Giant Intel Evolution: How To Promote AI New Track?

Intel, which focuses on data, is speeding up the landing of artificial intelligence. It has been integrated into Intel, intelligent driving, intelligent driving and other industries, such as intelligent driving, intelligent driving, etc.

Recently, Intel has launched a series of AI software and hardware products, including Intel's third generation Xeon scalable processor integrated with AI acceleration, and Intel's first artificial intelligence optimized FPGA Stratix 10 At the same time, the two sides announced that they would cooperate with robillyer, one of the largest taxi operators in Southeast Asia, and robileyi, one of the largest transport operators in Southeast Asia.

There is no doubt that artificial intelligence will accelerate into work and life. Since the beginning of this year, the epidemic situation of new crown pneumonia has become a common challenge facing the whole world. There is "opportunity" in "danger". From the front line of medical assistance and life science, to the orderly operation of economy and society, to public service and policy-making, intelligent technology represented by artificial intelligence plays an increasingly important role.

From the acquisition of Intel's AI, it is necessary to build a new brand of AI in all directions. Of course, Intel and Intel will face the competition from Intel and Intel.

Building AI Ecology: from XPU to quantum computing

In the process of expanding from PC track to data track, Intel reorganized six core technology capabilities to deal with intelligent business. Intel takes six technical pillars (process and packaging, XPU architecture, memory and storage, interconnection, security, software) as the engine, and integrates artificial intelligence into it. With the comprehensive product layout of CPU + GPU + FPGA + ASIC, Intel realizes intelligent deployment from cloud to end by combining software and hardware.

On this basis, Intel emphasizes its cloud, edge and end full stack advantages and "XPU" capability. In the field of artificial intelligence, "full stack" has become the focus of Intel, Huawei and other big companies, and everyone has great ambition.

Song Jiqiang, President of Intel China Research Institute, told the 21st century economic report: "X represents a lot of processing architectures, because the new market opportunities bring different types of data, different types of sensors, and different ways to obtain data. This leads to the fact that these data can not be processed with the same architecture, such as CPU or CPU + GPU, but must be heterogeneous The way. "

Although Intel can produce customized chips, it can be realized through the combination of heterogeneous chips. Song Jiqiang introduced, for example, Intel's movidius specializes in AI matrix operations, Habana Labs is specialized in accelerating matrix operation, and the flexible architecture of FPGA is used to accelerate sparse data processing. At the forefront, there are neural mimicry computing and quantum computing. "How to provide higher level and high parallel computing capability through quantum computing? It will also have a strong acceleration effect on AI. Intel is also doing research and development in this area."

Talking about the whole stack, song Jiqiang said: "many AI companies do algorithms and frameworks, and some even do some customized optimization in terms of hardware. Some have special accelerators for hardware. Intel is not only a full stack, but also has the support of the above framework and the underlying performance library. We have a wide range of hardware types for AI acceleration. We not only have the only DL boost technology that can embed special AI acceleration into the CPU, but also have chip hardware to support GPU, FPGA and special ASIC

Yao Jiayang, a consulting analyst with Jibang, said to the 21st century economic report: "in the data center or cloud business, Intel's processor business will not be easily shaken. This can be seen from the revenue performance of Intel's DCG (data center business). However, Intel will still have to compete with FPGA in terms of speed-up.

AI platform period: "weak AI is not weak, strong AI is not strong"

At present, Intel has tried AI applications in many industries. For example, Intel, Nanjing Economic and Technological Development Zone and a number of ecological partners built the "future science and technology smart center" in Nanjing, through the co construction of 5g + smart park to cultivate innovation ecology; in the field of industrial Internet, Intel Dalian factory needs to test the wafer in real time and accurately, so as to ensure the product yield. Through the artificial intelligence software and hardware based on Intel Compared with the manual detection method, the detection efficiency is increased by 100 times. In addition, Intel's chips are also used in ports and ships, airports, railways and vehicles, fleet management, road infrastructure and other fields.

However, there are still many problems in the process of AI landing. Zhang Yu, chief technical officer and chief engineer of the Internet of things division of Intel China, told 21st century economic news: "in vertical industries, such as retail industry, industrial manufacturing, intelligent transportation, medical treatment, etc., we can see that the problems faced by various industries are different. It needs different artificial intelligence To solve specific problems, that is, the fragmentation of various industries is very obvious. This requires us to have customized development for different industries when developing artificial intelligence, which increases the cost of artificial intelligence application development. How to reduce the cost is an important problem that we see in various industries.

The difficulty lies in the fact that we collect more than 4000 people and objects in the field of data collection, such as more than 4000 people and objects. But in some specific industries, the data collected or processed is often customized. This requires us to be able to quickly generate some new models that are suitable for specific fields and applications in a limited data set

In summary, how to reduce the cost of using AI function and get training results with a small number of data models is a difficult problem in practice. To solve these problems, we need to upgrade the application of AI technology.

From the perspective of the overall development of AI, song Jiqiang believes that deep learning has entered the platform stage from the development curve. This means that AI will also enter the platform stage.

Song Jiqiang explained: "artificial intelligence is not weak in the field of weak artificial intelligence. When you narrow down the application field of artificial intelligence, you can find a way to do it well. In general, if it doesn't have strong ability to resist attacks in the academic field, then it doesn't have a strong ability to move across the industry. According to the current development trend of the whole academic or industrial circles, artificial intelligence is moving from stage 2.0 to stage 3.0. Stage 2.0 is data-driven. It trains deep learning models through a large number of annotated data to help us complete tasks in specific fields. "

In his opinion, in order to develop AI in a more comprehensive way by moving to 3.0, firstly, it should be able to explain; secondly, it should be able to improve the ability of AI to use a small amount of data for continuous learning.

?