Shulou(Shulou.com)06/03 Report--

Last Friday, Huawei "see you on Friday" achieved the effect of brushing the screen.

In fact, our friends have been watching us, perhaps it is not difficult to speculate before that, Huawei such a solemn "Friday appointment" should be to release the location and specifications of the Penteng 910 chip that was disclosed at last year's HC conference.

Sure enough, Xu Zhijun, the rotating chairman of Huawei, released this AI chip, which belongs to the Ascend-max series and is claimed to be the most powerful on the surface, on Friday. For a while, the topic of AI and relationship chips returned to public view, boosting the morale of the society.

AI enthusiasts may no longer be strangers to the name Teng 910, so maybe our next question should be, what happens after Friday? In other words, everyone has learned from various information sources that the launch of the Teng 910 is Huawei's AI move, which is currently the most powerful AI chip in the world.

But you may wonder, what is the value of this big move to the industry, to you and to me? How will its capabilities be put into the broad market demand for enterprise intelligence?

So in the "post-Friday era", it may be necessary to get behind the Penton 910 and spy on the release of this domino, which could trigger a chain reaction in the AI and computing industry. It must be the public cloud market that bears the brunt.

After the landing of Teng 910, where is the wind of AI surging, and what ripples will occur in the cloud computing industry? This is a topic worth spending more time thinking about.

The change of Cloud and AI

Before we understand the industrial value of the Teng 910, we must review how the cloud computing industry has changed in recent years.

We know that from 2012 to 2018, the global demand for AI computing increased 300000 times. Today, with the development of AI computing power from scientific research and development to industrial applications, this curve is rising more straightly.

The third rise of AI represented by deep learning, the basic technical logic is to train the algorithm model with a certain goal, and then carry out reasoning deployment to achieve the desired coupling effect.

This means that the process of AI application includes two parts: training and reasoning. If enterprises want to acquire AI technology, they can either train themselves or use manufacturers to provide trained AI capabilities.

The training, development, capability invocation and scenario deployment of AI are all based on computing. And the AI task itself requires high throughput, high concurrency and high malleable computing power, so cloud computing is the main computing power solution with the highest efficiency and the closest fit mode of AI.

This brings a very powerful change to the cloud computing industry itself. In the past, the value of cloud computing is more to support the existing Internet tasks, but now cloud computing has become the main carrier to obtain AI capabilities and develop AI capabilities. From "support" to "acquisition", this has completely changed the value of the public cloud to customers and markets.

So we see that AI is becoming a major growth point for cloud computing vendors. The main trend in the global cloud computing market today is that traditional Big Brother AWS is slowing, falling below 40 per cent for the first time in the second quarter of this year. Microsoft Azure, which began to reorganize and integrate on a large scale last year, with a series of AI capabilities, AI tools, and deployment flexibility as the selling point, has suddenly emerged as the second most profitable business for Microsoft. Cloud + AI has become the second most profitable business. In order to continue to consolidate this advantage, Microsoft recently invested a billion dollars in OpenAI in order to continue to compete with Google in the future AI market.

Google Cloud, which is famous for its AI algorithm capabilities, ranks third with TPU and TensorFlow frameworks. Google expects to generate more than $8 billion in cloud computing revenue this year, which will be the most important growth point for Google's system.

In the domestic market, Baidu Cloud and Huawei Cloud, which are famous for AI technology, have become the two fastest rising clouds. It is not difficult to see that AI is becoming a well-deserved trump card in the global public cloud market. The strength of AI capabilities is also becoming the core competence standard for cloud growth.

So what is the core competence of Cloud + AI?

Judging from the development trajectory of cloud giants in Europe and the United States, special AI computing power, framework, capability and development ecology are the four connection points that constitute the combination of cloud service providers and enterprise AI requirements.

Among them, computing power and framework, which are the support of the whole system, are the basis for users to develop their own AI model, and the value is particularly important. For example, Google Cloud's rapid growth in the past two years is actually based on its core strategy of training AI chip TPU, and then bundling the development community and TensoFlow users into their own cloud services to achieve short-term rapid growth.

In order to deal with the AI developer wall set up by Google, Microsoft, Facebook and other giants have jointly opened up their development frameworks and released cross-framework tools such as ONNX.

Thus it can be seen that the special training of AI computing power and development framework is the core of building the future growth point of cloud computing enterprises. Of course, Huawei, which directly applies its value, has most directly benefited from Huawei's launch of Penteng 910 and its self-developed framework MindSpore. As other cloud computing vendors in the industry use Nvidia's board as a source of AI computing power, Huawei Cloud is about to become the only cloud service provider with its own training chip and framework besides Google Cloud.

