Why did the AI chip enter the shuffle period?
No chip and no AI, chip is the foundation to support artificial intelligence. In 2019, cloud AI chips welcome new players such as Amazon, Qualcomm, Alibaba, Facebook, etc. The trend of integration of software and hardware is strengthened; the power consumption of terminal chips is stronger than competition, and voice chips continue to be hot; the momentum of edge AI chips is emerging. In 2020, AI chips will gradually enter the reshuffle period, with both opportunities and challenges.
Edge AI chips enter the beach fight
AI is expanding from the cloud to the edge, and edge computing is seen as the next battlefield for artificial intelligence. Cambrian Vice President Liu Daofu said that in the category of edge computing, the edge is often connected to various sensors, and the sensor data is often unstructured, which is difficult to directly use for control and decision-making. Therefore, the need for edge artificial intelligence computing to unstructure Structured data is structured for control and decision making.
In 2019, the beachfront layout around edge AI chips has begun. On the one hand, manufacturers such as Nvidia, Cambrian, and Baidu have accumulated in the cloud and end, hoping to improve the cloud, edge, and end ecology with edge chips and create an integrated computing landscape. Nvidia released Jetson Nano, an edge computing device for embedded IoT, suitable for applications such as entry-level network hard disk video recorders, home robots, and intelligent gateways with comprehensive analysis capabilities. Later, it released the edge AI supercomputer Jetson Xavier NX, which can Provides up to 14 TOPS in 10W mode and 21 TOPS in 15W mode. The Cambrian released the SoC edge acceleration chip Siyuan 220 for deep learning, using TSMC 16nm process, maximum computing power 32TOPS (INT4), power consumption control at 10W, support for mainstream programming frameworks such as Tensorflow, Caffe, mxnet and pytorch. Baidu, together with the three major operators, ZTE, Ericsson, Intel, etc., launched the Baidu AI Edge Computing Action Plan, which aims to use AI reasoning, function computing, big data processing, and industrial model training to promote the support and platform of AI scenarios in edge computing. stand by.
On the other hand, the momentum of dedicated edge AI chips such as autonomous driving is gradually emerging. Horizon announces the mass production of China's first automotive-grade AI chip, "Journey II", using TSMC's 28nm process, which can provide an equivalent computing power of more than 4TOPS, with a typical power consumption of only 2 watts and a latency of less than 100 milliseconds. It can run more than 60 classification tasks at the same time, and identify more than 2,000 targets per second.
Multiple new players enter the cloud
The cloud is still the main battlefield of AI chips. In 2019, the cloud chip welcomes many new players, and the computing power war continues to upgrade. Qualcomm launched the Cloud AI 100, a cloud AI chip for data center inference computing, with peak performance exceeding 350 TOPS, which improves performance by 10 times compared to other commercial solutions. Cloud service leader Amazon has launched AWS Inferentia, a machine learning inference chip with a maximum computing power of 128 TOPS. Inf1, an AI inference instance, can carry 16 Inferentia chips, providing a maximum of 2000 TOPS computing power. Alibaba launched the so-called world's highest-performance AI inference chip containing Light 800, using its self-developed chip architecture and Dharma Academy algorithm, to achieve the highest single-chip performance in the Resnet50 benchmark test. Liaoyuan Technology invested by Tencent has released the AI Accelerator Card Cloud T10 for cloud data centers. The single-card single-precision computing power reaches 20TFLOPS. It supports mixed-precision calculations of single-precision FP32 and half-precision BF16. It has three modes: single node, single cabinet, and cluster. In the cluster mode, the 1024-node cluster is realized through inter-chip interconnection.
The chip is the carrier of AI, and software is the core of intelligent operation. As heterogeneous computing is gradually introduced into AI chips, software and hardware collaboration has become an important trend in cloud AI. Intel introduced the One API, a unified software platform for heterogeneous computing, to hide hardware complexity, and automatically adapt the acceleration method with the lowest power consumption and best performance according to the system and hardware to simplify and optimize the programming process. Xilinx has also launched the software platform Vitis AI, which provides users with an easy-to-access software interface that can automatically adapt to Xilinx hardware architecture based on software or algorithms.
Power consumption ratio is still the focus of the terminal
On the terminal side, the power consumption ratio is still the focus of competition. Especially for mobile phones and other terminals that must be compared with the battery life, the AI engines introduced by the main manufacturers have emphasized low power consumption. Kirin 990 5G's NPU adopts dual large core + micro core method. The large core is responsible for performance, and the micro core has ultra-low power consumption. According to reports, in the application scenario of micro-core, the energy consumption of the micro-core is reduced by 24 times compared with the large-core operation. Snapdragon 865 released by Qualcomm integrates the sensor hub, allowing the terminal to sense the surrounding environment with very low power consumption. Samsung proposes to implement AI processing on terminal devices through NPUs with lower power consumption, enabling more complex tasks to be performed directly on the device side.
