Not only 5G and Hongmeng, Huawei's latest big move is to throw the AI computing nuclear bomb.

Just now, Huawei's full connection conference opened and launched another heavyweight AI product Atlas900 。

In an interview with foreign media earlier, Ren Zhengfei had predicted that this is the "fastest AI platform in the world at present".

How fast is the fastest in the world?

Huawei gives data training ResNet-50 Only 59.8 seconds 。

At the same accuracy, it is 15% faster than the runner up.

Hu Houkun, vice chairman of Huawei, said that such computing power can also be widely used in scientific research and business innovation, such as astronomical exploration, weather forecasting, automatic driving, oil exploration and other fields.

At the same time, Huawei also announced that related cluster services of Atlas 900 will also Extremely favorable price Open to global scientific research institutions and universities.

With the release of this product, Huawei also disclosed its overall computing strategy, which is the first time Huawei has disclosed this strategy to the public.

Integrate thousands of AI chips with the strongest computing power

Altas 900 is composed of thousands of ascending 910, which can reach 256~1024   PFLOPS@FP16 。

The training of ResNet-50 only took 59.8 seconds of cluster size, and used 1024 Shengteng 910.

The performance in practical application is also very eye-catching. In the field of astronomy, the traditional 169 day task can be shortened to 10.02 seconds.

Shengteng 910 chip, the third AI training chip in the world and the first AI training chip in China, was unveiled at the Full Connect Conference last year and officially put into commercial use in August this year.

The Shengteng 910 adopts the Da Vinci architecture developed by Huawei and is known as the "AI processor with the strongest computing power". The maximum power consumption of the 7nm process is 350W, and the measured power consumption is 310W.

The chip has a clear score and benchmarking. Directly benchmarking Nvidia Tesla V100, focusing on in-depth learning training scenarios, and targeting AI data scientists and engineers.

How strong is the chip? Huawei also gives the data of its "hanging" friends:

With the same power consumption, the computing power of Shengteng 910 is twice that of the V100. The training speed is faster, and the time for users to obtain training output will be shorter. In a typical case, compared with V100, the computing speed of Shengteng 910 can be increased by 50% - 100%.

With the release of the Altas 900, Huawei's computing layout was officially completed and entered the landing stage.

How will Huawei land? How to drive the landing? Huawei has now made a full disclosure.

First disclosure of Huawei's computing strategy

On the whole, Huawei's computing strategy focuses on "chips" and is divided into four parts, including chip technology, overall planning, business strategy and ecological construction:

First, it is architecture innovation.

The core is investment oriented basic research, such as Da Vinci architecture, focusing on solving the architecture problem of full scene intelligence.

Secondly, it is the full scene chip, that is, the processor family.

At present, Huawei has four chip series, including Kunpeng series for general computing, Shengteng series for AI, Kirin series for intelligent terminals, and Honghu series for smart screens.

Hu Houkun said that there will be a series of processors in the future for more scenarios.

The third is business strategy.

Huawei has also clarified its chip business model once again, with open hardware and open software, enabling application development and migration.

However, chips are not directly sold to the outside world, but are oriented to customers in the form of cloud services and components, giving priority to supporting partners to develop complete machines.

Finally, build a development ecosystem.

In this step, Huawei advocates openness and updates the fertile soil plan. The goal is to invest in the next five years US $1.5 billion , bringing together 5 million developers to enable global partners to develop applications and solutions.

Behind these are new business opportunities and greater space. Huawei is no longer a supplier of communication equipment.

Hu Houkun stressed at the beginning that Huawei had been recognized as "connected" before, but faced new opportunities in the new era and new stage. Huawei now needs to face the intelligent industry and create a grand blueprint from connection to computing.

How does Huawei see the future?

Not only was the strategy released, Hu Houkun also disclosed the logic behind it and gave a prediction.

He said that with the continuous evolution of computing models, machine learning oriented computing has become the mainstream.

But now, computing is entering the intelligent era, which has three major characteristics: violent computing, ubiquitous computing from the center side to the end side, and collaborative computing with efficient interaction between the end and edge clouds.

The best solution is to use brute force computing at the central node, train the general model, and make personalized deployment at the edge.

However, in order to achieve this plan, we still face many difficulties, such as the new architecture that can support full scene intelligent computing. This is a huge challenge and opportunity for the development of the current computing industry.

How big is it?

Hu Houkun quoted Gartenr's report that the total space of the global computing industry is $2 trillion.

In his view, this is a big blue ocean. Huawei will firmly invest in the field of computing and promote its implementation.

Respond to the blow: AI landing progress

At the scene, Huawei also gave a reply to the cattle it had blown last year.

At the 2018 Full Connectivity Conference, Huawei released the full stack and full scenario AI strategy, and computing is the cornerstone and core of this strategy.

Today, Hu Houkun was in the attitude of defense and handed over the transcript of a year's landing: in August, with the release of Shengteng 910 and MindSpot, Huawei's AI solution has been built.

from Full stack It can be seen that Huawei has possessed Shengteng processor series IP and chips, chip enabled CANN (chip operator library and highly automated operator development tool), application enabled ModelArts, training and AI framework MindSpot.

from Full scene angle It can be seen that the products and solutions of Shengteng processor have covered the cloud, edge and end. On Huawei Cloud, corresponding reasoning and training services are provided.

Atlas series boards, modules and servers have been launched. Equipped with Shengteng series processors, it has been widely used in Huawei's smart phones.

In addition, Huawei will further introduce the implementation of general computing, which is divided into four steps:

First, build a competitive general-purpose computing processor.

Secondly, continue to invest in key technologies and products such as boards, servers, operating systems, databases, compilers, etc., open up the whole ecological chain, complete system level verification, and help partners better sell complete machines.

Third, Huawei has cooperated with partners to build a Kunpeng industrial ecological base. Currently, it has settled in Beijing, Shanghai, Chongqing, Shenzhen, Chengdu and other cities, and has carried out comprehensive cooperation in platform construction, talent training, application demonstration and other fields.

Moreover, these products are not only in China.

Huawei has previously disclosed that the series of products, solutions and services based on Kunpeng processors will face the global market, start in China and serve the world.

The vice chairman of Huawei said: Huawei has AI, and Huawei is pragmatic. Everyone can be assured to cooperate with Huawei.

How do you evaluate the "nuclear bomb" thrown by Huawei this time?