Chinese team develops world’s first AI training system that runs entirely on light

A groundbreaking advancement in artificial intelligence has emerged from a team at Tsinghua University in Beijing, which has developed the world’s first fully optical AI training system, the Taichi-II chip, South China Morning Post reported yesterday.

This innovative chip represents a significant leap in efficiency and performance over its predecessor, the Taichi chip, which already surpassed Nvidia’s H100 GPU in energy efficiency by over a thousand times.

The Taichi-II chip operates entirely on light, eliminating the need for electronic computers during the training process, thereby enhancing its efficiency. The researchers assert that this transition to optical computing marks a crucial step towards practical applications, addressing the increasing demand for computational power while minimizing energy consumption. This development is particularly timely, given recent restrictions on China’s access to advanced GPU technology.

In practical terms, the Taichi-II chip accelerates the training of optical networks with millions of parameters by an order of magnitude and boosts classification accuracy by 40%. Its energy efficiency in low-light imaging scenarios improved dramatically, showcasing its potential for complex applications.

The team also introduced a novel training method called fully forward mode (FFM) learning, which allows for parallel processing directly on the optical chip. This architecture not only enhances precision but also supports large-scale network training, potentially outpacing traditional GPUs in speed and efficiency.