The market for AI chips and technologies is growing rapidly
According to TrendForce, under the trend of global digitalization and intelligence, the Internet of Things devices continue to expand, such as industrial robots, AGV/AMR, smart phones, smart speakers, smart cameras, etc. In addition, the deepening application of medical PCBA, image recognition, speech semantic recognition, computing and other technologies in various fields catalyze the rapid growth of the AI chip and technology market. The global AI chip market is expected to reach 39 billion US dollars in 2022, with a growth rate of 18.2%, which drives the need for IC substrate.
Due to the majority of AI chip applications in cloud computing, security, robotics and vehicle use, 2023 will enter the high growth period, especially the rapid growth of the market led by cloud computing and vehicle use, the global AI chip market size is expected to reach 74 billion dollars by 2025, and the CAGR will reach 23.8% from 2022 to 2025.
TrendForce said the growing popularity of various consumer electronics and digital technologies around the world is driving the growth of the semiconductor market. It also allows the most scalable technologies such as AI and the Internet of Things to be widely used, enabling AI chips to process large amounts of data in a shorter time. Therefore, the rapid growth of AI/ML (Machine Learning), Internet of things and wireless communication equipment also stimulates the demand for semiconductors.
Moreover, AI chip in reasoning tasks, mainly through the terminal equipment of the sensor, the microphone array or lens for data collection, and will collect data generation into the trained model inference reasoning results, to minimize human errors, it also suggests that different terminal scene, demand for the performance of the work force, such as energy consumption also has difference, Therefore, special design for specific application scenarios is needed to achieve the optimal solution.
AI chips are rapidly expanding market share
From 2020 to 2021 new outbreaks leading digital wave, combined with the United States, China and the European Union countries have issued “several intelligent development strategy”, catalytic national and industrial digital transformation, such as the United States vigorously promote “made in USA” and “digital economy” process, the former mainly focuses on the semiconductor industry, lock in IC design, production process and core equipment.
The latter, which focuses on digital dollars, is currently in the process of developing a verification system and drafting a bill. In the EU, “Europe’s Digital Compass” program focuses on four categories: talent cultivation, ensuring security and sustainability, digital transformation of enterprises, and digitalization of public services. Digital transformation of enterprises is the top priority. By 2030, 75% of manufacturers are expected to make widespread use of cloud computing, big data and AI, and at least 90% of smes should be basically digital intensive.
National policies drive the industry to speed up the pace of digitalization, but also affect the growth of data centers, the rapid development of machine vision and autonomous driving technology, deepen the storage and logistics, logistics transportation of supply chain, self-driving and other fields of application, so that the global demand for AI chips is soaring, and rapidly expand its market share in the overall semiconductor market. Its market share expanded from 5.9% in 2020 to 6.4% in 2021, and the market size also grew by 26.9% from USD 26 billion in 2020 to USD 33 billion in 2021.
At present, the most widely used chip for Deep Learning (DL) belongs to GPU, which is good at parallel operation. With the increasing operation demand of Deep learning, and in order to meet the operation requirements of DNN (Deep Neural Network), NVIDIA, Cambrian, Google, Intel and other manufacturers actively explore the application and breakthrough of GPU in high-efficiency computing, and focus on the research and development of high-efficiency computing chips and related production plans.
Some vendors are seeking semi-custom chips based on FPGA (Field Programmable Logic Box Array) architecture, such as Google’s tensor processor TPU, Cambrian’s neural network processor NPU, and Intel’s Altera Stratix V FPGA. In addition, Intel also tries to promote XPU Programming for different brands and different types of hardware, and uses SYCL as a unified programming language and makes full use of Intel oneAPI, so that the original development framework from closed to open, in order to more effectively support developers to build high-performance heterogeneous applications.
AI leads the era of strong computing power and initiates multi-scenario applications in collaborative operation
AI chips can be roughly divided into CPU, GPU, FPGA and ASIC (customized chip). Among them, GPU is a large-scale parallel computing architecture composed of a large number of computing units, which is designed for processing multiple tasks at the same time. It is mainly used in workstations, personal computers, game devices, smart phones and other devices to process graphics and image-related computing.
Furthermore, the chip adopts unified rendering architecture, which can be used in the fields where the algorithm has not been finalized, so the degree of generality is high and the commercialization is mature.