This means that Huawei Cloud has joined Huawei Cloud in the cloud + AI international top-level competition between chips and frames. The application layout of Huawei Cloud in Teng Teng 310 and side scenarios even constitutes a more complete full-stack AI capability than Google Cloud. Behind this message, it indicates that the cloud computing industry, which has been surging now, will be even more restless.

After clarifying the industrial structure of cloud + AI, we need to look at the value of the combination of Teng Teng 910 and Huawei Cloud from a different perspective from the perspective of enterprise users.

Violent computing for AI training is an indispensable cornerstone of the industry.

For users of AI technology, AI is divided into two parts, known as training and reasoning. The relationship between the two is like cultivating a person. Training is like going to school and getting an education. Reasoning is starting to look for a job.

Everyone knows that it is easier to change jobs, and school education can be very expensive in the last few years. It is not only to hone students, but also to test parents and teachers. The same is true for AI, it is not difficult for enterprise users to apply AI capabilities, but it is difficult to train their own deep learning model and turn AI into their own exclusive competitiveness.

One of the most difficult is the lack of special training skills. We once interviewed a friend in the field of scientific research who designed an algorithm model for deep learning to be applied to the biomedical field. However, using the special math power purchased by the research institute, apply for four months at a time, train for two months at a time, and start all over again if you make a mistake. This kind of math situation directly made it impossible for him to complete research and development even after graduation.

Large-scale deep learning model training is the most time-consuming, energy-consuming and computing-intensive AI development link, but it is also the basis of AI development and innovation. If enterprise users and developers cannot train large models independently, AI will always stay on the surface, unable to produce differentiated innovation and in-depth industry applications.

And the solution is also very simple, using Xu Zhijun's description, is to carry on the violent calculation to the AI training. Only when AI is strong enough to calculate and crush data training with an unreasonable attitude can AI be really applied by thousands of industries and become a new highland of wisdom and innovation.

At present, the Teng 910 chip is in line with the "violent and unreasonable" chip "human setting". According to the published data, the test results show that the calculation power of Teng 910 reaches half precision (FP16): 256 Tera-FLOPS; integer precision (INT8): 512 Tera-OPS. And the power consumption required for compliance computing power is only 310W. A more direct comparison is that, when used in actual AI training tasks, in typical ResNet-50 network training, Tengten910 cooperates with MindSpore, and shows a nearly two-fold improvement in performance compared with the existing mainstream training single card with TensorFlow, and the number of pictures trained per second increases from 965 to 1802.

Perhaps we can put it more bluntly, the mainstream training card that Huawei refers to is the V100 of Nvidia. In fact, the AI training math currently available based on cloud services can only come from Google's TPU and Nvidia's V100. The latter is basically obtained through Amazon's AWS. According to the criticism of Nvidia, TPU2.0 only has the computing power of V100 1/2, and limited rental.

So we can see that the computing power of AI training based on cloud services is a very scarce resource, expensive and difficult to book. Such industrial conditions obviously can not promote the real development of the AI industry.

Now, after Penteng 910 is deployed to Huawei Cloud, global developers have a third choice, and China's public cloud market has ushered in the first full-stack cloud + AI solution. Coupled with the fact that Penteng 910 is more powerful than V100, Huawei Cloud always adheres to inclusive pricing strategy and adequate capacity supply, and training this ladder may really be able to embrace the solution of the computing layer.

Cloud service + AI training numeracy is a veritable pillar of the industry. But the corridor leading to this pillar used to be very narrow-companies and developers had to endure all kinds of cost restrictions, and when Pengteng 910 was connected to Huawei Cloud, the corridor was instantly widened and even opened a new thoroughfare.

From the widely distributed enterprise EI, to the application and ecological construction of Teng 310 and ModelArts, until today, Huawei Cloud's industrial intelligent capacity has also ushered in a qualitative improvement.

After mending AI heavy equipment, Huawei Yunning casting industry empowers Trident.

Looking further down the industrial logic of Teng 910, we will find that the differentiation ability that Huawei Cloud is about to dedicate to the market is the cultivation of AI with full chain and no compatible cost.

AI training computing is important for cloud services, on the one hand, it corresponds to the important needs of the development of AI industry, on the other hand, it is also a very important upstream support in the full-stack AI chain. When Teng 910 is applied to Huawei's cloud system, Huawei Cloud is fully based on Leonardo da Vinci architecture, providing comprehensive support from training and development to scenario deployment and systematic applications.

In the area of AI capability deployment, Huawei Cloud already provides image analysis services, OCR services, video analysis and other services based on Penton 310 chips. there are more than 50 API based on Teng 310, with an average of more than 100 million calls per day, and is growing rapidly.

In the area of developer enabling, Huawei Cloud, based on the ModelArts development platform, provides full-process model production services for AI developers, opening and covering the whole chain from data acquisition-model development-model training-model deployment. ModelArts has accumulated more than 30, 000 AI developers.

