Click to follow
Text丨Liu Yuqi, Hao Xin
In 2005, Huawei proposed "All IP" for the network era, "All Cloud" for the digital era in 2011, and "All Intelligence" for the intelligent era in 2023.
Up to now, Huawei's strategic upgrade has gone through three stages.
When entering the era of intelligence, there are still many difficulties that need to be faced. With the emergence of large model capabilities, multi-modal fusion, MOE and other trends, model parameters will soon exceed trillions from the current 540 billion, requiring massive stable computing power, large-scale parallel training, and overall architecture design. higher requirements. At the same time, the rapid growth of data set size, the difficulty of efficient data storage, cleaning and annotation, and the training of massive Tokens are full of challenges.
"It can be said that large models and related applications are the most complex software and hardware system projects to date," Huawei Cloud CTO Zhang Yuxin judged.
In the previous article "Comprehensive Intelligence, Huawei's "Hard Kung Fu"" , we summarized that at the hardware level, Huawei's strategy is to start from the bottom foundation, using hardware as the entry point, opening up the perception layer and connection layer, relying on characteristics Industry legions have entered all walks of life and issued intelligent solutions. It is precisely because of the particularity of Huawei's own integration of software and hardware that it has given rise to the need to not only connect the "hardware layer" and the "software layer" well, but also connect the software layer to thousands of industries.
At the recent Huawei Connect Conference 2023, Huawei Cloud announced a series of practices, which further confirmed the deep-seated logic of combining software and hardware:
For the underlying computing power, Huawei announced the official launch of Ascend AI cloud service, providing computing power clusters, computing engine CANN, AI development framework MindSpore and AI development platform ModelArts to build an AI cloud base;
For development training for customers and developers, the "Ascend AI Cloud Service Special Zone" is launched. It also integrates the development production line, low-code/no-code platform, AI application framework and new AI application engineering suite required for application development, reducing development costs. threshold;
For the implementation of large models, Huawei Cloud Stack 8.3 was released, providing 13 categories of 100+ cloud services, focusing on enhancing the four core capabilities of Pangu large models, industrial Internet, data element circulation, and software development production lines.
By sorting out the content of this conference, Guangcone Intelligence found that Huawei cloud services are generally divided into two major sections: one is computing-oriented, focusing on hardware delivery and providing computing power; the second is directly oriented to customers, focusing on software delivery. Mainly, the business of providing services.
If we visualize the comprehensive intelligence strategy proposed by Huawei, computing power is like an engine, and computing is the engine. Huawei Cloud Platform plays the role of a connector, calling the computing power platform downwards and being upwardly compatible with thousands of industries. Externally, it also provides tools for intelligent rendering of engineering drawings and tools for thousands of industries.
01 Shengteng AI Cloud Service: The “blood transfusion pump” of computing power
With the continuous advancement of large model manufacturers, the implementation of "large models in all walks of life" is gradually becoming a reality, which has also stimulated the demand for large-scale underlying computing power.
As the underlying computing power, in most cases it was completed through local deployment. However, the obvious change now is that large models are beginning to push the computing power infrastructure to continuously "upward" to serve customers in thousands of industries.
However, to directly deploy computing power on the enterprise side is an extremely high threshold. As Huawei cloud agent told Light Cone Intelligence, "The budget should start in the tens of millions, and secondly, a technical research and development team must be equipped." This creates a gap from the computing platform to the customer.
Who will fill this gap? The answer given by Huawei is "Shengteng AI Cloud Service". According to Huawei Cloud, as of now, it has built three major AI computing centers in Gui'an, Ulanqab, and Wuhu. The computing power center cannot provide services directly, but through cloud computing, computing power such as water and electricity can be converted into "ready-to-use" services and sold to users.
Just like building blocks, Shengteng AI Cloud Service combines the capabilities of the computing power platform and the cloud computing platform. The hardware exerts the capabilities of signal transmission and calculation, and the software exerts the capabilities of data transmission, storage, encryption, etc. The software and hardware are integrated to maximize efficiency. It is understood that it is currently based on Huawei's Ascend AI cloud service, and at the same time realizes operator fusion and mixed precision optimization through software and hardware collaboration, increasing training efficiency by 45%.
Specifically, Huawei's Ascend AI industry ecosystem consists of cloud-edge hardware, heterogeneous computing architecture, AI framework, application enablement, and industry application layers. These parts reveal the entire process of how Huawei's AI capabilities are implemented in industry scenarios.
We can understand the above picture as a factory being put into production. The heterogeneous computing architecture layer is like a universal production machine. It has strong compatibility and supports not only CPU engines, but also GPU and TPU engines. Only with computing power can the entire factory operate.
