CNCC2024
Brief introduction of the forum:
Can the application of multimodal large models on NPU create a new interaction paradigm?
Time: 13:30-17:30, October 26
Location: 1st Floor, West Corner Tower, Summer Garden
Note: If there is any change, please refer to the final information on the official website (https://ccf.org.cn/cncc2024).
In November 2022, OpenAI released an experimental chatbot called ChatGPT, which has driven the boom of generative artificial intelligence and large models, becoming an accelerator and catalyst for the popularization of artificial intelligence applications, and accelerating the intelligent transformation of all walks of life. The new AI experience will also be implemented at an accelerated pace on smart terminals. Large-scale model technology will be a watershed moment for the PC industry, bringing revolutionary improvements and huge growth opportunities for PC product experience in the next five years. Intelligent terminal devices, including AI PCs, are one of the ultimate carriers for AI to reach end users, and they are the computing power of PARFAITE in the true sense of the word.
In order to support more applications of generative AI, it is not only necessary to improve the computing power of the cloud, but also to cooperate with more powerful computing power at the edge and the end side to form a more balanced computing power distribution under the "end-edge-cloud" hybrid computing architecture. But at the same time, it also brings very practical problems, such as how to meet the needs of the device side to provide users with more distinctive large model services, how to provide a more powerful and low-power computing power solution for large model inference on the device side, what role can the neural network processor (NPU) play in this solution, and what kind of breakthroughs will there be in the corresponding model inference deployment software stack and model algorithm on top of the neural network processing unit? These topics have become cutting-edge research directions in the industry.
Forum Highlights:
What are the potential breakthroughs and trends of the new AI PC architecture?
Discuss what kind of large-scale service is the service with more advantages for terminal equipment?
How can terminal devices give full play to the advantages of CPUs, graphics processing units (GPUs) and neural network processors (NPUs) through the close integration of software and hardware?
Forum Agenda
Keynote Speech (Moderator: Zhu Wenwu)
| order | topic | Keynote speaker | unit |
| 1 | Artificial intelligence automatically designs processor chips | Chen Yunji | Institute of Computing Technology, Chinese Academy of Sciences |
| 2 | Graph neural network inference optimization for multi-core processors | Li Dongsheng | National University of Defense Technology |
| 3 | Energy-efficient circuit and system design for AI 2.0 | Wang Yu | Tsinghua University |
| 4 | Visual perception and embodied intelligence | Lu Jiwen | Tsinghua University |
| 5 | SRAM memory chip design for large computing power | Si Xin | Southeastern University |
| 6 | NPU-based personal computing system and interactive innovation | Yan Yiqiang | Lenovo Group |
Panel Session: (Moderator: Wang Zhepeng)
| topic | Keynote speaker | unit |
| The impact of large models and NPUs on smart terminal innovation | Chen Yunji | Institute of Computing Technology, Chinese Academy of Sciences |
| Li Dongsheng | National University of Defense Technology | |
| Si Xin | Southeastern University | |
| Li Yuan | Zhuhai Core Power Technology Co., Ltd | |
| Yan Yiqiang | Lenovo Group | |
| Guan Chaoyu | Qingli Intelligent Technology (Beijing) Co., Ltd |
Introduction of the chairman and guests of the forum
Chair of the Forum
Zhu Wenwu
He is a fellow of CCF and a professor at Tsinghua University
Professor of the Department of Computer Science, Tsinghua University, Deputy Director of the National Research Center for Information Science and Technology. He is a CCF Fellow, an ACM Fellow, an IEEE Fellow, an AAAS Fellow, and a Foreign Member of the European Academy of Sciences. Received the 2024 IEEE Circuits and Systems Society Charles A. Desoer Technical Achievement Award and the 2023 ACM SIGMM Technical Achievement Award. He has won the second prize of the National Natural Science Award three times.
Co-Chair of the Forum
Wang Zhepeng
Vice President of Lenovo Group
He is the head of the Personal Computer and Ecological Innovation Laboratory of Lenovo Research Institute, and the vice chairman of the Industry-University-Research Alliance of Intelligent Unmanned Systems. Since 2001, he has been engaged in the research and development of intelligent equipment, flexible display technology, intelligent retail, smart education and other fields, and has more than 160 patents. The innovative products developed by him have won more than 150 awards in the three major exhibitions of CES, MWC and IFA, and the visual algorithm has won more than 10 competition championships in the three major conferences of CVPR, ICCV and ECCV, and has won a Beijing Science and Technology Progress Award and a CCF Science and Technology Progress Award. At present, it undertakes one major project of scientific and technological innovation 2030 and one key special project of the national key research and development plan.
