Shulou(Shulou.com)11/24 Report--

On August 24, the 2023 wave information new product "Internet + AIGC" with the theme of "Smart Computing New Bureau Innovation Machine" and the OGAI press conference of the large model intelligent computing software stack opened in Beijing. Experts, bigwigs and ecological partners from the Internet industry, such as Baichuan Intelligence, Kuaishou, Jinshan Office, Jingdongyun, and so on, gathered together to discuss how the Internet industry can seize the opportunities of generative AI and share the ways of smart computing development in the big model era.

Peng Zhen, chairman of Chaochao Information, said in his speech: "artificial intelligence computing has become the core engine driving the evolution of large models, but the research and development of large models often face the dilemma of not being able to afford, building or being good. Only by constantly innovating the technology of smart computing products, combining the practical experience in the field of large models, and transforming the power of intelligence into business innovation, can we be at the forefront of the AI era."

At present, AI computing has become one of the core competitiveness of the development of large model industry. OpenAI public data show that the computing power equivalent from GPT-3 to GPT-4 has increased 68 times, and computing power is still the core driving force for the improvement of the ability of large models. Chen Weipeng, co-founder of Baichuan Intelligent Technology, said: "one of the laws of the development of the big language model is that the ability of the model expands with the amount of computation, and so far we have not seen the boundary of the model."

In order to meet the urgent needs of large models for computing power, it is necessary to build a strong intelligent computing infrastructure. However, the construction of intelligent computing system is not a simple accumulation of computing power. It is generally realized in the industry that large model training is a complex system engineering like building a rocket, which requires comprehensive consideration of the deployment, optimization and resource scheduling of the computing power system. A series of problems such as data collection, cleaning, processing, model training, reasoning, migration, tuning, application services, ecological construction and so on.

Round table forum

Solving the computing challenge of large models has become the direction of joint efforts of industrial innovation. Liu Jun, senior vice president of Chaochao Information and general manager of AI&HPC, said that in order to better release the productivity of the large model, first, to give full play to the efficiency of the computing system, the development of the large model is a race for time, and the efficiency is equal to time and money; second, scalability, for the computing cluster system that is prone to million cards, the parallel computing strategy is related to the performance expansion ratio. Third, it is necessary to ensure the sustainability of the supply of computing power in the training process, and fully consider the problems of training interruption and failure caused by hardware, software, algorithms and framework. We need the overall design from the system level to help enterprises cross the threshold of large model research and development. Chen Weipeng said that Baichuan Intelligence is exploring to improve the training efficiency of large models and the success rate of research and development. Liu Lingzhi, head of Kuaishou heterogeneous computing, said that when the number of model parameters exceeds a certain threshold, intelligence will also show a leap-forward improvement. In the face of the computing challenges brought by video and large models, Kuaishou deeply developed SL200 chips from audio and video, AI and data center technologies, achieving industry-leading video compression efficiency, and empowered infrastructure and AI, video atomic capabilities through the StreamLake brand.

At present, the large model empowered by smart computing has been widely used in the Internet industry. as a representative enterprise for the application promotion of the large model, Xiong Longfei, the technical director of Jinshan Office, shared the practice of the landing of the large model in the office field: "the development trend of AI in office scenarios is to meet the needs of public network users first, and then to privatize deployment, from single-mode to multi-mode, from highly dependent on the native capabilities of large models to richer technology integration." From conversational interaction to more mild click-and-select interaction. "

In order to promote the construction of large model intelligent computing system, modern data center is the key carrier. The large model brings multiple challenges to the data center in terms of energy consumption, computing structure and so on. Wang Guilin, senior hardware system architect at Jingdongyun, shared JD.com 's latest thinking on building a large-scale data center: "through a series of means, such as hardware technology innovation, software architecture optimization and the use of clean energy, building a low-carbon, environmentally friendly, sustainable and intelligent data center will become a powerful support for unleashing intelligence."

Participants pointed out that China's large models are still in the early stages of business ecological development, and extensive ecological cooperation is needed in the aspects of intelligent computing system construction, platform support, algorithm research and development, and application landing, so as to create more outstanding large models. And promote the scene landing of the model, so as to really enable the Internet industry to change.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.