The Tsinghua University team has made new breakthroughs in the field of ultra-high-performance computing chips, and related research has been published in Nature.
With the emergence of various large models and deep neural networks, how to create a model that meets the development of artificial intelligence and has both The next generation of AI chips with large computing power and high energy efficiency has become an international hot spot. Among the 2023 major scientific issues released by the China Association for Science and Technology, "How to realize low-energy artificial intelligence" was ranked first. Recently, the Tsinghua University team has made new breakthroughs in the superhigh-performance computing chip field. The relevant results were published in Nature under the title "All-analog photo-electronic chip for high-speed vision tasks". This chip is based on a pure analog optoelectronic fusion computing architecture. In the actual measurement of intelligent vision tasks including ImageNet, the same accuracy rate is higher than that of existing high-performance GPUs. The computing power is increased by 3,000 times, and the energy efficiency is increased by 4 million times. 
The content that needs to be rewritten is: Figure 1 Related papers (Source: "Nature")
Chen, Y. et al. All-analog photoelectronic chip for high-speed vision tasks. Nature https://doi.org/10.1038/s41586-023-06558-8 (2023) .
The future is already here? It is not easy to achieve a leap in computing power with light-based computing chips, especially the current traditional chip architecture, which is limited by the size of electronic transistors approaching the physical limit. A new computing architecture has become the key to breaking the situation. Optical computing, with its ultra-high parallelism and speed, is considered one of the most powerful competing solutions for future disruptive computing architectures. Optical computing, as the name suggests, changes the computing carrier from electricity to light, and uses the propagation of light in the chip to perform calculations. Faced with the attractive prospect of computing at the speed of light, well-known scientific research teams at home and abroad have proposed various designs in recent years. However, in order to replace existing electronic devices to achieve system-level applications, they still face major bottlenecks:
- The first is how to integrate large-scale computing units (controllable neurons) on one chip and constrain the accumulation of errors;
- The second is to achieve high-speed and efficient on-chip nonlinearity;
- The third is to be compatible with the current information society dominated by electronic signals, how to provide optical computing and electronic signal computing efficient interface. The current common analog-to-digital conversion power consumption is many orders of magnitude higher than the multiplication and addition operations of each step of optical computing, which masks the performance advantages of optical computing itself, making it difficult for optical chips to demonstrate their superiority in practical applications.
System-level computing power and energy efficiency exceed existing chips by 10,000 timesIn order to solve this international problem, the Tsinghua University team creatively proposed a computational framework for simulating electrical fusion and simulated light, constructed a large-scale multi-layer diffraction neural network under visible light to achieve visual feature extraction, and used photocurrent to directly perform Pure analog electronic calculation of Kirchhoff's law, the two are integrated in the same chip frame, completing a new computing system of "sensing before sensing, mid-range sensing". It greatly reduces the demand for high-precision ADC, eliminates the physical bottleneck of speed, accuracy and power consumption that mutually restrict the traditional computer vision processing paradigm in the analog-to-digital conversion process, and achieves breakthroughs in large-scale integration, efficient non-linearity, and high-speed on one chip. There are three key bottlenecks in the optical and electrical interface.
Figure 2. The computing principle and chip architecture of the optoelectronic computing chip Accel (Source: "Nature")
## Under actual measurement performance, the system-level computing power of ACCEL chips has reached thousands of times that of existing high-performance chips. At the same time, the system-level energy efficiency reaches 74.8 Peta-OPS/W, which is an improvement of two thousand to millions of times compared with the existing high-performance GPU, TPU, optical computing and analog electrical computing architecture. ACCEL running at ultra-low power consumption will helpsignificantly improve the heat problem and bring all-round breakthroughs to the future design of chips , and provide a computing power basis for ultra-high-speed physical observations. At the same time, it brings significant benefits to scenarios with high endurance requirements such as unmanned systems and autonomous driving.
## 表 Table 1. Accel and the existing high -performance chip system -level measurement performance indicators (Source: "Nature")
##Direct calculation of incoherent lightFurthermore, the ACCEL chip also supports non-coherent light Direct calculation of coherent light vision scenes, such as the traffic scene experiment demonstrated in the paper. It has significantly expanded the application fields of ACCEL and is expected to subvert the current idea of taking pictures and saving them in memory before performing calculations in fields such as autonomous driving, robot vision, and mobile devices. It avoids transmission and ADC bandwidth limitations and completes calculations during the sensing process. .
## 图 3. Accel can be used for electronic devices ultra -low power consuming face to wake up the movement of the movement (Source: Tsinghua University)
##################Open up a new path: the disruptive architecture is expected to be truly implemented##################Proposed by the Tsinghua team The new computing architecture is not only of great significance to the application and deployment of optical computing technology, but also deeply inspiring for the future integration of other high-performance computing technologies with current electronic information systems. ###############One of the corresponding authors of the paper, Academician Dai Qionghai of Tsinghua University, said, “It is a big mountain to develop a computing system using new principles, but to truly implement the new generation computing architecture into reality Life, and solving the major needs of the national economy and people's livelihood, is a more important task after reaching the peak."##############The special review of this research published in Research Briefing specially invited by Nature magazine also pointed out that, "###Perhaps the emergence of this work will allow a new generation of computing architecture to enter daily life much sooner than expected### (ACCEL might enable these architectures to play a part in our daily life much sooner than expected. )". ############## Academician Dai Qionghai, Associate Professor Fang Lu, Associate Researcher Qiao Fei, and Assistant Professor Wu Jiamin of Tsinghua University are the co-corresponding authors of this article; doctoral student Chen Yitong, doctoral student Maimeti Naza Dr. Maiti and Xu Han are co-authors; Dr. Meng Yao, Assistant Researcher Zhou Tiankuan, doctoral student Li Guangpu, Researcher Fan Jingtao, and Associate Researcher Wei Qi participated in this study. ###
The above is the detailed content of New breakthrough in Tsinghua Optoelectronics computing: chip performance increased by 10,000 times, research reaches the top Nature. For more information, please follow other related articles on the PHP Chinese website!
