A new achievement by a research team at Tsinghua University was unveiled in the latest issue of Science on the early morning of April 12th. They introduced a pioneering distributed breadth-intelligent optical computing architecture and developed the world's first large-scale interferometric diffraction heterogeneous integrated chip named "Taichi". This chip achieves a universal intelligent computing capability of 160 TOPS/W.
According to reports, the development of the "Taiji" photonic chip architecture was inspired by the ancient text "Yijing" (Book of Changes). Drawing from the phrase "易有太极，是生两仪" (Yi has Taiji, which generates the two polarities), the team established a novel computing model, harnessing the immense power of light to achieve formidable computational performance.
Photonics computing, as the name suggests, involves shifting the computing medium from electricity to light, utilizing light propagation within the chip for computation. With its exceptionally high parallelism and speed, photonics computing is considered one of the most promising contenders for future disruptive computing architectures.
Photonics chips offer advantages in high-speed, high-parallel computation and are hoped to support advanced artificial intelligence applications such as large-scale models.
According to Xu Zhiwu, the lead author of the paper and a doctoral student in the Department of Electronics, in the "Tai Chi" architecture, the top-down encoding-splitting-decoding-reconstruction mechanism simplifies complex intelligent tasks by breaking them down into multi-channel, highly parallel subtasks. The distributed 'large receptive field' shallow light network constructed deals with subtasks separately, overcoming the inherent calculation errors of physical simulator devices with multiple layers of deep cascading.
The paper reports that the "Tai Chi" optical chip achieves an area efficiency of 879T MACS/mm and an energy efficiency of 160 TOPS/N. For the first time, it enables optical computing to accomplish complex AI tasks such as recognizing thousands of objects in natural scenes and generating cross-modal content.
The "Tai Chi" optical chip is expected to provide computational support for large-model training and inference, general artificial intelligence, and autonomous intelligent unmanned systems.
