Shulou(Shulou.com)11/24 Report--

Published in 1968, can Biomimetic Dreams about Electronic Sheep? "it was boldly imagined that robots would also have human qualities. These humanoids could think, sleep and dream, opening up people's thinking about cold creation and life. [1]

Now, the world full of cyber punk doesn't seem far away. The chip structure has highly imitated the biological brain, began to have five senses, and has more and more human characteristics. It is a neuromorphological chip, a chip that consumes more than a thousand times more energy than the existing CPU or GPU. [2]

So far, neural morphological computing is still in the research stage, but it has been industrialized one after another, indicating that this technology will be the first batch of crab eaters. [3]

From biological brain to chip neuromorphological computing (Neuromorphic Computing) is also known as neural mimicry computing, which refers to the architecture built with reference to the biological brain neuron structure and thinking processing mode. It is an advanced computing form that jumps out of the traditional von Neumann architecture. The chip designed according to this architecture is the neuromorphological chip.

The simple explanation is to put the human brain into the chip. Although at first glance such a word is very obscure and difficult to understand, in fact, artificial intelligence (AI) technology, which also draws lessons from the human brain, has already entered thousands of households [4]. However, the neuromorphological chip is a device whose architecture is closer to that of the human brain.

A kind of neuromorphological computing in brain-like chip is also a kind of brain-like chip (Brain-inspired Computing, also known as brain-heuristic computing).

At present, the precise definition and scope division of brain-like chips have not been unified in academia and industry. Generally, it is roughly divided into neural morphological chip (based on pulse neural network SNN) and deep learning special processor (based on artificial / deep neural network ANN / DNN). The former approaches the biological brain from the structural layer and focuses on designing the chip structure according to the human brain neuron model and its organizational structure, while the latter designs the chip structure around mature cognitive computing algorithms. [5]

According to a simple explanation of the two principles, the special processor for deep learning is dimensionality reduction processing, which converts multi-dimensional problems into one-dimensional information flow; neuro-morphological chips are dimensionally enhanced processing, which is closer to the way of thinking of the human brain through multi-dimensional space-time transformation. in order to achieve better energy consumption, computing power and efficiency. [6]

Comparison between ▲ von Neumann architecture and neuromorphological architecture, figure source (Nature Computational Science [7])

Dedicated processors for deep learning belong to another sub-industry. As early as 2012, the Institute of Computing of the Chinese Academy of Sciences developed the Cambrian, the first international chip to support deep neural network processor architecture [8]. Current players include Mythic, Graphcore, Gyrfalcon Technology, Groq, HAILO, Greenwaves, Google, Horizon, Cambrian and so on.

The two major directions are not independent or mutually exclusive, but cross-integration. For those that are already very good at deep learning, such as simulating human vision or natural language interaction tasks, we will continue to use deep learning networks to deal with; for other things that are not suitable for deep learning, such as smell, robot control, multimodal and even cross-modal storage, a new architecture of neuromorphological chips will be used.

Both the research community and the industry are gradually blurring the line between the two, but in fact, the brain-like chips in many papers, reports or articles refer to neuromorphological chips. A more accurate description, neuromorphological chip, will be used in this paper.

Characteristics of two major platforms for ▲ brain computing [9]

The biological brain can realize various functions such as perception, movement, thinking and intelligence, but because of its complexity, the understanding and understanding of the brain is still very limited. this brings challenges to the neuromorphological chip from cognitive principle, hardware implementation, intelligent algorithm to dual-brain fusion and so on. [10]

According to the current research, in the biological brain, neurons transmit and adjust signals through dendrites and synapses, and at the same time, neurons communicate with each other in the form of pulse signals. In fact, the structure and function of a single neuron are not complex, but a variety of complex learning and cognitive functions can be realized through large-scale synaptic interconnected neural networks. [11]

In addition, the biological brain is very different from the mainstream man-made chip structure:

The information processing structure of neurons and synapses not only has higher efficiency, but also can realize large-scale parallel processing; [12]

In the von Neumann computing architecture of the traditional computing system, computing and storage are separated, but the biological brain is the integration of storage and processing, and there is no separate memory. In addition, there is no dynamic random access memory, no hash hierarchy, no shared memory, etc. [13] (see the history article of fruit shell hard technology, "memory and computing integrated chip, potential stock in the era of artificial intelligence")

Biological brain memory is not immutable, but there are both frequently repeated long-term memory and fast forgotten short-term memory. The conversion between the two is manifested in synapses, that is, the transformation of long-term / short-term plasticity; [14]

Computers are basically all-digital signal processing, while the biological brain is a mixed signal, brain communication uses digital signals for rapid transmission, and neuron and synaptic processing uses more effective analog chemical forms. [15]

▲ traditional circuit structure (left) compared with human brain structure (right) [16]

But it is not to say that artificial devices do not have any advantages. If CMOS can be used to construct devices of the same size as biological brains, the two will show different advantages. Neuromorphological chips must eventually be designed and manufactured by combining the respective advantages of biological brains and artificial devices.

