Chinese scientists have announced an optical computer chip with a speed and energy efficiency hundreds of times higher than Nvidia's leading AI hardware, especially in generative AI tasks such as photo and video creation. The chip is called LightGen, developed by a research team from Shanghai University of Transport and Tsinghua University.
LightGen exploits the speed of light to process complex artificial intelligence work. On a compact chip, the research team integrated more than 2 million optical neurons, allowing the system to create high-resolution images, including 3D scenes, as well as video synthesis.
The study was led by Chen Yitong, a professor at Shanghai University of Transport, and published in the journal Science.
Mr. Chen Yitong believes that LightGen can continue to expand and provide a new way to connect advanced chip architectures with complex daily AI applications, without reducing performance, while achieving outstanding speed and performance at a large digital level, towards sustainable AI.
In the context of rapid development of generative AI, the ability to create increasingly realistic images and videos requires great computing power and consumes a lot of energy. This makes scientists turn to optical computing as traditional electronic chips gradually reach the limit. Unlike conventional computers that use electrons, optical computers use laser pulses to process information, helping to respond quickly and reduce electricity consumption.
Previously, optical systems encountered difficulties with complex generative AI tasks due to limited architecture and training algorithms. The LightGen team focuses on solving three key problems: new architecture, new training algorithms and high integration density. They build an optical potential space that helps data circulate effectively, compress and quickly create information.
The research team also developed a supervisory-free training algorithm that does not require huge labeling Datasets, allowing chips to learn statistical rules similar to the human learning process.
The LightGen achieved a processing speed of 3.57×104 TOPS and an energy efficiency of 6.64×102 TOPS per watt, far surpassing leading electronic chips such as the Nvidia A100. Scientists say LightGen could create a major shift in generative AI hardware, while reducing the industry's growing energy consumption pressure.
