With the rapid development of ChatGPT-like artificial intelligence technology, AI large models as an important technical direction have made significant progress, application scenarios continue to expand and penetrate, and global technology companies have entered the competition. However, the resulting bottleneck of computing power is attracting more and more attention. The high computing power demand of AI large models is giving birth to the rapid iteration of AI chips, "no chips, no AI", and the computing power realized by AI chips as the carrier is becoming an important measure of the development level of artificial intelligence.
First, the background of the birth and development of AI chips
Since Dartmouth College first proposed the concept of artificial intelligence (AI) in 1956, AI technology has continued to make breakthroughs and rapid development, and the demand for computing power has also increased. In order to meet this demand, AI chips have been continuously iteratively upgraded and have become the core basic hardware for computing power improvement.
Before 2006, there was no breakthrough in AI algorithms, and the training data of AI was mainly based on small data. Therefore, the computing power demand of academia and industry for AI is mainly provided by CPU, and the development of AI chips is slower at this stage.
From 2006 to 2016, AI algorithms made breakthroughs in deep learning, while big data, cloud computing and other technologies developed rapidly during this period, further promoting the rapid development of AI in the "big data + deep learning" model, followed by the improvement of AI performance increasingly dependent on the size of computing power. The researchers found that GPUs have parallel computing characteristics compared to CPUs, making them more efficient in "brute force computing" scenarios required by advanced AI algorithms such as deep learning. By giving full play to the advantages of GPUs, the computational efficiency of artificial intelligence algorithms can be greatly improved, which has prompted researchers to widely adopt GPUs for research and applications in the field of artificial intelligence.
After 2016, with the development and commercialization of AI technology, AI chips have entered a stage of great development. In 2016, the AI system AlphaGo developed by Google's DeepMind team in the United States defeated South Korean chess player Lee Sedol, triggering a global AI boom. Since then, the demand for computing power in the AI field has increased. However, the high power consumption and high price of GPUs limit their application in different scenarios. To address these challenges, researchers have begun to develop customized AI chips to accelerate AI algorithm operations while reducing power consumption and cost. Since then, a large number of start-ups and traditional Internet giants have poured into the field of AI chips, promoting the rapid development of dedicated AI chips. In November 2022, OpenAI launched the AI large model ChatGPT, triggering a wave of global AI large model development, which further increased the demand for computing power in the AI field and promoted the investment and development of AI chips.
Second, the technical direction of AI chip development
Broadly speaking, AI chips refer to modules specifically designed to handle a large number of computing tasks in artificial intelligence applications, that is, chips for the field of artificial intelligence are called AI chips. AI chips in the narrow sense refer to chips that have been specially accelerated for artificial intelligence algorithms [1,2]. From the perspective of technical architecture, AI chips are mainly divided into four categories: graphics processing units (GPUs), field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), and brain-like chips. Among them, GPUs are general-purpose artificial intelligence chips, FPGAs and ASICs are semi-custom and fully customized chips for AI demand characteristics, and brain-like chips are processors that mimic the structure and function of the human brain nervous system.
(1) Graphics processing unit (GPU). GPUs were originally specially used for image processing, such as image rendering and special effects production, and were widely used in the AI field because of their excellent matrix computing capabilities and concurrent computing architecture. At present, GPUs have become the most mature and widely used general-purpose chips in the field of AI, favored in large computing facilities such as data centers and supercomputers, and occupy a dominant position in the AI chip market.
Among global GPU manufacturers, NVIDIA GPU technology has always been at the leading level, which integrates the unified computing device architecture CUDA to build an ecological barrier for software and hardware high-performance computing. In March 2022, NVIDIA released the high-performance GPU chip H100 based on the new generation of Hopper architecture at the GPU Technology Conference, which is equipped with the fourth-generation Tensor Core and Transformer engine, compared with the previous generation, H100's comprehensive technological innovation can increase the speed of AI large-scale language models by 30 times.
(2) Field programmable gate array (FPGA). FPGA is a flexible programmable hardware platform with high computing performance and customizability, which can accelerate and optimize AI algorithms. In the scenario of continuously iterative AI algorithms, FPGAs perform well in AI inference applications due to their technical advantages of flexibility, low power consumption and low latency. In November 2022, Intel released the Agilex FPGA chip based on the second-generation Intel Hyperflex FPGA architecture, which integrates the enhanced digital signal processing (DSP) function block that introduces the AI tensor module to better support AI/image/video processing and DSP-intensive applications that can perform complex number calculations.
(3) Application-specific integrated circuits (ASICs). ASIC is an ASIC designed for the user's needs for a specific electronic system, its computing power and computing efficiency can be customized according to the needs of the algorithm, is the product of the optimization design of fixed algorithm. In 2016, Google released the ASIC chip TPU v1, which is mainly used in the AI inference process. Since then, ASICs have overcome the shortcomings of expensive GPUs and high power consumption, and have gradually been applied to the AI field and become an important branch of AI chips. In May 2017, Google released TPU v2, compared with TPU v1, TPU v2's biggest feature is that it can be used for both AI training and AI reasoning. In May 2018, Google released TPU v3, which can achieve more than 100PFLOPS of processing power, almost 8 times that of TPU v2. In May 2022, Google launched TPU v4, which is up to 1.7 times faster processing speed and 1.9 times more energy-efficient than the NVIDIA A100 chip.