▲ biological brain contrast CMOS devices of the same size [16]

Three mainstream forms of implementation as early as 1952, there were studies modeling the nervous system as equivalent circuits [17] until the 1980s, when Carver Mead of California Institute of Technology, one of the inventors of very large scale integrated circuits (VLSI), coined the term Neuromorphic to describe devices and systems that mimic certain functions of the biological nervous system. [18]

Scientists, including Carver Mead, have spent more than 40 years working on the technology, with the ultimate goal of simulating analytical systems that simulate human senses and processing mechanisms, such as touch, vision, hearing and thinking. Now, the neuromorphological chip industry has developed an embryonic form.

Behind an ideal neuromorphological chip is the collision of many disciplines, including the pursuit of biomaterials in materials, the construction of neurons and synapses on devices, the connection of neural networks in circuits, and the algorithmic realization of brain thinking ability [19]. Different neural morphological chips involve a lot of materials, devices and processes, and it will drive algorithms and applications from materials, devices, circuits and architectures from the bottom up. [7]

The fields and opportunities involved in ▲ neuromorphological chip, figure source (Nature Computational Science), with changes [7]

So far, the structure of the neural morphology chip is basically the same, including neuron computing, synaptic weight storage, routing and communication, and the pulse neural network (SNN) model is adopted at the same time. [9]

However, according to the difference of materials, devices and circuits, it can be divided into three schools: analog circuit-dominated neuromorphological system (digital-analog hybrid CMOS type), all-digital circuit nervous system (digital CMOS type), and digital-analog hybrid neural morphological system based on new devices (memristor is a candidate technology).

The two CMOS-based methods can continue to make use of existing manufacturing technologies to build artificial neurons and connect artificial synapses, but to simulate the behavior of a single neuron or synapse, a circuit module is composed of multiple CMOS devices, so the integrated density, power consumption and functional simulation accuracy will be limited. From the bionic point of view of the underlying device, the new device simulates neurons and synapses at the device level, and has significant advantages in power consumption and learning performance, but it is still in the exploratory stage. [12]

Among them, digital CMOS is the easiest form of industrialization at present. On the one hand, the technology and manufacturing maturity is high, on the other hand, there are no concerns and limitations of analog circuits.

Three realization forms of ▲ neuromorphological chip, tabulation (fruit shell hard technology)

How to measure the quality of a neuromorphological chip? Its competitiveness is mainly evaluated from four indicators: computational density, energy efficiency, computational accuracy and learning ability. [20]

Four key indicators and current situation of ▲ neuromorphological chip, tabulation (fruit shell hard technology)

Source: Nature Electronics [20], New economy Guide [21]

To solve the urgent needs of the industry, why do we need to do neuromorphological chips? Its commercial value lies in its continuous self-learning under the condition of low power consumption and a small amount of training data, and ideally, in the same artificial intelligence task, the energy consumption of neural morphology chip is more than a thousand times lower than that of traditional CPU or GPU.

In the digital age, the computing speed of computers is getting faster and faster, and even playing chess, they can beat the world champions. Therefore, people affectionately call the computer a computer, but its energy efficiency and intelligence are far from reaching the level of the biological brain. [22]

For example, the AlphaZero is a giant made up of 5000 Google's dedicated machine learning processors (TPU), but consumes up to 200W per unit [23]. For example, IBM once simulated a cat's cerebral cortex model (equivalent to 1% of the human brain) on the Deep Blue supercomputer platform, requiring nearly 150,000 pieces of CPU and 144TB main memory and consuming as much as 1.4MW. [24]

The human brain, by contrast, is made up of about 85 billion neurons connected by 1000 trillion (1015) synapses and can perform 100 billion operations per second, but such a large system consumes only 20 watts of power to process daily tasks. At the same time, a two-year-old can recognize people he or she is familiar with without difficulty from many people at any angle, distance and light, far more intelligent than any existing computing system. [22]

Therefore, the chip into the brain, by imitating the biological brain structure, the neuromorphological chip does have the characteristics of energy-efficiency ratio. Its unique event-triggered operation mechanism does not trigger operation behavior when there is no dynamic information generation. At the same time, it is also good at analyzing complex spatio-temporal series. Although the rate of a single neuron is very low, it can perform large-scale parallel operations because it is similar to the mechanism of the biological brain, and the response speed will be much faster than the existing solutions.

