As a bridge between software and hardware, the instruction set is the foundation of the chip industry. RISC-V (fifth-generation reduced instruction set) has gained wide attention around the world, and the application scenarios are rapidly expanding from smart IoT to mobile phones, servers, and other fields. Recently, Dai Hongjun, a professor at the Institute of Intelligent Innovation and the School of Software of Shandong University, led the basic software team to quickly integrate the RISC-V architecture into the field of server industrialization. Professor Tian Yonghong's team at the School of Information Engineering, Shenzhen Graduate School of Peking University, built and open-sourced SpikingJelly (Chinese: sting), a deep learning framework for spiking neural networks, has been welcomed by researchers because of its simplicity and fast training speed.
Based on the one-week scientific and technological memory formed by the daily list of the Science and Technology Innovation Hot List of the International Science and Technology Innovation Center Network Service Platform, we have launched the column "Weekly Frontier Technology Inventory". Today, I would like to bring you the sixty-fourth issue.
Science Advances丨Benefiting researchers in multiple fields! "Sting" into a rookie
The overall structure of the SpikingJelly framework, the sample code, the simulation speed, the ecological niche, and the typical applications
Spiking neural networks (SNNs) are known as the third generation of neural networks, which use a lower level of abstraction of the biological nervous system, which is not only a basic tool for studying the principles of the brain in neuroscience, but also attracts the attention of computational science because of its sparse computing, event-driven, and ultra-low-power characteristics. With the introduction of deep learning methods, the performance of SNN has been greatly improved, and pulsed deep learning, as an interdisciplinary discipline of computational neuroscience and deep learning, has become an emerging research hotspot. Traditional SNN frameworks focus more on biological interpretability, are committed to building fine spiking neurons and simulating real biological nervous systems, do not support automatic differentiation, cannot make full use of the massively parallel computing capabilities of graphics processing units (GPUs), and lack support for neuromorphic sensors and computing chips, making it difficult to be used for spiking deep learning tasks.
In this regard, the team of Professor Tian Yonghong of the School of Information Engineering of Peking University Shenzhen Graduate School has built and open-sourced the spiking neural network deep learning framework SpikingJelly (Chinese name: sting). Stinger provides a full-stack pulsed deep learning solution that provides functions such as neuromorphic data processing, deep SNN construction, alternative gradient training, artificial neural network (ANN) SNN conversion, weighting, and neuromorphic chip deployment. According to the research team, the stinging framework has two significant advantages: it is easy to use, researchers can quickly learn and use it across domains, and easily build and train deep SNNs with just a few lines of code; Ultra-high performance, "Sting" up to 11 times faster training compared to other frameworks.
Therefore, since its launch in the winter of 2019, the "sting" framework has been favored by researchers, and a large number of research works based on "sting" have been published, extending the application of SNN from simple MNIST dataset classification to practical applications such as imageNet image classification, network deployment, and event camera data processing at the human level. Around this framework, there are also cutting-edge frontier explorations, including calibrated neuromorphic perception systems, neuromorphic memristors, event-driven accelerator hardware design, etc. There are currently more than 123 published papers experimenting with the "sting" framework. The above applications and studies show that the open source of "Sting" has greatly promoted the development of the field of pulsed deep learning.
RISC-V core open source code research and development, this university "on the big score"
Improved overall boot flow for RISC-V servers
As a bridge between software and hardware, the instruction set is the foundation of the chip industry. RISC-V is an open-source instruction set architecture based on the principles of reduced instruction sets. Due to its flexible, scalable, and open architecture, RISC-V has gained widespread attention around the world, and its application scenarios are rapidly expanding from smart IoT to more complex mobile phones, servers, and other fields. Recently, Dai Hongjun, a professor at the Institute of Intelligent Innovation and the School of Software of Shandong University, led the basic software team to successfully merge the UEFI startup solution of the first RISC-V CPU server into the open source community tianocore EDK2 mainline warehouse, completed the development of the first RISC-V server firmware that meets the UEFI standard, and realized the rapid integration of RISC-V architecture into the server industrialization field with the help of UEFI's mature ecosystem. This indicates that Shandong University has the world's leading ability in the research and development of basic software core code such as firmware and operating system kernel, and has become an important contributor to the key open source code of RISC-V.