FPGA mainly provides users according to their own needs repeated program design, solve the electricity shortage problem of route can be programmed component, and the efficiency is higher than the GPU, CPU, power consumption is relatively low, but when dealing with the task of repeatability is not strong, logic is relatively complex, the efficiency of the chip will be lower than the use of the von neumann architecture processor.
ASIC is a customized chip based on a specific algorithm and architecture. Its customization degree is higher than GPU and FPGA, and its specificity is strong. Therefore, its operation level is generally higher than CPU, GPU and FPGA. But as the amount of data grows and chip technology reaches its limits, the demand for computing power becomes harder to satisfy.
In particular, as the amount of data in some specific fields becomes larger and larger, the algorithms are gradually fixed, and the demand for DPU, TPU and NPU designed based on ASIC architecture increases. In particular, DPU can partially replace some functions of CPU and GPU and solve the data transmission problem between CPU and Memory caused by the sudden increase in data volume. Improve and accelerate the network data transmission operation speed.
At present, it is widely used in large data centers. Because the traffic processing of large data centers takes up nearly 30% of the overall calculation, and the switching efficiency between nodes and I/O switching efficiency within nodes of data centers is low, DPU can effectively solve the loose coupling problem by cooperating with CPU and GPU.
Therefore, with the continuous expansion of Internet of Things devices, such as industrial robots, AGV/AMR, smart mobile phones, smart speakers, smart cameras, etc., coupled with the deepening application and upgrading of automatic driving, image identification, speech semantic identification, computing and other technologies, the market for AI chips and technologies will grow rapidly.
In terms of the overall market in 2022, the demand for AI chips in the three fields of intelligent vehicles, robots and data centers will continue to increase, and the computing capacity and technical architecture will be continuously improved to meet the demand of these three fields. Among them, the smart car, the car electronic electrical architecture from distributed to centralized, MPU, the MCU demand increased year by year, plus electric cars fit closely now advanced driving assistance systems (ADAS) technology, make the depot with the conditions of the AI chip to solve complex operation, the power consumption of the vehicle running and data transmission, to improve the stability of the vehicle, security.
In addition, in recent years, the technology diffusion of robots is quite fast, and the application scenarios have extended from industrial environment to restaurants, restaurants, hospitals, warehousing and logistics, national defense and space exploration, etc. However, in order to enable robots to run image processing, face recognition and other functions, GPU and FPG are selected.
Among them, FPGA has a high utilization rate, because the chip has the characteristics of low power consumption, high performance, reconfiguration ability and self-adjustment, and can integrate the Robot Operating System (ROS) into the FPGA platform, so that the software and hardware inside the robot can effectively interact and play the best operation efficiency.
Moreover, robots are rapidly developing towards multi-dimensional processing capabilities such as 3D entities, Work Breakdown Structure (WBS) and time. In view of this, the GPU and FPGA architectures at the present stage will continue to innovate and make breakthroughs, and even design for special needs, which is bound to lead to the synchronous upgrade of manufacturing, sealing test and equipment, as well as the whole chain of materials and software.
Above all, smart cars, robot, and the Internet of things technology, rather than build intelligent transportation, intelligent new blue ocean, such as factories and smart cities for AI chip market demand and development period, also drives the further grow up of data center at the same time, its in the center of the training, to create a new operation model and maintain server operation, power control chips are used in a large number of AI.
For example, Tesla has expanded its GPU-driven AI Super, increasing the number of A100 GPU configurations to 7,360, ahead of deployment for DoJo. Baidu self-developed the second-generation Kunlun chip to enhance the efficiency of quantum computing, and launched the superconducting quantum computer in August 2022, as well as the full-platform quantum hardware and software integration solution, which can access various quantum chips through PC, mobile phone and cloud.
Under the rising demand of various parties, AI chips are bound to grow rapidly. It is expected that the market scale of AI chips is expected to reach 93 billion dollars in 2026, among which CPU and GPU still occupy the main share of the AI chip market and grow steadily, while the ASIC market has broad prospects. Its advantages and features can help users in data processing, consumer electronics, telecommunications systems, industrial and other industries to develop a range of products, shorten the innovation cycle of a product, service or system.
TrendForce research shows that CPU, GPU and ASIC chips will account for 33%, 34% and 26% of the AI market size in 2026, among which the ASIC chip market is growing the fastest. There are two reasons. First, the market demand for consumer electronic equipment is increasing, and most small and medium-sized device developers prefer 7nm ASics. Second, the workload and structuring needs of 5G, low-orbit satellite communications, cloud and edge computing are increasing, as telecommunications systems are the largest end-use market.