The Sora show opens a new era of large models, how do you view the challenges of computing power?

Visual China.

Text: New Voices of the Metaverse, written by Jia Guipeng.

According to a McKinsey report, generative AI will add 4.4 percent to the global economy each year$4 trillion in value. Generative AI automates 60 to 70 percent of employee time to increase productivity. Between 2030 and 2060, nearly half of all jobs will be intelligent.

However, it is undeniable that with the continuous development of AI, its demand for computing power will continue to increase, which also makes the entire industry in "anxiety".

At present, generative AI led by large models, especially the release of GPT-4, has allowed human society to see the dawn of the era of general artificial intelligence. This means that artificial intelligence technology, as a productivity tool or even the "new infrastructure" of the digital age, is no longer limited to a single or limited scenario, but can think and solve problems like humans in many fields, and carry out continuous and rapid self-evolution.

With the explosion of generative AI, the value for businesses and individuals has brought many new opportunities and advantages:

Creative Content Generation:Generative AI can help businesses and individuals quickly generate creative content, including text, images, audio, and more. These contents can be used for advertising and marketing, brand promotion, content creation, etc., saving time and cost for enterprises and individuals, and improving the efficiency and quality of creation.

Personalized service and experience:Generative AI can generate customized services and experiences based on the user's individual needs and preferences. Businesses can leverage generative AI technology to deliver personalized product recommendations, customized services, and more to increase user satisfaction and loyalty.

Innovation & Product Development:Generative AI can help businesses and individuals quickly generate new ideas and concepts, facilitating innovation and development of products and services. With generative AI technology, ideas and possibilities can be quickly validated and explored, opening up more possibilities for innovation.

Data augmentation and extension:Generative AI can help businesses and individuals generate large amounts of synthetic data for data augmentation and extension. These synthetic data can be used for model training, data annotation, testing, and validation to improve the performance and generalization ability of the model.

Personal creation and entertainment:For individuals, generative AI can be a powerful tool for creation and entertainment. Through generative AI technology, individuals can quickly generate a variety of artworks, creative content, and entertainment products, expanding the boundaries of their own creation and entertainment.

However, it's worth noting that generative AI typically has a high demand for computing power, especially when training large models and processing large amounts of data. Because generative AI models often have complex structures and a large number of parameters, they require significant computational resources for training and inference. As generative AI technology continues to evolve and be applied, the demand for computing power will continue to increase, requiring more computing resources to support it.

With the rapid development of AI technology, more and more people have begun to pay attention to AI computing power. The problem of AI computing power refers to the large amount of computing resources required to process complex AI models and algorithms, which in turn consume a lot of time and energy.

Here are some of the aspects of generative AI's computing power requirements:

Model Training:Training a generative AI model is a very computationally intensive process. These models are typically made up of millions or even billions of parameters and require multiple rounds of iterative training on large-scale datasets. Due to the large number of computationally intensive operations such as matrix operations and gradient calculations that need to be performed during the training process, powerful computing resources are required to accelerate the training process.

For example, ChatGPT's training parameters have reached 175 billion, training data is 45TB, and 4.5 billion words of content are generated every day, and its computing power requires at least tens of thousands of NVIDIA GPUs A100, and the cost of a single model training is more than $12 million.

Inference and Generation:Once the generative AI model is trained, the inference and generation process also requires a lot of computational resources. In the inference process, the model needs to generate corresponding outputs based on the input data, which involves complex matrix operations and neural network calculations, so powerful computing power is required to ensure the inference speed and generation quality.

The overall computing power of the inference process of creating a conflicting single token is 2 large model parameters, so the daily computing power requirement of the inference side of the large model = the number of calls to the large model per day The average number of query tokens per person 2 The number of parameters of the large model, taking the Google search engine as an example, the number of calls per year is at least more than 2 trillion, and once it is combined with the large model, its AI computing power demand will be very considerable.

Model Optimization and Adjustment:Optimization and tuning of generative AI models also requires significant computational resources. After the model is trained, it is usually necessary to tune, fine-tune, and optimize the model to further improve the performance and generation quality of the model. These processes also require significant computing resources.