The AI framework is like the drawings in the hands of workers. Only by comparing the drawings can we know the production steps of each step. Having machines and drawings is not enough. The factory also equips each worker with some packaged wrenches, hammers and other tools. This is the role of the "application enablement layer". When workers assemble the drawings into actual products, they will be immediately classified and put into different industrial lines for packaging processing. This is equivalent to the entry of AI into various industries.
Huawei has built the most comprehensive production factory from the bottom up and also provides a variety of services. For example, the heterogeneous computing architecture layer not only supports different types of processors, but also provides four computing power supply models. One is the bare computing power model of directly selling servers; the second is the rental computing power model of renting servers; The third is to create a cloud service space through servers to provide computing power support; the last one is to provide computing power in the form of MaaS services.
Throughout the world, cloud vendors such as Google and Amazon mostly provide computing power in the form of cloud computing power model and MaaS model. Hardware chip manufacturers such as Nvidia mostly provide computing power in the form of bare computing power and multi-tenant model. For a long time, the two There is a clear distinction between them.
However, the multi-computing power supply model is gradually becoming a trend. According to foreign media reports, NVIDIA has begun to sign GPU contracts with some cloud vendors and "force" them to rent NVIDIA servers in order to open up cloud computing power and the MaaS model.
02 Huawei Cloud: AI cloud base and production line
The advantage of the computing power layer is a new engine, and Huawei Cloud is the carriage that leads Huawei's intelligence to thousands of industries.
At the Full Connection Conference, in addition to releasing the Ascend AI cloud service, Huawei Cloud also upgraded the Stack 8.3 version to provide enterprises with one-stop tools and services for building exclusive large models. It is understood that the new Stack8.3 version provides a complete AI production chain, including AI computing power, computing structure, framework, development platform, development kit, basic large model and professional services, lowering the threshold for building large models.
"Large models are not the exclusive preserve of a few companies," said Shang Haifeng, President of Huawei Cloud Stack: "Huawei hopes that every company will have its own large model."
However, from an internal perspective, the real-time collection of data is limited by non-digital terminals, the real-time uploading of data is limited by low-speed networks, the real-time analysis of data is limited by data silos, and industry data is difficult to collect, transmit, and use, and many factors hinder it. the process of intelligence.
Looking at the current product layout of Stack 8.3 with this idea in mind, it feels like targeting enterprise pain points one by one.
For example, in version 8.3, the focus is on enhancing the circulation capabilities of data elements, which is the first difficulty that enterprises face in becoming intelligent. In terms of real-time data collection, Huawei has built perceptual layer hardware for support; in order to solve the problem of upload speed, Huawei's QingTian architecture is based on a high-speed new network protocol, breaking the boundaries of computing power, storage and network, and achieving parity with multiple computing powers. Internet to solve problems such as long data transmission time and data loss caused by previous low-speed networks.
On the other hand, data needs to be circulated to fully reflect its value, but the current industry still lacks a credible data circulation mechanism. This challenge is particularly prominent in the era of large models.
Data is the core asset and the source of competitive advantage for industry users. Some key sensitive data of industry users are difficult to share or "leave the factory", such as data related to urban development, public security and personal privacy in the government affairs industry; responsible persons in the financial industry Data related to rights and debt relationships; manufacturing asset details, production data, and data that are explicitly required not to leave the park, etc. At this time, the basic large model is difficult to adapt to the intelligent needs of the industry.
To this end, every cloud vendor is focusing on establishing a safe and trustworthy mechanism to ensure the circulation of data elements. Blockchain and privacy computing have become the core capabilities. This was originally the core technology in financial transactions, and now it has gradually become a basic capability.
In the data circulation process, blockchain technology can confirm data ownership and prevent tampering, and privacy computing technology ensures that data can be "available, invisible, computable, and unidentifiable" in applications. Huawei Stack 8.3 integrates the above-mentioned technologies with digital intelligence, allowing enterprises to develop high-quality data products such as algorithms, models, and data sets to meet data usage standards; and exchange data spaces through EDS to allow data to be trusted, circulated, and authorized. Operation to ensure that data does not leave the domain.
Other manufacturers are also actively deploying in the data layer. Ant has specially launched a platform "Mos" for data circulation security, which provides independent external services. Productized page operations include distributed and centralized computing modes to provide safe matching, anonymous query, and security models. , safety statistics and other products.
After ensuring data security and circulation, when faced with intelligent application development, many companies are like "monks six feet tall and confused". Stack version 8.3 is a one-stop development platform that provides more than 30 core capabilities from coding standards, distributed construction, vulnerability checking to collaborative development, and established CodeArts, a software development production line that integrates processes, tools and experience.
