Building AI-oriented infrastructure, Lenovo anchors the 'main channel' of AI computing power

There is no doubt that artificial intelligence is not a simple technological revolution for human society. It symbolizes the arrival of an era. Just like the industrial era to the agricultural era, it will bring about earth-shaking changes and affect human society. A process that takes hundreds or even thousands of years.

There is no doubt about the importance of AI computing power in promoting the application of artificial intelligence. Especially with the explosion of the entire generative AI market this year, it has brought huge demand for AI computing power. In order to ensure the further development of artificial intelligence applications, With rapid development, AI computing power can be said to have become a crucial factor.

Against this background, at the 2023 China Computing Power (Infrastructure) Conference held on August 18, Lenovo complied with the current demand for AI computing power and officially released the Lenovo AI computing power strategy, starting from "AI empowering intelligence and green empowerment." Starting from the two major characteristics of "power", we will comprehensively build AI-oriented infrastructure, and at the same time improve the overall layout of Lenovo's intelligent infrastructure with the dual brands of Lenovo servers, so as to promote the implementation of "universal intelligence" computing power in the future.

It can be seen that intelligence has become a new kinetic energy and driving force in China's current industrial upgrading and digital transformation of enterprises. Especially because the online and offline ecology and scenarios are far richer and more diverse than other countries, China's economic structure The differences in nature and the different application scenario requirements in subdivided industries have given huge imagination to innovations based on new technologies such as AI. This time, Lenovo proposed to build an AI-oriented infrastructure to promote "universal intelligence" in the future. "The implementation of computing power can not only help Chinese enterprises solve the problem of AI computing power demand encountered in the implementation of artificial intelligence applications, but also further accelerate the "go deeper and more practical" of AI, and better promote China's digital economy to move toward high quality. develop.

Two new trends in the evolution of AI computing power

At present, it has become the consensus of the entire industry that computing power is productivity. In particular, computing power, as a key productivity factor in the digital economy era, has become the core support for mining the value of data elements and promoting the development of the digital economy. With this year’s generative The emergence of AI applications has further promoted the demand for AI computing power to rise "exponentially".

Since 2018, with the launch of the Transformer large model, the computing power is no longer doubling every 18-20 months as in the past, nor is it doubling every 6 months during deep learning, but This is an increase of hundreds of times every year. According to an IDC report, AI computing power has become the main direction of China's computing power development, reaching 268 exascale operations per second (EFLOPS) in 2022, with a compound growth rate of 52.3% in the next few years.

According to Chen Zhenkuan, Vice President of Lenovo Group and General Manager of Lenovo China Infrastructure Business Group Server Division, the explosive growth of generative AI has also ushered in a new era of computing power for China’s computing industry—— The era of intelligent computing has arrived, and compared with the general computing in the past, intelligent computing has emerged with two new evolutionary trends, which are specifically reflected in:

On the one hand, unlike general computing power requirements, AI computing power infrastructure needs to be designed and optimized based on the characteristics of AI data and algorithms. The reason behind this is that there is a clear difference between AI computing power and general computing power, from tens of millions of parameters to hundreds of billions of parameters, from deep learning requiring medium-scale training and inference to now comprehensive To achieve ultra-large-scale training, AI computing power needs to be designed and optimized based on the characteristics of AI data and algorithms. Similarly, the AI ​​computing power infrastructure level also needs to be targeted and optimized based on the characteristics of "advancing with the times."

On the other hand, under the background of the country's vigorous development of green and low-carbon, it also poses new challenges to AI computing power, among which liquid cooling is a key technology that enables the development of green computing power. Indeed, the move towards green, low-carbon and sustainable development of data centers has become an "irreversible" trend, especially with the promulgation of policy documents such as national integrated big data centers and new data centers, and "numbering in the east and counting in the west". "With the official launch of the project, the scale and number of data center construction are showing a rapid growth trend, which also means that the "double carbon" task faced by the data center industry is more arduous and urgent, and it also has a greater impact on the "energy saving and reduction of computing power products". Platoon" puts forward higher requirements.

