks and large-scale generative models has led to extremely high computing power and energy consumption demands, creating a significant performance gap in traditional chip architectures, which has drawn attention to new architectures like optical computing [1] - Optical computing utilizes light propagation within chips instead of electrons in transistors, leveraging the inherent speed and parallelism of light to address bottlenecks in computing power and energy consumption [1] Group 2: LightGen Chip Performance - The research team introduced the LightGen chip, which demonstrated a performance improvement of two orders of magnitude in computing power and energy efficiency compared to top digital chips, even when using relatively outdated input devices [2] - LightGen overcomes three key bottlenecks: integration of millions of optical neurons on a single chip, all-optical dimensional transformation, and a light-based generative model training algorithm that does not rely on true values, enabling end-to-end implementation for large-scale generative tasks [2] - LightGen can complete a closed loop of "input-understanding-semantic manipulation-generation," achieving high-resolution (≥512×512) image generation, 3D generation (NeRF), high-definition video generation, and semantic control, while also supporting denoising and feature transfer tasks [2]