It can be said that the neural morphology chip has the potential to become the current savior to solve the three major problems facing the industry: first, the magnitude of data is large; second, the digital form is becoming more and more diversified, and a lot of data can no longer be solved by manual editing input or manual processing. Third, applications have increasingly strong requirements for delay, and the traditional single computing architecture will encounter performance and power consumption bottlenecks.

In addition, the neuromorphological chip also fits the concept of green computing. Computing power becomes another economic index after electric power, the calculation method which consumes a lot of energy is difficult to take the lead, and the way of energy optimization is the best solution to solve the problem. It is estimated that data centers consume about 200 terawatt hours (TWh) of electricity a year, which is already equivalent to the national electricity consumption of some countries in a year. [26]

How to go to large-scale commercialization although neuromorphological chips are good everywhere, they only play a special role in specific areas and will not replace traditional computing platforms. Traditional digital computing chips such as CPU and GPU are good at accurate computing, while neuromorphological chips are good at unstructured data, image recognition, classification of noisy and uncertain data sets, new learning systems and inference systems.

Quantum computing, which can subvert a particular field of computing, is actually the same logic, and it cannot be separated from existing computing systems. The advanced computing system in the future will inevitably require traditional digital chips, neuromorphological chips and quantum computing to cooperate with each other. [27]

At present, neuromorphological chips are difficult to design and manufacture, and have not yet formed a large-scale market. At the same time, the unanimous conclusion of the industry is that the money invested in neuromorphological chips lags far behind artificial intelligence or quantum technology. [27]

A small chip contains knowledge of semiconductor manufacturing technology, brain science, computational neuroscience, cognitive science, statistical physics and other disciplines [28]. To involve physicists, chemists, engineers, computer scientists, biologists and neuroscientists, it is undoubtedly challenging for so many characters to do the same thing and speak the same language.

But its subversive value has led to accelerated commercialization around the world. Data show that the neuromorphological chip market will increase from US $22.743 million in 2021 to US $550 million in 2026, with an annual compound growth rate of 89.1%. [29] in addition, if basic technical problems are solved in the next few years, the global market for neuromorphological chips will account for 18% of the overall artificial intelligence market, reaching US $22 billion by 2035. [30]

So, what problems should be solved if we want to promote large-scale commercialization?

First, the design problem: the brain consumes very little energy while processing complex information in real time, so it is difficult to better understand this efficient working mechanism and apply these mechanisms to the chip. Take only the digital CMOS type with the closest commercial path, the chip interconnection of multi-chip all-digital asynchronous design, the effectiveness and timeliness of chip connection, as well as software layer interconnection computing, distributed computing and flexible partitioning are all difficult to bridge the gap.

Second, the manufacturing problem: using the silicon-based transistor route can reuse the existing manufacturing technology, while the non-silicon-based route has to solve the problems of the underlying process, manufacturing yield and supporting large-scale production, even if the problems are solved and the experimental chip has been made. we should continue to consider the problem of stable supply on the production scale.

Third, software and ecological problems: the neural morphology chip is completely different from the existing architecture, and many developers in the community construct their own pulse neural network algorithms at the bottom and burn the software into the hardware to do experiments through the underlying library. this is obviously not a large-scale scheme. Large-scale commercialization, software tool chain is very important.

Fourth, the lack of killer applications: whether it is robot, autopilot or industrial large-scale optimization, its own logic should be application-driven technology development, and then continue to build the ecosystem on this basis. At present, it is a generally accepted view that neuromorphological technology will be the first to find applications in consumer electronics, mobile terminals and the industrial Internet of things.

