On May 13, 2024, OpenAI released a new flagship model, GPT-4o, which can deduce cross-format information in real time. A step further in the field of natural human-computer interaction, you can input and process any combination of text, audio and images, and imagine generating any combination of text, audio and image output. GPT-4o (the "o" stands for "omni") has an audio input response time of just 232 milliseconds, and its average (320 milliseconds) is similar to that of a human conversation. The GPT-4o's English text and encoding capabilities are on par with the GPT-4 Turbo, and non-English text features are significantly improved, while the application program interface (API) is faster and costs 50% less. In particular, it is excellent in visual and audio understanding, which is better than previous models. The average latency of speech modes in ChatGPT models was 2.8 seconds (GPT-3.5) and 5.4 seconds (GPT-4), respectively. Such speech patterns are piped by three independent models: first a simple model converts audio to text, then the text is received and output by GPT-3.5 or GPT-4, and then a third simple model converts text back to audio. In this process, the intelligent agent GPT model cannot directly perceive tones, multiple speakers, or background noise, nor can it output laughter, singing, or expressing emotions, so it loses a lot of information.

GPT-4o is a new AI model trained end-to-end in text, vision and audio, combining all existing patterns, all inputs/outputs processed by the same neural network, and both capabilities and limitations are being explored.

Most of the large AI models represented by GPT-4o have the following life cycle links:

• Research and development stage: The research and development department conducts a lot of experiments to determine the model characteristics (parameters) and effective algorithms.

• The training phase: A large amount of data is fed into the model and output is generated, and the machine learns and creates the internal structure of the data

• Fine-tuning stage: Fine-tune the training results to further improve the performance of the model

• On-line stage: After on-line, the model processes user requirements in real time and processes massive data

AI model belongs to High Performance Computing (HPC), in order to achieve low time delay and high speed, powerful, reliable and controllable computing power runs through every link of its life cycle. Therefore, large computing power is the basis, and since the amount of data required for video training is much larger than that for text training, the continuous development, updating and widespread use of AI models will further increase the demand for computing power upgrading.

In order to meet the centralized training needs of AI models and accelerate the large-scale model training of China's transportation, medical, education, energy, finance and other industries, the 2024 China Mobile Computing Network Conference, which closed on 4/29, showed the new infrastructure system of integrated computing network, and officially launched three autonomous controllable intelligent computing centers (with a total scale of nearly 60,000 Gpus) in Harbin, Hohhot, and Guiyang. At the same time, the first nine other intelligent computing centers were officially commissioned, providing a total capacity of 11ExtraFLOPS. Among them, Hohhot intelligent Computing Center is the largest single liquid-cooled intelligent computing center of global operators, which has attracted wide attention. At the same time, the world's largest 400G OTN backbone network based on independent research and development technology, that is, the computing power optical network is the core pillar of the above computing power network functions. It includes highly reliable and high-quality connection based on optical fiber and cable, super bandwidth transmission and extensive access based on optical signal, and high coordination, intelligent management and control and security guarantee based on optical transmission.