NIE 2023 | Roundtable Forum: Infrastructure Requirements in the AI Era

Yuan Xucong, Chief AI Industry Analyst at LeadLeo Research Institute

Yuan XucongIt is stated that in recent years, AI technology has been continuously upgraded and iterated. We have witnessed breakthroughs in intelligent voice, natural language processing, and other areas. The influx of generative AI and large models has led to a closer integration of artificial intelligence with various industries. With the breakthroughs in new-generation intelligent technologies and the arrival of large models, they have now become a force leading the trend.

AI infrastructure provides a solid foundation for the success of large models, including data, computing power, and algorithms. At today's forum, we will delve into the latest trends in AI infrastructure development, technological breakthroughs, and industry patterns, share insights from industry leaders, and jointly consider the intersection of AI infrastructure and the future of artificial intelligence. We hope that this exchange can inspire new thinking, promote innovative applications of AI technology, and contribute our wisdom and efforts to building an even more intelligent and convenient future society.

Liu Liang, Director of Business Strategy Research at Shangtang Technology

Q1:The development of artificial intelligence is a historic strategic opportunity that is crucial for promoting the transformation and upgrading of China's economic structure. With the urgent need for intelligent transformation across various industries, what new trends are emerging in the AI infrastructure industry that deserve attention in promoting industrial intelligent transformation? What new changes within the industry can further drive the development of artificial intelligence?

Liu LiangIt is indicated that the current era is AI 2.0, and the rapid development of AI is driven by three important forces. Firstly, policy promotion has become a key factor since 2020. Secondly, user scenarios have undergone significant changes. Before 2020, users of AI facilities were mainly trying out applications of AI, whereas now, from an enterprise perspective, this has shifted from pilot projects to large-scale applications. Finally, AI technology itself is evolving, especially with the rise of large models and generative AI. AI can not only automate, classify, and support decision-making but also achieve generative capabilities. AI has become creative, which has a profound impact on the entire industry. In the process of embracing the AI 2.0 era, the support of AI infrastructure becomes crucial. This includes the full-stack foundational dimensions of hardware and software, covering key links from AI development, deployment, services to actual applications, as well as subsequent iterative operations, forming a complete lifecycle. This infrastructure must be a comprehensive development of computing power infrastructure, AI development platforms, AI development tools, and data governance platforms.

In anticipation of the arrival of the AI 2.0 era, SenseTime has proactively laid out new AI infrastructure, namely the SenseCore, an AI supercomputer. The supercomputer encapsulates intelligent computing centers, algorithm optimization services, and data governance as a 'force field', continuously outputting it to the market, customers, and partners.

Q2:The craze for large models has swept across China for over half a year, and we are entering an era of ubiquitous large models. Every company that possesses the capability to develop its own large models wants to become an infrastructure operator in this era. In the era of large models, what role does AI infrastructure play, and how should it be developed?

Liu LiangIt is indicated that AI infrastructure for large AI models is relatively complex. It cannot simply be supported by piling up chips to train large models, nor is it a simple matter of brute force aesthetics. Behind it are numerous software engineering system issues. Ensuring the stability of AI infrastructure during training and improving parallel efficiency across multiple machines and cards are fundamental. High-quality training data is also an important factor in competition, as the quality of underlying data is crucial for the training and performance of large models. Precisely screening and processing underlying data can not only improve the efficiency and accuracy of model training but also ensure the stability and credibility of models in complex scenarios. To create a powerful and truly competitive large model, hundreds or thousands of model iterations are required to improve various algorithms and engineering knowledge accumulation, refine data cleaning methods, and improve data ratio recipes, all of which depend on a large number of GPUs and supporting software and systems.

The AI infrastructure of SenseTime - SenseCore, the large-scale AI system of SenseTime, has supported the training and iterative upgrade of its own billion-parameter models. To this end, SenseTime has established an engineering system to support rapid iteration of large models. It coordinates R&D efforts in software, systems, and hardware with the goal of serving rapid iteration of large models, achieving very agile, low-cost, and mass model iterations. Only in this way can the best and most effective production recipes for large models be discovered and practiced.

Additionally, the Shangtang Large Model Architecture is also serving external clients. Since 2023, over 1,000 large models with tens of billions to hundreds of billions of parameters have been trained on the Shangtang Large Model Architecture, supporting dozens of generative AI applications. These extensive project experiences have allowed us to accumulate and build a valuable knowledge system, enabling the Large Model Architecture to become an even better AI infrastructure for the era of large models.

In the era of large models, enterprises can adopt a series of strategies to enhance competitiveness, including strengthening algorithm research, improving data quality, continuously updating computing power, and actively participating in interdisciplinary and international collaborations. These measures will help promote the widespread application and continuous development of AI technology in various fields.

