Domestic FPGA enters the acceleration period

We believe that with the rise of 5g, aiot and other industries, the competitiveness of domestic FPGA manufacturers is improved, and it is expected to usher in an explosion of performance. Compared with other general logic devices or ASICs, FPGA has advantages in flexibility and small-scale deployment cost, which just meets the needs of flexible deployment of 5g communication services and long tail fragmentation in aiot market. 2020 is the first year of China's 5g scale construction, which is in a leading position in the world, promoting the vigorous development of relevant domestic aiot, industrial control and other industries and stimulating FPGA demand. We believe that from the perspective of products, domestic manufacturers have gradually approached foreign leading enterprises in the field of medium and low-end products. With the continuous breakthrough of high-end products and the continuous iteration of application scenarios and cases, it is expected to quickly improve their competitiveness.
abstract
FPGA has a good balance among flexibility, performance, power consumption and cost. 1) FPGA is a hardware reconfigurable chip structure, which is more applicable to the links where the technology changes quickly or needs to meet the flexible changes of application functions. 2) FPGA is one of the devices with the shortest development cycle and the lowest application risk in special circuits. Some customers can obtain applicable FPGA chips without investment in R & D. therefore, FPGA has higher cost performance in areas where the demand is temporarily low and the scale is not significant. 3) The FPGA chip can build the corresponding pipeline according to the number of data packet steps, so as to realize data parallelism and pipeline parallelism. It has strong computing power and has the advantages of power consumption and delay in some application scenarios.
Solve the pain points of downstream applications and the demand grows rapidly. 1) Network communication is the most widely used scenario downstream of FPGA, which effectively realizes a large number of high-speed digital signal processing functions. At the same time, 5g flexible deployment and business diversification increase the demand for FPGA. 2) The integration and security requirements in the field of industrial control are constantly improving. FPGA can help to achieve more intelligent, high-performance and low-energy solutions. 3) Relying on the advantage of delay, the data center acceleration layer is formed, and the remaining resources can be deployed with neural network. We see that Microsoft has begun to use this scheme. 4) According to Frost &;; Sullivan, in 2020, the global FPGA market scale will reach US $6.08 billion, the CAGR will be 16.4% from 2021 to 2025, and the market scale will be US $12.58 billion in 2025; China's FPGA market has increased from 6.55 billion yuan in 2016 to 15.03 billion yuan in 2020. In the future, China's FPGA market demand is expected to continue to expand, and the market scale is expected to reach 33.22 billion yuan in 2025.
At present, domestic FPGA products are mainly low-density products, and high-end products continue to make breakthroughs. At present, the domestic FPGA products active in the market are mainly low-density products, and most of the architecture adopts the LUT + wiring concept. There is still a gap between the technical level of medium and high-density FPGA and the international leaders. At the same time, some manufacturers have made breakthroughs and formed a catch-up trend, such as Ziguang Tongchuang, Gaoyun semiconductor, Fudan micro and so on.
risk
5g construction was less than expected, cloud data center deployment was less than expected, and FPGA R & D progress was less than expected.
text
FPGA chip: semi customized integrated circuit with flexibility and performance
FPGA: field programmable gate array based on programmable devices
FPGA (field programmable gate array), namely field programmable gate array, originally originated from Xilinx company, which launched the world's first FPGA chip "xc2064" in 1985. FPGA chip is developed based on programmable devices such as pal, gal and CPLD, which not only solves the disadvantage of limited number of traditional programmable gate arrays, but also has larger scale capacity and improves performance; At the same time, the field programmable feature helps it overcome the defect of insufficient flexibility of customized circuits (such as ASIC chips).
Chart: FPGA development history
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: electronic enthusiast network, research department of CICC
In terms of structure, FPGA chips are arranged in matrix by logic unit blocks, and different functions are designed through programmable interconnection resources. FPGA chip mainly includes three parts: programmable logic block (CLB), programmable input / output unit (I / O) and interconnection wiring resources.