Based on such market changes, Lenovo proposed two major features of the AI ​​computing power strategy, namely "AI empowering intelligence and green empowerment". Among them, "AI empowers intelligence" refers to having 100% of computing power infrastructure products support AI, 50% of infrastructure R&D investment in the AI ​​field, and comprehensive AI technology covering all AI applications. After several years of forward-looking intelligent transformation, Lenovo has now achieved support for AI large model and ultra-large model training, as well as AI model training and reasoning capabilities, with parameter scales ranging from tens of millions to hundreds of billions. At the same time, Lenovo has launched more than 60 infrastructure products that support artificial intelligence applications, including servers, storage, etc.

"Green empowerment" refers to Lenovo's comprehensive layout and leadership in liquid cooling technology, creating a "lighthouse" of liquid cooling cluster solutions, and promoting the sustainable development of the intelligent computing industry with green computing power infrastructure. It is understood that Lenovo’s liquid cooling technology currently covers mainstream liquid cooling technologies such as air-liquid cooling hybrid, full liquid cooling, whole cabinet liquid cooling, and single-phase immersion liquid cooling. At the same time, Lenovo continues to develop and lead future liquid cooling technology to solve future kilowatt-level chip heat dissipation problems, improve memory liquid cooling coverage efficiency, and explore the ultimate in energy efficiency.

It should be said that Lenovo’s proposal of the concept of “AI Empowering Intelligence and Green Empowerment” can be described as “at the right time”. From the perspective of "AI empowering intelligence", under the general trend of industrial intelligence, the diversification of AI application fields and the continuous expansion of market scale have created huge demand for AI computing power, and Lenovo has continued to invest in AI computing power. Investing in building AI-oriented computing infrastructure can accelerate the implementation of artificial intelligence in thousands of industries, thereby promoting the intelligent transformation process of various industries and even the entire society; from the perspective of "green empowerment", in green Under the background that low carbon has become a "certain" trend in the future, Lenovo focuses on the most urgent needs of customers and creates new computing infrastructure with innovative products and solutions in the field of liquid cooling. It can also support the digital transformation and development of various industries. The development of the digital economy provides stable and reliable green computing power.

On this basis, Lenovo adheres to the value proposition of "universal computing power" and hopes to make computing power a public resource like electricity, so that everyone can use universal and intelligent computing power. Chen Zhenkuan pointed out that "universal computing power" includes "universal" and "wisdom." "Universal" means that computing power, like water and electricity, can be used by everyone without seeing it - people can use it at any time, but do not need to see it; "wisdom" means that everyone can use it at any time Need to use without asking - People do not need to ask whether computing power can meet the requirements before using it, they can use it whenever they need it.

It is not difficult to see that with the explosion of generative AI, the development of the entire computing power industry has entered a new stage, which also forces the entire computing power technology innovation to evolve accordingly, and Lenovo takes the lead The release of a new AI computing power strategy in the industry is inseparable from Lenovo's forward-looking insights into the development of AI technology. It can be said that it has made sufficient preparations for this outbreak of generative AI; and based on a large number of technological innovations and business practices, It also lays the foundation for Lenovo's two strategic features of "AI empowering green empowerment", which can also point the way for the continued innovation and evolution of AI computing power in the future, and lead the entire computing power market into a new era of intelligent computing. era.

What’s even more commendable is that Lenovo has always adhered to the value proposition of “universal computing power” during this process, which will undoubtedly accelerate the promotion of AI computing power as the infrastructure of the digital economy to play a greater role in hundreds of industries. Value, faster opening up the "last mile" of the AI ​​industry, not only providing key technical support for very large models represented by generative AI, but also accelerating the entry of AI into various core production links in the industry, better empowering Thousands of industries.

The precipitation and release of Lenovo AI computing power

It is worth mentioning that the Lenovo Wentian WA7780 G3 AI large model training server and Lenovo Wentian WA5480 G3 AI training and promotion integrated server released by Lenovo at the 2023 China Computing Power (Infrastructure) Conference are Lenovo’s AI-based The latest “representative” product in a oriented infrastructure strategy.

