Reporter Ni Yuqing Shanghai Report
[Introduction] The release of the AI strategy is a new upgrade for Huawei. It has entered the intelligent stage. This not only expands the boundaries of Huawei, but also means that Huawei and the international AI giant compete with each other, except for the basic hardware and software business. In addition, there is an opportunity to surpass in more areas in the future.
From October 10th to 12th, the third HUAWEI CONNECT 2018 (Huawei Full Connection Conference) arrived as scheduled. This is Huawei's most important developer conference and the window for Huawei to display the ecosystem. Like 2017, ICT (Information and Communication Technology) is still the key word, which is Huawei's birthplace and the frontier of touching other fields. However, unlike the previous conference, AI became the protagonist of this year.
Xu Zhijun, the chairman of Huawei's rotating company, systematically expounded Huawei's AI strategy and full-stack full-site AI solution, including the artificial intelligence Ascend series chips.
The concept of a full-stack solution is enormous. On a global scale, there are no companies that can provide full-stack solutions, especially in Europe and the United States, often emphasizing their expertise in one area. Therefore, Huawei's "full stack" is more unique.
During the exchanges during the meeting, Huawei executives repeatedly told reporters that the “full stack” was proposed for multiple scenarios. From the cloud to all types of terminal devices, such as cars, watches, mobile phones, etc., AI needs to be assisted, and the scenes must be related to each other. Partners need equipment and need computing power.
Xu Wenwei, director of Huawei's director and strategic marketing, said in an interview with the 21st Century Business Herald: "The future is definitely a smart world. It turns out that Huawei is doing infrastructure, and then doing cloud is actually doing services. This is also the trend. Although there is a bubble in artificial intelligence, the big trend is unchanged. The primary artificial intelligence has improved efficiency, just as Ren Zong (Huawei founder Ren Zhengfei) said, in the case of roughly correct direction, a shot must be fired first. ”
So what does this gun mean for Huawei? Looking back at Huawei's development process, from the initial B-side server and switch business, to the C-side mobile phones, notebooks and other terminal products. In this process, Huawei began to approach consumers and open up a broader space. Huawei's image is more comprehensive. Nowadays, the release of the AI strategy is a new upgrade for Huawei. It has entered the intelligent stage, which not only expands Huawei's borders, but also means that Huawei and international AI giants compete with each other, except for the basic hardware and software business. In the future, there will be opportunities to surpass in more fields.
Decrypt AI chip
According to Xu Zhijun, Huawei's full stack is a technical function perspective, which refers to a full stack solution including chip, chip enable, training and reasoning framework and application enablement. The solution consists of four parts. One is the chip-level Ascend, which is a series of AI IP and chips based on a unified and scalable architecture, including Max, Mini, Lite, Tiny and Nano. During the conference, the Shengteng 910 (max) and the Shengteng 310 (mini) were released. The Shengteng 910 will be launched next year, and the Shengteng 310 has been mass-produced.
The second is CANN, the chip operator library and highly automated operator development tools, which can improve development efficiency; the third is MindSpore, which supports end, edge, cloud independent and collaborative unified training and reasoning framework; Yes, provide full process services (ModelArts), layered APIs and pre-integration solutions.
Among these four parts, the AI chip and the Mindspore framework are the most concerned, both of which represent computing power and algorithms, and are two core elements of artificial intelligence. First look at the AI chip, the rising 310 is more used in edge computing products, the rising 910 is mainly used in the cloud to provide training capabilities. Among the data provided by Huawei, the performance of the Shengteng 910 is twice as high as that of the NVIDIA V100. Huawei calls it “the AI chip with the highest density calculated by the global single chip.”
