Article**: Titanium**
No industry is looking forward to large models more than the Internet, just like Bole is looking forward to wearing a thousand miles of horses.
As a representative of emerging technologies and business models, the Internet naturally relies on new technologies and new waves, and Internet companies are always at the forefront and quickly transform technology into products and services.
Past history has proved the correctness of the Internet to make good use of new technologies, the transition from PC Internet to mobile Internet, communication networks from 2G to 5G, Internet terminals from computers to mobile phones, and digital infrastructure from traditional IT to cloud computing.
The large model is a new technology with greater potential energy, and the Internet industry must not miss it, especially in today's industry situation, the Internet industry has collectively encountered growth bottlenecks, and the dividends brought by the last wave of technology have gradually faded, and Internet companies are exposed to a more severe external environment, and the large model is the best answer that the Internet industry can find.
But this time, Internet companies are faced with a very different proposition. In the chain from technology to business, the key link of AI computing power is missing, and AI computing power is no longer as readily available as general computing powerAt the same time, the landing of large models and the implementation of commercial innovation should also be cautious and cautious. The Internet industry has begun to rethink, how to complete the application of large models
HUAWEI CLOUD's Ascend AI cloud service has recently attracted a lot of attention, and the author notes that the iterative upgrade of cloud services and the dual requirements of large-scale AI computing power are redefining the cloud services required by the Internet industry.
Internet companies are flocking to HUAWEI CLOUD
Over the past two decades, the Internet has had two distinct faces, alternating between the two faces. On the one hand, the Internet is almost a supporter of all new technologies, no matter how the technology changes, the Internet can always superimpose new technologies on top of the old technology, e-commerce, games, ** and other new formats, in the final analysis, are the application of technology.
Cloud computing is the digital infrastructure of Internet enterprises, the Internet is a generation of enterprises native to the cloud, cloud computing enables enterprises to achieve higher efficiency and flexibility at lower costs, compared with non-cloud native enterprises, Internet companies are always able to launch products one step faster, and Internet cloud service providers have also become the leading manufacturers in the first half of cloud computing.
A cloud vendor sales line person said, "The Internet is a bit of a 'buffet' feeling, if you like to eat, you can eat, if you don't like to eat, they will leave, they pay more attention to efficiency and direct resource exchange." ”
In 2019, HUAWEI CLOUD officially entered the Internet track, at that time, HUAWEI CLOUD was not gaining momentum in the Internet industry, and Internet companies had basically adopted the products of Internet cloud service providers.
For example, an Internet customer had investment from an Internet giant and used all Internet cloud services in the early days, but in the follow-up service process, the Internet cloud service provider was unable to help the customer solve the problem.
Cases such as "tooth extraction" are commonplace in HUAWEI CLOUD's view, which is also an important reason why HUAWEI CLOUD has been able to gain a firm foothold in the Internet industry, which is not dominant. HUAWEI CLOUD's approach is destined to be slow to come by, but once the cooperation is reached, HUAWEI CLOUD will be able to serve customers for a long time.
From 2019 to 2021, HUAWEI CLOUD has embarked on a long industry battle, from 5G commercial use, to multi-cloud options, to escalating data security issues, the Internet has begun to choose more and more reliable HUAWEI CLOUD, and HUAWEI CLOUD has also changed from an Internet cloud to a new business cloud choice.
2021 is undoubtedly a watershed for the Internet, after 20 years of "rapid growth", most of China's Internet companies have experienced a painful process of shrinking market value and slowing or decreasing revenue growth, but on the other hand, the market value of American Internet companies continues to soar. The academic community condemned: "The backwardness of core technologies, the difficulty of Internet companies to go to sea, and the lack of support for the real economy are the shortcomings of China's Internet." ”
For a time, China's Internet began to transform from rapid expansion to industrial empowerment, however, the industry-oriented Internet cannot be quickly replicated, because each industry has a threshold set by professional knowledge, and the empowerment of the Internet needs to slow down. Both the simple and efficient consumer Internet model and the extensive management model need to be changed.
At this time, HUAWEI CLOUD was keenly aware of the new changes in the market and proposed new value propositions for the Internet: "realistic", "new", and "outward". First, the direction of Internet innovation returns to technological innovation, penetrates innovative application scenarios into the industrial field, and empowers the real economyThe second is to adhere to hard-core technological innovation and enhance the resilience of sustainable development of enterprisesThird, Internet companies are actively expanding the global market and passing on their experience to the global market to achieve new growth.
Even without today's large model, the Internet industry is slowly migrating to HUAWEI CLOUD, placing some of its core businesses on HUAWEI CLOUD to obtain a secure and stable experience and new business growth.
