Beijing, October 13 (Reporter Jin Haotian) – After more than half a century of digital computing dominating the field of computer science, Chinese scientists have achieved a major breakthrough in a new computing architecture. A team led by Sun Zhong from the Institute for Artificial Intelligence at Peking University, in collaboration with a research team from the School of Integrated Circuits, has successfully developed a high-precision, scalable analog matrix computing chip based on resistive random access memory (RANM). This achievement marks the first time that the precision of analog computing has been improved to 24-bit fixed-point accuracy. Performance evaluations indicate that when solving key scientific problems such as large-scale MIMO signal detection, the chip's computational throughput and energy efficiency are hundreds to thousands of times higher than current top-tier digital processors (such as GPUs).This achievement marks a breakthrough in my country's efforts to overcome the century-old challenge of analog computing and represents a major breakthrough in the paradigm shift of computing in the post-Moore's Law era. It opens up a new path for addressing the computing power challenges in fields such as artificial intelligence and 6G communications. The relevant findings were published on the 13th in the international academic journal Nature Electronics.

Sun Zhong, the corresponding author of the paper, told reporters that analog computing was the core technology of early computers, directly calculating

based on physical laws, and possessing inherent advantages such as high parallelism, low latency, and low power consumption. However, due to the low precision and limited scalability of traditional analog computing, it has gradually been replaced by high-precision, programmable digital computing, becoming an "outdated technology" relegated to textbooks. This represents a bottleneck in its development."How to make analog computing both high-precision and scalable, so as to give full play to its inherent advantages in modern computing tasks, has always been a 'century-old problem' that has plagued the global scientific community. Digital computing, although highly accurate, is slow and suffers from the 'memory wall' problem of the von Neumann architecture, which has become a bottleneck for the development of artificial intelligence, scientific computing and 6G communication."

Faced with this challenge, the research team chose a path of integrated innovation. Through the collaborative design of novel information devices, original circuits, and classic algorithms, they achieved for the first time an analog computing system with accuracy comparable to digital computing, improving the accuracy of traditional analog computing by five orders of magnitude. "Our new solution maintains the low complexity advantage of analog computing while achieving computational accuracy comparable to the digital FP32 processor. The team also proposed a block matrix analog computing method, which decomposes large problems into multiple chips for collaborative solving, like a jigsaw puzzle, successfully breaking through the scale limitations of analog computing. Experiments have shown that a 16×16 matrix equation can be solved."

Sun Zhong revealed that through rigorous experimental testing and benchmark comparisons, the technology has demonstrated outstanding performance. In terms of computing power, when solving the 32×32 matrix inversion problem, the solution's computing power surpasses that of a single core of a high-end GPU. When the problem size increases to 128×128, the computational throughput can reach more than 1000 times that of a top-tier digital processor. "This technology also demonstrates extreme energy efficiency. At the same level of precision, its energy efficiency is more than 100 times higher than that of traditional digital processors, providing crucial technical support for addressing energy consumption issues in computing centers."

Sun Zhong revealed that through rigorous experimental testing and benchmark comparisons, the technology has demonstrated outstanding performance. In terms of computing power, when solving the 32×32 matrix inversion problem, the solution's computing power surpasses that of a single core of a high-end GPU. When the problem size increases to 128×128, the computational throughput can reach more than 1000 times that of a top-tier digital processor. "This technology also demonstrates extreme energy efficiency. At the same level of precision, its energy efficiency is more than 100 times higher than that of traditional digital processors, providing crucial technical support for addressing energy consumption issues in computing centers."

"The significance of this breakthrough goes far beyond a top-journal paper. Its application prospects are broad, enabling diverse computing scenarios and potentially reshaping the computing power landscape," said Sun Zhong. He explained that in the future 6G communication field, it will enable base stations to process massive antenna signals in real-time and with low energy consumption, improving network capacity and energy efficiency. For the rapidly developing field of artificial intelligence, this research is expected to accelerate computationally intensive second-order optimization algorithms in large model training, thereby significantly improving training efficiency. "More importantly, the low-power characteristics will also strongly support the direct operation of complex signal processing and AI training and push integration on terminal devices, greatly reducing reliance on the cloud and thus propelling edge computing into a new stage."

"The greatest value of this work lies in its demonstration that analog computing can solve core computational problems in modern science and engineering with extremely high efficiency and accuracy. It can be said that we have explored a highly promising path for improving computing power, which is expected to break the long-term monopoly of digital computing and usher in a new era of ubiquitous, green, and efficient computing power." Sun Zhong revealed that the team is currently actively promoting the industrialization of this technology and will bring the laboratory results to the market as soon as possible.