【Chinese Scientists Achieve New Breakthrough in Next-Generation Optical Computing Chip Research】Cailian News, December 19th. From a report by Shanghai Jiao Tong University, the university's research team has made breakthroughs in the field of next-generation optical computing chips, achieving for the first time a fully optical computing chip that supports large-scale semantic media generation models. The related achievements were published in the journal "Science" on December 19th. According to the information, with the rapid evolution of deep neural networks and large-scale generation models, there is an extremely high demand for computing power and energy consumption. The performance growth of traditional chip architectures has already shown serious gaps, and new architectures such as optical computing have attracted widespread attention. "Optical computing can be understood as not letting electrons run in transistors, but letting light propagate in the chip, using changes in the light field to perform calculations. Light naturally has the advantages of high speed and parallelism, so it is considered an important direction to break through the bottlenecks of computing power and energy consumption," said Chen Yitong, assistant professor at the School of Integrated Circuits of Shanghai Jiao Tong University and author of the paper. He added that applying optical computing to generative AI is not simple. Existing fully optical computing chips are mainly limited to small-scale, classification tasks, and photonic cascading or multiplexing would severely weaken the speed of optical computing. How to make the next-generation optical computing chips run complex generation models is a recognized challenge in the global intelligent computing field.
Chen Yitong's research group proposed and implemented a fully optical large-scale semantic generation chip called LightGen. Experimental results based on extremely strict computing power evaluation standards show that even with performance-lagging input devices, LightGen can still achieve two orders of magnitude improvement in computing power and energy efficiency compared to top digital chips. The team stated that the performance leap of LightGen lies in its simultaneous breakthrough of three key bottlenecks on a single chip: "integration of millions of optical neurons on a single chip," "fully optical dimension transformation," and "optical generation model training algorithm independent of ground truth." This makes it possible to realize a fully optical end-to-end solution for large-scale generation tasks.
According to the introduction, LightGen can complete the closed loop of "input-understanding-semantic manipulation-generation," achieving high-resolution (≥512×512) image semantic generation, 3D generation (NeRF), high-definition video generation, and semantic control, while supporting multiple large-scale generation tasks such as noise reduction, local and global feature transfer. "LightGen opens up a new path for next-generation optical computing chips to support cutting-edge artificial intelligence, and also provides a new research direction for exploring faster and more energy-efficient generative intelligent computing," said Chen Yitong. (Xinhua News Agency)
Original: toutiao.com/article/7585577159444431386/
Statement: This article represents the views of the author.