《Science》发文！上海交大集成电路学院首次实现全光大规模语义生成芯片

Core Viewpoint - Shanghai Jiao Tong University has developed the world's first all-optical computing chip, LightGen, which supports large-scale semantic media generation, overcoming the bottlenecks of computing power and energy efficiency. This achievement is published in the prestigious journal Science [1][10]. Research Background - The rapid evolution of deep neural networks and large-scale generative models has led to unprecedented advancements in AI, but the explosive growth in model size has created a significant gap in computing power and energy demand compared to traditional chip architectures [2]. Breakthrough in Research - The research team has introduced the LightGen chip, marking the first international achievement in large-scale all-optical generative AI chips. This chip integrates over one million optical neurons, achieves all-optical dimensional conversion, and addresses the recognized bottlenecks in training algorithms without ground truth [5]. Functionality of LightGen - LightGen has been experimentally validated for high-resolution (≥512×512) image semantic generation, 3D generation (NeRF), and video generation, among other large-scale generative tasks. It enables an end-to-end process where the chip can fully understand and manipulate semantics to generate new media data [7]. Performance Metrics - LightGen employs stringent performance evaluation standards, achieving comparable generative quality to leading electronic neural networks like Stable Diffusion and NeRF while significantly reducing system-wide time and energy consumption. Even with less advanced input devices, LightGen demonstrates a two to three orders of magnitude improvement in computing power and energy efficiency compared to top digital chips. Theoretically, with advanced input devices, it could achieve a seven orders of magnitude increase in computing power and an eight orders of magnitude increase in energy efficiency [9]. Implications for Future AI Development - The research emphasizes the necessity of developing chips capable of executing real-world tasks, particularly for large-scale generative models that are sensitive to latency and energy consumption. LightGen paves the way for next-generation computing chips, opening new research directions for faster and more energy-efficient generative intelligent computing [10].