World Frontier Technology Development Report 2023 "World Information Technology Development Report" (5) Advanced Computing Technology

Data comes from: "World Frontier Technology Development Report 2023" and the Internet

1 Overview

The global digital transformation has entered the stage of multiplication and innovation, and the proportion of digital economy in various countries continues to increase. In this context, computing power has become an important driving force for the sustainable development of digital technology and the core productivity in the digital economy era. Advanced computing is the carrier of the integration of multiple fields and technologies, and is the main driving force for the sustained and rapid growth of computing power. With the continuous development of advanced computing, computing performance and energy efficiency will usher in new breakthroughs, and the computing power system will undergo disruptive changes.

2. Supercomputer

Supercomputer is a type of computer with the most powerful functions, fastest computing speed and largest storage capacity among classical computers. It is mostly used in national high-tech fields and cutting-edge technology research. It is an important symbol of national scientific and technological development level and comprehensive national strength. In 2022, the competition among the world's major economies in the field of supercomputers will still be fierce. At the same time, the application scenarios of supercomputers continue to expand and the difficulty of tasks continues to increase.

2.1 Crusher, the first exascale supercomputer in the United States, goes online for trial operation

In March 2022, the United States' first E-class (exascale) supercomputer Crusher went online for trial operation. The Crusher supercomputer runs at the Oak Ridge Leading Computing Facility led by the U.S. Department of Energy and is a scaled-down version of the Frontier supercomputer. Crusher has 10.82 million AMD cores, a total of 192 nodes, and a total area of ​​only 44 square feet (approximately 4 square meters). Compared with the Titan supercomputer launched in the United States in 2013, Crusher uses only one percent of the volume to achieve better performance. Currently, the United States has three routes in building E-class supercomputers: Frontier with AMD processor + AMD accelerator card; Aurora with Intel processor + Intel accelerator card; and Polaris with AMD processor + NVIDIA accelerator card. The launch of Crusher is seen as a preview for researchers to prepare in advance for the deployment of more powerful exascale supercomputers.

2.2 Europe’s most powerful supercomputer completed

In June 2022, the European High Performance Computing Joint Undertaking (EuroHPC JU) built Europe's most powerful supercomputer LUMI in Finland and opened access to the system to European users in September 2022. . The supercomputer consists of HPE Cray EX nodes with 2,560 child nodes, including a 64-core AMD EPYC "Trento" central processor and four AMD Instinct MI250X graphics processors. When LUMI was built, it was the third fastest supercomputer in the world, behind Frontier developed by Oak Ridge National Laboratory (ORNL) in the United States and Fugaku in Japan.

2.3 NVIDIA and Microsoft jointly develop artificial intelligence supercomputer

In November 2022, the American company Nvidia and Microsoft reached a cooperation to jointly develop artificial intelligence supercomputers. The artificial intelligence supercomputer is built on tens of thousands of NVIDIA H100, A100 and other high-performance chips and operates on Microsoft's Azure cloud. In addition, Nvidia will also cooperate with Microsoft to develop artificial intelligence models and provide artificial intelligence services to customers. Nvidia said Azure will be the first public cloud to use Quantum-2 InfiniBand networking technology, which can deliver processing speeds of up to 400 gigabits per second.

3. New computing technologies

The amount of information in human society is growing explosively, and the limitations of Von Neumann architecture computers in certain special application scenarios are gradually becoming apparent. Emerging computing models and structures such as brain-inspired computing and probabilistic computing are constantly emerging, which will meet the needs of artificial intelligence, data centers and other applications for high-load, low-energy computing and become a breakthrough in future intelligent computing.

3.1 Beijing University of Aeronautics and Astronautics in China proposed a “mixed probabilistic logic calculation” mechanism

In February 2022, researchers from Beijing University of Aeronautics and Astronautics in China proposed a "mixed probabilistic logic calculation" mechanism and created a prototype. In probabilistic computers, arithmetic operations are performed to perform probability calculations with the help of random and uncorrelated correlations representing the logical levels of data. Binary computing has huge advantages in hardware consumption, but its principle of representing probability values ​​based on pulse frequency brings about a large computing delay problem. Based on this, the Beihang research team proposed the idea of ​​replacing the original single-bit stream probability calculation with mixed probabilistic logic calculation, which solves the problem of large delay in traditional probability calculation and will be used in future scenarios such as brain-inspired computing, artificial intelligence, deep learning and information processing. has broad application prospects.

