Huawei believes that the future of statistical-based computing may account for 80% of the computing power of the entire society, we will enter an era of intelligent computing.
In the outside of the Huawei-connected conference hall, there are two huge billboards. One wrote “AI peak, will be announced soon”, and one wrote “Yu Peng spread wings, rises thousands of miles, and strives to calculate the future.”
Huawei’s ambition is self-evident.
This is the fourth Huawei-connected conference. It is also the first time that Huawei has integrated enterprise BG, carrier BG, and cloud BU into ICT organizations.
In 2016, Huawei held the first full-connect conference and announced the cloud computing strategy. In the next year’s fully-connected conference, Huawei announced the establishment of cloud BU, which has long been invested in public cloud. Huawei Cloud is the foundation for Huawei to build a smart Internet of Things. After laying the foundation, the full-stack AI strategy was released in 2018, and the 910 and 310 chips were released.
To this year, Zeng Peng and Shengteng’s server-oriented chips have been released. The theme of this conference will focus on the ecological construction of 鲲鹏, Shengteng chip and the upgrade of AI calculation.
What is surprising is that Huawei has elevated its calculations to a position that is “connected”. The connection is the most important vocabulary for defining Huawei’s entire company from the business level.
Two years ago, Huawei released a new vision: to build a better connected world. But Hu Houkun said that Huawei is not just a connected company. Computing and connectivity are just as important. They are like twin brothers. Where there is a connection, there should be calculations wherever there is, where there is a calculation, and there is a connection.
Moreover, the scale of the computing industry is also very impressive. According to Gartner statistics, the computing industry is expected to exceed $2 trillion by 2023.
Based on these two points, Huawei released the overall computing strategy for the first time.
The strategy includes four aspects: Architecture innovation, investment in full-process processor families, and business strategies that do something different.
In terms of architectural innovation, Huawei has released a self-developed DaVinci architecture with a 16x16xN scalable Cube. From the perspective of the industry, after the failure of Moore’s Law, the industry needs to find a new structure to make breakthroughs. From Huawei’s own perspective, as the business expands, it also needs an architecture to cover multiple scenes of the edge cloud. At present, Da Vinci architecture is the only architecture in the industry that can cover end, edge and cloud at the same time.
On the processor, Huawei has long insisted on self-research, and has released a series of new products such as Kirin, Shengteng, Kunpeng and Hongjun. Among them, Shengteng and Kunpeng are server-oriented.
In business strategy, Huawei’s keyword is the boundary. Hu Houkun said that Huawei will not sell processors independently, and will open and sell them in the form of servers, motherboards and boards. Huawei will open the motherboard and modules of the server to the outside world. On the software, the server’s operating system, database, and AID data framework are open source. Huawei does not use applications, but invests resources to help developers migrate applications and applications.
AI computing needs to be supported by developers, that is, to build an open ecosystem. Huawei has released 1.3 million developers since its release of the Fertile Land Project. Hu Houkun revealed that Huawei will release a new fertile plan at the conference, which will invest 1.5 billion US dollars. The goal is to increase the number of developers to 5 million, but the cycle of unspecified investment will be released on the 20th.
With the advent of the AI era, is unable to describe more and more problems with certain rules. CPU-based computing cannot be satisfied, and statistical-based intelligent computing will become more and more popular.
Huawei believes that future statistics-based calculations may account for 80% of the computing power of the entire society, and we will enter an era of intelligent computing.
The era of intelligent computing will have different characteristics from the previous ones: First, intelligent computing is a violent calculation, and we have a high degree of dependence on computing power. If you want your computer to know a cat, you need a million photos to train, and the computing power of more complex scenes such as autonomous driving, astronomical exploration, and weather prediction can be imagined. Second, on the terminal, on the edge, on the cloud, in any corner, the calculation is everywhere. This leads to the third feature that the edge cloud needs to work together.
Therefore, Huawei needs a strategy for AI calculations in addition to general-purpose computing. Last year, Huawei released a full-scenario AI strategy. Today, Huawei announced that announced that the full-stack full-scenario solution has been fully implemented. The Soaring 310 and 910 are available, and the cloud service based on the Shengteng 310 has been launched. The 310-based ATLAS series products have been launched.
At this conference, Huawei also released the training cluster ATLAS 900, which consists of several thousand rising processors. Huawei claims that this is the fastest AI training cluster in the world. According to Faster cluster statistics, during the training of ResNet-50 v1.5, Huawei ATLAS 900 took 59.8 seconds, which is more than 10 seconds faster than the second one, saving 14% of the time. Hu Houkun made an analogy. This is equivalent to the sprint champion running over the finish line. After drinking a bottle of water, he will wait until the second place.
Huawei shared a case. SKA (square-square array radio telescope) brings 200,000 stars on a starry sky map. Under current conditions, astronomers need to find a certain feature from these 200,000 stars, which takes 169 days. The amount of work. Now with Atlas 900, in just 10 seconds, the stars with corresponding characteristics are retrieved from 200,000 stars.
At present, Huawei has deployed it to Huawei Cloud and is open to global research institutions and universities.
Started as a communication device, and later entered the mobile phone market, becoming a first-line manufacturer, and has become the number one manufacturer in China and the second largest in the world in terms of shipments. This time, Huawei put its eyes on the AI computing market and competed with NVIDIA and Google. The results are equally expected.