Players of neuromorphological chips worldwide, there are mainly three types of organizations involved in the development of neuromorphological computing chips: technology giants represented by Intel, IBM, Qualcomm, universities / research institutions represented by Stanford and Tsinghua University, and start-ups. [31]

▲ neural morphological calculation players are not complete statistics, the source of the picture is intelligent [32]

There are many famous universities such as Massachusetts Institute of Technology, Stanford University, Boston University, Manchester University and Heidelberg University. Technology giants are represented by Intel, IBM, Qualcomm and Samsung, while startups include BrainChip, aiCTX, Numenta, General Vision, Applied Brain Research, Brain Corporation and so on. [32]

From the perspective of research and implementation, the experimental chips of Intel and IBM are the most representative.

▲ currently known neuromorphological chip detailed parameters comparison, tabulation (fruit shell hard technology)

Reference material: IEEE [33]

Intel's Loihi is an all-digital neuromorphological chip. In 2017, Intel developed the first Loihi. In 2021, Intel launched the second generation Loihi2, which is produced by Intel 4 process, and the number of neurons on a single chip reaches 1 million.

Intel neuromorphological chip has made great progress in the field of perception, including gesture recognition, visual reasoning and odor sensing with up to 3,000 times learning data. In addition, Intel has developed an Pohoiki Springs data center rack system that integrates 768 Loihi chips into five standard server-sized chassis.

To make neuromorphological chips easier to use, Intel has also launched an open source software framework called Lava, which builds applications without specialized hardware, runs seamlessly on the heterogeneous architectures of traditional and neuromimic processors, and allows researchers and application developers to further develop on each other's achievements. [34]

Intel is in no hurry to commercialize neuroform chips, and unlike small companies maintaining specific applications, Intel treats it as a general-purpose technology and treats all business opportunities at a level of more than $1 billion.

Brief introduction of ▲ Intel Loihi and Loihi2

TrueNorth is the experimental chip that IBM has been working on for nearly 10 years, and the DARPA program in the United States has funded this project since 2008. In 2011, IBM launched the first generation of TrueNorth.

By 2014, the number of second-generation TrueNorth neurons in IBM has increased from 256to 1 million, the number of programmable synapses has increased from 262144 to 256 million, and it can perform 46 billion synaptic operations per second, with a total power consumption of 70mW (20mW per square centimeter), and the overall size is only 1/15 of that of the first-generation brain-like chips. [5]

It is worth mentioning that IBM also launched a neuromorphological supercomputer called Blue Raven in 2019, which has the processing power of 6400 million neurons and 16 billion synapses and consumes only 40 watts, the equivalent of a household light bulb. [35]

Domestic research on domestic development includes top universities and institutions such as Tsinghua University, Zhejiang University, Fudan University and Chinese Academy of Sciences. At the same time, start-ups have been emerging in the past two years, such as Lingxi Science and Technology, time-aware Science and Technology, Chinese Science nerve Morphology and so on. Among them, the celestial movement of Tsinghua University and the Darwin chip of Zhejiang University are the most representative.

▲ domestic neural morphology chip start-up enterprises incomplete statistics, tabulation (fruit shell hard technology)

Source: company website, New economy Guide [21], Quantum bit [36]

The neuromorphological chip of Tsinghua University is the most representative experimental chip in China. The first generation of celestial movements developed in 2015 used 110nm technology, which was just a small sample at the time. In 2017, the second generation of celestial movements began to achieve advanced results, based on the 28nm process, composed of 156 functional core FCore, containing about 40, 000 neurons and 10 million synapses. Compared with the first generation, the density is 20% higher, the speed is at least 10 times higher, and the bandwidth is at least 100 times higher. [37]

In order to make the neuromorphological chip more practical, Tsinghua University has also independently developed a software tool chain to support automatic mapping and compilation from the deep learning framework to the celestial movement. According to the plan of Tsinghua University, the next generation of celestial movement will be 14nm or more advanced technology, and the function will be more powerful. [38]

Another representative in China is the brain-like computer jointly developed by the United Zhijiang Laboratory of Zhejiang University, whose number of neurons is similar to that of mouse brain neurons. The computer contains 792 Darwin II chips and supports 120 million pulsed neurons and 72 billion synapses. At such a large scale, the typical operating power consumption is only 350W~500W. [39]

In fact, the domestic technical strength in the field of neuromorphological chips has been in the global leading level, and the global players are on the starting line in this field.

It can be said that this is a technology worth laying out. But it is also a difficult bone, a huge track of materials, devices, technology, architecture and algorithms, and a field that goes to the end in the dark. at the same time, it is bound to encounter the problems of application scenarios and input-output ratio in the future.