Large-scale model computing power, redefining cloud services
In 2023, OpenAI will open the era of large models. Large models require large computing power, and AI computing power is inevitable in the reshaping of cloud services, which has the effect of "one force reduces ten meetings", which is AI computing power.
Large-scale pre-trained models require a large amount of AI computing power to process massive amounts of data and perform a large number of calculations to learn the internal rules and manifestations of Xi data, and this training process requires high-performance AI chips and large-scale computing resources to complete the training of the model quickly and accurately.
The reality is that Internet enterprises often do not have the conditions to build large-scale computing power clusters, first of all, the hardware investment is huge, such as GPU, TPU and other chips are expensive, and secondly, the technical threshold is high, from the hardware selection, system architecture design, cluster management, performance optimization, etc., professional and technical personnel are required, in addition to data storage, management, transmission, performance stability and efficiency of scalable clusters, as well as follow-up operation and security guarantees, large model training is a complex system engineering.
Cloud service providers have simplified the complex operation of large model computing power clusters to the greatest extent, so that Internet enterprises can focus on the large model application layer, but this does not mean that the complexity has disappeared, but cloud service providers to solve the hardware and software problems of large models.
Both overseas and domestic are trying to find a solution to replace Nvidia's GPU. At the recent Re:Invent conference, AWS CEO Adam Selipsky announced its next-generation GR**ITON 4 and Trainium 2 chip hardware, with a focus on meeting the needs of machine Xi (ML) training and generative artificial intelligence (AI).
Domestic cloud service providers have also followed up, but the extent is far less than AWS, which is due to the limitation of investment resources and the lack of strategic awareness. Due to overseas sanctions and other reasons, Huawei's investment in AI computing power has been earlier and more resolute, and when the popularity of large models has raised AI computing power to a new level, HUAWEI CLOUD Ascend AI computing power combined with cloud services has become the preferred application model for Internet enterprises.
At the lowest level, Huawei has built an AI computing cloud platform based on Kunpeng and Ascend, as well as a heterogeneous computing architecture such as Cann, an all-scenario AI framework MindSpore, and an AI development production line, ModelArts, to provide key capabilities for the development and operation of large models, such as distributed parallel acceleration, operator and compilation optimization, and cluster-level communication optimization. Based on Huawei's AI root technology, the training performance of large models can be tuned to 11 times.
Based on the Ascend cloud service, HUAWEI CLOUD also provides a full set of tool chains, such as the large model engineering suite, covering three major links: data engineering, model development, and application development. Compared with traditional annotation platforms, HUAWEI CLOUD's data engineering suite provides prompt-assisted composing functions for SFT training, and multi-person RANK** annotation and task assignment functions for RLHF training. Compared with offline, the measured efficiency of these two tasks can be increased by 3 times.
The development of cloud services to the underlying chip is not only the ticket to the next stage of cloud service providers, but also the life and death line of survival.
The hard things that Huawei has done are when it comes time to realize the value
The reason why Huawei is a respectable company is that it not only considers the maximization of business interests, but also does some things that cannot be of value in the short term, or even that may not have value in the long term.
The good thing is that when the pace of history falls at a certain point, Huawei's previous accumulation can play a great role. For example, for example, the Ascend AI cloud service, for a long time, domestic enterprises still had the illusion that AI computing power would be as available as general-purpose computing power, and they only needed to stand on the shoulders of giants and innovate to replicate new successes.
When technology is dressed up in politics, Huawei does not give in, but attacks the problems that the outside world does not dare to solve, from the software tool chain to break through the Wujiang River, to MetaERP to forcibly cross the Dadu River, GaussDB to establish the first brand of full-stack independent database, and then to the Pangu model of "not writing poetry, only doing things".0 releases, all of them.
At first, these products and services may seem to have little to do with the Internet, but the technological superposition of the Internet industry must be in a new direction, rather than a sense of existence in an already rotten field.
China's cloud computing industry has been developing for more than ten years, and there has never been a shortage of people to do simple things, but the barriers built by the first-mover advantage and scale will always be broken by new needs little by little, such as the changes in the cloud of Internet enterprises.
The theory of constraints states that any complex system will be determined by a few constraints, and once these few constraints or even the core constraints are found, they can be used to leverage the whole system.
When the cloud service reaches the era of large models, the decisive factor is the AI computing power cloud service, and the cloud service without AI computing power will be eliminated, and there is no lasting AI computing power cloud service, and there is also a lack of imagination space in the future.
The Internet industry has reached a turning point in updating its technology stack, and HUAWEI CLOUD has done a lot of difficult things, and it is time to realize the value.