Big model of hot war between cloud and edge AI chips! Essential information on the first day of the 2023 Global AI Chip Summit

Core West reported on September 18 that the 2023 Global AI Chip Summit (GACS 2023) was successfully held in Nanshan District, Shenzhen from September 14th to 15th . The venue was very popular on the first day of the opening. Not only was the venue full of seats, but the three floors inside and outside the venue were also packed with enthusiastic audiences. Many of the audience even stood and listened to the whole process. The entire summit was broadcast live on the cloud, with 1.53 million viewers across the entire network.

This high-profile industrial conference was sponsored by Zhiyi Technology's subsidiary Core East West and Chiangutan. Under the guidance of the Nanshan District Science and Technology Innovation Bureau, it has seven major sections with the theme of "Competing in the Core World in the AI ​​Era". The main venue includes the opening ceremony and Three special sessions were held on AI chip architecture innovation, AI large computing power chips, and high-energy-efficiency AI chips, and the results of the 2023 Chinese AI chip enterprise "Top 30 Pioneer Enterprises" and "Top 10 Emerging Enterprises" were announced; sub-venues included the first newly added integrated circuit policy exchange Conference, AI Chip Analyst Forum, Intelligent Computing Center Computing Power and Network Summit Forum.

In response to industry changes, this summit set an agenda around the academic research of AI chips, GPGPU, ASIC, Chiplet, storage and computing integration, and intelligent computing centers. It invited 46+ guests to talk about generative AI and large model computing power needs, investment and financing Opportunities, architectural innovation, commercial implementation and other key topics of AI chips.

At the main venue on the first day, Wei Shaojun, a professor at Tsinghua University, vice chairman of the China Semiconductor Industry Association, and IEEE Fellow, gave a keynote report "Re-discussing the Development of Artificial Intelligence Chips", with contributions from Nvidia, Suiyuan Technology, Qualcomm, Yizu Technology, The founders and technologies of 12 top domestic and foreign AI chip companies and cutting-edge companies, including Arctic Xiongxin, AMD, Kuixin Technology, Houmo Intelligent, Anmou Technology, Kunyun Technology, Zhuhai Core Power, Xinzhi Technology, and Deep Thought Technology Decision makers and executives delivered keynote speeches to share cutting-edge research and judgment on AI chip industry trends and the latest practices.

On behalf of Shenzhen Nanshan District Science and Technology Innovation Bureau, Cao Huan, Secretary of the Party Leadership Group and Director of Shenzhen Nanshan District Science and Technology Innovation Bureau, expressed warm congratulations on the convening of the summit and delivered a speech for this summit. He said that the 2023 Global AI Chip Summit (GACS 2023) is positioned as a publicity platform for integrated circuits and high-tech industry policies, a gathering and exchange platform for the global AI chip industry, with a strong guest lineup and rich content, and a dedicated Shenzhen integrated circuit policy The exchange meeting is of great significance to promoting the high-quality development of Shenzhen’s AI chip industry.

Big model of hot war between cloud and edge AI chips! Essential information on the first day of the 2023 Global AI Chip Summit ▲Cao Huan, Party Secretary and Director of the Shenzhen Nanshan District Science and Technology Innovation Bureau

Gong Lunchang, co-founder and CEO of Zhiyi Technology, delivered a speech as the organizer, announcing that the Shenzhen headquarters of Zhiyi Technology has officially settled in Xili Lake International Science and Education City, Nanshan District. The Zhiyi East-West Open Course brand has been officially upgraded to "Zhiyi Orangutan", and the Shenzhen headquarters will assume the responsibility of Xili The construction of the chimpanzee AI industry service platform in Hu International Science and Education City hopes to have more cooperation with enterprises in the Greater Bay Area in the future. Intelligent Orangutan is newly upgraded from the Intelligent East-West Open Course. It is positioned as a new technology service platform, focusing on artificial intelligence and emerging technologies, and providing value services to enterprises and users.

