AI chip "hashing power burst", how to break through the layout of industrial applications?

As the underlying computing power support of the AI ​​industry, AI chips have become the absolute focus.

The AI ​​industry is booming, and AI chips are absolute stars.

"The Chip Forum is extremely hot. More than 1,000 people have made appointments, but there are only 200 on-site locations. The organizer recommends that we come to the site as soon as possible." Before the World Artificial Intelligence Conference "AI Smart Chips Define Industry Future Forum", some vendors asked for Feedback from interface journalists.

As the underlying computing power support of the AI ​​industry, AI chips have become the absolute focus. Generally speaking, chip-driven artificial intelligence and autonomous driving are the focus of the exhibition. Whether it is the exhibition area of ​​the conference or the industry forum, it is a topic that visitors stop to watch and discuss eagerly from time to time.

Domestic AI chips accelerate to catch up

Inside and outside the WAIC venue, AI chip companies have a strong sense of existence. Several chip companies released AI chip products during the exhibition and stepped onto the world stage.

AI calculation is roughly divided into two levels. The first is to train the model. The whole process may take several days; then, the trained model responds to actual requests and makes inferences. At present, NVIDIA's GPU (graphics processing unit) occupies the training market, and most inference tasks are still undertaken by traditional Intel CPUs. For training and reasoning scenarios, domestic AI chip manufacturers have launched related products.

The domestic AI chip start-up company Suiyuan Technology released the second generation of artificial intelligence training products-"Yunsi 2.0" chip, "Yunsui T20" training accelerator card based on Yunsi 2.0 and "Yunsui T21" training OAM module, comprehensive The upgraded "TopsRider" software platform and the brand-new "Yunsui Cluster", at the WAIC exhibition venue, the Suiyuan Technology booth attracted much attention from the audience.

The founding team of Suiyuan Technology comes from NVIDIA's long-term competitor AMD. The founder and CEO of the company, Lidong Zhao, used to serve as the vice president of Ziguang Group. He had worked for AMD for 7 years as a senior director of the computing division. In the 1990s, he worked for S3 Graphics, a veteran display chip company.

"Intel’s chip technology development strategy is the Tick-Tock model, the Tick stage upgrades the process, the Tock stage upgrades the processor architecture. For us, our first-generation chip is a brand-new architecture, and for the second-generation we have an option, architecture Keeping roughly the same, the process evolves forward, and then the architecture and then the process. This is to follow the traditional Tick-tock road. The first generation of chips is in the process of landing and running in with customers. We feel that the evolution of the architecture is more than that of the process. Evolution is more important." Zhao Lidong said.

In addition to Suiyuan Technology, Tianshu Zhixin demonstrated the domestic first fully self-developed, 7-nanometer process cloud training chip B1 and GPGPU (General Purpose Computing GPU) product card under the GPU architecture. This chip can hold 24 billion transistors and adopts 2.5D CoWos wafer packaging technology, supports FP32, FP16, BF16, INT8 and other multi-precision data mixed training, supports inter-chip interconnection, and single-chip computing power per second is 147T@FP16. The reason for choosing the new 7-nanometer process is that AI chips have higher process requirements, and the pursuit of more advanced chip manufacturing processes has also become the focus of AI chip manufacturers.

In terms of reasoning chips, Denglin Technology has demonstrated its own innovative GPGPU chip, dedicated to solving versatility and high-efficiency problems. Based on providing CUDA/OpenCL hardware acceleration capabilities, it fully supports various popular artificial intelligence network frameworks and underlying computing. child.

Zhang Linglan, a co-founder of Biren Technology, told the outside world recently that computing power chips are the engine of the artificial intelligence era. Compared with decades of accumulation in the West, my country’s chip manufacturing industry is still in its infancy. According to reports, Biren Technology will be released next year. The first general-purpose intelligent computing chip product.

Not surprisingly, NVIDIA, which has a huge advantage in artificial intelligence, has become the target of comparison among various manufacturers. At present, competition in the AI ​​chip field is fierce. At present, GPU (graphics processing unit) manufacturer NVIDIA is the market leader in graphics computing and AI computing.

New architectures and products of AI chips are emerging endlessly. Server and cloud computing vendors are open to the choice of new chips. Liu Jun, general manager of Inspur AI&HPC, once told Jiemian News that competition in the AI ​​chip field is very fierce, with a large number of companies, including in China, the United States, and the United Kingdom. There is a certain degree of homogeneity.

Chip giant AI has its own "unique skills"

The major chip giants also announced their own AI application solutions at the venue.

In his speech at the 2021 World Artificial Intelligence Conference, Apple’s vice president and managing director of Greater China Ge Yue introduced in detail the AI ​​applications, architecture and models of Apple’s chips.

"Apple's chips are not designed to be sold to other companies, but are specifically designed for Apple's specific products, or even for specific functions. This optimization runs through the CPU, GPU, image signal processor and more components. "Ge Yue said.

The persistence in self-research is also reflected in machine learning. Apple's chips are equipped with a "neural network engine." The so-called "neural engine" is Apple's hardware dedicated to machine learning in the chip, which can be used for image processing, face recognition, etc. At present, the neural network engine in the A14 Bionic chip on the iPhone 12 can perform 11 trillion operations per second, making it one of the most powerful mobiles AI chips.

Ge Yue further introduced that the M1 chip adopts the Arm architecture design, is a 5-nanometer process, and is the most advanced chip in current computer products. There are 16 billion transistors on this chip, which integrates the central processing unit, graphics processor, neural network engine, various connection functions, and many other components on the chip.