In addition to mobile phones, another popular fried chicken on the terminal side is the AI voice chip. IFlytek, Alibaba, Prospecting Technology, and Qingwei Intelligent have all released AI voice chips for smart homes, reflecting the professional and customized trend of AI chips in specific fields. Ali Dharma Academy announced the first AI FPGA chip technology Ouroboros dedicated to speech synthesis algorithms, using on-board custom hardware acceleration technology to reduce dependence on cloud networks, support real-time speech synthesis and AI speech recognition, and is expected to be in Tmall Genie Piggyback.
Opportunities and challenges coexist in 2020
From 2019 to 2021, the size of China's AI chip market will still maintain a growth rate of more than 50%. By 2021, the market size will reach 30.57 billion yuan. CCID think tank predicts that the growth rate of cloud training chips will slow down in 2019-2021, and the growth rate of cloud inference chips and terminal inference chips will continue to increase. It is estimated that by 2021, the market scale of China's cloud training chip will reach 13.93 billion yuan, the market scale of cloud inference chip will reach 8.22 billion yuan, and the terminal inference chip will reach 8.41 billion yuan.
Yaobang Yang, an analyst at Jibon Consulting, pointed out to reporters that in 2019, AI chips have generally taken a clearer path, and the chip specifications of the end, edge, and cloud are relatively clear. In 2020, major chip makers will continue their product development path in 2019, and continue to deepen the chip's cost performance and power consumption performance. From the training point of view, it is worth paying attention to the integration of HBM (high-frequency wide-band memory) and related packaging technology yields, which will affect the changes in the cooperative relationship between chip manufacturers and memory and packaging and testing manufacturers. The decisive point of inference is in the field of INT8, and the focus is on how to further improve the performance and power consumption of the chip itself.
New technologies such as 5G and VR / AR will also provide more room for AI chips, especially AI chips on the edge side. Dennis Laudick, vice president of business and marketing of the Arm ML business group, told reporters that 5G communication technology has changed the way data is processed, so that edge AI workloads also have processing requirements. It can be said that 5G brings more innovation at the edge of the network. Yao Jiayang also said that AI has opportunities in the 5G core network. As 5G brings more spectrum combinations, AI can assist the core network to more effectively schedule network resources and maximize the use of bandwidth resources. At the same time, 5G also covers the Internet of Vehicles, and AI will have a great opportunity in autonomous driving. On the VR / AR side, AI is also being introduced, mainly focusing on applications such as eye tracking or scene recognition, which is expected to improve the smoothness and real-time performance of VR / AR.
Wei Shaojun, the director of the Institute of Microelectronics at Tsinghua University, said that from the perspective of industrial development, AI chips will continue to be hot in 2019-2020, and enterprises will gather to enter; The card starts. Since the AI algorithm is still in the process of continuous evolution and aggregation, the ultimate success or otherwise will depend on the choice of each technology path and the speed of product implementation.
Pain points yet to be overcome
In the past two years, AI has made breakthroughs in applications such as speech recognition and image recognition. However, to move from a single point of breakthrough to full bloom, it is necessary to create a general AI computing chip like the CPU in the AI field. Yin Shouyi, associate professor of the Department of Micro-Nano Electronics of Tsinghua University, pointed out that AI chips are mainly heterogeneous computing in the short term, and self-restructuring, self-learning, and self-adaptation will be developed in the medium term, and general computing chips will be developed in the long term.
Specifically, if AI is to adapt from hardware to hardware, it will require AI chips to have a programmable and dynamically variable computing architecture to cope with new algorithms and applications. Wei Shaojun said that the AI chip must adapt to the evolution of the algorithm, and it must have a structure that adapts to all applications, which requires the architecture to have efficient transformation capabilities. In the field of cost-sensitive consumer electronics, it is also necessary to pay attention to the computing efficiency of AI chips, to achieve low power consumption, small size, and easy development, all of which need to explore architectural innovation.
The global AI chip industry is still in the early stages of industrialization. Domestic processor manufacturers and international manufacturers are on the same starting line in the new field of artificial intelligence. Liu Juncheng, the founder and CEO of Nengeng, said that China has a huge market for smart phones, smart homes, and smart security. For Chinese AI companies, they not only have the advantage of localization when serving domestic customers, they can also take advantage of these customers' The manufacturing advantage enters overseas markets and realizes the business layout of “based on China and global vision”.