The team also completed the "RISC-V CPU+TPU" intelligent computing fusion solution, realized the world's first enterprise-level TPU BM1684X to drive and optimize on the SG2042 CPU, and successfully ran the AI image generation model StableDiffusion, large model inferllm, etc. This opens up a new path for the deployment of high-performance AI applications on the RISC-V server platform, and promotes the development of large computing power combining RISC-V and AI technologies.
During the research and development process, Professor Dai Hongjun practiced the concept of "organized scientific research and integrated development", led the teachers and students of the School of Software, the School of Integrated Circuits, and the Institute of Intelligent Innovation to form a joint research team, adhered to the "integration into the ecology" and "open source development", and successively joined the RISC-V Foundation, the UEFI community, the OpenKylin community, the Open Atom Open Source Foundation, etc., and established the China UEFI on RISC-V Working Group and the RISC-V Open Source Club of Shandong University.
"International Journal of Fatigue"丨The strategy of doing a quick "physical examination" for fatigue engineering materials is here
Schematic diagram of a high-throughput symmetrical bending cantilever fatigue testing system
Fatigue failure is an important problem faced by engineering components in long-term reliable service. In order to evaluate the fatigue reliability of engineering components and various service materials, the stress amplitude-life curve (S-N curve) of materials is often established by measuring the fatigue fracture life of the material under different stress amplitudes (S) or the service cycle (N) corresponding to the failure of cyclic loading. The maximum stress amplitude of a material when it is loaded without failure after an infinite number of cycles is called the fatigue limit. Fatigue limits are often used to evaluate the fatigue properties of materials at low stress levels, and some metallic materials have significant fatigue limits. In order to obtain the S-N curve and fatigue limit of the material, according to the current test standards such as ASTM and GB, it is usually necessary to use a sufficient number of fatigue samples to carry out a large number of long-term fatigue tests, which is a time-consuming and consumable fatigue test method that has been used in industry and laboratories for nearly 100 years.
With the rapid development of high-tech fields such as aerospace, information, energy, biomedicine, and artificial intelligence, there is an increasingly urgent need to evaluate the fatigue properties of engineering materials and predict the fatigue life of in-service components at low cost, high efficiency, and rapid development. For example, the fatigue properties of non-detachable in-service engineering components in nuclear power and integrated shapes with complex geometries formed by additive manufacturing would be of greater engineering significance if they could be evaluated quickly and at high throughput. At present, although some new materials have made corresponding progress in the high-throughput preparation and high-throughput characterization of the single properties of materials, how to establish high-throughput fatigue testing methods and characterization techniques to achieve low-cost and rapid evaluation of material fatigue reliability is still a key problem to be solved.
Recently, Zhang Guangping's team from the Shenyang National Research Center for Materials Science, Institute of Metal Research, Chinese Academy of Sciences proposed the idea of high-throughput and rapid evaluation of material fatigue properties, and designed and established a method that can simultaneously evaluate multiple small and micro samples (>7) The test system for symmetrical bending fatigue loading was carried out, and the high-throughput fatigue test was carried out on several typical engineering materials used in nuclear power, high-speed rail, automobile and other fields, which were verified by comparison and computational simulation, and the high-throughput testing technology and method of material fatigue performance were established. This technology can not only simulate the fatigue limit of the material specified in the standard to quickly obtain the fatigue limit, but also obtain the stress amplitude/strain amplitude-fatigue life curve at one time. Fatigue data for materials can be obtained quickly within a week, which takes only 1/4 of the time using the standard test methods described above.
The establishment of the test system, technology and principle not only provides a low-cost, efficient and fast new method for the fatigue performance testing of key engineering components in service such as nuclear power, but also provides an effective evaluation strategy for the intrinsic fatigue performance evaluation of complex shape components, material surface coatings, corrosion layers and modified layers, weld areas, material structural units and stress/strain concentration areas and other small areas of additive manufacturing. The efficient establishment of component fatigue performance database and the rapid prediction of fatigue life of engineering components driven by physical model and data.
"Science Advances"丨The world's first! Breaking the previous record for the upper bandwidth limit of pure silicon modulators
Silicon-based modulation structure based on the slow-light effect
With the large-scale application of a new generation of information technology such as artificial intelligence, big data, cloud computing, and the Internet of Things, the total amount of global data is growing exponentially, and the optoelectronic integration technology represented by silicon-based optoelectronics will become an important development trend of optical communication systems. In silicon-based optoelectronic chip systems, silicon-based modulators are the key active devices to complete on-chip information transmission and processing. Limited by the slow carrier transport rate of silicon materials, the typical bandwidth of pure silicon modulators is generally 30-40GHz, which is difficult to adapt to the needs of more than 100Gbaud communication rate in the future, which has become one of the bottlenecks for the further development of silicon-based optoelectronics in the high-speed field.
