With the explosion of ChatGPT, local governments, research institutions, and technology enterprises in China have expressed that they have started researching and promoting ChatGPT like large model technology, which is a matter of long-term strategic significance.
However, as everyone knows, there are still many problems to be faced in this process. For example, at present, Nvidia is the main training model in China. At an AI conference a few days ago, Yang Fan, co-founder of SenseTime and president of the large device business group, said that the A100 on the Chinese market can support 100 ChatGPT like large model networks. It sounds a lot, but some large companies may need to connect thousands of cards together and conduct more forward-looking internal testing, such as three thousand cards for a cluster and two or three such clusters. Overall, the volume of computing power is far from sufficient.
What are the challenges faced by the domestic ecosystem in developing the Chinese version of ChatGPT?
The development of the Chinese version of ChatGPT is a very promising thing for the entire artificial intelligence industry, but it also reflects that there are still many challenges in this industry that need to be solved together. For example, whether domestic chips and software can support large-scale clusters with more than kilocalorie and can be comparable to Nvidia in throughput.
In the past few years, it seems that China has been running very fast in the application of artificial intelligence industry, but it lags behind the United States in basic technology. If we want to develop in the long term, it is necessary to build a domestic ecosystem. The launch of ChatGPT this time highlights the importance of building a domestically produced ecosystem.
So what are the problems facing domestic ecology at present? Yang Fan talked about several points: First, there are indeed diversified AI chips on the market to try to challenge Nvidia's dominance. However, diversified AI chips also bring its unique challenges. In other words, there are a lot of chips and frameworks in China, which forms a many to many network relationship, while Nvidia has formed a strong cooperation ecosystem with PyTorch.
The chaotic state of wolf warfare in China has resulted in a large workload and high cost of adaptation, including the adaptation cost of operators and frameworks. In Yang Fan's view, the industry needs to establish industry standards for software and hardware adaptation to reduce the development costs of the entire ecosystem in this process.
Secondly, chips are located at the upstream of the artificial intelligence industry chain, which includes various manufacturers such as boards, servers, operating systems, frameworks, data, AI computing platforms, software development, industry system integration, and so on. Currently, artificial intelligence is still an immature industry as a whole, with algorithms updated almost daily and software updated every month. In this situation, it is difficult for each layer to have a stable connection relationship to serve downstream vendors well.
Yang Fan believes that the development cycle of AI chip itself is very long, and it is difficult to compete with Nvidia in computing power. Therefore, chip manufacturers should be able to design and optimize the next generation of chips around the needs, so as to at least form a fast overtaking in the subdivision of the track. Of course, this requires a good understanding of the application, and from the perspective of the industrial chain
Shang Tang proposed corresponding solutions to the two problems faced by ecological construction mentioned above. According to Yang Fan, the company hopes to work with industry partners to promote standardized interfaces for diverse frameworks and chips; For the problem of long links and weak understanding of AI scenarios, SenseTime itself has a deep accumulation of algorithms and applications. It hopes to work with industry partners to get through the downstream scenarios point-to-point for some important scenarios with large-scale demand, and truly realize the landing application of domestic chips.
How to build a computing ecosystem?
Zhang Yalin, founder and COO of Suiyuan Technology, believes that the elements of AI ecological explosion include algorithms, computing power, data, platforms and applications. How to turn the flywheel of the five elements of ecology? He believes that it is necessary to use underlying computing power as application support, data and algorithms as the foundation of large models, cloud services and the entire distributed framework as a platform, and ultimately lead to application scenarios.
In the current context, Chinese AI chips and large computing power enterprises need to work together to build a heterogeneous computing ecosystem. Seen from the figure below, this computing ecosystem, from the system perspective on the left, includes software and hardware. Software has many solutions, cloud services, Toolchain, and system software; Hardware includes chips, peripherals, hosts, etc. The three circles on the right are an expansion of the ecosystem from the inside out. Any computing power alliance in China that wants to do this well must create its own core software stack, which is a set of underlying software stacks similar to CUDA; Then, on top of it, grow an extension library and framework interface to collaborate with the open source community and ecological partners; The solutions and applications in various industries, including universities, research institutes, and industry pilot projects, are drawn from above.
Zhang Yalin believes that the only way for China's computing power and ecological development is to expand from the bottom of the computing power and software stack to the open source community and partners, and then pull out more applications.
To be able to approach Nvidia, China's software ecosystem needs to go through four processes, from usability, ease of use, win-win cooperation, to ecological closed-loop. At present, the vast majority of AI chips and software stacks in China are still in the stage of being usable or moving towards ease of use. Enterprises need to constantly iterate their products and application scenarios, making the products from usable to usable. After it becomes useful, the customer base will begin to expand, form partnerships and ecological alliances, and then achieve an ecological closed-loop.
Summary
ChatGPT is not only an enhanced version of search engine, but also a Chatbot. Its artificial intelligence services provided by SaaS (Software as a service) will reshape all digital applications and all industries, bringing a new industrial revolution to people.
The emergence of ChatGPT has brought opportunities and also reflects the lack of ecological construction in the artificial intelligence industry in China. How to leverage the advantages of scientific research institutions and large technology enterprises, combined with the power of various sub sectors of the industry chain, to build a comprehensive ecosystem is a matter that the industry needs to jointly explore and solve.
