AIGC Computing Power Panorama and Trend Report Released! One article interprets AIGC's computing power composition, industry chain, and five new trend judgments

Source: Qubit

AI computing power has never been more eye-catching than it is now.

Since the emergence of the trend of large models, the number and scale of large models have increased dramatically in just a few months.

The tens of billions of hundreds of billions of large models have soared to dozens, and the trillion-parameter large model has been officially born.

Under such magnitude changes, the demand for computing power shows drastic changes.

Model-level companies are snapping up computing power services at almost all costs. Nvidia’s market value once exceeded one trillion U.S. dollars, and the cloud computing market is being reshaped at an accelerated rate…

There is no doubt that computing power is the same basic energy as hydropower and oil for the development of AIGC industry.

After the prelude to the AIGC era, how to understand the computing power industry is particularly important.

What kind of computing power does the enterprise need? What changes will happen to the computing power industry due to the rise of AIGC? What is the current composition of the computing power market?

“AIGC Computing Power Panorama and Trend Report” is here to help you understand these issues.

In the report, the qubit think tank systematically analyzed the composition of AIGC computing power and the industrial chain, and further pointed out the five new trends of AIGC computing power** and the development forecast of three stages.

Core ideas include:

• Driven by AIGC, chips compete for high performance and large computing power, and introduce a new computing architecture;
• The AI server has sprung up suddenly, and the bonus curve is trained first and then reasoned;
• MaaS reshape cloud service paradigm, AIGC business model closed loop;
• The AI model all-in-one machine is ready to come out, and the traditional industry is “out of the box”;
• Intelligent Computing Center escorts the operation of AIGC, and the computing power leasing model has become a new solution; ……

Let’s look at the details one by one.

Domestic server manufacturers’ business growth exceeds 30%

From the analysis of the current industry status quo, the main body of the industry mainly includes:

• chip
• AI server (cluster)
• cloud computing

Chip layer: AIGC two routes provide computing power

In computing chips, there are currently two mainstream routes in the industry to meet the computing power needs of the AIGC industry.

One is the GPU route represented by Nvidia, which is called universal chip.

The other is the ASIC route represented by Huawei and the Cambrian, which is called special chip route.

At present, these two routes bring together different types of players, and the computing tasks they undertake are also different.

Under the general-purpose chip route, it can complete diverse computing tasks and is suitable for large-scale parallel computing.

That is, general-purpose chip (GPU) is more suitable for AIGC’s current computing power.

The advantage of the dedicated route is reflected in the better energy-efficiency ratio in specific scenarios. Since special-purpose chips are designed to perform specialized or customized tasks, they can achieve better energy-efficiency ratio and computing efficiency than general-purpose chips in specific scenarios**.

Just because dedicated chips can release greater computing efficiency in specific scenarios, it has also become the technical route chosen by Internet and other cloud vendors when developing their own chips.

Usually, the self-developed chips of Internet cloud manufacturers mainly serve their own products, emphasizing the maximum release of chip performance in their own ecology.

### Server layer: business growth is obvious, mainly based on Internet customers

AIGC’s demand for high-performance computing has made AI servers the fastest-growing segment in the server field.

Large model training, such as GPT-3, requires a lot of computing resources and memory, and usually involves using thousands or even tens of thousands of GPUs to speed up training.

Because these calculations have very high requirements on chip performance, specialized hardware and software are required to support massively parallel computing and high-speed data transmission.

AI servers are servers specially designed to handle artificial intelligence workloads, using specialized hardware accelerators (such as GPUs, TPUs, etc.), as well as high-speed network connections and storage to provide high-performance computing capabilities.

In contrast, CPUs (general-purpose servers) usually cannot meet AIGC’s demand for extreme computing power, and their computing power, memory and storage capacity are usually low. In addition, CPUs usually do not have dedicated hardware accelerators to provide high-speed computing.

Therefore, large-scale model training needs to rely on AI server clusters to provide computing services.

According to the research of Qubit Think Tank, after the outbreak of AIGC this year, domestic server manufacturers have generally increased their business by more than 30%**.

Recently, TrendForce also raised the compound annual growth rate of AI server shipments from 2022 to 2026 to 22%. Behind the surge in AI server business, the biggest buyers are still Internet companies.

