At the opening ceremony of the 2023 Shanghai World Artificial Intelligence Conference today, Ken Hu, Huawei's rotating chairman, delivered a speech, saying that two key measures or starting points are: on the one hand, to take root computing power and build a strong computing power foundation to support the development of China's artificial intelligence cause; On the other hand, it is to combine the research and innovation of large models, from general large models to industry large models, and truly let artificial intelligence serve all walks of life and scientific research.
The following is the transcript of the speech:
The two key measures or starting points are: on the one hand, take root computing power and create a strong computing power base to support the development of China's artificial intelligence cause; On the other hand, it is to combine the research and innovation of large models, from general large models to industry large models, and truly let artificial intelligence serve all walks of life and scientific research. Therefore, we propose "AI for industry" and "AI for science".
I will report to you in detail below.
First of all, regarding computing power, the development of artificial intelligence is inseparable from computing power. In the current situation in China, we face challenges such as availability of computing power and cost. Over the years, Huawei has focused on the development of computing power. We focused on the foundations of Kunpeng and Ascend and made a breakthrough. At present, we are building China's computing power base through architecture innovation, ecological development and flexible co-construction methods. We hope that through joint efforts with everyone, we can eliminate the constraints of computing power on the development of artificial intelligence.
First, we improve computing efficiency through architectural innovation. At the compute node level, we launched a revolutionary peer-to-peer evaluation architecture, which breaks through the performance bottleneck caused by traditional CPU-centric heterogeneous computing, improves the bandwidth of the entire computing, reduces latency, and improves node performance by 30%.
On the other hand, at the data center level, we launched the Ascend AI Computing Cluster in 2019. Through the cluster approach, we bring together the advantages of computing, storage, networking, and energy, and design and manage the AI data center as a supercomputer, which greatly improves performance.
At present, the largest AI computing cluster we have built in China is Pengcheng Cloud Brain Phase II in Shenzhen, with a current computing power of 1000P. According to our planning, by 2024, the scale of the third phase of Pengcheng Cloud Brain will reach 16,000P. In addition, we have deployed thousands of cards in the computing center in Ulanqab, and have found that through clustering, efficiency can be improved by more than 10% under the same computing power.
Below I will report to you on ecological development, because ecology is a key means for the development of the computing power industry, and it is also a bottleneck that is difficult to overcome.
Four years ago, Huawei proposed a strategy of hardware openness, software open source, and talent training, focusing on the development of the computing industry. By working with all parties, we have achieved certain results. In terms of hardware, we insist on further opening up, launching more diversified modules and boards this year, working with more than 30 hardware partners to launch hundreds of artificial intelligence hardware based on Ascend to meet the needs of different industries.
On the software side, we strengthen the basic software through open source, especially for the innovation of current large models. We provide a full-process enabling platform to better support scientific research institutions and enterprise customers. At present, our ecosystem has incubated more than 20 basic large models and adapted to more than ten mainstream large models in the industry.
Statistics show that half of China's current large models are supported by AI ascend computing power. Therefore, I am especially grateful to those companies and institutions that use AI Ascend computing power, and thank them for their trust in us, which makes us more confident in the future development of the computing power ecosystem. Of course, we also know that there are no shortcuts to ecological development, and we must move forward in a down-to-earth manner. We hope that more companies will join us in the future and build this ecosystem together.
In terms of developing computing power, we also take into account the flexibility of the body and diversified means and models, which is based on the actual situation in China. Therefore, we use a variety of modes for computing power construction. For example, in the construction of computing power infrastructure in cities, we have supported local governments to build 25 city-level AI computing centers, including Shanghai.
On the other hand, we have noticed that many large enterprises have the need to build their own AI computing power centers, so we actively cooperate and help these enterprises to build their own AI computing power centers. For example, China Mobile, iFLYTEK, China Southern Power Grid and other large enterprises are all under construction, and we are actively participating in them.
At the same time, we realize that there is a strong demand for AI from SMEs. To meet these needs, we provide computing power services on HUAWEI CLOUD through cloud services, combining multiple means. We hope that in this way, computing power will no longer become a bottleneck in the development of artificial intelligence. While Huawei is deeply engaged in computing power, it is also committed to enabling artificial intelligence to serve thousands of industries and scientific research.
In this regard, we believe that on the one hand, we must continuously improve the ability of general large models, and at the same time, we must explore the combination of general large models and industry large models. For example, if I live in Futian District and I have a 78-year-old in my family, I wonder if the government can provide subsidies. If we use a generic large model, what kind of answer will it give? At the same time, let's see what kind of answer it will give if we use the industry big model.
This industry model is a big model of government affairs that we developed in cooperation with Futian District of Shenzhen. Obviously, the industry big model can give more accurate and valuable answers on the basis of the general big model. This is what we are working towards.
To this end, Huawei has launched a three-layer large model structure, the bottom layer is the basic large model, which we liken to "reading ten thousand books", and its goal is to learn a large number of basic knowledge. On top of this layer, we created industry models and scenario models, which we likened to "traveling thousands of miles". There are still many challenges to overcome from "reading 10,000 books" to "traveling 10,000 miles", and one of the keys is to fully match and integrate knowledge from all walks of life with big models. In this regard, Huawei is working with various industry partners.
Currently, Huawei's Pangu model has supported the implementation of AI applications in more than 400 service scenarios in more than 10 industries, including finance, manufacturing, government affairs, electric power, coal mining, healthcare, and railways. In the future, we hope to work with more industry partners to expand this aspect, deepen cooperation, and truly go deeper and do more practically.
In addition to making artificial intelligence serve thousands of industries, we believe that there is another important task at present, that is, to make artificial intelligence serve scientific research. We came up with the slogan "AI for science". We have found that by learning massive amounts of historical data and scientific knowledge, and encoding mathematical equations into our large models, AI can facilitate the integration with basic disciplines (such as molecular dynamics, fluid mechanics, heat transfer, biology, etc.) to help us discover more scientific laws.
At present, Huawei has released Pangu scientific computing big models, including drug molecular models, Pangu meteorological models, and ocean wave models. We have worked with scientists and achieved good results. For example, in meteorological research, our Pangu model can complete weather forecasts for one hour to seven days in the future in seconds, quickly and accurately. This requires more than 40 years of study of meteorological data, combined with industry knowledge for training and correction. We are constantly training and correcting in various areas of scientific research. Through our efforts, we hope to bring more new ideas, methods and tools to scientists and researchers, and inject new impetus into the development of human society.
Finally, I would like to make a small advertisement, because tomorrow we at Huawei will hold a global developer conference, where we will introduce Huawei Pangu Model 3.0 in detail, and how we can provide services for thousands of industries and scientific research. I hope everyone will pay attention.
Dear friends, we are very fortunate to witness several waves of scientific and technological revolution, from the Internet to mobile to cloud computing to artificial intelligence, each round of transformation has brought far-reaching impacts to society. There is no doubt that general artificial intelligence has brought us unlimited imagination and led us into the next golden decade. We hope to work with you to innovate so that artificial intelligence can better serve thousands of industries and better serve scientific research. AI for industry, AI for science, let's win the new era of artificial intelligence together.