Between the Romanesque architecture and the technological AI scene, Intel’s AI helmer Naveen Rao talked about Intel’s artificial intelligence software and hardware combination, and the most important information was the Nervana neural network chip's announcement, according to the plan. , Intel's latest AI chip Nervana NNP L-1000, will be officially introduced to the market in 2019, this is Intel's first commercial neural network processor products.
Two years ago, Naveen Rao was the chief executive and co-founder of deep learning startup Nervana Systems. After the company was acquired by Intel, Nervana became the core battleship of Intel artificial intelligence, Nervana NNP series came into being, Naveen Rao Appointed as head of the artificial intelligence product division.
Carey Kloss, vice president of Intel’s artificial intelligence product group, and an Nervana team member during an exclusive interview with a 21st Century Business Herald reporter, said: “We started researching and developing Lake Crest (Nervana NNP series initial chip code) at the beginning of our business. At that time, our entire team was about 45. People, are building the largest Die (silicon chip), we have developed Neon (deep learning software), also built a cloud stack, these are done by small teams. But this is also the challenge, small team growth will have pain , It took us a long time to get the first products out. Nervana was founded in 2014. Until last year the chip was really introduced.
However, after joining Intel, Nervana can use Intel's various resources, 'Of course, calling resources is not an easy task, but Intel has extensive experience in the marketization of products. At the same time, Intel has so far I have seen The best post-silicon bring-up and architecture analysis. Carey Kloss told the 21st Century Business Herald reporter, 'We have hundreds of systems running at the same time in the production chip, Nervana's employees and 6 months The members who had just joined were also working day and night to work together for new products. 'In his view, Nervana is now at a reasonable pace and has all the elements for success next year.
In addition to Nervana, Intel’s acquisition of artificial intelligence flagship companies also includes Movidius, FPGA (Field Programmable Gate Array) giant Altera, SmartDrive related Mobileeye, etc. that focus on visual processing. In fact, since 2011, Intel has been investing continuously. AI-related companies, including China's Cambrian, Horizon. At the same time, Intel's competitors are also growing. Nvidia's GPU in the field of artificial intelligence has made great strides; Google recently released the third generation of AI chips TPU, the chip optimized for Google's deep learning architecture TensorFlow, and Google provides developers with TPU and other underlying services; last year, Baidu and ARM, Ziguang Zhuangrui and Hanfeng Electronics released DuerOS smart chip, mainly to provide voice interaction solution Programs; Facebook and Alibaba have also entered the chip field, among them, Alibaba Dharma is developing a neural network chip called Ali-NPU, which is mainly used for images, video recognition and cloud computing scenarios.
How will Intel respond to the 'experience' of this artificial intelligence chip?
Three factions fighting for hegemony On the whole, the current global artificial intelligence landscape is not yet clear, and it belongs to the local warfare for technological exploration. It has not entered the overall battle for crowds. Artificial intelligence is a general concept. The specific application scenarios vary widely. The emphasis is different. If classified according to technology and business genre, global companies can be divided into three factions. The first is system application school. The most typical representatives are Google and Facebook. They not only develop a system-level framework for artificial intelligence. , such as Google's famous artificial intelligence framework Tensorflow, Facebook's Pytorch, but also put into large-scale application. For example, Google spends heavily on research and development of automated driving, launching translation and other 2C services. Facebook also applies artificial intelligence technology widely used in social networks. Image processing, natural language processing and many other fields.
The second category is the chip school. Currently, it mainly provides computing support. The biggest players are Intel and Nvidia. Nvidia's GPU captures the critical timing of computing device demand, and the computing performance in graphics rendering, artificial intelligence and blockchain fields. Prominently, it also puts pressure on Intel in these businesses. At the same time, Nvidia seems to be different from Intel's 'Intel Inside'. It also hopes to become a real computing platform and has successfully launched its own CUDA platform.
On May 30th, Nvidia released the world’s first computing platform that combines artificial intelligence and high-performance computing—HGX-2. This is also the largest GPU—the computing platform behind DGX-2. As a traditional computing field. Intel, the boss, is naturally not to be outdone. The 50-year-old company is full of old meanings. In recent years, it has repeatedly launched heavy-handed mergers and acquisitions in the field of artificial intelligence: In 2015, it acquired the Field Programmable Gate Array (FPGA), a field programmable gate array (FPGA) giant. Altera, lays the groundwork for the future development trend of computing power. FPGA has great potential in cloud computing, internet of things, and edge computing. In 2016, Intel acquired Nervana and plans to use the company’s ability to engage in deep learning to combat it. GPU; also acquired visual processing chip startup Movidius in the same year; in 2017, Intel acquired Israel to help drive the company Mobileye with 15.3 billion US dollars, designed to enter the field of automated driving.
In addition to the system application school and the chip school, the third category is the technology application school, and most of the remaining companies belong to this category. Although different companies all claim to be studying in depth, the field of artificial intelligence has deep or even unique Accumulation of technology, but in fact most of them are based on system application and chip technology platform. Only technical applications send more C-end users, including autopilot, image recognition, enterprise-level applications, etc. Objectively speaking, technology applications The faction belongs to 'the gentleman is good at things and things are also'.