Big model of hot war between cloud and edge AI chips! Essential information on the first day of the 2023 Global AI Chip Summit ▲ Gong Lunchang, co-founder and CEO of Zhiyi Technology

Through the guest sharing and viewpoint collision at the AI ​​Chip Summit on the first day of the summit, we discussed the computing power pressure and opportunities brought by generative AI and large models, and saw how the industry and academia can bypass Moore's Law through an increasingly wide range of innovative ideas. Bottlenecks that are on the verge of reaching the limit have witnessed the contention of various innovative schools such as storage and computing integration, reconfigurable computing, and "unified computing power field". Through this informative AI chip industry feast, we have summarized the following five major trends:

1. Generative AI and large models are dominating AI applications, and those with computing power win the world. The industry is required to optimize AI chip development and implementation from aspects such as chip design, communications, clustering, and software to achieve higher performance and lower costs. latency, lower cost.

2. There is a gap between actual computing resources and ideal computing resources. On the one hand, the AI ​​chip industry continues to pursue large computing power; on the other hand, it breaks through the bottleneck of computing power through architectural innovation and new software-defined methods.

3. As chip manufacturing technology reaches the ceiling, architectural innovation is expected to usher in a "golden decade". For example, the integrated sensing-memory-computing architecture is expected to reduce the burden of massive data handling by breaking down the "storage wall"; the reconfigurable computing architecture replaces design with process by improving chip utilization; chiplets achieve heterogeneous integration through splitting and combination to improve chip performance. , entrepreneurial pioneers have emerged in these fields.

4. Transformer is building a "big tree" unified concentrated ecosystem through a unified large model to solve the fragmentation pain points of multiple "small saplings" AI ecosystems in the past. Therefore, AI chip companies can promote the commercialization of AI chips through multi-industry linkage of hardware, software, systems, and solutions.

5. From the perspective of the capital market, regulatory policies in the primary market will objectively promote the reallocation of AI chip industry resources, and IPO shrinkage may be short-term; the "hype" phenomenon in the secondary market objectively attracts investment in the primary market .

1. Wei Shaojun: Let’s talk about the development of AI chips in the era of large models. Large computing power is not the final challenge.

The opening keynote speaker of this summit is Wei Shaojun, a representative figure in academic research on AI chips - Professor of Tsinghua University, Vice Chairman of China Semiconductor Industry Association, and IEEE Fellow. He quoted the 2017 Turing Award winners John L. Hennessy and David A. Patterson's golden words "Now is a new golden age of computer architecture" and talked about entering the era of ubiquitous computing, where computing is everywhere and the requirements for intelligent computing The urgency is growing.

After evolution, artificial intelligence is now divided into two major categories: brain-inspired computing and deep learning. Brain-like computing imitates the calculation method of the human brain. Following a similar idea, what is gradually gaining momentum is the integration of storage and computing, which can increase energy efficiency by 10 times compared to the traditional separation of storage and computing architecture. However, this architecture can currently only imitate the surface of the human brain mechanism. Another direction is deep learning. The "Three Deep Learning Giants" who won the Turing Award in 2018 use neural networks to continuously perform calculations to achieve computing goals, forming a deep neural network. This has prompted today’s artificial intelligence to include three elements: algorithms, data, and computing power. The challenge of computing power determines that we need to find a new computing architecture.

Big model of hot war between cloud and edge AI chips! Essential information on the first day of the 2023 Global AI Chip Summit ▲ Wei Shaojun, professor at Tsinghua University, vice chairman of China Semiconductor Industry Association, and IEEE Fellow

When participating in the GTIC 2019 Global AI Chip Innovation Summit, Professor Wei Shaojun once shared a diagram dividing the AI ​​chip (AI Chip) into three stages: stage 0.5 is implemented using existing chips, and stage 1.0 develops application-oriented fields. For dedicated chips, the 1.5 stage emphasizes versatility and flexibility. The final conclusion of the 2.0 stage is whether intelligence can truly be achieved. Looking at this picture now, he believes that a stage should be added in the middle - AI Chip 1.7.

In his view, large computing power may be the biggest challenge currently faced, but it is not the ultimate challenge. The 1.7 stage emphasizes the understanding of versatility. Today, the versatility of AI chips has been alienated into large computing power, and computing power has become the pursuit goal, but it is only an intermediate stage in the evolution process. Finding complete intelligence in the AI ​​Chip 2.0 stage is our ultimate goal.

To achieve this goal, it is necessary to give the chip the ability to self-learn and receive education to continuously improve performance and efficiency. He proposed to find a system that is both software and hardware programmable, that is, a software-defined chip. Nvidia, Intel and some of the top scientific research teams in the United States are studying this technology.