AI hashrate is the core competence of future countries, cities, and enterprises.
In the era of artificial intelligence, AI hashrate is like electricity, and AI computing centers are the power plants.
01 No hash rate is just like no electricity for AI
Recently, what most disturbs many local governments and business leaders across China is lack of power. And around the world, many countries are beginning to face a key element that determines the national economy and people's livelihood like electricity, AI hashrate.
AI hashrate is the computing power that supports AI. The calculation here is not addition, subtraction, multiplication and division, but calculation of everything in the world, a highly complex and ubiquitous calculation under the interconnection of everything and artificial intelligence.
Different from traditional computing power, in order to support the development, training, and reasoning of AI models, AI hashrate has particularly high requirements for parallel processing capabilities, and therefore requires special AI chips and frameworks.
For example, only AI chips with powerful floating-point computing capabilities can provide high-quality AI models that meet the intelligent transformation of industry enterprises through training and continuous iterative optimization. In complex model training, hundreds of billions of floating-point parameters need to be fine-tuned by hundreds of thousands of steps, and fine floating-point expression capabilities are required. If there is no powerful training chip, it is difficult to guarantee the efficiency of the output of the algorithm model. The release of the hundreds of billions of Chinese NLP (Natural Language Processing) model "Pengcheng Pangu" and the release of the "Pengcheng Shennong" platform for the field of biomedicine are inseparable from the support of AI chips.
It is with the AI hashrate released by these AI chips and AI frameworks that we can accelerate our entry into the era of Internet of Everything and artificial intelligence.
Today, from the mobile phone in everyone's hand, to the enterprise cloud platform, to the urban brain...Our production and life are increasingly dependent on AI, and we are increasingly relying on AI to gain power.
In the world of artificial intelligence, no AI hashrate is like no electricity. AI computing power has penetrated into all aspects of our life and production. Take medicine from a hospital pharmacy as an example:
The robot that uses artificial intelligence technology for fully automatic replenishment of medicines uses 3D vision positioning, intelligent robot grasping, and intelligent vision review technology to ensure 100% replenishment accuracy and higher efficiency. The dispensing speed can reach hourly 2500 boxes, one order can be processed in 8 seconds. In the process of drug distribution, the system can autonomously dispatch and transport drugs without manual intervention.
The most intuitive change it brings is that the time for drug sorting can be shortened from the original 50 seconds to 3 seconds, and patients can get the drugs in only one minute.
In this process, how to identify the prescription and how to sort and distribute medicines accurately? To realize these functions, it depends on how fast, how good, and how accurate the AI system used by this robot can be calculated. This is the AI hashrate.
02 To have electricity, there must be a power plant and a power grid
AI Hashrate is so important, but many companies lack sufficient funds to build their own AI hashrate. So how can the demand for AI computing power be met, and how should countries and cities provide sufficient AI computing power support to promote the development of the AI industry and win the competitiveness in the AI era? The answer is to make AI hashrate a public resource, and to build a new type of infrastructure.
This new type of hashrate infrastructure is the artificial intelligence computing center. To use the analogy of the electrical age, it is to build power plants and power grids.
The first is to accelerate the construction of artificial intelligence computing centers.
The artificial intelligence computing center is based on an artificial intelligence computer cluster built on artificial intelligence chips, covering a complete system of infrastructure, hardware and software infrastructure. Its core function is to provide the release of hash rate from the underlying chip to The artificial intelligence full-stack capability enabled by the top-level application is the output of AI hashrate.
In addition to being a platform for providing public hash rate services, the artificial intelligence computing center should also be an incubation platform for application innovation, an industry aggregation development platform, and a research and innovation talent training platform. Only when these roles are played well at the same time can we get through the "government, industry, education, research and application", concentrate the most power, form industry convergence and enhance AI competitiveness.
At present, the world is accelerating the construction of artificial intelligence computing centers. In particular, the United States, on the one hand, is doing everything possible to suppress the development of other countries, on the other hand, it has invested heavily in strengthening the development of artificial intelligence in its own country. The Biden administration once announced an investment plan of 300 billion US dollars to defend the United States’ leading position in the field of artificial intelligence. , And a very important investment is to strengthen the new infrastructure of data centers and intelligent computing centers.
Of course, China will not easily miss the opportunities brought about by the development of the artificial intelligence industry. As early as 2017, the State Council issued the "New Generation Artificial Intelligence Development Plan" and emphasized the need to "build a high-performance computing infrastructure." During the epidemic last year, the central government further clearly proposed a new infrastructure strategy, and strengthening the construction of data centers and artificial intelligence computing centers is the top priority of the entire new infrastructure.
Without a strong hash rate, the seven major areas of new infrastructure that focus on digitization can hardly achieve their construction goals.
Under the pressure of national strategic guidelines, market prospects, and even economic transformation and upgrading, many local governments have taken active steps to take the lead in the construction of artificial intelligence computing centers, and on this basis, improve the level of local hash rate and build core competitiveness in the digital age.
On May 31 this year, among the 15 national artificial intelligence innovation and development experimental zones approved by the Ministry of Science and Technology, Wuhan’s artificial intelligence computing center has been completed and put into operation; Xi’an’s future artificial intelligence computing center has also been online, and artificial intelligence in other provinces and cities The construction of the computing center is also being planned.
After the Wuhan Artificial Intelligence Computing Center was put into operation, it provided computing power support for economic development, scientific research innovation, and enterprise transformation in Wuhan and even Hubei.
For example, Wuhan University has created the world's first dedicated remote sensing framework Wuhan.LuojiaNet based on the Wuhan Artificial Intelligence Computing Center. For "large-format, multi-channel" remote sensing images, it has achieved significant results in whole image analysis and minimal data set reading and processing. breakthrough.