Forum Speaker
Chen Yunji
Director of the CCF Architecture Committee, Deputy Director of the Institute of Computing Technology, Chinese Academy of Sciences
Deputy Director of the Institute of Computing, Chinese Academy of Sciences, Director of the National Key Laboratory of Processor Chips, and Director of the CCF Architecture Committee. He has been engaged in the interdisciplinary research of processor chips and artificial intelligence for a long time, and has developed the world's first deep learning processor chip (Cambrian No. 1), which has been evaluated as the "pioneer" and "leader" of deep learning processors by Science magazine. One of the main architects of Loongson No. 3 CPU, he opened the first intelligent computing system course in China in 2019, and the textbook "Intelligent Computing System" compiled in 2020 has been reprinted many times. He has won the second prize of the National Natural Science Award, the National Science Foundation for Distinguished Young Scholars, the National May Day Labor Medal, the May Fourth Medal of Chinese Youth, and the Ho Leung Ho Lee Science and Technology Innovation Award, and has been named one of the world's 35 outstanding young innovators by MIT Technology Review.
Title: Artificial Intelligence Automatically Designs Processor Chips
Abstract: In 1957, Turing's doctoral advisor, Alonzo Church, proposed the Church's Problem, which is whether a machine can design circuits automatically. This question is regarded by many computer and artificial intelligence scholars as a "holy grail" of computer science. However, there has been no outstanding progress in decades of exploration in the industry, and the machine can only design a toy circuit with a thousand doors. Through long-term research, we have made a breakthrough in this problem, forming the world's first CPU chip automatically designed by AI, "Qimeng-1". The chip contains more than 4,000,000 logic gates and was reported by Nature News as "good news for China's chip development".
Li Dongsheng
He is the deputy director of the CCF Architecture Committee and a professor at the School of Computer Science, National University of Defense Technology
He is an outstanding member of CCF and a member of the Big Data Committee. Professor of the School of Computer Science, National University of Defense Technology, Deputy Director of the National Key Laboratory of Parallel and Distributed Computing. He was selected into the New Century Excellent Talent Program of the Ministry of Education. He has published more than 100 academic papers in academic journals and conferences such as Science China and IEEE/ACM Transactions, and the system developed by him has been applied in important national fields. He has won the second prize of the National Science and Technology Progress Award, the first prize of the Military Science and Technology Progress Award, the first prize of the Natural Science Award of Hunan Province, the special prize of the Teaching Achievement Award of Hunan Province, the China Youth Science and Technology Award, and the special government allowance of the State Council.
Title: Graph Neural Network Inference Optimization for Multi-Core Processors
Report Summary: Graph neural networks (GNNs) have attracted attention in recent years. The increasing scale of graph datasets and the diverse graph computing modes of graph neural networks have brought challenges to the inference of graph neural networks on computing platforms. The report will analyze the performance bottleneck of graph neural network inference on multi-core CPUs, and effectively improve the inference performance of GNN on multi-core CPUs by optimizing the memory access mode of graph data, designing an efficient load balancing strategy, and using vectorization instructions of multi-core processors to reconstruct the core operators of graph neural networks. On multi-core processors with various architectures such as Intel, AMD, and ARM, this method can significantly improve the inference performance of graph neural networks compared with the current mainstream graph neural network frameworks such as DGL and PYG.
Wang Yu
He is a tenured professor and head of the Department of Electronic Engineering, Tsinghua University
He is a tenured professor and head of the Department of Electronic Engineering at Tsinghua University, an IEEE Fellow, a recipient of the National Natural Science Foundation of China for Distinguished Young Scholars, the deputy dean of the School of Information Science and Technology of Tsinghua University, and the dean of the Tianjin Institute of Electronic Information of Tsinghua University. Professor Wang Yu has been engaged in the research of smart chips, energy-efficient circuits and systems for a long time, and has published more than 60 IEEE/ACM journal papers, more than 270 conference papers, and more than 20,000 Google Scholar citations. He has presided over a number of national and enterprise joint projects, and won the first prize of technological invention of CCF Science and Technology Award, the Innovator Award under 40 of the International Design Automation Conference, and the CCF Qingzhu Award. He has won 4 Best Paper Awards and 12 Best Paper Nominations at International Academic Conferences. In 2016, the knowledge transformation was invested in Deejian Technology to build a world-class deep learning computing platform, and in 2018, it was acquired by Xilinx (now AMD), a leading company in the industry. In 2023, we will promote the establishment of Wuwen Core Dome to form a joint software and hardware optimization platform for large models, and achieve industry-leading large model inference performance on more than 10 kinds of chips at home and abroad.
Title: Energy-efficient circuit and system design for AI 2.0
Abstract: Large language models based on Transformer architecture have achieved excellent performance in a variety of applications, marking the advent of the AI 2.0 era. With the rapid increase in the number of model parameters, the computing, storage, and memory access overhead of large models have increased by 4-5 orders of magnitude compared with traditional deep learning models, making it difficult for existing device-side hardware platforms to achieve efficient deployment of large language models. This report will first focus on the key challenges and development status of software and hardware systems in the era of large models, and introduce energy-efficient circuit and system design methods for AI 2.0, including a series of software and hardware collaborative optimization methods, such as algorithm model compression, software operator optimization, and hardware architecture design. Secondly, this report will introduce our work on the collaborative optimization of software and hardware of large models in the AI PC scenario. Finally, this report will provide an outlook on the future development of software and hardware on AI PCs.