3.2 Graz University of Technology in Austria and Intel Corporation prove the superiority of neuromorphic chips

In June 2022, a study conducted by researchers at the Graz University of Technology in Austria and Intel Corporation in the United States showed that neuromorphic computing hardware inspired by the structure and biology of the human brain is effective in supporting complex deep neural networks ( Deep Neural Networks (DNN) has more advantages. The researchers noted that deep neural networks are crucial for machines to achieve higher-level cognitive functions, such as finding connections between sentences and answering questions. By comparing the power consumption and efficiency of neuromorphic chips and standard computer chips when performing the same tasks, researchers found that neuromorphic chips can run large deep neural networks 4 to 16 times more efficiently than traditional computing hardware.

3.3 The University of Texas at Austin uses graphene to create brain-inspired computer transistors

In August 2022, researchers at the University of Texas at Austin used graphene to develop synaptic transistors for brain-like computers. The researchers combined graphene and Nafion polymer film materials to form the skeleton of the synaptic transistor. Together, these materials demonstrate synaptic-like behavior, with the system's capabilities increasing over time with increased frequency of use, exhibiting neuromuscular memory. This means devices will be able to do tasks like recognizing and interpreting images better and faster. The research is expected to spur the development of brain-inspired computers.

3.4 The University of Chicago uses elastic semiconductors to manufacture wearable neuromorphic chips

In August 2022, researchers at the Pritzker School of Molecular Engineering at the University of Chicago developed a flexible and stretchable computing chip that functions like a human brain and can store and process information in an integrated manner. Researchers are using stretchable and bendable polymers to create chip devices that can be worn on the body and collect data from multiple biosensors. The chip can also use cutting-edge machine learning methods to draw conclusions about a person's health. The research was published in the journal Matter.

3.5 Chinese and American research teams have developed a new neuromorphic chip for edge AI computing, which can take into account both versatility and high efficiency

In August 2022, a joint research team composed of Tsinghua University in China, the University of California, San Diego (University of California, San Diego), and Stanford University developed a neuromorphic impedance random access memory chip for edge artificial intelligence computing. NeuRRAM combines the advantages of multi-function, high efficiency and low energy consumption. The NeuRRAM chip uses 48 cores that can perform parallel computing. Cores can be selectively shut down via power gating during periods of inactive use, while model parameters are retained by a non-volatile resistive memory device. At the heart of each core is a bidirectional transposable synaptic array, which consists of 256 × 256 RRAM cells and 256 complementary metal oxide semiconductor neuron circuits to implement analog-to-digital converters and activation functions. The bidirectionally transposable synapse array architecture is designed to provide flexible control over the direction of data flow, which is critical to enable different model architectures with different data flow patterns. NeuRRAM chips support a variety of neural network models and architectures and can be applied in scenarios such as computer vision, speech recognition, and natural language processing.

3.6 A research team from Harvard University in the United States developed an ion circuit that can perform neural network calculations in water

2022年10月，美国哈佛大学(Harvard University)工程与应用科学学院研究团队开发出一种可在水中进行神经网络计算的离子电路，并执行了神经网络计算的核心过程。该团队受大脑的启发，构建了一种由醌(Quinone)分子的水溶液组成的新型离子晶体管，其中2个同心环形电极和1个中心圆盘电极连接在一起。研究人员设计了液体局部PH值门控离子晶体管，使圆盘电压与代表局部PH门控晶体管的“权重”参数的算术乘积恰好等于圆盘电流。未来，美国哈佛大学研究团队将使用更加多样化的离子种类来处理信息内容。相关成果发表在《先进材料》(Advanced Materials)期刊。