Recently, Associate Professor Fang Lu from the Department of Electronic Engineering at Tsinghua University led a research team to collaborate with academician Dai Qionghai from the Chinese Academy of Engineering and his team from the Department of Automation at Tsinghua University, achieving a new breakthrough in the field of intelligent optical computing chips. The relevant achievements have been published in the journal Science. They pioneered the interference-diffraction distributed breadth optical computing architecture and developed the world's first large-scale general-purpose intelligent optical computing chip - "Taichi," with a system-level energy efficiency of 1.6 quadrillion operations per joule per second, surpassing mainstream commercial AI chips by three orders of magnitude, opening up a new path for high-performance intelligent computing in the post-Moore era. The Taichi chip has enabled intelligent optical computing to achieve image classification of over 1000 natural scene categories and intelligent tasks such as cross-modal content generation, providing powerful computational support for AI large models, intelligent unmanned systems, and general artificial intelligence (AGI). In the early stages of this research, the team encountered significant difficulties. What impressed the researchers was that they ultimately overcame the research bottleneck by revisiting classic achievements from the 1980s, 1990s, and even earlier. They stated that it was this pursuit of original intentions that allowed the research team to transcend the limitations of the times, unaffected by trends, and always maintain a focus and passion for the essence of scientific problems, leading to new breakthroughs.
Screenshot of a Science Paper
One Plus One Greater Than Two
Light, with its fast propagation speed, multiple representation dimensions, and low computational power consumption, has become the international cutting-edge of intelligent optical computing, with broad application prospects. In addressing the challenge of large-scale general intelligent optical computing, the research team has abandoned the traditional paradigm of deep electronic computing in optical computing and proposed a distributed breadth computing architecture. They have constructed a deep but wide optical neural network, with a reconfigurable and reusable overall architecture. Unlike the traditional method of stacking layers of deep computing, Taiji simplifies complex intelligent tasks by breaking them down into multi-channel, highly parallel subtasks. It organizes clusters and allocates computing resources separately for each subtask, thereby achieving efficient processing of complex tasks. Inspired by the traditional Chinese philosophical concept of "Taiji is born of the two principles," the research team represents "the two principles" through the interference and diffraction of light, establishing an interference-diffraction joint propagation model on the interference-diffraction chip. "By integrating the flexible and reconfigurable characteristics of interference with the large-scale and highly parallel characteristics of diffraction, we achieve interference-diffraction intelligent optical computing with a dialectical unity concept. This unity enables the Taiji optical chip to not only have reconfigurable general computing capabilities but also high-throughput parallel computing capabilities, achieving the effect of one plus one greater than two," explained Xu Zhihao, co-first author of the paper and a doctoral student in the Department of Electronic Engineering at Tsinghua University.
Taiji Chip. Image provided by the interviewee.
The Most Primitive Approach
In the early stages of research, the team adopted the deep learning architecture of electronic computing to build a large-scale intelligent optical computing system. However, after six months, they encountered a bottleneck: as the number of layers increased, there was an irreconcilable contradiction between computing scale and computing accuracy. "In the past, we mostly built network structures based on electronic computing architecture, but we found that the advantages and potential of light could not be fully utilized in the electronic architecture, like a trapped beast in a cage. Through theoretical modeling and analysis, we found that the electronic architecture 'imprisoned' the capabilities of light, meaning that the existing architecture of deep neural networks is not suitable for intelligent optical computing," said Fang Lu in an interview with Science China Press. To break free from the bottleneck, the research team decided to step out of the comfort zone of traditional electronic architecture thinking and seek a new architectural breakthrough. They turned their attention to classic research results from the 1980s, 1990s, and even earlier on machine learning and neural networks. By reexamining and drawing inspiration from these possibly forgotten ideas and classic wisdom, they found the key to overcoming the current dilemma: returning to the most primitive approach—making it wide and shallow. In Fang Lu's view, revisiting classics is also a return to the original intention of scientific research, letting go of blind adherence to trends. This pursuit of the original intention enables the research team to remain focused on and passionate about the essence of scientific problems, free from the limitations of the times and trends, thus discovering new ideas and proposing new theories. The birth of the Taiji Optical Chip is the result of interdisciplinary collaboration, and research in neuroscience has also provided important insights for the architectural development of the Taiji Optical Chip. Studies in neuroscience have proposed the "shallow brain theory," suggesting that the brain forms large-scale parallel computing units with a shallow and flat architecture. From perception to movement, and even consciousness, each brain region plays a crucial role in this shallow network. "The series of achievements in neuroscience have provided us with a lot of inspiration for our research," Fang Lu added. The "Hardship" of the Chip
However, overturning the architecture was just the beginning. What followed was another daunting challenge: chip development.
The Taiji Optical Chip is the result of three years of relentless effort by the research team, after countless failures and challenges.
In the long journey of chip development, tape-out is a crucial milestone, marking the point where the research team's theoretical concepts are about to be transformed into chips that can be manufactured. For Fang Lu and her research team, this stage is filled with anticipation and anxiety. "The tape-out cycle usually takes 3 to 6 months, and waiting for the chip to be processed is a very anxious time. The team hopes the chip will come back quickly to move on to the next testing phase, but at the same time, they worry about the effectiveness of the tape-out. This would mean starting over from scratch, which comes with a high time cost," Fang Lu said. When the research team tape-out for the first time, they waited for 4 months, but the results were not as expected. They had to start over, reexamining every detail and looking for possible issues. After 2 months of adjustments and optimizations, followed by another 6-month wait, when the results of the second tape-out came back, the team members showed happy smiles. In this research, the experimental results of the Taiji Optical Chip were the result of the team's collective "hardship." "In order to achieve experimental results that meet the expectations of theoretical simulations, we continuously adjusted and optimized the experimental system. Each experiment was a long and tedious process. Similar processes were repeated over a hundred times. The goal of the research team was to achieve a 90% accuracy rate for thousands of intelligent tasks, ultimately surpassing expectations to conclude this 'prolonged battle' with outstanding results," Xu Zhihao told Science China Press.
Fang Lu (left) and the research team's student (Xu Zhihao in the middle) in the laboratory. Photo by Du Shanni. Challenging Tradition
In September 2023, the team submitted their paper to the Science editorial office. The paper was sent for review three days later, and over a month later, the research team received the first round of review comments. "The reviewers had differing opinions on the architecture of the Tai Chi light chip. With the development of deep learning, deep neural networks have become the mainstream intelligent computing architecture," Fang Lu said. The research team stuck to their viewpoint, using more theories and experimental evidence to persuade the reviewers. After the second round of review, their paper was successfully accepted. This year marks Fang Lu's 17th year in scientific research: she graduated from the University of Science and Technology of China with a bachelor's degree, obtained her Ph.D. from the Hong Kong University of Science and Technology, and is currently a tenured associate professor in the Department of Electronic Engineering at Tsinghua University. Large-scale optoelectronic intelligent computing has always been the research goal of Fang Lu's team, and the entire team has constructed a vertical parallel and horizontally interconnected path planning to achieve this goal.
"Each team member has their own independent research questions. While exploring paths in parallel, they also engage in cross-collaboration, exchanging original ideas. 'The road ahead is long,' the research team will not forget their original intentions and will persistently seek progress on the path of large-scale intelligent light computing. Currently, we are building application systems for chips to provide feasible solutions for industrialization," Fang Lu added.
Paper Link: https://www.science.org/doi/10.1126/science.adl1203
