A team of scientists from Beijing has announced a groundbreaking development in artificial intelligence (AI) with the creation of the world’s first fully optical AI chip, named Taichi-II. This innovative chip marks a significant leap in both efficiency and performance, surpassing even NVIDIA’s renowned H100 GPU in energy efficiency.

Led by Professors Fang Lu and Dai Qionghai from Tsinghua University, the research team revealed their findings, showcasing the Taichi-II as a major advancement over its predecessor, the Taichi chip. Earlier this year, the original Taichi chip was reported to have exceeded the energy efficiency of NVIDIA’s H100 GPU by more than a thousand times, according to the South China Morning Post (SCMP). The new Taichi-II chip has further elevated this benchmark, demonstrating superior performance across a range of applications.

The study highlights the Taichi-II chip’s potential to revolutionize AI training and modeling. Unlike traditional methods that rely on electronic computers, the Taichi-II leverages optical processes, significantly enhancing efficiency and performance. The chip has shown remarkable improvements, including a tenfold increase in the speed of training optical networks containing millions of parameters and a 40 percent boost in the accuracy of classification tasks. Additionally, it has improved energy efficiency in low-light imaging scenarios by six orders of magnitude.

One of the key innovations behind the Taichi-II chip is its use of Fully Forward Mode (FFM) learning, a novel approach that allows for intensive training processes to be conducted directly on the optical chip. This technique enables parallel processing of machine learning tasks, supporting high-precision training and making it ideal for large-scale network training.

Xue Zhiwei, the lead author of the study and a doctoral student, emphasized the chip’s architecture as a foundation for the future of optical computing power in AI model construction. The FFM learning method, utilizing high-speed optical modulators and detectors, has the potential to outperform traditional GPUs in accelerated learning scenarios. This breakthrough opens new possibilities for moving optical computing from theoretical concepts to practical, large-scale applications.

By Impact Lab