On July 4-6, the 2024 World Artificial Intelligence Conference and High-level Conference on Global Governance of Artificial Intelligence (WAIC 2024) was successfully held in Shanghai.
Figure | Tang Wenkan, deputy director of the Shanghai Municipal Commission of Economy and Information Technology, source: VeriSilicon
Tang Wenkan, Deputy Director of the Shanghai Municipal Commission of Economy and Information Technology, said in his opening speech: "Shanghai has cultivated a series of RISC-V enterprises, based on the results formed in the early stage, Shanghai will accelerate the construction of the RISC-V patent pool, promote the joint construction and sharing of the industry, strengthen the research of cutting-edge technologies and the cultivation of talents, strengthen the cooperation with the RISC-V Foundation, actively participate in the expansion of new instruction sets, the exploration of new technologies, study the establishment of a new industry-university-research cooperation mechanism, and promote the inclusion of RISC-V in university textbooks." As well as the training plan for talents in short supply of integrated circuits in Shanghai, we will further optimize the business environment, continuously improve the supply of services and systems, and work together with you to build a world-leading industrial highland for green development. ”
Figure | Zhang Xiaodong, Chairman of Wuzhen Think Tank, source: VeriSilicon
Zhang Xiaodong, chairman of Wuzhen Think Tank, pointed out in the theme report on "The Principles and Development and Application of Large Language Models": "The Church-Turing thesis is the cornerstone of computer science, the origin of artificial intelligence comes from the Turing machine, and learning is the inverse of the Turing machine, and learning is Solomonoff's induction. To talk about the difference between the two, BERT is bidirectional, for example, give X1 to XN, then remove X2, and then guess X2; GPT gives X1 to XN and predicts XN+1, so the mechanism of GPT is simpler than BERT. In addition, Zhang Xiaodong believes that there are three main development routes of artificial intelligence, namely: first, logicism from expert system to knowledge graph; 2. The nexus of neural networks and deep learning; 3. From cellular automata to reinforcement learning. As for whether the future will be dominated by large models or small models, Zhang Xiaodong said: "When artificial intelligence matures, will it be a few super large models monarchy or multiple models of democracy? It will take time to verify, and it is not yet possible to draw a conclusion. ”
Figure | Chairman of China RISC-V Industry Alliance; Dai Weimin, Founder, Chairman and President of VeriSilicon, source: VeriSilicon
In recent years, ChatGPT and Sora have successfully made artificial intelligence out of the circle, and have generated a global "big model and AIGC fever". In this regard, Chairman of China RISC-V Industry Alliance; Dai Weimin, founder, chairman and president of VeriSilicon, said: "Before GPT came out, artificial intelligence was still at the level of the mentally handicapped; By 2025-2026, artificial intelligence will surpass the level of university students; By 2027, AI models will be able to do the job of AI researchers/engineers, and a leading AI lab will be able to train a GPT-4 model in less than 1 minute. It is expected that by 2028, the number of basic large models in China will be less than 10, and the sales of device-side fine-tuning cards and inference cards will exceed those used for cloud-side training cards. Focusing on the hardware level, a start-up team has recently launched the first Transformer-specific ASIC chip. "By 2030, generative AI will nearly triple semiconductor revenues in the smartphone space," Dai stressed. "VeriSilicon's ultra-low-power IP and FLEXA IP provide a more competitive base for the lightweight wearable and consumer electronics markets. In addition, Dai Weimin also pointed out: "In the future, Chiplet, 2.5D/3D packaging, and SIP will become the mainstream solutions for large chips, and the current M series chips cooperated by Apple and TSMC are the best case show. ”
Figure | Dai Lu, Chairman of RISC-V International Foundation, source: VeriSilicon
The RISC instruction set is not new, ARM is one of them, and RISC-V is the fifth generation and the latest RISC instruction set. Dai Lu, Chairman of RISC-V International Foundation, said: "Although most of the companies that have joined RISC-V International Foundation are small companies, their size is extremely objective. Why the RISC-V instruction set is very suitable for the AI field, Dai Lu gave several reasons: "First of all, RISC-V is open source and free; The second is the flexibility of RISC-V, AI companies do have a lot of flexibility when using the RISC-V architecture, they can add custom or featured instructions according to their specific needs, while still maintaining their compatibility, and if the custom instruction set developed by the company is also valuable to other users, they can submit proposals through the RISC-V community process and strive to incorporate them into future official standards; In addition, the RISC-V ecosystem is thriving today, and engineers are embracing the RISC-V ecosystem. According to the statistics of the RISC-V International Foundation, AI accelerators are currently the field with the highest market proportion and the fastest growth rate in the downstream ecosystem of RISC-V. According to SHD Group, the size of the AI acceleration market reached $2.6 billion, accounting for 42.6% of the RISC-V SoC market, and it is expected that the proportion of AI accelerators in the RISC-V SoC market will continue to rise to 49.2% by 2030, which also supports the conclusion of the RISC-V International Foundation.