In 2022, big manufacturers such as ByteDance, Tencent, Alibaba, Baidu, etc. will become the main purchasers in the proportion of AI server procurement. This year, the enthusiasm for large-scale model research and development has driven the purchase demand of downstream Internet companies, making it still the largest buyer of AI servers.

### Cloud Computing: MaaS Reshape Service Model, Old and New Players Restructure Competitiveness

The MaaS model was first proposed by Ali, and then major Internet companies and artificial intelligence companies (such as SenseTime) have introduced the MaaS model.

In addition, companies such as Internet giants and Huawei have already used self-developed chips in the construction of MaaS bases.

In 2023, leading domestic cloud manufacturers will successively launch their own MaaS platforms, based on large model bases, to provide one-stop MaaS services for enterprises with limited computing resources and lack of professional experience.

For cloud vendors, the main purpose of MaaS services is to help customers quickly build industry-specific large models. Based on this, the competition dimension between cloud vendors has changed to computing power infrastructure, general large-scale model capabilities, and AI platform/tool capabilities.

### Status of Intelligent Computing Center: infrastructure-level AI computing power supply, creating a new engine for regional economic growth

From the perspective of computing equipment distribution, in the server and AI server market, Beijing, Guangdong, Zhejiang, Shanghai, and Jiangsu rank in the top five, with a total market share of (server and AI server) of 75% and 90% (2021 data) .

From the perspective of supply, most of the intelligent computing centers are located in Eastern and Central provinces, and the AIGC business needs to process massive data, which leads to the high cost of computing power resources in the East**.

Moving tasks with high computing requirements such as large-scale model training to the western region forms “East Data Training”, which can effectively reduce costs and achieve the optimal comprehensive cost of computing network resources.

Specifically, in order to solve problems such as the unbalanced supply and demand of computing power demand, it is necessary to transfer the computing power and data processing tasks in the east to the western region with lower costs through computing power scheduling. Among them, optimizing the interconnection network between the east and the west and the direct connection network between hub nodes is the key to improving the level of computing power scheduling.

From the perspective of demand, the demand for AIGC computing power mainly comes from manufacturers who develop AIGC large models, mainly distributed in the Beijing-Tianjin-Hebei region, the Yangtze River Delta and the Greater Bay Area.

Chip server revolution is emerging

The trend of large models puts forward new requirements for computing power and leads to new changes in the underlying hardware. Let’s look at two levels of chips and servers:

chip level

At present, among the high-performance chips, Nvidia A100 has an absolute advantage, and A100 only has stock in China but no increase. The domestic market will give domestic GPU manufacturers more opportunities**.

In addition, at the chip foundry level, there is currently no domestic foundry that can undertake 7nm and above processes, and most GPU manufacturers choose mature process + advanced packaging solutions to achieve higher performance indicators.

At the server cluster level, high-performance computing is realized through multi-card multi-machine parallel computing and high-performance network.

Since the hardware product + CUDA ecology built by Nvidia is difficult to break through in 10 years, in the future when the high-performance GPU is limited, analysts predict that there will be two main solutions for the hardware layer, the last one is to develop GPU + inter-chip Interconnect technology to achieve massively parallel computing**.

The other is to jump out of the von Neumann architecture and develop an integrated storage and computing architecture to integrate computing units and storage units to achieve an order of magnitude improvement in computing energy efficiency.

At the software level, sparse computing and building a high-performance network are currently two solutions.

The innovation of sparse computing is reflected in the algorithm level. By deleting invalid or redundant data (such data is usually huge), the amount of data calculation is greatly reduced, thereby speeding up the calculation.

The purpose of building a high-performance network is to reduce the training time of large models. By building a high-performance network, each computing node has an ultra-high communication bandwidth, bringing several times the improvement in traffic performance, thereby shortening the training time of large models.

### Server level

The demand for computing power of AI large models is increasing exponentially, making AI servers with higher configurations the main carrier of AIGC computing power.

Compared with traditional servers, the computing, storage and network transmission capabilities of AI servers can reach a higher level.

For example, the configuration of NVIDIA DGX A100 server with 8 GPUs and 2 CPUs is much higher than that of traditional servers with 1~2 CPUs.

In my country, the Intelligent Computing Center is a public infrastructure platform that provides computing power resources for artificial intelligence (large models), and its computing power units are mainly AI training servers and AI reasoning servers.

With the evolution of large models, the main demand for future AI servers will shift from training to inference. According to IDC’s forecast, by 2026, 62.2% of AIGC’s computing power will be used for model reasoning.