For another example, the Institute of Automation of the Chinese Academy of Sciences used the computing power of the center to develop the world's first multi-modal video generation model-Zidong Taichu. As the industry’s first 100 billion-level three-modality model, Zidong Taichu’s video understanding and description performance has become the world’s first. It not only has multi-task joint learning capabilities, but can also realize AI-based graphic search through learning. And audio, short video, MV production, greatly shortening the creation time of audio and video.
In September, the first artificial intelligence computing center in the northwestern region was located in Xi'an, a major city in the northwest. The first phase of the planned Xi’an future artificial intelligence computing center with 300P AI computing power will provide accurate and reliable models based on the construction of the Asend AI basic software and hardware platform. Training and reasoning.
Xi’an Future Artificial Intelligence Computing Center has signed the remote sensing project of Xidian University, the voice model project of Northwestern Polytechnical University, and the "MindSpore Research Laboratory" of Shaanxi Normal University. It is supporting Xi’an "6+5+6+1" modern As the industrial system develops, it will also strengthen the artificial intelligence industry clusters in Xi'an and the entire Northwest region, and provide computing power support for the development of the artificial intelligence industry in the Northwest region.
The Xi'an Future Artificial Intelligence Computing Center, which has just launched, has almost reached its full capacity utilization rate. The thirst for computing power of local enterprises, scientific research institutions, and universities in various industries can be seen.
Secondly, efficiently use the hashrate resources of the artificial intelligence computing center.
When more and more artificial intelligence computing centers are built and put into operation, how to make their computing power more efficient and serve more industries and enterprises? How to avoid uneven distribution of computing power and inconsistent use efficiency in various places? How to make places where there is no plan to build an artificial intelligence computing center can also enjoy the enablement of AI computing power? The interconnection, collaboration, and sharing between artificial intelligence computing centers has become a problem that needs to be considered by all walks of life.
This requires an artificial intelligence computing power network, just like the power grid is to power plants and power consumers.
With the computing power network, we can connect the artificial intelligence computing center nodes distributed in various places, dynamically and real-time according to the state and demand of computing power resources, realize the overall distribution and scheduling of computing tasks, forming a nationwide perception, distribution, and scheduling The computing power network of the artificial intelligence center then gathers and shares computing power, data, and algorithm resources on this basis.
The most important thing is that with this network, more industries and companies can use AI computing power just like electricity.
So, what effect will the computing power network have on the entire artificial intelligence industry?
The first is the convergence of computing power, that is, the high-speed interconnection of computing power resources in different regions and different cities to achieve reasonable scheduling of computing power across nodes and flexible resource allocation. This is conducive to improving the utilization rate of each artificial intelligence computing center. For the overall energy saving, the follow-up can support cross-node distributed learning to provide super computing power for the research of large models.
The second is the aggregation of data. The government takes the lead in cooperating with enterprises in various industries. After reaching the openness of public data in the field of artificial intelligence, it can rely on the artificial intelligence computing center to gather high-quality open source and open artificial intelligence data sets, which can promote the development of artificial intelligence. Algorithm development and industry landing.
Finally, there is the convergence of ecology. Between each artificial intelligence computing center, unified interconnection standards and application interface standards, realize the opening of large model capabilities in the network and the sharing of application innovation results, strengthen cross-regional scientific research and industrial collaboration, and implement artificial intelligence for users across the country. Application innovation provides more resource choices and more convenient ways of cooperation, accelerates industrial aggregation, and activates industrial inclusiveness and symbiosis.
The simple total settlement network is to gather big data + big computing power, enable large models and major scientific research innovations, and incubate new applications. In turn, the computing power will be networked, the computing power cost will be reduced, and the computing energy efficiency will be improved.
The Ministry of Science and Technology pointed out in the three-year action plan that it is necessary to "lay out a number of artificial intelligence computing centers to form a wide-area collaborative artificial intelligence platform." Under the guidance of this plan, artificial intelligence computing centers have successively landed in many cities. At the just-concluded HC2021, more than 20 artificial intelligence computing center construction cities jointly lit up the "artificial intelligence computing power network".
This computing power network in the artificial intelligence industry has begun to weave and build.
03 AI computing power construction is not a long-term plan but an urgent issue
In 2020, Charles Leiserson, a computer scientist at the Massachusetts Institute of Technology and a pioneer in parallel computing, wrote an article in the "Science" magazine:
Deep learning is approaching the computing power limit of existing chips.
In fact, in the past ten years, the demand for computing power for the best AI algorithms of mankind has almost increased by 1 million times, doubling on average every 3.4 months.
In contrast, the growth of global AI computing power is very limited.
The huge gap between demand and supply has prompted governments of various countries, especially regions with leading AI technologies such as China, the United States, Europe, and Japan, to vigorously build AI computing power.
Without strong AI computing power, a country or region will inevitably be at a disadvantage in the future technological competition.
Judging from the current progress in the construction of computing power infrastructure, Shenzhen, Wuhan, Xi'an and other cities have built artificial intelligence computing centers and put them into operation, Chengdu, Henan and other cities are under construction, and artificial intelligence computing centers in Beijing, Nanjing, and Shanghai are under construction. Speeding up construction is also gaining momentum.
In the future, once the artificial intelligence computing center is fully completed and an artificial intelligence computing power network is formed, it will not only provide society with a continuous flow of super computing power across regions. Moreover, it can also realize cross-regional scientific research and industrial collaboration, enable large models and major scientific research innovations, and incubate new applications for thousands of industries.
In the end, artificial intelligence will empower more industries and scenarios, allowing us to remain invincible in the future industry and technological competition between countries.
Source: https://36kr.com/p/1416826462541192
No comments:
Post a Comment