Comparing the logic of ModleArts mentioned above, the logic is the same. Developers can find answers in CodeArts for almost all processes, tools, and services in software development. Haizhou Technology Company, a subsidiary of China State Shipbuilding Corporation, specializes in software business and development of ship systems. Based on CodeArts' structured process and contracted R&D, it ensures zero deviation in delivery, enables efficient collaboration of thousands of people, and creates a "source" of original technology in large-scale industrial fields.
Whether it is passive integration or active one-stop, it can be seen that the overall thinking of Huawei Cloud is not only to be in-depth but also to be broad, to create an AI cloud base for the intelligent era.
03 5+N+X, decoupled large model solution
There is no doubt that although large models are not everything about intelligence, they have become the soul of enterprise intelligence.
In the early stages of technology development, sometimes establishing standards is more important than developing the technology itself. This is also the thinking of Huawei Cloud. For example, autonomous driving technology experienced chaos and confusion in the early stages of its development. It was not until the industry classified autonomous driving technology into L0-L5 and the framework was clear that specific technological breakthroughs and implementations were achieved.
Huawei Cloud believes that thinking about large models should also be like this. Based on thinking, Huawei Cloud divided the large model into three levels: L0, L1, and L2, forming a three-layer decoupling architecture of 5+N+X.
Among them, 5 refers to 5 basic large models, including natural language (LLM), vision (CV), multi-modal, predictive decision-making (reasoning), and scientific computing. N is built based on a general large model. The industry's large model uses specific industry data and is based on unsupervised autonomous learning of industry knowledge to form a large industry model. It is also the main form of large model application in the industry at present; while X is L1 combined with scene data to form a scene large model. , to meet the needs of the industry.
From L0, L1 to L2, it follows the hierarchical model from "general" to "specialized". The general large model is used for qualitative, the large industry model is used for quantitative, and the large scenario model represents infinite possibilities. In such a hierarchical architecture, the rapid development process from L0 general model to L1 industry model to L2 special model can be completed.
According to Light Cone Intelligence, first of all, the Pangu large model adopts a complete layered decoupling design, which can be flexibly adapted to meet the changing needs of the industry. Enterprises can either load independent data sets for their own large models, or independently Upgrade the model.
But at the same time, the three-level models of large models can be interactively optimized. The L0 model can provide initialization for the L1 model to accelerate convergence. L1 can produce a stronger L2 model through model extraction and distillation. L2 can also feed back L1 by accumulating difficult case data or industry experience in actual problems.
Perhaps this is why Huawei Cloud can quickly launch nine major industry models at the same time. At the Full Connection Conference, Huawei Cloud "launched" nine major industry models in one breath, including large mining models, large meteorological models, large pharmaceutical models, etc., providing practical solutions for specific industries and specific businesses.
Different from the industry large models provided by other cloud vendors, Huawei's industry large models have already "run through a round" in actual business, and can play a key role with practical experience.
For example, as extreme weather has increased in recent years, there have been further requirements for the accuracy of weather forecasts. The Pangu Meteorological Model cooperated with the China National Meteorological Administration to make the accuracy of medium- and long-term weather forecasts exceed traditional numerical methods for the first time, and the speed increased by more than 10,000 times. . Specifically, the model was trained using 39 years of global weather data, and completed global 24-hour weather forecasts in just 1.4 seconds. It also improved accuracy for extreme weather predictions such as typhoon track forecasts.
Similarly, the coal industry faces challenges such as complex mining geological conditions, frequent disasters (such as coal dust, water, fire, gas, roof and other natural disasters), low production efficiency, harsh operating environment, and personnel shortages. To cope with these challenges, large-scale The model is trying to increase safety and reduce people as much as possible. Taking only one scenario of gas hazard prediction as an example, large models are used to conduct intelligent fusion analysis of underground coal mining, equipment, mine pressure, ventilation, safety monitoring, geology, gas extraction and other system data, and predict the gas concentration in key underground locations to achieve gas Early warning of hidden dangers.
Of course, this is due to Huawei's in-depth layout in the intelligent sensing layer, which can once again play a role and value in specific application scenarios.
In the future, hundreds of models will face increasingly difficult industries and more and more specific scenarios. Whether it is Huawei Cloud or other cloud vendors, they need to constantly think about the landing point and penetrate into countless capillaries to usher in qualitative breakthroughs. Change.
Welcome to follow Light Cone Intelligence and get more cutting-edge knowledge of science and technology!