Figure | Chief CPU Architect, Tenstorrent Inc.; Weihan Lian, Senior Academician of Machine Learning Hardware Architecture, source: VeriSilicon
AI is undoubtedly an important technology to change the future, since 2012, large models have gradually developed, and today several large model architectures have become mature, but the problem of computing power has not been well solved, and this challenge must be overcome. In this regard, Chief CPU Architect of Tenstorrent Inc; Lian Weihan, a senior academician of machine learning hardware architecture, said: "The demand for computing power will appear in all aspects, and will be distributed in various hardware demand stages. At the same time, the company has developed an AI accelerator based on RISC-V architecture and Chiplet technology, which is an open-source instruction set architecture (ISA) that has become the core of Tenstorrent technology, enabling the company to develop highly efficient and flexible AI processors, from Grayskull in 2021 to Aegis in 2024, and has currently cooperated with many international manufacturers such as Samsung, LG, and LSTC. In the future, the open-source RISC-V still has a lot of room for imagination, especially in computing hardware, which will play a more enabling role. ”
图 | MIPS CTO Durgesh Srivastava，来源：芯原
As we all know, MIPS was one of the early chargers in the RISC camp, and also tried the open source strategy in 2018, so what else can MIPS play in the AI era? In this regard, MIPS CTO Durgesh Srivastava said in a speech entitled "The Unleashing of Generative AI: The Transformative Power of MIPS and RISC-V": "AI is driving the future and empowering thousands of industries, but the support of computing power cannot keep up with the demand for computing power, and it is imperative to improve computing power, and who will lead the development of hardware in the AI era?" MIPS will be part of the AI revolution and play a role in real-time adaptive processors, hybrid AI/ML and electrostatic processors, AI-driven dynamic power management, accelerated product customization, and AI-powered security chips. At the same time, Durgesh Srivastava called on more professionals to participate in the RISC-V AI community, innovate with open standards, and promote the growth of AI computing power through the architecture innovation of hardware chips, so as to further promote the industrialization of AIGC and promote the revenue growth of enterprises. The forum also featured a roundtable session with the theme of "Convergence and Innovation of Generative AI and RISC-V", moderated by Dai Weimin, who recorded the key discussion topics and insights of experts below.
