Shanghai Jiao Tong University announced a breakthrough in computing power chips on Friday by introducing LightGen, an all-optical computing chip capable of running large-scale generative artificial intelligence models, which provides insights into addressing the immense computational and energy demands in the generative AI era.

The research team claimed that it is the first time an all-optical computing chip has been developed to support large-scale semantic and visual generative models. A paper about the study was published on the website of the journal Science on Friday, where it was highlighted as a featured paper.

Generative AI is increasingly being applied to complex real-world scenarios, such as generating images from text in seconds and creating videos in a matter of moments. As these models grow larger and more sophisticated, the demand for computational power and energy efficiency becomes more pressing. In the post-Moore's law era, global research efforts are shifting towards next-generation computing power chips like optical computing.

Currently, optical chips excel at accelerating discriminative tasks, but fall short of supporting cutting-edge large-scale generative models. The challenge lies in enabling next-generation optical computing chips to run complex generative models, a recognized problem in the field of intelligent computing globally.

Scientists explained that optical computing involves processing information using light instead of electrons within transistors. Light naturally offers high speed and parallelism, making it a promising direction for overcoming computational and energy bottlenecks. However, applying optical computing to generative AI is complex due to the large scale of these models and their need to transform across multiple dimensions.

Chen Yitong, a leading researcher on the team, said that LightGen achieves its performance leap by overcoming three critical bottlenecks: integrating millions of optical neurons on a single chip, achieving all-optical dimensional transformation, and developing a training algorithm for optical generative models that does not rely on ground truth.

"Any one of these breakthroughs alone would be deemed significant. LightGen achieves all three simultaneously, enabling an end-to-end, alloptical implementation for large-scale generative tasks," said Chen, who is also an assistant professor at Shanghai Jiao Tong University's School of Integrated Circuits.

LightGen is not merely about using electronics to assist optics in generation. It achieves a complete "input-understanding-semantic manipulation-generation" loop on an all-optical chip, according to the research team. After an image is fed into the chip, the system can extract and represent semantic information, generating new media data under semantic control, effectively enabling light to "understand" and "cognize" semantics.

Experiments in their research demonstrated that LightGen can perform high-resolution image semantic generation, 3D generation, high-definition video generation, and semantic control, supporting various large-scale generative tasks like denoising and feature transfer.

In performance evaluations, LightGen adhered to rigorous computational standards. It achieved comparable generation quality to leading electronic neural networks, such as Stable Diffusion and NeRF, while measuring end-to-end time and energy consumption. Tests showed that even using relatively outdated input devices, LightGen achieved computational and energy efficiency improvements of two orders of magnitude compared to top digital chips. With advanced devices, LightGen could theoretically achieve computational power improvements of seven orders of magnitude and energy efficiency improvements of eight orders.

The research team said that the study emphasizes that as generative AI becomes more integrated into production and daily life, developing next-generation computing power chips capable of executing cutting-edge tasks required by modern AI society becomes imperative.

"LightGen opens a new path for advancing generative AI with higher speed and efficiency, providing a fresh direction for research into high-speed, energy-efficient generative intelligent computing," said Chen.