Judging from the current competitive landscape, the System Application School has gradually taken up its overall advantage and has the most core competitiveness in the field of artificial intelligence. In the era of traditional computers and mobile phones, systems and chips are more cooperative, and chips are even more Take a dominant position. Specifically, for example, in the computer market, Intel completely dominates in the field of computing power and spans PCs and Apple’s MAC machines. On the system side, Windows and iOS each have their own merits and cannot replace each other, but their common Intel But it cannot be replaced. In the era of mobile phones, although the protagonist of computing has changed from Intel to Qualcomm, the chip is still at the core, and its importance and operating system are equally divided.
In the last 1-2 years, the situation has changed rapidly. Apple has released its own tactics to develop and produce MAC chips. Intel’s share price has been falling for a while. In the field of artificial intelligence, this trend is even more pronounced due to the different requirements of computing scenarios. Big, Google needs to develop mature chips according to their needs, and it is technically more feasible. If Intel wants to customize chips for different scenarios, it means that Intel will fully transfer to 2B, compared with the previous 2B2C model, pure 2B's business will obviously be more like that of Party B, and the complexity of the business line will increase dramatically. Historically, a company's shift from 2C to 2B is generally due to the loss of its core dominance in the industry and has to retreat. Seeking times.
Betting on Nervana NNP So, in the fierce competition, how does Intel further increase its chip business?
After Naveen Rao joined Intel, he became Intel's vice president and head of the AI Business Unit (AIPG) and led the launch of the Intel Nervana NNP series of chips. This time at AIDC, he proposed to provide developers with software tools. , Ecology. In the industry view, with Intel's technological capabilities, software tools and hardware are not a problem, but the ecology is yet to be discussed. In the PC era, the core of the ecology is the chip, so around Intel’s chip-building ecosystem, Intel can be entrenched, but In the age of artificial intelligence, artificial intelligence systems are the core of the ecosystem. Providing chips for computing power is part of the ecosystem. CPUs can provide computing power. GPUs can provide them. Intel can produce them. Nvidia can also produce it. Even Google, Apple itself. Can produce. Currently in the field of data science and deep learning computing, Intel's chip layout mainly Xeon (Xeon) chip series, Movidius' vision chip VPU, Nervana NNP series, and FPGA (field programmable gate array). The product lines correspond to several different subdivided application scenarios.
Nervana NNP series is a neural network processor. In the deep learning training and inference phase, Nervana NNP is mainly for the calculation of the training phase. According to Intel's plan, deep learning (hereinafter referred to as 'DL') will be implemented by 2020. ) The effect is 100 times better. This neural network processor was designed by Intel and Facebook together. It can be predicted that this chip will have a great support for Facebook's machine learning framework Pytorch. After all, Facebook's Pytorch ambitions. Certainly it is to compete with Google’s Tensorflow. However, the latest chips will be officially launched in 2019. How will the pattern of deep learning change unpredictably?
Naveen Rao wrote in his blog: 'We are developing the first commercial neural network processor product Intel Nervana NNP-L1000 (codenamed Spring Crest), planned to be released in 2019. Compared to the first generation of Lake Crest products, we Intel Nervana NNP-L1000 is expected to achieve 3-4 times the training performance. Intel Nervana NNP-L1000 will also support bfloat16, which is widely used in the industry for a numerical data format for neural networks. In the future, Intel will be in artificial intelligence. The product line expands support for bfloat16, including Intel Xeon processors and Intel FPGAs. 'In fact, the rumors of the Spring Crest launch at the end of 2018 already existed, but at present, the official announcement of this point in time in 2019 There is a delay. In response, Carey Kloss explained to reporters: 'Into more modern process nodes, we integrate more Die (silicon chips), can get faster processing speed. But it takes a certain amount of time to manufacture silicon wafers. It also takes time for the silicon chip to become a new neural network processor, which is the reason for the delay.
For the difference between the two generations of chips, he analyzed that: 'Lake Crest as the first generation of processors has achieved very good computational utilization on GEMM (matrix operations) and convolutional nerves. This is not just about 96% throughput. The amount of utilization, but in the absence of full customization, we have also achieved a computational utilization rate of more than 80% for GEMM in most cases. When we develop next-generation chips, if we can maintain high computing utilization Rate, the new product has a performance improvement of 3 to 4 times. '
Talking about competition, Carey Kloss said: 'I don't know what our competitors' roadmap is, but our response is relatively fast, so I don't think we will be at a disadvantage in neural network processing. For example, bfloat16 has been around for a while. Recently, it has become more popular. Many customers have asked for support for bfloat16. We have also gradually turned to support bfloat16. 'Comparing to Google's TPU, he thinks that the second generation of TPU is similar to Lake Crest. The third generation of TPU is similar to Spring. Crest.
Attack on all sides In addition to the highly regarded Nervana NNP, Intel's Xeon chips are targeted at servers and large computing devices. For example, China's supercomputers Tianhe No. 1 and No. 2 use Intel Xeon six-core processors.