Figure: main composition structure of FPGA
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Performance Analysis of nanoelectronic relay based field programmable gat, Tian Qin et al., research department of CICC
In the internal structure of FPGA, programmable logic block (CLB) is the most important resource in FPGA, which can reflect the structural differences of different FPGAs and help FPGA realize programmability. Taking Xilinx 7 series as an example, its CLB contains two slices. The basic unit of each slice consists of a look-up table (LUT), which is essentially random access memory (RAM). Therefore, FPGA can be regarded as a static memory (SRAM). When the user describes a logic circuit through HDL language, the FPGA development software calculates all the results of the logic circuit and stores them in RAM. When the user inputs the signal for logic operation, it looks up the table similar to the input address, so as to output the operation results. From this point of view, the basic reason why FPGA can realize programmability is that in theory, as long as the storage capacity is expanded and the input signal line is increased, the look-up table can realize the input and operation of any number of logic functions.
Chart: each CLB in Xilinx 7 Series FPGA contains 8 lookup tables (LUTS)
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Xilinx official website, CSDN, CICC Research Department
FPGA chip mainly has the following characteristics:
·Flexible design, FPGA is a hardware reconfigurable chip structure, with a rich number of programmable input and output unit pins and triggers;
·It is applicable and convenient. FPGA is one of the devices with the shortest development cycle and the lowest application risk in special circuits (some customers can obtain applicable FPGA chips without investment in R & D);
·Parallel computing, the FPGA chip can build the corresponding pipeline according to the number of data packet steps, so as to realize data parallelism and pipeline parallelism;
·High compatibility, FPGA can work together with CMOS and other large-scale integrated circuits.
Based on the above characteristics, FPGA chip was used in some scenarios as a semi customized circuit substitute for ASIC chip in the early stage. In recent years, with the increasing construction of data center by leading Internet enterprises such as Microsoft and Amazon, the application scope of FPGA chip has also been broadened.
FPGA has a good balance among flexibility, performance, power consumption and cost
Compared with CPU, the parallel computing power of FPGA can improve the operation speed and reduce the delay. The essence of CPU is to reuse processing units in time dimension by using large-scale memory, and realize more application logic with the support of powerful logical database, but it will also lose the parallel processing ability of processing units.
Compared with GPU, FPGA has advantages in power consumption and flexibility. On the one hand, because GPU adopts a large number of processing units and accesses a large number of off chip storage SDRAM, its calculation peak is higher and its power consumption is also higher. The average power consumption (10W) of FPGA is much lower than that of GPU (200W), which can effectively improve the heat dissipation problem; On the other hand, GPU cannot change the hardware resources after the design is completed, and FPGA has more flexibility to program the hardware according to specific applications. Machine learning uses multiple instructions to process single data in parallel, and the customization ability of FPGA can better meet the requirements of low accuracy, decentralized and unconventional deep neural network computing.
Compared with ASIC chips, FPGA has the advantages of short cycle and high cost performance in the early stage of the project. The ASIC needs to be designed from the standard unit. When the functional and performance requirements of the chip change or the process progress, the ASIC needs to be re cast, resulting in higher sunk cost and longer development cycle; FPGA includes prefabricated gate and trigger, which has the advantages of programming, debugging, reprogramming and repeated operation, and can realize the reconfiguration of chip functions. Therefore, early FPGA often appeared as a semi custom circuit in the field of customized ASIC, which is considered by the industry as one of the faster ways to build prototype and develop design.
Chart: economic cost comparison between FPGA and ASIC
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: semiconductor industry alliance, CICC Research Department
Figure: comprehensive comparison of FPGA and ASIC
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: field programmable gate arrays for military application, research department of CICC
Special EDA software to build core barriers
FPGA chip needs the coordinated development of software and hardware to realize function configuration, and the programming complexity and specificity of EDA software tools for FPGA build differentiated competitive advantages. From the perspective of design process, the front-end design process of FPGA is the same as that of other chips, and generalized general EDA is also required. However, when the chip design is basically completed, FPGA needs to compile the user written program into FPGA through the process in software tools, so as to realize the corresponding functions. FPGA special software tool is a combination of a series of EDA. From the perspective of design process, it mainly includes the following steps: circuit design, logic synthesis, design implementation (process mapping, layout and wiring, etc.), simulation and programming download.
