2023 Zhongguancun Forum|Zhou Bin, CTO of Huawei Ascend Computing Business: Without a powerful computing power base, large models have no roots

#FMCG Eight Talks##2023 Zhongguancun Forum#ChatGPT's popularity has led to the fire of the big model track, and the domestic AI big model has also heated up rapidly. Including Baidu, Ali, Huawei, Jingdong, etc., statistics show that at least 30 large models have appeared in China. However, the large model "who belongs to the hero" has not yet been revealed, and the computing power field, one of the AI "troikas", has taken the lead in setting off a round of battle. At the Zhongguancun Forum on May 26, Zhou Bin, CTO of Huawei's Ascend computing business, gave his answer to the development of computing power in the era of artificial intelligence.

Q: In the era of artificial intelligence, what kind of existence is computing power?

A: For ChatGPT, the circle often ridicules as "vigorous miracles", this "force" is engineering ability, the most intuitive embodiment is computing power, which is the basic ability to incubate or train large models. From this point of view, what we need to provide is a very powerful computing base, otherwise the large model above will "have no root".

At present, Huawei has provided Ascend AI-based computing power infrastructure for more than 20 cities in China, and provided full-process software enablement from model development, fine-tuning adaptation, and inference deployment to support the construction of large models.

So in the era of artificial intelligence, what kind of AI computing system do we need? In the future, the computing system of artificial intelligence is likely to become a new infrastructure, if artificial intelligence is ubiquitous, the infrastructure of computing power will also become ubiquitous, for such an infrastructure, it is naturally required to be stable, reliable, safe and efficient.

Security There is no doubt that in the future, large models will become the core of a large number of business systems, and it is necessary to ensure the security of the model itself, including the security of data. Secondly, to ensure high efficiency, large models consume a lot of energy, and improving its efficiency is a very important aspect. In addition, the running time of large models is usually calculated in "months", which also requires that the entire system be very reliable throughout the life cycle of the model.

Q: The explosion of domestic large models has caused an obvious gap in computing power, what practices do you think can effectively fill this gap in the technical field?

A: Different from Moore's Law, which doubles the demand for computing power every 18 months, in the era of artificial intelligence, there is a new law in the industry: about every 4 months, the computing demand of AI will double, which we call the computing power growth curve of AI. In the era of stand-alone systems, the scale of computing power can be increased through some violent means, but in the era of artificial intelligence, the requirement for us is to solve the demand for computing power in a systematic way, and we have a lot of basic software that needs to be innovated.

For example, from the computing architecture, the tensor computing used in artificial intelligence is different from the traditional computing mode, and the basic computing architecture should be innovated around the core of tensor computing, such as how to increase innovation in actuarial accuracy and computational bits, so as to better accelerate tensor calculation, so as to ensure computing power requirements.

On the other hand, you can work the distributed cluster system, there is currently a saying that AI Super Computer, in addition to the computing unit, the artificial intelligence supercomputer must also contain a large amount of storage, this storage space is hierarchical, taking the large model as an example, including memory space, corpus storage space, cloud storage space, etc., focusing on the ability of distributed interconnection.

To sum up, following the growth of computing power demand for large models requires distributed large-scale AI parallel training, in which there are a large number of computing nodes that work together towards the same task. These nodes may be distributed on a large number of physical institutions, and how to achieve interconnection and unified scheduling between these institutions is also a very important point.

Q: At the application level, what will be the future direction or hotspot of artificial intelligence?

A: In thousands of industries, artificial intelligence has a very large landing prospect, the most obvious of which is the Internet industry, people have been able to enjoy the dividends brought by a new generation of artificial intelligence, whenever we open the mobile phone to access the Internet, there are a large number of artificial intelligence systems behind the support. In addition, in the financial field, artificial intelligence is also landing on a large scale to help enterprises better serve customers, improve operational efficiency, and prevent risks. In addition, the new generation of artificial intelligence, including autonomous driving, education and even manufacturing, will become a very important productivity tool, empowering all walks of life and improving the overall level.

Q: Enterprises are rushing to enter the market, and there is also a grand situation of "100-model war" in China. What do you think are the advantages of China's bigger model?

A: First of all, "more" itself is an advantage of us, or we have the advantage of concentrating on big things, so that more resources can be invested in it. Of course, this kind of resources is not disorderly and chaotic, but has a certain degree of coordination and supervision, and in this way, we can gather greater forces to overcome difficulties in the short term.

Another level of "many" lies in the application scenario, that is, the advantage of the scene. The research of large models is continuous, and some phased results have appeared, such as some models with good results in specific fields and even general fields, and these models are still evolving. I think the application scenario will drive the research on the big model, because the big model must be combined with the scene to play its value, and our advantage is that there are many scenes, coupled with our great attention, from the national level to concentrate on promotion, we can promote its rapid application in the industry.

Once the large model generates value, a positive flywheel will be formed, which in turn drives the research of the model. The value of model research will gain positive gain in the application, and once this flywheel is started, it will be unstoppable and quickly generate more value in all walks of life.

Q: Open source is considered to be a major trend in the development of artificial intelligence, how is the open source ecosystem in the field of computing power created?

A: Open source and openness are a fundamental philosophy we uphold. Computing power is a very abstract concept, which must rely on specific software and hardware systems. Taking Huawei Ascend as an example, it is a software and hardware system used to build the foundation of computing power, including our own computing system, accelerator cards, servers, large-scale artificial intelligence clusters, etc., these physical forms should be opened to the industry through certain means, including the Ascend C language, acceleration libraries, APIs, and operating systems that are open to the bottom layer.

There are also a lot of basic software at the upper level, such as artificial intelligence frameworks, and large model suites in different industries, etc., and we have also adopted a lot of open methods, directly open to the entire community.

In the future, computing power will definitely become an important production resource like hydropower, and it must be turned into an achievement that everyone can enjoy, so that everyone can benefit through computing power.

Beijing Business Daily reporter Yang Yuehan