In addition to scenario reasoning deployment and simple development, begin to complete the large-scale AI computing power, as well as the top-level links of model training, which means that enterprise users will soon be able to complete all the production and application links from training-development-reasoning-deployment based on Huawei Cloud. The main beneficiaries of this chain are more enterprise users who have a lot of trainable data and require AI technology to meet the needs of differentiation and verticalization of the industry.

If we frame users in the market according to this demand, we will find that Huawei Cloud will provide full-stack AI enablement to three types of customers after filling the AI heavy device of Penteng 910:

1. Internet enterprises and high-tech enterprises with heavy demand for AI research and development, as well as scientific research institutions and developers with AI innovation desire.

2. Large-scale political, enterprise and research institutions that need to train a large amount of data, carry out stable deployment, and have high requirements for AI security.

3. Traditional industry + AI scenarios with in-depth R & D and training requirements for industry AI capabilities, especially focusing on industrial scenarios.

In today's fierce competition in the public cloud market, these three kinds of customers are likely to constitute the next Huawei Cloud Enterprise Intelligence Trident, thus prying artificial intelligence into social productivity and achieving a real inclusive AI.

The Great Coffee Machine: the Road from full Stack AI to Pratt & Whitney AI

The above may be seen as the short-term changes and goals of Teng 910 chip and Huawei Cloud. If we look longer, we may be able to answer the question: why does Huawei still insist on doing so many challenges in infrastructure such as chips and frameworks?

For some reason, what Huawei Cloud did to AI today reminds me of a device that we have become accustomed to: coffee makers.

Do not underestimate this machine, the coffee from grinding to boiling, brewing the whole process, integrated in a device, is a very great invention of human beings. It has changed the original rules and threshold of coffee production, so that coffee is readily available, and cafes are open all over the street.

With more flavors, less waiting time and easier operation, the coffee maker has made the drink truly universal-even from an aristocratic elegance of a few people to a productivity tool to improve social and work efficiency.

What HuaWeiYun is doing today is actually concentrating various AI processes such as grinding and brewing in the public cloud. So that developers and users do not need to go to three streets to grind coffee beans, ask someone to help make coffee on Fifth Street, and then go home and wait for three days and nights-- and the premise of integration is that every ability to AI coffee must be complete.

This completeness, what Huawei calls full-stack AI, refers to a full-stack solution that includes chips, chip enabling, training and reasoning frameworks, and application enabling. In terms of software and hardware architecture, Huawei's full-stack AI includes a series of operator libraries for Teng chips and IP; chips based on unified and scalable architecture, and a highly automated operator development tool CANN;, an independent and collaborative unified training and reasoning framework for end, edge, cloud, MindSpore; developer enabling tool ModelArts, and a large number of multi-level API.

When these capabilities and technologies are built into a complete closed-loop structure, Huawei Cloud can continue to reduce excessive and compatibility costs, iterate and provide more capabilities. Based on mechanized and efficient AI coffee, it can finally be brought to the front of thousands of industries.

Some people may wonder why Huawei wants to make its own framework MindSpore when there are so many strong frameworks. The answer is that MindSpore mainly strives for friendly development and efficient operation, and can adapt to different scenarios of end, edge and cloud. In other words, MindSpore is a deep learning framework strongly aimed at reducing development costs and improving the availability of development results, which is in line with Huawei Cloud's idea of providing a ModelArts platform for developers.

Reduce costs and improve computing power through self-developed chips; reduce development difficulties and improve application efficiency based on development frameworks and development tools; reduce compatibility threshold based on full-stack architecture; and provide flexible, controllable and highly integrated AI infrastructure based on all-cloud services. The ultimate direction is to reduce the cost and improve the efficiency as much as possible, so that AI can be transformed from the theory in the scientific research environment to the industrial feasibility in the industrial production environment.

When Xu Zhijun answered the reporter after the 910 press conference, he made it clear that Li Teng would go to the "road of no return" like Kirin. In other words, the full stack and full scene AI capability will continue to develop iteratively, which means that Huawei Cloud will push AI to Pratt & Whitney, which is also a road that will not stop until the goal is reached.

Therefore, the real goal of Penteng 910 is Huawei Cloud's inclusive AI strategy, the determination of "+ intelligence, see the future"; the market and service reconstruction between user-cloud computing-AI capabilities; and the rewriting of the rules of the AI threshold.

Pratt & Whitney AI has a long way to go, and the landscape is the meaning of the topic. But based on Huawei's full stack and full scene AI creation, the basic rules of cloud + AI have changed.

There is no precedent and possibility for human beings to retreat in terms of technology and productivity.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.