Figure: FPGA software tool design flow chart
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Ziguang Tongchuang official website, research department of CICC
In the FPGA software tool design process, except that some front-end processes can use third-party tools, other links basically need independent R & D coverage by FPGA chip manufacturers, and there are high technical barriers. In the layout and wiring, the circuit needs to be described as chip configuration information through FPGA special EDA software, so this link must use the internal structure information of FPGA chip.
From the research and development of software tools of FPGA manufacturers at home and abroad: for foreign manufacturers, Xilinx recently launched Vitis unified software platform to help users with software development; Altera provides Quartus II, a design tool for programmable logic; Lattice software tool is diamond. In terms of domestic manufacturers, Ziguang Tongchuang has developed a Pango design suite that supports tens of millions of gate FPGA designs.
Chart: research and development of software tools of some FPGA manufacturers at home and abroad
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Official Websites of companies, research department of CICC
FPGA Software and hardware drive each other, especially in the field of high-end large-scale FPGA chips, the design process of FPGA is often software first. It can even be said that the R & D level of software tools has a great impact on the height of hardware scale and performance. In order to maximize the hardware performance at the application end, FPGA manufacturers need to consider not only the matching and consistency with the chip hardware architecture, but also the diversity of application scenarios, the complexity of timing model, user experience and efficiency, especially for the core links represented by logic synthesis and layout and wiring, There is still a gap between domestic FPGA manufacturers and overseas leading enterprises.
FPGA application field: "3 + 3" application scenario, expanding downstream
Three core application scenarios: communication, industry and data center
Communication field: FPGA strengthens the parallel processing capability of communication equipment, and 5g opens the emerging incremental space
The field of network communication is the most widely used scenario downstream of FPGA, which effectively realizes a large number of high-speed digital signal processing functions. At present, FPGA is widely used in wired and wireless communication equipment to realize the functions of interface expansion, logic control, data processing, single chip system and so on. In the wireless communication system, many functional modules usually require a large number of filter operations, and these filter functions often require a large number of multiplication and accumulation operations. The internal distributed logic and operation unit structure of FPGA chip makes it easier to realize the distributed algorithm structure, so these multiplication and accumulation operations can be realized effectively, It can realize a large number of high-speed digital signal processing functions in the communication process.
Massive MIMO technology innovation highlights the flexibility of FPGA in multi-channel beamforming. In order to enhance signal coverage and spectral efficiency, massive MIMO has become one of the key technologies introduced by 5g. Under the massive MIMO technology, the number of transmitting and receiving channels is increased from 16t16r to 64t64r or even 128t128r. At the same time, the signal received by each antenna unit needs to be digitally processed, which produces a lot of computational load in adaptive beamforming. FPGA has more advantages than CPU and DSP in I / O, operation speed and delay, and FPGA is more flexible in multi-channel beamforming.
Chart: FPGA application in 5g field
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Master Xin, research department of CICC
Industrial control: integration and safety requirements are improved, and FPGA provides cost-effective solutions
FPGA can help industrial applications achieve more intelligent, high-performance and low-energy solutions. Taking the motor of NC machine tool as an example, compared with the traditional special chip that can only control a single motor, the multi-channel motor control based on FPGA can be realized: 1) without adding additional devices while increasing the Ethernet connection capacity, 2) flexible methods can be used to control multiple motors and differentiated products, 3) Realize the single chip scheme supporting motor control and Fieldbus connection; At present, 2 / 3 of the world's industrial power is used to drive motors. We believe that with the development of intelligent unmanned factories with automation as the core element, all kinds of FPGAs that can accurately control motors and realize multi motor control on a single chip will continue to penetrate in the future.
Chart: Altera cyclone series: Industrial Ethernet solution based on FPGA
CICC semiconductor Computing Series I: in line with the new scenarios of downstream applications, domestic FPGA has entered the acceleration period
Source: Lianquan technology, research department of CICC