In March 2022, China Cambrian Company launched the MLU370-X8 integrated AI acceleration card, equipped with dual-chip four-core Siyuan 370, integrated Cambrian MLU-Link multi-core interconnection technology, which can be applied to AI training tasks such as YOLOv3 and Transformer, and each acceleration card can obtain 200GB/s communication throughput performance, which is 3.1 times the bandwidth of PCIe 4.0, which can efficiently perform multi-chip multi-card AI training and distributed AI inference tasks.
(4) Brain-like chips. Brain-like chips are AI processors designed by combining microelectronics technology and new neuromorphic devices to imitate the human brain nervous system, aiming to break through the "von Neumann bottleneck" and achieve ultra-low power consumption and parallel computing capabilities. Brain-like chips are considered to be one of the important development directions in the post-Moore era and may become a breakthrough in intelligent computing in the future.
In 2017, Tsinghua University developed the second-generation heterogeneous fusion brain-like chip "Tianji Movement", which has the characteristics of high speed, high performance and low power consumption, and the process is 28 nanometers. Compared with IBM's TrueNorth chip, which was the world's most advanced at that time, it had more functions, better flexibility and scalability, 20% higher density, at least 10 times faster speed, and at least 100 times higher bandwidth. In 2019, the paper "Towards Artificial General Intelligence with Hybrid Tianjic Chip Architecture" based on the research results of "Tianji Movement" appeared in Nature magazine as a cover article. In April 2021, Intel released the second-generation neuromorphic chip Loihi 2, which has 1 million integrated neurons, 7.8 times that of the previous generation, and 10 times faster processing.
Third, the development trend of AI chips
The development and upgrading of chips has always relied on the promotion of three aspects: process, architecture and application. In terms of application, with the in-depth development and wide application of AI technology, different AI application scenarios are promoting the development of AI chips in the direction of specialization to meet the needs of performance, power consumption and cost in specific scenarios. In terms of technology, with the development of mimicry neurons, quantum and other technologies, AI chips are constantly breaking through the constraints of traditional architectures and processes on performance, exploring and innovating on different technical paths, showing diversified development directions.
(1) AI scenarios and algorithms promote the specialization of AI chips
Driven by AI algorithms and application scenarios, GPUs, FPGAs and ASICs are showing a development direction characterized by meeting specialized needs. (1) GPUs perform well in processing a large number of parallel computing tasks, and can better realize the potential of AI through accelerated design, but they also have disadvantages such as high power consumption and high cost. At present, GPUs are still the main hardware choice for the computing power required for AI training. (2) FPGAs have strong computing power, low trial and error costs, and sufficient flexibility, but their disadvantages are high prices and complex programming, so they have advantages in semi-customized AI scenarios. (3) ASICs have higher processing speed and lower energy consumption, and can be optimized for specific AI tasks, so as to have better comprehensive quality in performance and energy consumption, which makes them excellent in fully customized AI scenarios.
(2) Brain-like and quantum technologies promote the diversification of AI chips
With the development of cutting-edge technologies such as mimicry neurons and quantum, AI chips have gradually developed new chips with diversified technical paths such as brain-like and quantum, and brain-like chips have begun to move towards commercialization. (1) Brain-like chips have the potential of large-scale parallel computing, ultra-low power consumption and ultra-low latency, which make them play an important role in future AI application scenarios. In the future, an important development direction of brain-like chips is to build a more efficient storage-computing integrated computing system around AI algorithms, such as developing more efficient chip architectures and chips with more neurons, so as to continuously iteratively upgrade the comprehensive performance of AI chips. (2) Quantum chips are chips built based on the principles of quantum mechanics, which can promote the exponential growth of human computing power and form "quantum superiority". Some experts believe that quantum chips are expected to completely solve the problem of AI computing power bottlenecks. In the future, with the widespread application of AI, the demand and power consumption of AI computing power in the whole society will increase significantly, and quantum chips are a potential solution to the above series of problems. However, the current development of quantum computers is still facing problems such as decoherence, resulting in the current quantum chips still mainly exist in the laboratory stage, far from commercialization. In general, brain-like chips and quantum chips, as new chip technologies, have great potential and will play an important role in the future field of AI and computing, bringing us more efficient and powerful computing power.
Bibliography:
[1] James A P., Towards strong AI with analogneural chips[C]. International Symposium on Circuits and Systems, 2020.
[2] Wang Xin, Concept and application analysis of artificial intelligence chip[J]. China New Communications, 2020.
About the author
Liu Jicheng, Research Office II, Institute of International Institute of Technology and Economics, Development Research Center, State Council
Research interests: strategy, technology and industrial frontiers in the field of information
Contact: [email protected]
Editor丨Zheng Shi
About the Institute
Founded in November 1985, the International Institute of Technology and Economics (IITE) is a non-profit research institution affiliated to the Development Research Center of the State Council, whose main functions are to study major policy, strategic and forward-looking issues in the development of the mainland's economy, science and technology, and economy, track and analyze the development trend of world science and technology and economy, and provide decision-making consulting services for the central government and relevant ministries and commissions. "Global Technology Map" is the official WeChat account of the International Institute of Technology and Economics, dedicated to conveying cutting-edge technology information and scientific and technological innovation insights to the public.
Address: Building A, Building 20, Xiaonanzhuang, Haidian District, Beijing
Tel: 010-82635522
WeChat: iite_er