Lu Jiwen
Deputy Director and Tenured Professor of the Department of Automation, Tsinghua University
He is a tenured professor at Tsinghua University, a doctoral supervisor, deputy director of the Department of Automation, a winner of the National Science Fund for Distinguished Young Scholars, an IEEE/IAPR Fellow, the editor-in-chief of the international journal Pattern Recognition Letters, the project leader of the National Key R&D Program, the director of the Visual Computing and Simulation Professional Committee of the Chinese Simulation Society, and the deputy director of the Expert Advisory Committee of the Chinese Society of Automation. He has published more than 140 IEEE Transactions papers (including 40 T-PAMI papers), more than 160 CVPR, ICCV, ECCV, NeurIPS papers, more than 32,000 Google Scholar citations, more than 60 national invention patents, presided over 2 key projects of the National Natural Science Foundation of China, and won 1 first prize of the Natural Science Society of China (ranked 1) and 1 second prize of the National Teaching Achievement AwardHe serves as an editorial board member of journals such as T-IP, T-CSVT, T-BIOM, PR, Acta Automatica Sinica, and the chair/program committee chairman of ACCV2026, FG2023, ICME2022, VCIP2022, AVSS2021 and other conferences.
Title: Visual Perception and Embodied Intelligence
Abstract:Embodied intelligence is a research hotspot in the field of artificial intelligence and robotics, and has important application prospects in industry, agriculture, and service industries. The report will review the research progress of visual perception and embodied intelligence in recent years, mainly including autonomous environment perception, 3D scene representation, visual positioning and cruise, and other methods, as well as their applications in tasks such as multimodal scene understanding, robot grasping and packaging, and end-to-end deployment of large models, and finally look forward to future development trends.
Si Xin
Associate Professor/Ph.D. Supervisor, School of Integrated Circuits, Southeast University
He has been engaged in the research of in-memory computing and high-computing AI chips for a long time, and has published a total of 45 high-level papers in recent years, including 12 ISSCC papers known as the "Chip Olympiad" and 8 top integrated circuit journals JSSC and Nature Electronics. He has been nominated for the IEEE CICC Best Student Paper Award and the MCSoC Best Paper Award. He has presided over a number of national, provincial and ministerial scientific research projects. Serves as a member of the Technical Committee of IEEE VLSI-TSA, ICTA and MCSoC conferences.
Title: Design of SRAM memory and computing chips for large computing power
Report Summary: With the vigorous development of the big data era, the edge-oriented AIGC scenario has attracted more and more attention, which requires a large number of frequent access and computation of data, so there is an urgent need for energy-efficient intelligent processor chips. In the processor design based on the traditional von Neumann architecture, the data exchange between the storage unit and the computing unit must pass through a limited data bus, and the performance of the system is largely limited by the bandwidth of the bus and the read and write power consumption of the storage cell. In order to break the bottleneck of this "storage wall", in-memory computing has received extensive attention, and the system architecture based on in-memory computing can support different logic or matrix multiplication and addition operations while retaining the storage and read-write access functions of the storage unit itself, thereby greatly reducing the frequent bus interaction between the computing unit and the storage unit, and further reducing a large amount of data migration and the resulting power consumption, thereby greatly improving the energy efficiency of the system. This report will focus on the challenges and latest development trends of in-memory computing chip design based on static random access memory (SRAM) at home and abroad.
Yan Yiqiang
Principal Researcher of Lenovo Group
Chief Researcher of Lenovo Group, he has been engaged in the development of smart devices for a long time, leading the development of a number of industry-first devices, including: wireless display devices, smart TVs, flexible devices, smart retail, and tablet PC 2-in-1 devices. The products developed by him have won more than 120 awards at CES and MWC, and won the CCF Science and Technology Progress Award in 2021. At present, he is in charge of the personal computer computing architecture and interactive innovation for large models. A total of 78 invention patents have been published, including 19 overseas patents.
Title: NPU-based Personal Computing System and Interaction Innovation
Report Summary: The inference requirements of large models on the device side pose a huge challenge to the traditional personal computers with CPU+GPU as the core. The report will explore whether CPU+GPU+NPU can become the next mainstream architecture? How to realize the rapid deployment, inference optimization, and heterogeneous computing of large models based on NPU, what are the common model optimization methods, and what are the application prospects on the device side? The report will further discuss that the improvement of inference capabilities of device-side models will bring opportunities for PC interaction innovation, and summarize the expectations for the future development of AI PCs.
About CNCC2024
CNCC2024 will be held on October 24-26 in Hengdian Town, Dongyang City, Zhejiang Province, with the theme of "Developing New Quality Productivity, Computing Leads the Future". The three-day conference included 18 invited reports, 3 conference forums, 138 thematic forums, 34 thematic activities and more than 100 exhibitions. More than 800 speakers, including Turing Award winners, academicians of the Chinese Academy of Sciences and the Chinese Academy of Sciences, top scholars at home and abroad, and well-known entrepreneurs, looked forward to cutting-edge trends and shared their innovative achievements. More than 10,000 people are expected to attend.