How can Intel turn the tide by AI when it misses mobile?

The former chip giants are embracing the transformation of AI in an all-round way. However, under the pressure of Microsoft, an old rival, NVIDIA's interception, and Google and other Internet upstarts' pursuit, how can Intel use AI to illuminate the future?

[Video can be viewed directly with traffic, time: 14min 48s]

[The flow limit can be viewed in pictures and texts. 2880 words take about 10 minutes]

◆ Which is more suitable for AI, CPU or GPU?

To answer this question, we should start with the architecture of CPU and GPU. This is a very complex and academic problem. Despite some technical hurdles, this picture can give us a more intuitive understanding of the difference between CPU and GPU.

Architecture diagram of CPU and GPU (picture from electronic engineering)

In the picture, the green is the computing unit, the orange red is the storage unit, and the orange yellow is the control unit. GPU uses a large number of computing units and super long pipelines, but only very simple control logic. CPU has complex control logic and many optimization circuits, but its computing power only accounts for a small part. Therefore, GPU is better at large-scale concurrent computing, and is suitable for computing intensive programs with high data parallelism (the same algorithm is repeated for different data). However, CPU has low requirements for parallelism and low dependence on data, and is suitable for processing various tasks.

In 2016, when the AI concept came to the public again, the popularity of Alphago quickly heated up the concept of deep learning. For deep learning that requires parallel computing and processing of a large amount of data, the advantages of GPU were directly amplified. However, from the perspective of the complete path of AI, CPU and GPU must perform their own duties and cooperate with each other.

For a complete AI application, the CPU is good at obtaining data in the early stage and making final information decisions, while the GPU is better at doing a lot of data calculation in the middle. "For example, for face recognition applications, the early image acquisition and data storage run best on the CPU, and then the GPU is responsible for a large number of parallel recognition calculations. When the final recognition results need to be associated with other information to make decisions, or do information integration with other sensors and channels, the CPU is better at it." Song Jiqiang, president of Intel China Research Institute, gave an example when receiving ZOL video interview.

Song Jiqiang, President of Intel China Research Institute

However, GPU is not the only platform for parallel computing. If there is no clear data algorithm and a large number of experiments are needed to determine which algorithm is more suitable, the GPU will be more efficient and flexible. however Once the algorithm is fixed, the work of GPU can be replaced by an acceleration chip specially optimized for AI.

At present, Intel has two kinds of hardware acceleration chips, FPGA and Nevana, which do not need CPU instruction scheduling, and can be directly optimized for hardware and algorithms with higher execution efficiency. This kind of special acceleration chip can solve the problems of computing density, memory storage bandwidth, and optimize the performance to the best. "The front-end equipment should consider the size, power consumption, price and other issues, and the special acceleration chip is the best choice. The acceleration chip is particularly suitable for computer vision applications in front-end equipment, small-scale deep learning network, so it will appear on UAVs, artificial intelligence cameras." said President Song Jiqiang.

◆ About three future acquisitions

Intel has reduced the impact of parallel computing with special acceleration chips, which is inseparable from three key acquisitions.

Intel acquired Altera for $16.7 billion

On June 2, 2015, Intel announced that $16.7 billion cash acquisition of Altera, a programmable logic chip giant This transaction has become the largest acquisition in Intel's 49 year history CEO Ke Zaiqi exclaimed that "the amount of this M&A is too large". Indeed, at that time, Intel's quarterly revenue just reached 13.2 billion dollars. Now it seems that Intel was right to break the ice at that time. After the Altera FPGA was included in Intel's product line, it became a crucial weight for Intel to develop special acceleration chips. Wendell Brooks, vice president of Intel M&A, said in an interview at that time that compared with traditional processors and FPGA independent components, the integrated chip packaged by FPGA and Intel processors will initially bring 30% - 50% performance improvement, while the final performance improvement will reach 2-3 times.

Intel acquires Nevana System for $400 million

Although FPGA can also complete high-performance parallel computing, it is not enough to make up for the lack of GPU, so Intel continues to buy. On August 10, 2016, Intel announced that US $400 million acquisition of deep learning startup Nevana System 。 Nevana is a company with independent intellectual property rights in the chip field. Its Engine chip can compete with GPU in terms of power consumption and performance. With the help of the Nevana acceleration chip, Intel has extended its processor business system and gained more voice in the field of artificial intelligence.