Figure | "Convergence and Innovation of Generative AI and RISC-V" forum, source: VeriSilicon
1. Aschenbrenner (former security researcher at openAl) released "Situation Awareness" in June 2024, analyzing the path and potential impact of AI evolving from its current state to AGI and further to super-artificial intelligence (ASI). Based on the development status and trend of the AI industry at home and abroad, when is it expected to reach superintelligence? What are the potential impacts on energy, the environment, and human security? Dean of the Institute of Political Science, East China University of Political Science and Law; Gao Qiqi, President of the Institute of Artificial Intelligence and Big Data Index: At present, artificial intelligence is developing very rapidly, and at the same time, 2027 may bring some "super-intelligence" problems. If this happens, it will definitely bring about an unemployment crisis; In addition, it will bring about information bubbles, lose trust in information, and affect order, including political order; In addition, AGI will also bring a risk of getting out of control, including the risk of malicious actors using AI to make superweapons and super viruses. Therefore, the large model must be classified, the current super model with 10 million H100 cards, must use the same method to control nuclear weapons, there can not be too many actors with such superpowers, otherwise it will be very difficult to control. In addition, we can also refer to the governance model of carbon emissions, use the method of paying taxes to adjust the speed, and solve the problems it generates. 2. Regarding the "autocratic" problem of a few large models, will there be a concentrated use of a few very powerful large language models (LLMs), such as Google's PaLM 2 and Gemini models; Trained on large-scale data and compute resources, these models are able to handle a wide range of tasks and perform well across multiple benchmarks? Zhang Xiaodong, Chairman of Wuzhen Think Tank: I think there may be two outcomes for the development of global AI: one is dominated by a very small number of giant models; The other is the coexistence of multiple larger models. If, after four years, there are only a few large models, as is the case with the global means of controlling nuclear proliferation, the situation of only a few large models is actually manageable. In addition, there is a high possibility that super artificial intelligence will surpass human intelligence in the future, and whether there is a way to control super artificial intelligence is not optimistic at present. Dean of the Institute of Political Science, East China University of Political Science and Law; Gao Qiqi, President of the Institute of Artificial Intelligence and Big Data Index: Large models are the foundation of knowledge, which will create a problem, adding only a few very popular large models in the world, all of which are expected to be trained in English, which will form marginalization for non-English speaking countries and eliminate the diversity of knowledge, so the concept of "sovereign AI" has recently emerged, so how to balance is a difficult problem. 3. How do AI and RISC-V combine? What exactly is the relationship? Meng Jianyi, CEO of Zhihe Computing: RISC-V starts with CPU, but it is better than AI. Today's architecture can no longer meet the rising speed of AI computing power demand, so architecture innovation must be done. In the past 10-20 years, the architecture of one CPU and many accelerators has been the mainstream, and today's prosperity of "X" PU, only the architecture of RISC-V can meet its future development trend, which is also a demonstration of the vitality of RISC-V. Startup Etched Al's Transformer ASIC "Sohu" delivers 20 times the performance of Nvidia's products, which is also a very good attempt, and there will be more breakthroughs if they use RISC-V. 4. [Voting] Will ASICs for specific AI models replace GPUs and become the development trend of AI inference chips in the future? Voting Results:
Source: VeriSilicon
5. The triangular strength of x86, Arm, and RISC-V, which architecture has more advantages for CPU, and is there an opportunity for ARM and RISC-V in the era of AI PC? Dai Lu, Chairman of RISC-V International Foundation: CUDA is the most advantageous AI ecosystem, and many companies make GPUs, not because they really want to make GPUs, but for AI. Both Arm and RISC-V belong to the RISC line, and both have their own unique advantages. Compared to x86, Arm and RISC-V have power advantages. As for the difference between Arm and RISC-V, RISC-V can achieve the same base model, while still retaining the characteristics of each company's own model; In addition, in terms of security, the AI protection of each country does not want to be cracked by other countries, so this is also something that can only be achieved by RISC-V. Voting Results:
Source: VeriSilicon
6. Is there an advantage for server-grade RISC-V processors to compete with Arm and x86 for future automotive architectures? Peng Jianying, CEO of Xinlai Zhirong Semiconductor Technology (Shanghai) Co., Ltd.: We have been making processors for many years, but in fact, the architecture has not changed much, and it is indeed because AI, whether it is on the server or on the device side, such as automobiles, has brought a large incremental market. One of the things that most startups are doing these days is to strike a better balance between so-called DSA and general-purpose CPUs or general-purpose GPUs. As far as Xinlai is concerned, we believe that the automotive future is a better growth market, of course, the rapid development of intelligent driving and intelligent cockpit is obvious to all, this change is not only the enhancement of computing power, but also needs a better experience, better APP empowerment. However, from these differences, we can find some commonalities, such as general security AI software, functional safety, information security, and timely response, etc., and the integration of these chips requires a unified software ecosystem. So, why was RISC-V given such an opportunity? This is because the RISC-V architecture can meet all the needs of automotive electronics from low to high, and meet the change of automotive electronics from traditional to software-defined, so I think RISC-V is a trend in the future development of automotive electronics.