However, the market is like this, who dares to challenge the blank area, who can become the first batch of people to receive dividends.

Referenc

[1] Philip K. Dick Do Androids Dream of Electric Sheep [M]. Doubleday, 1968. [First Edition].

[2] Davies M, Srinivasa N, Lin T H, et al. Loihi: A neuromorphic manycore processor with on-chip learning [J]. Ieee Micro, 2018, 38 (1): 82-99. Https://doi.org/10.1109/mm.2018.112130359

[3] Hello, Zhang Jiang: time-aware science and technology Joning: rectify the name of brain-like intelligence and become the "crab eater" .2021.11.9. Https://mp.weixin.qq.com/ s / IIxy8E9tfqt6XHk_bfFidg

[4] Arrow: neural morphological computing and artificial intelligence: neural morphological computing chip and artificial intelligence hardware. 2020.4.16. Https://www.arrow.com/ zh-cn / research-and-events / articles / neuromorphic-computing-chips-and-ai-hardware

[5] Tao Jianhua, Chen Yunyi. Development status and thinking of brain-like computing chip and brain-like intelligent robot [J]. Journal of the Chinese Academy of Sciences, 2016. 31 (07): 803-811.

[6] Xinhua News Agency: Shi Luping, Chinese forerunner of brain-like computing: exploring the way in AI "no man's land". 2021.9.26. Https://www.tsinghua.edu.cn/ info / 1662/87321.htm

[7] Schuman C D, Kulkarni S R, Parsa M, et al. Opportunities for neuromorphic computing algorithms and applications [J]. Nature Computational Science, 2022, 2 (1): 10-19. Https://doi.org/10.1038/s43588-021-00184-y

[8] Chen Y, Luo T, Liu S, et al. Dadiannao: A machine-learning supercomputer [C] / / 2014 47th Annual IEEE/ACM International Symposium on Microarchitecture. IEEE, 2014: 609-622 https://doi.org/10.1109/micro.2014.58

[9] Deng Lei. Research on computing model and key technologies of heterogeneous fusion brain computing platform [D]. Tsinghua University, 2017.

[10] Li F, He Y, Xue Q. Progress, challenges and countermeasures of adaptive learning [J]. Educational Technology & Society, 2021, 24 (3): 238-255. Https://www.jstor.org/stable/27032868

Deng Yabin, Wang Zhiwei, Zhao Chenhui, Li Lin, he Shan, Li Qiuhong, Shuai Jianwei, Guo Donghui. A review of brain-like neural networks and neural morphological devices and their circuits [J]. Computer Application Research, 2021 Magi 38 (08): 2241-2250

Wang Zongwei, Yang Yuchao, Cai Yimao, Zhu Tao, Cong Yang, Wang Zhiheng, Huang Ru. Intelligent chip and device technology for neural morphological computing [J]. China Science Foundation, 2019. 33 (06): 656-662.

[13] Wan C, Cai P, Wang M, et al. Artificial sensory memory [J]. Advanced Materials, 2020, 32 (15): 1902434. Https://doi.org/10.1002/adma.201902434

[14] Wu Y, Wang X, Lu W. Dynamic resistive switching devices for neuromorphic computing [J]. Semiconductor Science and Technology, 2021 https://doi.org/10.1088/1361-6641/ac41e4

[15] Sterling P, Laughlin S. Principles of neural design [M]. MIT press, 2015.

[16] Schuller I K, Stevens R, Pino R, et al. Neuromorphic computing-from materials research to systems architecture roundtable [R]. USDOE Office of Science (SC) (United States), 2015. Https://doi.org/10.2172/1283147

[17] Hodgkin A L, Huxley A F. A quantitative description of membrane current and its application to conduction and excitation in nerve [J]. The Journal of physiology, 1952 (4): 500. https://doi.org/10.1113/jphysiol.1952.sp004764

[18] Mead C. Adaptive retina[M] / / Analog VLSI implementation of neural systems. Springer, Boston, MA, 1989: 239-246. Https://doi.org/10.1007/978-1-4613-1639-8_10

[19] Wu Chaohui. Brain-like computing-- to build an artificial super brain [N]. People's Daily, 2022-1-11 http://ent.people.com.cn/ N1 32328291.html 2022 C1012

[20] Zhang W, Gao B, Tang J, et al. Neuro-inspired computing chips [J]. Nature electronics, 2020, 3 (7): 371,382. Https://doi.org/10.1038/s41928-020-0435-7

Liu Xing, Li Xingyu. Analysis of the development prospect of neural morphological computing chip industrialization [J]. New economy Guide, 2021, (03): 31-34.

