China launches a breakthrough optical chip in AI training

The study was published by LightGen Science December 19, introducing a breakthrough in the field of high-performance computing chips, selected as a featured article in this famous scientific magazine. The chip, developed by Shanghai Jiaotong University and Tsinghua University, is an “all-optical microprocessor capable of running large-scale generative AI models, providing great power in addressing huge computing and power needs”.

“For the first time, a fully optical computing chip has been developed to support large-scale image and semantic generative AI models,” Associate Professor Chen Yitong, representative of the research team, told China Daily.

Generative AI is increasingly widely applied to complex real-life situations, such as creating images from text in seconds or creating videos in a split second. However, as these models become increasingly powerful and sophisticated, the need for computing power and energy efficiency becomes urgent. In the post-Moore’s Law era, global research efforts are shifting toward next-generation computing chip areas such as optical computing.

Optical chips are considered superior in accelerating specialized tasks, but are not capable of supporting large-scale generative models with advanced technology. The challenge lies in the architecture within them. Optical computing involves processing information using light instead of electrons inside transistors. Light’s inherently high speed and parallel processing capabilities make it a promising direction to overcome the computing and energy “bottleneck”, but applying optical computing to AI is complex due to its large scale and variability across many dimensions.

According to Yitong, LightGen achieved a “leap” in performance by overcoming three critical bottlenecks: integrating more than two million photonic neurons on a single chip; performs all-optical dimensional conversion; and developing training algorithms for photogenetic models that are not based on real data.

“Each of these three breakthroughs has been considered significant,” Yitong emphasized. “LightGen achieves all three at once, enabling all-optical deployment from start to finish for large-scale generative tasks.”

This expert further explained that LightGen does not simply use electronic systems to support optics in generating data. Instead, it takes it to the next level by forming a complete loop of “input – understanding – semantic manipulation – generation” on an optical chip. After the image is inserted into the chip, the system can extract and represent semantic information, create new multimedia data under semantic control, and finally enable light to effectively “understand” and “perceive” semantics.

In terms of power, LightGen achieves a computing speed of 3.57×10^4 TOPS (each TOPS is equivalent to 1,000 billion calculations per second) and energy efficiency of 6.64×10^2 TOPS per watt, far surpassing leading electronic chips such as Nvidia A100. Through a series of experiments, the team demonstrated that LightGen can perform high-resolution image semantic generation, 3D rendering, high-definition video generation and semantic control as well as support many other large-scale generation tasks.

In performance tests, LightGen adheres to rigorous computational standards, achieving imaging quality comparable to leading neural networks such as Stable Diffusion and NeRF, while measuring execution time and power consumption. Tests also show that, even when using a relatively outdated input device, LightGen improves computational and power efficiency by up to two orders of magnitude compared to leading digital chips. With advanced equipment, LightGen can theoretically achieve improvements in computing power by up to 7 orders of magnitude and energy efficiency by up to 8 orders of magnitude.

“LightGen is opening new avenues to advance generative AI with greater speed and efficiency, opening up new directions in research on high-speed and energy-efficient generative intelligent computing,” Yitong added.