Thanks to this, the M1 chip has completed seemingly impossible tasks: significant performance improvements, including machine learning speed up to 15 times, and battery life up to 2 times. “Apple’s chips are tailor-made for its own products and are specifically optimized for machine learning. Since machine learning can provide users with a more unique experience, Apple has long started to develop complex machines that can effectively run on the device side. A chip for learning algorithms."

The veteran chip giant Intel also has its own understanding of artificial intelligence. "In the past two years, we have discovered that there are a lot of autonomous devices and a lot of digital systems that need to be processed by artificial intelligence. Artificial intelligence has become a superpower to promote digital transformation." Vice President of Intel Research Institute, Intel China Research Institute Dean Song Jiqiang said at the WAIC meeting that he pointed out that in terms of data changes, processing massive data can only rely on artificial intelligence, and diversified data forms also require many new algorithms to sort out.

In this regard, Intel believes that in the new era, it is necessary to master more different architecture combinations to meet the needs of exclusive specific fields. Therefore, architecture innovation is the key driving force. "CPU is suitable for processing scalar architecture, one by one, such as control flow, is very easy to process, and can be concurrent; GPU is suitable for processing vector operations, and a lot of data is calculated together; AI is more of block operations and requires specialization. Matrix acceleration, data access also needs to be optimized; FPGA is particularly suitable for some sparse calculations, which can greatly reduce I/O and calculation consumption. These can be integrated to get what they need. We say that we often play combination punches. It will be better than using only one weapon to solve all problems." Song Jiqiang explained to the interface reporter.

Qualcomm has a certain amount of research on integrating AI functions on its mobile chips. Meng Pu, chairman of Qualcomm China, introduced that AI technology is driving a new generation of smart edge terminals and applications including ultra-fast 5G connections, high-performance and low-power computing, and terminal-side applications. With the development of cloud computing, everything can be connected to the cloud in real-time, allowing terminals, experiences and data to benefit from the ever-increasing content, processing power and cloud storage space.

Meng Pu explained that to achieve the scale of AI, it is first necessary to ensure that intelligence is widely distributed throughout the network. The industry predicts that from 2020 to 2026, monthly mobile data traffic will increase by more than 500%, and billions of new connected terminals and things will be distributed on the edge. Effectively responding to the rapid growth of data requires not only the transmission of data to the cloud, but also the integration of AI capabilities on the terminal side and the direct operation of algorithms to provide a powerful supplement to cloud intelligence.

The terminal-side AI has several key advantages, including higher immediacy, reliability and security. These advantages are essential for delay-sensitive and business-critical applications, such as autonomous vehicles, smart grids, and networked infrastructure.

“The rich data generated by the smart edge terminal can be shared to the cloud in real-time, so that AI can fully play its role and realize the large-scale application of AI from the cloud to the edge. Nowadays, AI is integrated into almost every aspect of the smartphone experience, from images to Voice recognition and security, and soon AI will also bring experiences that include a high degree of personalization, interactivity and contextual relevance." Meng Pu said.

The scene becomes the AI ​​chip powerpoint

If the AI ​​chip does not have a landing scene, it may be empty talk. The scene-oriented and actual needs have become the publicity focus of many AI chip manufacturers.

For example, Chen Tianshi, the founder and CEO of Cambrian, announced on WAIC the car smart chip that Cambrian is designing, saying that "has more than 200TOPS computing power, adopts 7nm process, has car-level standards and independent safety islands, and It will inherit the Cambrian unified software toolchain that integrates the edge and end of the cloud."

Cambrian is the first startup company to launch cloud smart chips in China. It has a complete product line of the cloud training chip Siyuan 290 and the inference chip Siyuan 270, as well as the Siyuan 220 chip for edge computing. "Today, the Cambrian has formed a comprehensive range of AI chips covering the three fields of cloud, edge and end, including training and inference. The current Cambrian product line is characterized by the integration of cloud, edge and end." Chen Tianshi said.

Currently, Journey 3 of Horizon Automotive Grade Chips sells more than 45,000 pieces, and is expected to sell more than 200,000 pieces throughout the year, mainly for models such as Ideal One. As an upgraded version of Journey 3, the Horizon Journey 5 chip was lit up in May of this year for L4 high-level autonomous driving.

In addition to autonomous driving, AI chips for smart security, industrial vision, and vehicle vision scenarios have become popular. During WAIC, Coolcore Microelectronics released a new high-definition AI camera chip AR9341. The founder of Coolcore Microelectronics, Shen Bo, said at the press conference that the chip integrates Coolcore’s self-developed deep learning processor (NPU) with 4TOPS (equivalent to 4TOPS). (4 trillion times per second) peak computing power, the actual computing power after architecture optimization can be equivalent to 8-10TOPS of competing products.

As a mid-to-high-end camera chip product, the target application areas of the AR9341 chip include high-end intelligent IPC, vehicle-assisted driving, edge computing boxes, and intelligent robots. Its engineering samples are expected to be available in September this year, and mass production time is December this year.

But not everyone is optimistic about such chips, especially AI chips for specific subdivision scenarios. This type is actually sacrificing versatility in exchange for higher efficiency in specific tasks such as AI. Some people in the industry believe that a very prominent problem with AI chips on the market is the lack of versatility of the products. At present, the only truly universal is the GPU. He believes that AI chips will flourish, and new products will appear in different fields.

In this regard, Ren Xiang, director of the Integrated Circuit Evaluation Center of the China Electronics Standardization Institute, reminded that while the AI ​​chip industry is developing upwards, it is also facing many challenges: first, AI chips are facing more extensive and diversified needs, and the AI ​​chip industry needs ecologicalization. It is becoming more and more obvious; at the same time, when facing different scenarios, the utilization and compatibility of AI chips need to be improved, and it is difficult to coordinate various heterogeneous devices based on different AI chips.