Gao Ping, Vice President of Suiyuan Technology

Q1:With the continuous vigorous development of computing power infrastructure, we have witnessed a series of profound changes in the field of computing. In this process, the state has continuously improved the layout of public computing power, accelerated the construction of edge computing power infrastructure, and laid a solid foundation for future development. What will be the future trends and main themes of computing power infrastructure development?

GaopingIt is indicated that, on the one hand, general-purpose computing power continues to develop steadily, and the development of AI technology will continue to drive the growth in demand for general-purpose computing power. In the future, we may see the emergence of more hardware and software tools optimized for AI workloads to achieve more efficient AI computing. This will further deepen the integration of AI with computing power and promote the continuous development of intelligent applications.

On the other hand, future cloud-edge collaboration will become another development trend. More and more devices need to perform local computing at the edge to reduce data transmission latency and improve response speed. This will drive the migration of computing infrastructure to the edge to meet the growing demand for edge computing.

In addition, sustainability and green computing will become focal points. Environmental sustainability is attracting widespread attention, so future data centers and computing facilities will pay more attention to energy efficiency and environmental protection. It is expected that more green energy and efficient cooling technologies will be adopted to reduce carbon footprints and achieve more sustainable computing.

Q2:In the field of AI infrastructure, enterprises' business models may include hardware sales, technology licensing and collaboration, software and hardware integrated solutions, AI platform-as-a-service, etc. Each of these business models has its own advantages, depending on the enterprise's technical strength, market positioning, and customer needs. Regarding the business models for AI infrastructure, which model do you think performs better in current market competition? How does it meet customer needs and drive enterprise development?

GaopingIt is indicated that, first and foremost, the hardware sales model is the most common business model, and it is also the most important business model from the perspective of revenue generation. This model can provide customers with high-performance hardware devices such as GPUs or specialized AI chips. This meets customers' high-performance needs for computing resources. In the era of large models, AI platforms and services are the most promising business model, providing powerful AI development and deployment platforms for enterprises and developers and reducing the burden on customers in terms of underlying infrastructure. This encourages innovation and application development, enabling customers to focus on building and deploying their own AI models and applications. Technology licensing and cooperation provide customers with greater flexibility, allowing them to choose and pay according to their own needs. Moreover, the software-hardware integrated solution model is also very attractive in certain cases. This model provides a complete software-hardware solution, from hardware to application scenarios. This reduces customers' integration and management work and improves convenience.

Founder and CEO of TopCoder, Feng Lei

Q1:The large-scale application of artificial intelligence requires the use of massive amounts of data to train models. It can be said that without high-quality datasets, there will be no large-scale application of artificial intelligence. In the large-scale application of artificial intelligence, how can we ensure the acquisition, processing, and management of high-quality datasets to support model training and development, thereby promoting the continuous development and application expansion of artificial intelligence technology?

Feng LeiIt is indicated that artificial intelligence mainly involves model generation, with data and computation at its underlying level. Parameters need to be fed into the data. To ensure the quality and effective utilization of data, multiple aspects need to be considered, including data collection, cleaning, annotation, security and privacy, storage and management, sharing and collaboration, automated processing, monitoring and maintenance, etc. These steps and strategies can guarantee the high quality and reliability of data, providing solid support for the training and application of artificial intelligence models.

We believe that in the future, data will form a data network, which also implies the consumption of data resources. This optimistic scenario requires time to settle and regulatory norms to be implemented. Currently, data products have a strong repurchase ability. In the short term, business models can support companies by providing better data management and processing methods, thereby meeting the growing demand for artificial intelligence and promoting the development of large AI models.

Q2:Currently, China has made remarkable progress in chip, cloud services, and data resources. It possesses vast data resources that support AI infrastructure, and the industry has seen a proliferation of participants. What are the distinctive strengths of mainstream participants? Can you share some specific cases or key areas? And how do various manufacturers collaborate with each other to drive the development of the entire AI infrastructure ecosystem?

Feng LeiIt is stated that there are numerous AI infrastructure vendors, which can be broadly categorized into cloud computing vendors, data services and data governance providers, as well as AI infrastructure hardware providers. Each type of vendor plays a unique role and contributes to the AI infrastructure field. It is predicted that in the future, one-fourth of China's GDP will be related to AI. China's huge AI industry chain is taking shape, with technology giants among cloud infrastructure vendors performing quite well; there are also companies that excel in vertical scenarios, such as SenseTime Technologies on-site; as a service provider in the field of data computing, the company has launched a large model data computing system, providing solid data computing support for business decision-making and ensuring data quality, security, and compliance.

Currently, AI-related manufacturers in China have their own unique advantages in the domestic market and have collaborated to build a relatively complete AI infrastructure ecosystem. This ecosystem provides diverse options and support for enterprises and developers, promoting the rapid development and widespread application of AI technology. In the future, these manufacturers will continue to leverage their respective strengths to drive further innovation and development in the field of AI infrastructure.