For Huawei's chip performance, Yao Jiayang, an analyst at Jibang Consulting Tuoba Industrial Research Institute, said in an interview with the 21st Century Business Herald: "In the case of the Shengteng 910, the computing performance of the FP16 has surpassed that of the Tesla V100. It is not difficult to imagine the power of other computing precision, it should also have more performance than Tesla V100. The 910 adopts the 7nm process, which will help the performance of the chip to a certain extent. Therefore, in terms of performance advantages, it can really suppress the Tesla V100 using the 12nm process. At present, it can be confirmed that these two products will optimize the competitive advantage of Huawei's terminal products. ”
However, an AI chip expert analyzed the 21st Century Business Herald reporter: “(Unit Computational Density) is just one of these indicators and does not represent the final actual performance. Computational density is primarily related to manufacturing process nodes and architecture. For example, if you use 7nm, it is definitely higher than 10nm because the transistor is small. In addition, if all the chips are stacked on the chip, there is no complicated data link and a large number of other parts such as on-chip memory cells, and the area calculated for calculation is large, and the natural density is high. ”
Behind the two AI chips, Huawei's DaVinci project was announced for the first time. In this project, Huawei also researched the DaVinci architecture. The AI chip is based on the unified DaVinci architecture. A Da Vinci project's R&D staff told 21st Century Business Herald: "This project was officially launched last year, but it has been designed for several years, and is led by Huawei's internal experts. In order to fully research the chip, it was developed overtime last year. ”
Previously, some of Huawei's Kirin chips were based on the Cambrian structure. After that, they should all use the self-research structure. They will not use the Cambrian, and the Kirin will change. ” A Huawei insider told reporters.
Xu Zhijun said at the media meeting: “The new architecture is based on our understanding and demand for artificial intelligence. The Cambrian is also very good, but there is no way to support our full scene. We need cloud to the edge to the end, as well as IoT terminals, the needs of the full scene, need to create an architecture. The market has not yet seen the corresponding architecture, you can do this. ”
In fact, Huawei has been making chips since more than 20 years ago. In Xu Wenwei's view, Huawei's chip history has four stages. “The most difficult thing is to decide whether to do it or not, or to do it technically. of. ”
He recalled to reporters that in 1991, Huawei made chips, which required high integration, which was the first stage; the second stage was because ordinary commercial chips could not meet business needs, but commercial chips were either expensive or bought. No, so you need to do it yourself; the third stage is based on business judgment. Huawei started to make 3G mobile phone chips in 2005. At that time, Huawei did not have a mobile phone business, only data cards, but the data cards were in demand at the time. Trend, so I decided to do it.
“Now the unicorn is determined to do it. It can be said that it is based on business decisions, but it is actually a vision drive. Now it can be said that it is the fourth stage, that is, artificial intelligence chips. There are not many companies on the market to buy, and we are cloud, Artificial intelligence chips are required for edges, ends, pipes, and the like. We need artificial intelligence solutions to innovate on the basis of scenarios. Our chips are full-scenario coverage, scalable, high performance, and low power consumption in some application scenarios. The main purpose is to reduce costs and increase integration. & rdquo; Xu Wenwei said.
AI framework challenge
Beyond the chip, Huawei's artificial intelligence framework Mindspore is also attracting attention.
As we all know, AI is not a single technical system, but the intelligent integration of many systems. The AI system mainly includes the following aspects: one is the business application layer, the packaged third-party application, or the customized application and technical solution; the second is the intelligent framework layer, and the deep learning framework intelligently processes the data based on the cloud computing layer. It is the brain of the AI system; the third is the cloud computing layer, which provides computing power, and handles various traditional forms of data on the basis of the big data layer; the fourth is the data acquisition of the big data layer, including the physical world and the online world. Storage, such as production data, financial data, social life data, autopilot road data, etc.
At present, most domestic enterprises are at the application level, and a small number of giants have made achievements in the cloud, such as Alibaba Cloud and Huawei Cloud. For the computing power of artificial intelligence, there are very few top players in China. The giants include Huawei, Ali’s Pingtou and Dharma. The vertical field has Cambrian and horizon.
An artificial intelligence industry insider told the 21st Century Business Herald: "The chip is the blood of artificial intelligence, but the smart frame is the brain and soul of artificial intelligence. The intelligent framework is the operating system in the field of artificial intelligence. The chip can still be foundry, but the system represents the ecology and is the key to the entire intelligent world. ”
There are many vendors and technology genres in the chip, but there are only a handful of operating systems in the world. The ecosystem built on the operating system determines the life rhythm of all hardware and software. At present, the top-level artificial intelligence frameworks come from the United States. The most famous ones are Google's TensorFlow, Facebook's pytorch, Amazon's Mxnet, and even Microsoft has developed the smart framework CNTK. It can be said that almost top technology giants are competing for champions in this field. .
