The research team, headed by Chen Yitong from the School of Integrated Circuits at Shanghai Jiao Tong University, has achieved a remarkable milestone in the realm of next-generation computing chips. For the first time ever, they've successfully crafted a fully optical computing chip capable of supporting large-scale semantic media generation models. The team's related findings have been published in the prestigious journal Science.

With the swift advancement of deep neural networks and large-scale generative models, traditional chip architectures are now facing significant bottlenecks, particularly in terms of computing power and energy consumption. The question of "How can we equip next-generation computing optical chips to run complex generative models?" has emerged as a global conundrum.

The research team has introduced LightGen, a fully optical large-scale semantic generation chip, which shatters several widely acknowledged technical barriers. Experimental outcomes reveal that the chip excels in multiple large-scale generative tasks, demonstrating optical "understanding" and "cognition" of semantics.

LightGen adheres to stringent computing power evaluation criteria. Actual measurements indicate that its computing power and energy efficiency far surpass those of top-tier digital chips. Theoretically, the performance enhancement is even more staggering.

Paper link: https://www.science.org/doi/10.1126/science.adv7434