Large-scale data processing:Generative AI models often require training and inference on large-scale datasets, so they require powerful computing resources to process and manage this data. This includes data storage, reading, preprocessing, and other processes, which require efficient computing resources to support.

Therefore, we see that the demand for computing power is increasing with the development of AI technology and the expansion of application scenarios. AI model training and inference require a large amount of computing resources, especially complex models such as deep learning models, which require more computing power to train and run.

As this trend deepens, the semiconductor industry's revenue has grown exponentially. Gartner expects AI semiconductor revenue to continue to grow by double-digit growth during the ** period, growing by 25 percent in 20246%, reaching $67.1 billion (about 488.4 billion8.8 billion yuan), by 2027, AI chip revenue is expected to more than double the market size in 2023, reaching $119.4 billion (currently about 8692.).RMB 3.2 billion).

Therefore, in the face of AI's demand for computing power and the development of the computing power market, the computing power industry also has some measures to deal with it:

Improve hardware performance and efficiency:The computing power industry should continue to improve the performance and efficiency of hardware devices, including CPUs, GPUs, and TPUs, to meet the growing demand for computing power in AI technology. At the same time, it can also improve the energy efficiency ratio of computing equipment and reduce energy consumption costs by optimizing hardware design and manufacturing processes.

Promote innovation in computing technology:The computing industry should continue to promote the innovation of computing technology, including algorithm optimization, parallel computing, distributed computing, etc. Through continuous innovation and technical improvement, we can improve the performance and efficiency of computing devices, reduce computing costs, and meet the needs of AI technology.

Expand cloud computing services:The computing power industry can provide enterprises and individuals with flexible computing resources to meet the computing power needs of different scenarios by expanding cloud computing services. By providing cloud computing services, the flexible allocation and shared utilization of resources can be realized, and the utilization rate and efficiency of computing resources can be improved.

Strengthen industry cooperation and sharing:The computing power industry can strengthen cooperation and sharing with enterprises, research institutions and other related fields to jointly promote the development and application of computing technology. Through cooperation and sharing, we can make better use of the resources of all parties and improve the overall competitiveness and service level of the computing industry.

In summary, with the development and application of AI technology, the demand for computing power will continue to increase, and the computing industry should respond to this trend by improving hardware performance and efficiency, promoting technological innovation, expanding cloud computing services, and strengthening industry cooperation and sharing, so as to achieve the sustainable development and progress of the industry.

The scale of China's computing power industry has grown rapidly, with an average growth rate of more than 30% in the past five years, ranking second in the world. According to the characteristics of their own resource endowments and industrial advantages, all localities have formulated plans to accelerate the layout and development of the computing power industry.

On the one hand, we will speed up the opening of channels, promote the fast entry and exit of data, and successively build national Internet backbone direct connection points, dedicated channels for international Internet data, root server mirror nodes and national top-level domain name nodes, and accelerate the construction of the national "Eastern Data and Western Computing" southern route main road. On the other hand, through the construction of a computing power supermarket and a computing power scheduling platform, the computing power provider, the demand side and the upstream and downstream enterprises are organized to enter the docking.

Relevant data show that up to now, the total scale of data center racks in use across the country has exceeded 7.6 million standard racks, and the total scale of computing power has reached 197 trillion petascale floating point operations (197eflops), ranking second in the world, the world's first output of computing products such as servers, computers, and smart phones, and the construction of 130 trunk optical cables around computing hub nodes, computing power applications are widely used in government affairs, industry, transportation, medical and other fields, constantly giving birth to new technologies, new models, and new business formats, and computing power empowers thousands of industries to advance in depth, becoming an important fulcrum for the transformation and upgrading of traditional industries, and injecting strong impetus into high-quality economic development.

At present, the digital economy, as a new economic form, is becoming a key force in reorganizing global factor resources, reshaping the global economic structure, and reconstructing the global competition pattern. At present, China's computing power industry has begun to take shape, and is accelerating its expansion and deepening to the fields of government affairs, industry, transportation, and medical care, promoting the deep integration of the Internet, big data, artificial intelligence and the real economy, and continuing to empower thousands of industries.