From the cloud to the edge AI promotes FPGA application development
In recent years, one of the most interesting trends in FPGAs is the continuous expansion of application fields. Communication, industrial control, etc. are the traditional application markets of FPGAs. However, in recent years, with the development of cloud computing, big data, artificial intelligence, and Internet of Things, FPGA Start to expand into new areas.
Zhang Wei, associate researcher of the Heterogeneous Computing Group at Microsoft Research Asia, pointed out: 'The rise of artificial intelligence in these years relies on large-scale multi-layer network models of deep neural network algorithms, typically represented by cyclic neural networks and convolutional neural networks. Such a model inference usually requires billions or even billions of operations, while the online service system has a response time of the order of milliseconds. This means trillions of times per second (TFLOPS) or even millions. Million times of computing performance, as well as the size and power consumption of the device, there are certain constraints. This makes us expect the high-performance computing hardware of the data center should have the following advantages: low latency, low overhead and scale. In the data center, FPGA Compared to the core advantage of GPU is low latency, using FPGA to accelerate, only need micro-level PCIe delay (Microsoft's current FPGA is used as a PCIe accelerator card), for communication-intensive tasks, FPGA compared to CPU, GPU The advantage is even greater. 'Today, Microsoft has deployed FPGAs on hundreds of thousands of cloud servers. Microsoft is also currently using FPGAs in data centers.
The rapid rise of network edge computing is also increasing the demand for FPGA applications. Chen Yingren, senior business development manager for Lattice Semiconductor Asia Pacific, said: 'Drivers are being deployed more and more by delays, privacy and bandwidth limitations. Among the IoT system devices, the need for ultra-low-power flexible reasoning needs to be met. 'The data from Semico Research Market Research shows that the number of network edge devices using artificial intelligence will explode at a compound annual growth rate of 110%.
Huang Jun, vice president of Gaoyun Semiconductor Market, divided the emerging market into three parts: First, in the intelligent terminal equipment represented by consumer electronics, the interface is more and more abundant, and at the same time, it needs to maintain small size and low power consumption, which requires The underlying devices can more flexibly support interface conversion and glue logic, making low-density FPGAs more useful. Second, more and more AI operations are beginning to be performed in the field of edge computing, such as audio and video acquisition and preprocessing. They need to support simple or medium complexity algorithms and IP, which in turn gives opportunities for low- and medium-density FPGAs. Finally, the cloud needs to support high-speed SerDes, and high-speed FPGAs are becoming a necessity.
Generalized, Artificial Intelligence, Big Data, Cloud Computing, ADAS, Internet of Things, Robotics, Wearables, 5G Communication Equipment, Network Switching, Industrial Control, Financial Equipment, Security Monitoring, Video Drive, Medical Instruments, Automotive Electronics, Home Electrical appliances, information security, etc. will be important application markets for FPGAs in the future. In 2017, the global FPGA scale market is about 5 billion US dollars. In the report released by Goldman Sachs, it is predicted that the FPGA market will reach 10 billion US dollars in the next 5 years, to 2025. The year will grow to 12 billion US dollars.
From technology to brand, domestic FPGA catches up huge space
FPGAs have always been a highly monopolistic industry. Xilinx, Intel (Altera), Microsemi and Lattice have monopolized most of the world's markets. Despite the emergence of emerging markets in theory, new entrants Bring more development opportunities, but the challenges in the actual operation process are still very large.
'Some emerging markets are not expected to be available, especially for high-density, high-performance FPGAs. Not only is hardware design difficult, Chinese companies are lagging behind in hard IP resources, soft IP IP reserves are lacking, and software performance is The gap between foreign manufacturers is also very big. 'Huang Jun pointed out.
Wang Haili, CEO of Jingwei Qili, also believes that from the perspective of the product itself, there is still a long way to go before Chinese companies want to develop. 'The five years are less, and the more may take 8 to 10 years. 'Wang Haili said.
In addition to the technical gap, the gap in brand awareness is also very obvious. 'Zhongxing incident only gave the domestic companies some touch in the big environment, making them aware of the importance of supply chain management and control. But it is implemented in specific products. At the time, domestic suppliers still have many difficulties. The customer's brand recognition for domestic enterprises is not high, and they are worried about whether they can continue to supply. Even for the learning of product design tools, there is a process. FPGA, I feel that the pressure is still very large. ' Huang Jun said.
From differentiated competition to autonomously controllable FPGA 3.0 era requires long-term efforts
So how can we achieve a breakthrough in domestic FPGAs? Huang Jun believes that differentiation is a must-have for Chinese FPGA companies. 'Domestic manufacturers should start from the small point based on the characteristics of the chip itself and the difficulty of entering the industry, from low-density FPGAs. The products began to break through and gradually developed into medium and high-density FPGAs. In short, we can't simply copy foreign products, but we must carefully study the market segments, what are the special needs of new products, and then optimize the local characteristics of our products to meet the needs of users. For example, the video capture of many smart devices now requires a high-speed interface. We will do high-speed MIPI on the chip, which is faster than other products at the same density. At the same time, we will do more similar reference designs. Application is more convenient. These are examples of differentiation. ' Huang Jun said.
It is also very important to establish an independent industrial ecology. Wang Haili pointed out: 'More than 50% of the patented technologies of foreign companies are patents that protect the core architecture. Chinese companies should also pay attention to the development process, users are easy to use, friendly, and support application-oriented soft IP. The performance of the library and mature EDA tools is an important part of the product landing.
Whether it is a differentiated market strategy or the establishment of an industrial ecosystem, it is a compulsory course in the development of China's FPGAs. Only the establishment of an independent controllable + independent innovation FPGA product system is first embodied in the independent research and development of programmable logic. The ability of the core architecture of the device is secondly reflected in the ability to independently develop software tools for FPGA devices. The Chinese FPGA is truly a foothold. This is the key to the development of FPGAs in China.
According to Wang Haili, China's FPGA industry can be roughly divided into four stages: From the 1990s to 2005, it can be regarded as the Pre-1.0 era of Chinese FPGAs. At that time, the Chinese FPGA industry was just getting started, only in reverse. Design-oriented, chip design / software developers only a hundred people, no own structure, backward technology, no software tools and application development talents, can only gradually explore the progress. After accumulation from scratch, 2006 to 2016 China's FPGA industry has entered the 1.0 era. China FPGA has hundreds of chip design/software development teams, has its own prototype, uses general technology, has basic software tools and basic application soft IP, has the ability to conduct positive To design, began to cultivate the application talent of domestic FPGA.
Now it is in the era of China FPGA development 2.0, which will be from 2017 to 2025. This is a good development period. The Chinese FPGA industry already has thousands of chip design/software developers. After hard work, it is possible Form a complete architecture design, have the ability to adopt advanced technology, easy to use software tools, have some industry soft IP and solutions, form a complete forward design, and gradually establish an FPGA application ecosystem, and form a domestic production of thousands of people. Application talent team developed by FPGA.
As for the 3.0 era (2026 2035), it will be the development stage of China's FPGA from good to strong: the formation of thousands of chip design / software developers, using innovative architecture design and the most advanced manufacturing processes, with leading software tools, Complete industry soft IP and solutions form a relatively complete FPGA application ecosystem, and the application talents of domestic FPGA development reach tens of thousands.
I hope that China's FPGA industry will be able to take off as soon as possible.