The artificial intelligence framework and platform touched almost no domestic enterprises. Previously, only Baidu's Paddle-Paddle and Ali's PAI, but overall it was not strong enough. Today, Huawei's Mindspore is a new artificial intelligence framework, and there is no open source yet. .
“The whole strategy, in the full stack, full scene solution, the biggest challenge is our Mindspore as a computing and reasoning framework, whether we can build an ecology in the future, whether you will use it or not, this is the only challenge. "" Xu Zhijun told reporters, "We analyzed that now all the frameworks are difficult to adapt to the full scene, for example, we will have a learning framework deployed on the side, this is not before. For example, Google mainly promotes artificial intelligence on the cloud. He trains and reasoning in the cloud, but this framework cannot be used on the end side of mobile phones. You have to do training on the end side, do reasoning, the framework is small, the architecture is flexible, and now Tensorflow can't fit in. At the edge, in the future network, there is a need to have a framework for learning and reasoning. We can't find a framework that really implements the full scene, so we've done one ourselves to meet my own needs. ”
What is more cruel than the Internet era operating system is that the giant's smart framework is basically an open source model. In the PC and Internet era, Microsoft and Apple are relatively closed operating systems. Only Android is open source, which makes other countries that do not have operating system technology have the motivation to develop operating systems. But in the era of artificial intelligence, Google, Facebook, etc. have opened up their top-level deep learning framework, and freely give computing power, build a community for free, and even spend a lot of effort to train developers. The fundamental purpose is to build a very high migration. Cost, if the top developers are attracted to Google, the Facebook camp, it is very difficult to let them learn the domestic smart framework.
The same is true for the company. If there is a ready-made free open source technology solution, it takes great strength and courage to invest huge amounts of manpower to develop the underlying technology. At present, China can only expect that giants such as Huawei and BAT can have such a determination to regain a game on the artificial intelligence operating system after losing the PC and the Internet operating system.
Ecology and borders
There are three kinds of ecology in the computer world: one is the PC ecology led by Intel and Microsoft. It is characterized by improving hardware performance, using Moore's Law to continuously upgrade firmware for the baton to gain market growth. Second, the Internet ecology represented by Google and Apple. It mainly provides infrastructure to attract developers to enter and obtain user value through Internet services. The third is artificial intelligence ecology. This ecological pattern is undetermined, and each family is doing early-stage technology investment. At present, Google's artificial intelligence project is the most comprehensive and powerful, and its deep nerves have drawn a lot of developers.
Now, the artificial intelligence industry is in the stage of building an ecological environment, and the business form is the second. It is the first place to attract more developers and upstream and downstream manufacturers. Huawei also built its own technology ecology with high-profile, and held a full-connect conference to improve its voice in the AI field. On October 11, Huawei announced that it plans to develop 1 million developers and partners within three years, and also promote the establishment of the Global Industry Organization Summit (GIO), which aims to promote exchanges and collaboration among industry organizations, including 3GPP, 5GAA, 16 industrial organizations such as 5GSA. In the application case, Huawei Airport, China Merchants Bank, Audi, etc. all applied Huawei's AI technology.
Yao Jiayang told reporters: "In terms of Huawei's approach, Huawei seems to be interested in building its own AI ecosystem. However, in the past, Huawei itself was also a server supplier, such as Alibaba, Baidu, etc., can be regarded as Huawei. Customer base, now Huawei provides the Mindspore framework, which will form a competitive relationship for CSP (cloud service provider). How does Huawei maintain the relationship with the domestic BAT leaders? The cooperation is greater than the competition & rsquo; Will be the next observation indicator. After all, BAT used to invest resources and build its own ecosystem in AI. BAT should not allow other competitors to influence their AI ecosystem. ”
For the challenge of building an ecology, Xu Wenwei told the 21st Century Business Herald: “The first thing to be a platform, and as a platform must be technically leading. The second point must be open and the third must be fair and equitable. In other words, when working with partners and building an ecosystem, the core is that the business must have boundaries. ”
It can be seen that Huawei's AI strategy has covered all aspects from cloud to end, from algorithm to computing power, from training to reasoning. This also raised questions about Huawei's borders. Xu Wenwei said that Huawei will continue to adhere to the "do not touch the application, do not touch the data", the future industry is a competitive relationship, not vicious competition, the most important thing is to make the industry bigger.