In terms of visual chips, Intel's business volume has grown rapidly. Movidius VPU chips have long been used in the emerging hardware markets such as cars, drones, such as DJI UAVs, Tesla, and Google Clips cameras using Movidius. The visual chip.
Gary Brown, market leader at Movidius, told the 21st Century Business Herald reporter: 'At Movidius, the chip we developed was called the visual processing unit VPU. VPU is a chip that combines both computer vision and a smart camera processor. So our chip There are three types of processing done: ISP processing, which is image signal processing, processing based on camera capture technology, and computer vision and deep learning.
He cited examples of specific use scenarios including VR products and robotics, smart homes, industrial cameras, AI cameras, and surveillance and security. Among them, 'monitoring and security is a huge market, especially in China, surveillance and security cameras The market is particularly large, and some large companies are developing surveillance cameras such as Hikvision and Dahua.
Gary Brown also mentioned that the smart home sector is currently growing rapidly. Although the market is small, it is developing rapidly. 'There are many companies developing smart devices such as smart home security, personal home help, smart doorbells, and apartment and home visits. Control. But in the home field, low cost, low energy consumption, long battery life, and very high precision are very challenging. Because outdoor shade, for example, moves, it may trigger a burglar alarm and is therefore very low. The false alarm rate is very important and should have good accuracy.
One of the company's challenges is how to continue to create high-performance chips, 'We have some strategies, such as using a front-end algorithm to reduce power consumption, so that we can turn off most of the chips, only a small part of the optimized face detection function When a face comes up, other chips will be activated. This will keep the facial monitoring system on at all times. We also have a lot of energy-saving algorithms to make the home smart camera last for about 6 months.' Gary Brown explained.
In addition, Altera is in charge of this line of FPGA. With the arrival of the 5G wave, the data analysis and computing requirements of the IoT Internet of Things will surge. The number of access nodes of the Internet of Things is at least tens of billions of scale, which is larger than the mobile phone scale. To be 1-2 orders of magnitude higher. The typical requirement of the Internet of Things is the need for flexible use of algorithmic changes. This is the strength of FPGAs. FPGAs can adapt to the needs of customized computing scenarios through their own structural changes, which also makes Intel in the future. It has become possible to provide efficient chips for more different types of devices. From the $16.7 billion acquisition amount, it can be seen that Intel’s purchase is obviously not just immediate value.
Fast-breaking enterprise-level scenarios According to a recent Intel survey, over 50% of U.S. enterprise customers are turning to existing cloud solutions based on Intel Xeon processors to meet their initial demand for artificial intelligence. Multiple Intel executives are In an interview, they all told reporters that there is no single solution for all artificial intelligence scenarios. Intel will match technologies and businesses according to customer needs. For example, Intel will configure Xeon and FPGA, or Xeon and Movidius, together. In order to achieve higher performance artificial intelligence function.
For Intel, these enhanced artificial intelligence capabilities will be widely used in enterprise-level scenarios. Naveen Rao said: 'We need to provide a comprehensive enterprise-class solution to accelerate the transition to artificial computing-driven computing in the future. This means that our solution provides the widest range of computing power and can support multiple architectures from milliwatts to kilowatts.
Carey Kloss further explained to the 21st Century Business Herald reporter the application scenario of the artificial intelligence chip: 'Spring Crest can be said to be the highest level of Nervana's neuron processor architecture. So its customers include super-large-scale computing centers and already have quite powerful data. Large companies with scientific work, governments, etc. If you need low-end and small models, Xeon can help you, it can open the data from the cloud to the end.
Specifically, Intel has also explored scenarios such as medical care, driverlessness, new retail, and the Internet of Things. For example, in the medical field, according to reports, Intel is working with Novartis to use deep neural networks to accelerate high content. Screening - This is a key element in early drug development. The cooperation between the two sides reduced the time needed to train the image analysis model from 11 hours to 31 minutes - efficiency increased by more than 20 times.
In terms of unattended stores, Intel provided 'computing brains' to Jingdong unmanned convenience stores, and it has been deployed and deployed in multiple smart stores (Sinopec Express Store, Jingdong Home) and smart vending machines. Algorithmically, JD.com stated that the machine learning algorithms used by unmanned shops are mainly focused on knowledge, knowledge, and knowledge. In three directions, the need to use unstructured data, such as video, is converted into structural data due to the need to open and close data online and offline. The popular CNN (convolutional neural network) algorithm in the field of machine vision, the traditional machine learning algorithms used in the intelligent supply chain, such as SVM, linear regression of statistics, logistic regression, etc. In the case of better network conditions Most of the video data can be completed in the cloud using a larger model. In the case of a poor network, through edge computing such as mobile, edge computing is done using a small network. The hardware used includes Intel's edge servers.
Despite Intel's strong enemies, the transition was very firm. From the perspective of R&D value, according to IC Insights statistics, the total R&D expenditure of the top 10 semiconductor manufacturers in 2017 was US$35.9 billion. Intel ranked first. According to the report, Intel’s R&D spending in 2017 was US$13.1 billion, accounting for 36% of the Group’s total expenditure, which is approximately one-fifth of Intel’s sales in 2017. With huge investments from various companies, the battle of AI chips will also continue. Intensified.