Intel acquired Mobileye for $15.3 billion

On March 13 this year, Intel announced that US $15.3 billion acquisition of Mobileye This is also Intel's second largest transaction in 49 years after the acquisition of Altera with US $16.7 billion 。 In 2016, Intel's net profit in the whole fiscal year was only $10.3 billion, and it took Intel almost one and a half years to complete the acquisition. Why? Mobileye holds a number of key technologies to realize automatic driving, such as cameras and sensor chips. Some analysts even said that it has occupied more than 70% of the global market share in the field of driver assistance and anti-collision systems. This acquisition will further strengthen Intel's position as a technology supplier in the field of autonomous vehicles. President Song disclosed: "After the acquisition, the customers of Mobileye have directly become the customers that Intel can support. Intel will be able to provide a complete vehicle solution in the field of automatic driving from algorithm to chip, from multisensor integration to storage and communication technology. The driverless technology is not a certain technology, but needs to integrate many technologies. Intel will be very good at creating an efficient and reliable driverless platform. "

◆ Artificial intelligence will lead to new hardware forms

In the past three decades, human-computer interaction technology has accounted for a large proportion from PC to mobile. In the era of artificial intelligence, machines will have better subjective initiative. They can sense tasks independently through sensor data and do not need to be initiated by people. However, President Song believes that AI technology has many limitations at present, and it can better guarantee user experience by combining with human-computer interaction. "AI can train a good model to achieve 99% reliability, but in the process of application, it may encounter new situations that have never been encountered in previous training, and the accuracy of AI may not be able to maintain that high. AI and human-computer interaction should be combined, so that even if AI predicts wrong, it will not cause errors. AI can further reduce the time for people to actively operate, but it can still interact with people to ensure the best effect. "

Siri Voice Assistant It is billed as an early AI model, but many users feel that they are not useful. In President Song's opinion, due to the limited equipment form, there is not much room for AI to play. "Many operations on mobile phones are actually not complicated, and human operation costs are not high and must be more accurate than machine operation, so you will not experience revolutionary and disruptive experiences. But if the product form is not mobile phones, but smart cars, drones, and intelligent robots, the value will be unprecedented." In the era of intelligent robots, You can buy a good intelligent robot at the price of a high-end smartphone. It will develop user profiles for all family members, understand the characteristics and needs of different people, and provide differentiated customized services. If the intelligent robot has a mature ability of space movement and autonomous operation, it is possible to achieve intelligent pension.

Song Jiqiang Dean talks about Intel's AI strategy

Today, Intel's AI solutions have quietly appeared in many application scenarios. On the Double 11 in 2016, during the peak visit period of JD users, Intel's artificial intelligence technology monitored the upload of illegal pictures in real time to ensure the operational security of JD during the peak traffic period, while helping JD analyze the passenger flow behavior in detail and provide more accurate recommendations; stay IFLYTEK In the hearing intelligent conference system, all voices can be converted into words in real time, with an accuracy rate of more than 95%. In addition to the core voice technology of iFLYTEK, it also relies on Intel's Xeon processor and real-time voice transcription solutions; By 2020, it will only take 24 hours to complete a series of precision medical processes from gene sequencing to medical analysis, and then to develop medical plans. The Intel GTX One Biocomputing Acceleration Platform hopes to rely on the computing power of AI to promote the rapid development of precision medical treatment

Next, we will Look To the technologies and solutions released by Intel AI in the field of virtual/augmented reality and driverless driving. In the VR/AR field, besides users can use their own hands to interact with the virtual environment, Intel AI technology will enable users to wear helmets and people who interact with them can still see their faces. Intel's unmanned driving technology will provide intelligent interactive experience in the cockpit. The driver can release his hand to do other things, but AI will monitor you, prevent you from wandering, and can pull you in to take over corresponding tasks in an emergency.

Once, Intel's first microprocessor became an important milestone in computer history. However, when mobile Internet gradually obscured the light of PC, Intel failed to get out of the fog, and the loss of mobile dividends gave Intel heroes the determination to cut their wrists. Today, Intel has removed the label of "chip giant" and replaced it with“ drive Intelligent interconnection of all things ". Looking at the key deployment areas of Intel in 2017, whether it is artificial intelligence, driverless, virtual reality, 5G, Intel hopes to turn the tide. Fortunately, it is not too late this time.

This article is an original article. If it is reproduced, please indicate the source: How can Intel turn the tide by AI when it misses mobile? //news.zol.com.cn/634/6349806.html