A large-scale photonic chiplet to power artificial general intelligence
Scientists report that the chip, called Taichi, for the first time experimentally realises on-chip large optical neural networks for thousand-category-level classification and artificial intelligence-generated content (AIGC) tasks, with improved area efficiency and energy efficiency compared to current AI chips
Researchers from Tsinghua University have reported the development of a new photonic AI chiplet, called “Taichi,” in the journal Science. The team, led by Lu Fang, an associate professor in the Department of Electronic Engineering and Qionghai Dai, a professor in the Department of Automation, says that Taichi can empower artificial general intelligence (AGI) at 160 tera-operations per second per watt (TOPS/W).
Rapid development in the field of AGI imposes stringent energy efficiency and area efficiency requirements on next-generation computing. Integrated photonic neural networks have shown promise as a technology that could break the plateauing of Moore’s Law and achieve superior processing speed and high energy efficiency. However, the technology faces challenges including limited computing capability and scalability, meaning that only simple tasks and shallow models (such as 4-class vowels classification, 10-category digits classification, with up to thousand-parameter-level networks) have been experimentally realised in photonic integrated circuits.
However, the research team at Tsinghua University has innovated Taichi, a large-scale photonic AI chip along with a distributed optical computing architecture, which they say has billions-of-neuron on-chip computing capability with an energy efficiency of 160 TOPS/W. According to the scientists, Taichi exploits the high parallelism and high connectivity of wave optics to implement very high-density computing, as well as exploring a general and iterative encoding-embedding-decoding photonic computing to effectively increase the scale of the optical neural network to the billion-neuron level.
The researchers say that, for the first time, Taichi experimentally realises on-chip large optical neural networks for thousand-category-level classification and artificial intelligence-generated content (AIGC) tasks, with up to 2-3 orders of magnitude improvement in area efficiency and energy efficiency compared to current AI chips.
Describing their work, the research team say they proposed a universal and robust distributed computing protocol for complex AGI tasks. Instead of going deeper as electronic computing, the Taichi architecture sought to go broad for throughput and scale expansion. A binary encoding protocol is proposed to divide challenging computing tasks and large network models into sub-problems and sub-models that can be distributed and deployed on the photonic chiplet. The team says this enables large-scale tasks to be adaptively solved with flexible scales, achieving on-chip networks with up to 10 billion optical neurons.
The scientists say they developed the largest-scale photonic chiplets to support input and output dimensions as large as 64 x 64. By integrating scalable wavefield diffraction and reconfigurable interference, the team says that entire inputs are encoded passively and modulated in a highly parallel way, achieving 160 TOPS/W on-chip energy efficiency and 879 T MACS/mm² area efficiency (up to 2 orders of magnitude improvement in both energy and area efficiency than existing AI chips).
According to the researchers, on-chip experiments have demonstrated Taichi’s versatility and flexibility with state-of-the-art performances, including: an accuracy of 91.89 percent in 1623-category Omniglot characters classification and 87.74 percent in 100-category mini-ImageNet classification; and on-chip high-fidelity AIGC models such as music composing and generation of high-resolution styled paintings.
The scientists add that Taichi not only breaks the scale limitation towards beyond-billion-neurons foundation model with large-scale high-throughput photonic chiplets, but also achieves error-prone robustness through information scattering and synthesising. They also say that Taichi has experimentally solved the complex on-chip AGI tasks (thousand-category-level classifications and various-modality AIGC), leading to a giant leap in scalability, accuracy, and efficiency. The development could pave a viable route to real-world photonic computing and represents an important step for next-generation high-efficiency AI chips, supporting applications in large machine learning models, AIGC, and robotics, among others.