In this regard, Chen Zhenkuan said that the two new AI servers will continue Lenovo Wentian's brand value proposition of "local innovation, agility and efficiency" and inherit the Lenovo server's "three highs and one" of high performance, high reliability, high scalability and low power consumption. low" characteristics.

Among them, Lenovo Wentian WA7780 G3 is a server specially built for AI large model training. This server is designed to break the dilemma of AI computing power supply and can provide strong computing power support for training. In terms of performance, compared with the previous generation product, Lenovo Wentian WA7780 G3 AI large model training server can increase the speed of large model AI training by up to 9 times, and can increase the speed of large model AI inference by up to 30 times.

In terms of flexible architecture, in order to meet the needs of high-speed data communication between GPU servers in AI very large model training scenarios, Lenovo Wentian WA7780 G3 AI large model training server also supports a variety of external network connection solutions such as IB and RoCE. It can support up to 8 RDMA high-speed network cards and provide 3.2Tb/s aggregate bandwidth, which fully meets the cross-node communication needs during parallel computing of very large model training; in the design of local storage, GPU Direct Storage (GDS) technology is adopted . Through GDS technology, GPU and NVMe achieve fast and low-latency interconnection, which can effectively increase the speed of training data loading and improve the performance of large model training.

In terms of energy saving and high efficiency, this product adopts a triple independent air duct design, and systematically optimizes the heat dissipation characteristics of different components, effectively reducing the heat dissipation power consumption caused by the fan. Compared with products of the same level, the power consumption of Lenovo Wentian WA7780 G3 AI large model training server is reduced by about 10%. At the same time, in order to meet the requirements of future intelligent computing centers for lower PUE values, this product also reserves a liquid cooling cold plate design, which will support liquid cooling in the future, which will greatly save the system heat dissipation during the training process of large AI models. Due to the additional energy consumption, the PUE value is expected to be lower than 1.1.

AndLenovo Wentian WA5480 G3 AI training and push all-in-one server is a 4U rack-mounted AI server under the Lenovo Wentian brand. It supports multiple computing power and rich ecology, and can provide training and reasoning for AI models. Provides extremely reliable computing power. In terms of multi-computing power, this product supports multiple types and brands of AI accelerator cards, and can be flexibly used in a variety of application scenarios such as AI general model training, large model reasoning, AI generation, cloud games, and scientific computing.

Not only that, Lenovo Wentian WA5480 G3 AI training and promotion all-in-one server also adopts a "flexible and changeable" design concept. Based on different AI workloads, it can provide enterprises with a variety of CPU-GPUs including pass-through, Balance, and Common. The interconnection method avoids potential performance bottlenecks and reduced system efficiency caused by a single CPU-GPU data communication method that does not match the workload. At the same time, combining different types and numbers of accelerator card options can achieve perfect matching for various complex AI scenarios.

In addition, this product also takes full consideration of design redundancy to provide a safe and stable operating environment for AI computing power. Supports N N redundancy of power supply. The power supplies within the nodes are mutually backed up. Dual-channel mains power can be connected externally to ensure that the server operates online 24 hours a day. Supports power consumption monitoring of key components such as CPU, GPU, memory, etc., and caps the power consumption of the entire machine to dynamically ensure that the server runs within a reasonable power consumption range to ensure stable performance, as well as multiple safety designs such as fan N 1 redundancy to ensure high Ambient temperature and continuous operation.

Chen Zhenkuan emphasized: "These two products respond to different scenarios and needs respectively. The difference between the two can be seen from the application and strategic levels. At the application level, Lenovo Wentian WA7780 G3 AI large model training server has the ultimate The computing power to support ultra-large-scale training; Lenovo Wentian WA5480 G3 AI training and promotion all-in-one server supports more diverse computing power, suitable for algorithm training and reasoning. These two products will help Lenovo use comprehensive AI technology to support the entire AI domain. Application. On this basis, the new AI server can also cooperate with inference servers, edge AI devices, etc. to further improve and enrich Lenovo’s AI computing power infrastructure product portfolio."