Industry changes spawn new business opportunities

Furthermore, the trend of AI large-scale models brings new opportunities to the computing power industry, and new paradigms, new products, and new infrastructures are emerging.

New game rules: MaaS reshape cloud service paradigm, AIGC business model closed loop

MaaS (Model as a Service) embeds large models in computing power, algorithms, and application layers, integrates applications with intelligent bases, and unifies external output.

The essence of MaaS is to refine and integrate common basic technologies in the industry into services to meet the needs of various application scenarios.

In the process of commercialization, large-scale model capabilities and supporting middleware tools will become new dimensions for enterprises to consider for cloud computing vendors.

The discriminant of cloud computing service capabilities has shifted from the computing power level to the “cloud-intelligence integration” capability. In addition to the computing power infrastructure, the core competitiveness has changed to the ability to build computing power, models, and scene applications into standardized products.

### **New species: AI model all-in-one machine is ready to come out, traditional industries “out of the box” **

The AI model all-in-one machine deeply integrates software and hardware. According to the different needs of enterprises, the corresponding products or solutions are deployed on the AI server in advance and packaged to form a complete set of solutions.

The cost advantage of the AI model all-in-one machine is mainly reflected in the following three points:

• 1. The overall purchase price is lower than the price of separate purchase of software + hardware;
• 2. It takes a long time for the enterprise to purchase the server separately and hand it over to the AI enterprise to deploy the software. The AI model all-in-one machine can be used out of the box, reducing the delivery cost;
• 3. The number of required servers is greatly reduced, saving space costs for customers.

### New Infrastructure: Intelligent Computing Center escorts AIGC operations, computing power leasing mode becomes a new solution

The computing power leasing model can effectively lower the threshold for large-scale model development. For small model companies in vertical industries that do not have the strength to purchase enough AI servers, the public computing power basic platform will help small and medium-sized enterprises build their own required models.

Enterprises do not need to purchase servers, but can access the computing power center through a browser and use computing power services.

For small and medium-sized enterprises, there is no need to rely on the large model base built by cloud vendors for secondary development, but to develop small models of vertical industries by renting computing power resources from public computing power platforms.

Industry Development Forecast

To sum up, the qubit think tank predicts the future development of AIGC, which can be divided into three stages:

• AIGC infrastructure period
• AIGC development period
• AIGC business period

AIGC Infrastructure Period

At present, most companies in the AIGC model layer are in the pre-training stage, and the main source of demand for chips is GPU.

In the initial stage, high-performance GPU manufacturers will become the biggest beneficiaries.

However, there is a big gap between domestic GPU manufacturers and Nvidia at present, and the first beneficiary is the dominant party in the stock market.

Therefore, domestic AI server manufacturers are strong suppliers at this stage. At present, the domestic AI server field has been in short supply.

### AIGC development period

In the mid-term stage (within 5 years), the computing power layer is a process of leaning from training to reasoning.

At this stage, reasoning chips will become the main demand side. Compared with the high computing power and high power consumption of the GPU and the corresponding waste of computing power, the inference chip pays more attention to the chip’s computational efficiency ratio, and has better control over power consumption and cost. In addition, this stage will also be an opportunity for innovative chips.

Analysts expect more market opportunities for memory-computing integrated chips, brain-inspired chips, and silicon-optical chips.

In the inference phase, edge computing will have more opportunities than cloud computing in the training phase.

First of all, the applications corresponding to the inference stage tend to be diversified, and the diverse requirements make cloud computing generate more waste of computing power and low computing efficiency.

Second, edge computing can provide sufficient computing power for large model reasoning.

At this stage, the dividend period of AI server manufacturers will gradually peak, and demand will shift to lower-cost general-purpose servers; chips will also shift from GPU to NPU/ASIC/FPGA/CPU and other forms coexist. Among the routes of domestic innovative chips, ** is optimistic about the development of integrated storage and computing architecture**.

### AIGC Business Period

The underlying innovation advantages of chips have begun to emerge, and chip manufacturers with truly innovative technologies such as storage and computing integration, photonic chips, and brain-like chips have increased their role in the market.

The types of chips required for AIGC computing power are more diverse.

Enterprises at this stage have a more comprehensive consideration of computing power, not only considering the size of computing power, power consumption and cost may exceed the size of computing power and become the indicators that companies at each model layer care about.

Report Get

Click on the link below: