In the history of semiconductor, FPGA plays an important role in the design and implementation of electronic systems. This is why we added the history of FPGA to fabless: the transformation of the semiconductor industry, and added a new chapter on the history of achronix in the 2019 edition.
In a recent blog post “FPGA in the 2020 – the new old thing”, achronix said that FPGA has a history of 35 years, and the future era of cloud computing artificial intelligence represents a new FPGA growth opportunity. In fact, in our first webinar last year, the achronix ml webinar broke the analytical record.
FPGA is a semi custom circuit, which is mainly used in application specific integrated circuits. It has an irreplaceable position in the fields of Aerospace / national defense, consumer electronics, electronic communication and so on. In the downstream application of FPGA, communication occupies the largest market segment, up to about 60%. Among them, FPGA chip is indispensable for 5g development.
During this time, the hot monthly trade war pushed the 5g industry to the forefront of the storm. However, China’s civil FPGA supply depends on the four chip giants of Xilinx, Altera, lattice and MICROSEMI in the United States, and the localization rate of FPGA in the civil field is only 4% (mrfr data). 5g will bring new requirements for FPGA, and the construction of base station is close at hand. At present, FPGA is used in 5g macro base stations.
Whether at the edge (efga) or in the cloud (FPGA), programmable technology will play a key role in the initial 5g data explosion. Since the fourth quarter of 2015, we have tracked AI on semiwiki, published 182 blogs and obtained nearly 1 million views, which is a very good achievement. In short, artificial intelligence is everywhere, and companies large and small are consuming artificial intelligence design support information at the fastest speed.
Back to achronix’s blog post, it is full of interesting data and links, which will be very useful if you are studying the use of FPGA in the 5g era. I also spent a lot of time studying artificial intelligence and completed several artificial intelligence projects in cooperation with some large companies and semiwiki.
FPGA is an antique in today’s semiconductor field. Although FPGA has a history of 35 years, the next 10 years represent a growth opportunity that has not appeared since the early 1990s. Why is this happening now?
The global data volume continues to surge. IDC predicts that by 2025, the world will generate more than 175zb data every year. With so much data, there is a great opportunity to analyze it and gain insights that can change and affect the world. Artificial intelligence will play an important role in this data mining operation. Enterprises are increasing employees with deep skills in machine learning and data analysis to meet future challenges
In addition, due to the significant increase in the number of 5g channels, the consumption of single station FPGA increases accordingly. China’s 5g commercial progress leads the world, and China’s large-scale capital expenditure on each generation of mobile communication technology is generally concentrated in the first few years of commercial use. Therefore, FPGA will occupy a more important position than 4G era. Due to the high frequency band of 5g applications, the number of 5g base stations may reach 1.5 times that of 4G. In addition, with the maturity of “5g second half” millimeter wave technology after 2022, the number and scale of small base stations are expected to reach tens of millions. The business scenarios that 5g needs to meet will far exceed 1g ~ 4G. 5g devices will face more complex physical protocols and algorithms, and have higher requirements for logic control and interface rate.
Therefore, experts estimate that the market value of single base station side FPGA in 5g market will reach several times that of 4G. FPGA has high threshold and low localization rate. In the future, it will become a key field of localization substitution. The FPGA demand in China market is the largest in the world (more than 30%), but the localization rate in China’s FPGA market is very low. At present, in the civil field, the market share of domestic FPGA in China is less than 5%, and domestic FPGA is expected to flourish in the future.
New block floating point arithmetic unit for processing AI / ml workload in FPGA
Block floating point (BFP) is a mixture of floating-point and fixed-point algorithms, in which data blocks are assigned a common exponent. We describe a new arithmetic unit that performs block floating-point operations and creates floating-point results for common matrix arithmetic operations. The BFP arithmetic unit supports a variety of data formats with different precision and range. Compared with traditional floating-point arithmetic units, BFP saves a lot of power and area by sacrificing some accuracy. This new arithmetic unit has been implemented in a new 7Nm FPGA family from achronix. In a demonstration, artificial intelligence and machine learning workloads were benchmarked, and the performance and power consumption of BFP were improved compared with half precision (fp16) operation.
Can FPGA dominate?
As a large FPGA (field programmable gate array) manufacturer, Xilinx believes that Moore’s law is close to the end. In the future, we can’t rely solely on the periodic update of chip design to meet the needs, which leads to the need to rely more on chip architecture innovation in the future. Gilles Garcia, marketing director of Xilinx communication department, said that the common ground of these challenges is that they require higher performance, more bandwidth and stronger computing power. The 5g era means greater development opportunities for Xilinx.
Different from the previous communication period, 5g means more fragmented processing nodes, and end edge pipe cloud constitutes a huge computing network. Garcia further pointed out that 5g will bring three main challenges to FPGA manufacturers: how to manage the complexity of radio for large-scale antenna arrays, how to manage the bandwidth requirements of forward transmission, and how to meet the requirements of more data return.
Specifically, in terms of forward transmission, while meeting the use of traditional 4G radio devices, 5g protocol device processing is added, which means that it should be able to support the integrated access of different protocols; For multiple interfaces between metropolitan areas, the maximum rate of each interface is 100gbps, which means that the core trunk transmission interface needs to have at least 100-400gbps transmission capacity. At the same time, the data backhaul throughput of the terminal will increase by 10 times. Coupled with the huge user group, the data volume will show an exponential upward trend.
From the perspective of network characteristics, 5g has further increased the requirements for higher bandwidth and lower delay. Gilles Garcia said that the 16 nm process technology adopted by Xilinx can save 60% power consumption; RF SoC technology can combine digital and analog on a single-chip SOC (system level chip), and reduce the package size by 70%.
Where is the next competitive point?
An important change in the 5g era is that telecom operators carry out edge computing in the data center, which involves the function of network virtualization. In the core network, manufacturers’ requirements for bandwidth are increasing. Therefore, Xilinx believes that the data center will be an important market. Large manufacturers, including Xilinx, have spent a lot of money on R & D to provide artificial intelligence, machine learning and heterogeneous computing solutions. In the FPGA market, Intel had previously acquired Altera, a competitor of Xilinx, and a major force of Intel also lies in the data center. Previously, Intel acquired easic to transition FPGA to ASIC (application specific integrated circuit) level with easic. There are also views in the industry on whether ASIC will take over FPGA and become a popular processing scheme of the next generation.
Compared with the past, there are obvious differences in the Internet of things era. The deployment needs of edge computing business show diversified characteristics. Only companies that accurately grasp customer needs can gradually become bigger and stronger. In the field of artificial intelligence learning reasoning and image intelligence demand, GPU is popular, which promotes NVIDIA and other companies to usher in high-speed growth. In the era of 5g edge computing, FPGA will become a new industry focus. Compared with other heterogeneous processors, FPGA is more suitable for edge computing scenarios. It can be deeply customized according to its algorithm structure for each specific application to achieve high computing efficiency and energy efficiency; At the same time, compared with GPU, FPGA architecture can greatly optimize bandwidth and improve computing efficiency; And it has natural advantages in low delay and stability. It can be seen that the diversification of 5g edge scenarios will drive the computing industry into the FPGA era.