According to IDC data, Lenovo Group is the fastest growing AI hardware infrastructure (server and storage) provider in 2022, with a year-on-year growth of 139% and ranking third in the world. This also confirms Lenovo’s strong capabilities in the field of AI infrastructure. The strength and capabilities of , while supporting the country in building green computing power while reducing the cost of obtaining computing power.

"Puhui" computing power realizes new value

With the launch of Lenovo's AI-oriented computing power infrastructure product portfolio, Lenovo's "full stack intelligence" strategy has also formed a complete "closed loop", which is also important for Lenovo's promotion of "universal intelligence" computing power. will have a greater superposition effect. Based on the forward-looking layout of the intelligent transformation strategy, Lenovo has now built AI-embedded smart terminals, AI-oriented infrastructure and AI-native solution services, thus forming a full-stack intelligent product and solution service layout.

Objectively speaking, the release of Lenovo’s AI-oriented computing infrastructure product portfolio and the complete display of products and solution services with a “full-stack intelligence” layout can not only continue to promote the upgrading of China’s industrial intelligence , and at the same time, it has laid a key foundation for AI to truly "take root" in thousands of industries, and behind it it also reflects three new values.

First, it provides solid technical support for the integration of training and promotion of large AI models. With the continuous breakthroughs in key technologies and basic theories, big computing power, big data, and big models will become the key development directions of future AI development, which will lead to the subdivision of natural language processing, translation models, generation algorithms, and data sets. Elements continue to improve, promoting output with richer content details, types, and higher quality.

And Lenovo has built an AI-oriented computing infrastructure that fully integrates the two strategic features of "AI empowering and green empowerment". It can be said that it will effectively provide technical support for the integration of AI super-large model training and promotion. In specific applications, you can also see that Lenovo’s independently developed Poseidon warm water water cooling technology reduces the PUE value to 1.1, and the heat dissipation efficiency can be as high as 98%. Benefiting from this, the computing density of Shanghai Jiao Tong University's "Siyuan No. 1" liquid-cooled cluster reaches five times that of the general computing center, while the required computer room area is only 1/10 of the general computer room.

Second, it will greatly promote the implementation and implementation of AI in various industries in the future. With the explosion of generative AI, the demand for AI computing power continues to show an "exponential" upward trend. More and more industries have emerged with clear artificial intelligence application scenarios, which has also led to the rise of AI computing power. The demand for computing power is higher than the supply, forming a "gap" in supply. Only by providing solid support for AI computing power through AI-oriented computing power infrastructure can the contradiction between supply and demand be better resolved.

Third, it accelerates the “going deeper and more practical” of AI in industries and industries. In the era of digital economy, computing power as the "underlying power" has a wide-ranging and profound impact on economic and social development, and Lenovo's product portfolio is based on AI-oriented computing power infrastructure, as well as products and solution services with a "full-stack intelligence" layout , it can also accelerate the "going deeper and more practical" of AI in industries and industrial fields.

In fact, as China's leading intelligent solution service provider, Lenovo has currently empowered multiple vertical industries and fields, including smart cities, smart parks, smart manufacturing, smart transportation, smart education, smart energy, smart finance, etc. .

As Chen Zhenkuan said at the end: "In the future, Lenovo will use its AI-oriented infrastructure and the advantages of 'full-stack intelligence' to help the popularization and penetration of artificial intelligence in thousands of industries. At the same time, it will use the Lenovo Wentian brand to localize The advantages of innovation empower local Chinese customers. Through leading AI technology, covering all areas of application, accelerating the realization of 'universal intelligence' computing power, leading China's intelligent transformation."

Generally speaking, the transformation and upgrading of China's industrial intelligence and the implementation and development of China's digital economy have a long way to go. Lenovo's continuous innovation of its AI-oriented computing infrastructure product portfolio, coupled with the "full It is believed that the leading advantages built by the products and solution services laid out by "Stack Intelligence" will be able to further accelerate the "going deeper and more practical" of AI and release the "multiplier effect" of intelligent productivity. In the process of intelligent transformation of thousands of industries in China, , and continue to contribute better value.

The above is the detailed content of Building AI-oriented infrastructure, Lenovo anchors the 'main channel' of AI computing power. For more information, please follow other related articles on the PHP Chinese website!