[22] learn from the Han people. The Wisdom and Beauty of Science-- numerical Metaphysics and the face of complexity [M], higher Education Press, 2010 Physics 103-105

[23] Furber S. Large-scale neuromorphic computing systems [J]. Journal of neural engineering, 2016, 13 (5): 051001. Https://doi.org/10.1088/1741-2560/13/5/051001

[24] Ananthanarayanan R, Esser S K, Simon H D, et al. The cat is out of the bag: cortical simulations with 109 neurons, 1013 synapses[C] / / Proceedings of the conference on high performance computing networking, storage and analysis. 2009: 1-12. Https://doi.org/10.1145/1654059.1654124

[25] Wu K J. Google's new AI is a master of games, but how does it compare to the human mind [J]. Smithsonian, 2018, 10.

[26] Jones N. How to stop data centres from gobbling up the world's electricity [J]. Nature, 2018 (7722): 163167. Https://doi.org/10.1038/d41586-018-06610-y

[27] Mehonic A, Kenyon A J. Brain-inspired computing needs a master plan [J]. Nature, 2022, 604 (7905): 255-260. Https://doi.org/10.1038/s41586-021-04362-w

[28] Zhiyuan Community: the past life and present life of brain computing in one article | Li Guoqi, researcher, Institute of Automation, Chinese Academy of Sciences. 2022.5.7. Https://mp.weixin.qq.com/ s/9-FW5ZxJt4fthNhQI10qHA

Markets And Markets:Neurom orphic Computing Market by Offering, Deployment, Application (Image Recognition, Signal Recognition, Data Mining), Vertical (Aerospace, Military, & Defense, Automotive, Medical) and Geography (2021-2026). Https://www.marketsandmarkets.com/Market-Reports/neuromorphic-chip-market-227703024.html

[30] Yole:Neuromorphic Computing and Sensing.2021.5. https://s3.i-micronews.com/uploads/2021/05/Yole-D%C3%A9veloppement-Neuromorphic-Computing-and-Sensing-2021-Sample.pdf

Shi Luping, Pei Jing, Zhao Rong. Neural morphological computation for artificial general intelligence [J]. Artificial Intelligence, 2020 and 4 (01): 6-15.

[32] Wisdom: scraping the global bionic chip project! 15 + company has entered, artificial brain is the door to future computing. 2020.4.26. Https://mp.weixin.qq.com/ s / VW21Q7IZOiSX9lCmw2jEFA

Paul A, Song S, Das A. Design technology co-optimization for neuromorphic computing [C] / / 2021 12th International Green and Sustainable Computing Conference (IGSC). IEEE, 2021: 1-6. Https://doi.org/10.1109/igsc54211.2021.9651556

[34] know IN: Intel launches Loihi 2 and new Lava software framework, working with more partners to promote the further development of neural mimicry computing. 2021.10.1. Https://mp.weixin.qq.com/ s / gDdTwxqmhj7wudEgqCib-g

[35] ARFL:World's largest Neuromorphic digital synaptic super computer. https://afresearchlab.com/technology/artificial-intelligence/worlds-largest-neuromorphic-digital-synaptic-super-computer/

[36] Quantum bits: start-ups raise hundreds of millions of dollars, why isn't this "artificial super brain" track a gimmick? .2022.8.18. Https://mp.weixin.qq.com/ s / sXNK2piMFa3QI17DijNcJg

[37] Tsinghua University: professor Shi Luping of the brain Computing Center of Tsinghua University published the cover article in Nature. 2019.8.1. Https://www.tsinghua.edu.cn/ info / 1181/34641.htm

[38] Tsinghua University: Tsinghua University general artificial intelligence chip "Heavenly Machine" appears on the cover of "Nature"-- "brain-like" to find "Heavenly Intelligence" .2019.8.12. Https://www.tsinghua.edu.cn/ info / 1182/50148.htm

[39] Xinhuanet: the largest Neuron-like brain computer came out. 2020.9.11. Http://www.xinhuanet.com/ science / 2020-09 Universe 11 / c_139360379.htm

This article comes from the official account of Wechat: fruit Shell hard Technology (ID:guokr233), author: Fu Bin, Editor: Li Tuo

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.