Recently, the joint team of Professor Wang Xingjun, Professor Peng Chao and Shu Haowen of the School of Electronics of Peking University has made a record breakthrough in ultra-high-speed pure silicon modulators, designing and fabricating the world's first pure silicon modulator with an electro-optical bandwidth of 110GHz.
In order to solve the problem of limited bandwidth of traditional silicon-based modulators, the research team used the silicon-based coupled resonator optical waveguide structure to introduce the slow-light effect, constructed a complete theoretical model of silicon-based slow-light modulator, and comprehensively balanced the optical and electrical index factors by reasonably adjusting the structural parameters to achieve in-depth optimization of the modulator performance. The pure silicon modulator has the advantages of ultra-high bandwidth, ultra-small size, ultra-large passband and CMOS process compatibility, which meets the needs of ultra-high speed, high integration, multi-wavelength communication, high thermal stability and wafer-level production in future ultra-high-speed application scenarios.
On the premise of not introducing heterogeneous materials and complex processes, the research team has achieved a leap in the bandwidth performance of silicon-based modulators, and low-cost wafer-level mass production is possible, demonstrating the great value of silicon-based optoelectronics in the next generation of ultra-high-speed applications, which is of great significance to the development of next-generation data centers.
Nature Communications丨Robot Easy Tofu: Who Says High Sensitivity and Wide Linear Range Can't Be Combined?
Schematic diagram of the structure of DPyCF@SR flexible pressure sensor (left); DPyCF@SR flexible pressure sensor manufacturing process (right)
Flexible pressure sensors are one of the indispensable core functional components of intelligent robots and wearable devices. Over the past two decades, researchers have developed a variety of flexible pressure sensors with high sensitivity or wide linear ranges. Enabling robot dexterity, such as gripping objects of unknown weight or delicate and fragile, requires pressure sensors with high sensitivity and a wide linear range. However, for most existing flexible pressure sensors, high sensitivity and wide linear range are like a "fish and a bear's paw". This has greatly hindered the development of intelligent robots.
Recently, the team of Professor Zhou Wei of the Institute of Micro and Nano Science and Technology of Xiamen University and the team of Professor Yao Haimin of the Department of Mechanical Engineering of the Hong Kong Polytechnic University proposed a new "force-electricity" coupling nonlinear co-design strategy to solve the problem that the current flexible pressure sensor cannot have both high sensitivity and wide linear range, and then explored a special manufacturing process and defined it as a DPyCF@SR.
Accordingly, the flexible pressure sensor manufactured by the company has the performance of high sensitivity (24.6 kPa-1) and ultra-wide linear range (1.4 MPa), and its linear influencing factor (sensitivity × range) is much higher than that of most other piezoresistive flexible pressure sensors reported so far.
In this study, systematic research was carried out on the structural design and manufacturing, sensitive mechanism, performance test and application of flexible pressure sensor, and good practical application results were achieved in the aspects of micro pressure resolution under high load preload, intelligent grasping of tofu and iron by robots.
This study provides a new method and new ideas for the design of flexible pressure sensors, and has important guidance and reference significance for the design and manufacture of other similar high-performance flexible pressure sensors.
For more cutting-edge scientific and technological advancements
You may wish to pay attention
Network service platform of the International Science and Technology Innovation Center
（www.ncsti.gov.cn）
Science and Technology Hot List
Scan the QR code or click "Read more" below
Quick login
Network service platform of the International Science and Technology Innovation Center
Additional reading:
About "Science and Technology Hot List - Frontier Technology"
Based on the research results released by nearly 200 scientific research institutes and units, including the Chinese Academy of Sciences, the Chinese Academy of Engineering, the Chinese Academy of Medical Sciences, the Academy of Agricultural Sciences, 985 universities and new R&D institutions, the network service platform (www.ncsti.gov.cn) of the International Science and Technology Innovation Center dynamically extracts and dynamically extracts and selects multi-dimensional weights such as field dimensions, journal levels, innovation carriers, scholar information, and time gradients, and forms a recommendation list through artificial intelligence calculation and analysis, which is updated daily.