AI computing chip patent competition: Taiwan behind the United States, mainland China and Japan

The rise of artificial intelligence and the Internet of Things has led to a huge increase in demand for massive data computing. The traditional von Neumann computer architecture, which uses data processing rules and data transfer mechanisms between the CPU and memory, has hindered the overall efficiency of operations. The processing time can no longer meet the needs of the smart device for real-time processing.

In recent years, scientists have begun to put forward many technological solutions on R&D issues of hardware and software, such as artificial intelligence computing chip solutions inspired by research on human brain neurons and synaptic mechanisms, that is, using neural network decentralized architecture. Perform massive data computation and learning, and then meet the needs of the massive data computing efficiency and shortened processing time in the era of artificial intelligence.

The literature shows that the main types of artificial intelligence computing chips can be divided into: neuromorphic chips, artificial intelligence system chips, cloud central processing system.

Neuromorphic chip: It is composed of biological nerve system, signal transmission, operation and memory function, electronic circuit materials, components, circuit simulation, circuit design, operation architecture, algorithm, system engineering simulation, such as artificial neuron, artificial nerve Synapses, pulsed neural networks, etc.

Among them, artificial neural synapses are the bridges that connect artificial neurons and allow signals to pass from one artificial neuron to another. Neuromorphic chips treat processors as artificial neurons, and memories as artificial neurites. Touch, and integrate the memory and processor together, which is obviously different from the traditional computer computing architecture. The main technology development direction includes: computational architecture circuit design, algorithm, memristor, etc., typical representatives developed for IBM. TrueNorth.

Artificial Intelligence System Chip: A customized ASIC chip designed to provide artificial intelligence and machine learning operations and processing for smart devices. Its internal computing architecture is usually equipped with CPUs, GPUs, digital signal processors (DSPs) and other traditional computers. The computing architecture, and more specifically, the additional configuration is dedicated to the neural processing engine or neural network processing unit of machine learning and artificial intelligence applications. The typical representatives are Qualcomm's Xiaolong 845 and Huawei's Qilin 970.

Cloud Central Processing System: A custom ASIC chip designed to provide artificial intelligence and machine learning operations and processing for intelligent devices. It is deployed in the hard disk drive slot of cloud data center cabinets to perform specific computing tasks. TPU published for Google.

Industry information shows that most of the scientific and technological community’s R&D results will be protected by patent applications, systematic patent analysis and research will be conducted to grasp the scientific and technological community’s layout of R&D achievements in the artificial intelligence computing chip, and also to learn about Taiwan and other countries. Or what are the opportunities and challenges in the artificial intelligence computing chip compared with the international giants?

Based on the analysis of US patents for artificial intelligence computing chips, it was found that the American assignee included the most patents, followed by Korean rights holders, Japanese rights holders, etc. Among them, 35 domain items covered by patents, found US rights holders. The largest number of computing technology patents, followed by the order of basic communications programs, semiconductors, etc.; Korean rights holders are mainly located in computing technology, semiconductors; Japanese rights holders are concentrated in computing technology.

The data also shows that Chinese rights holders have obtained 2 U.S. patents, and the 35 domain items covered by the patent are mainly distributed in computing technology.

On the other hand, of the 37 technical subcategories involved in patents, American rights holders, Korean rights holders, and Japanese rights holders placed emphasis on the design of neural network hardware - using electronic means, neural network models - system construction, and The U.S. patents on digital computing components - learning machines, neural network models - learning methods, are significantly more than South Korea and Japan.

In addition, the 37 technical items involved in the 2 U.S. patents obtained by Chinese rights holders are mainly distributed in the neural network hardware design - electronic, neural network mode - system architecture, image data processing or processor architecture/ Processor Configuration, Digital Computation Element - Method or device for learning machine, application electronics for identification.

At present, it has not yet been discovered that Taiwan rights holders have obtained US patents for artificial intelligence computing chips. However, Chinese rights holders have already obtained US patents. Taiwan's deployment of this emerging technology patent has obviously fallen behind China. It is worth noting. Research also finds that IBM has received most artificial intelligence. Computing chip US patent, followed by Google, Qualcomm, Samsung and so on.

Artificial intelligence computing chips US patents related to the 35 fields of distribution, found that IBM has the most patents in computing technology, followed by semiconductors, measurement, audio-visual technology, Google and Qualcomm focused on computing technology, Samsung is mainly deployed in computing technology And semiconductors.

On the other hand, in the 37 patents involved in the distribution of technology, it was found that IBM, Qualcomm, Google, and Samsung relatively focused on the implementation of the hardware implementation of the neural network - using electronic, neural network model - system construction, and IBM in the neural network Modes - Learning Methods, Digital Computation Elements - The patented layout of learning machines is more than Qualcomm, Google and Samsung.

It is worth mentioning that the neuromorphic chip involves digital storage memory - components that mimic nerve cells (such as memristors). Technology has the most patents, followed by Samsung, the French Alternative Energies and Atomic Energy Commission, Seoul, South Korea. According to statistics from the University, there are currently no U.S. patents for artificial intelligence computing chips in Taiwan.

In addition, the neuromorphic chip developed by IBM was submitted to Samsung for development. The cooperation between the two parties started from the R&D stage, thus resulting in Samsung’s superiority in the development of neuromorphic chips compared to Taiwanese manufacturers. It is suggested that Taiwan’s semiconductor industry Manufacturers should invest in R&D as soon as possible to shorten the gap between the development technology of Korean factories.

2016 GoodChinaBrand | ICP: 12011751 | China Exports